Distributed Typhoon Track Prediction Based on Complex ...
Transcript of Distributed Typhoon Track Prediction Based on Complex ...
Research ArticleDistributed Typhoon Track Prediction Based on ComplexFeatures and Multitask Learning
Yongjiao Sun1 Yaning Song 1 Baiyou Qiao1 and Boyang Li2
1School of Computer Science and Engineering Northeastern University Shenyang 110819 China2School of Computer Science and Technology Beijing Institute of Technology Beijing 100081 China
Correspondence should be addressed to Yaning Song 1901780stuneueducn
Received 30 April 2021 Accepted 4 July 2021 Published 13 July 2021
Academic Editor Guanfeng Liu
Copyright copy 2021 Yongjiao Sun et al is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Typhoons are common natural phenomena that often have disastrous aftermaths particularly in coastal areas Consequentlytyphoon track prediction has always been an important research topic It chiefly involves predicting the movement of a typhoonaccording to its history However the formation and movement of typhoons is a complex process which in turn makes accurateprediction more complicated the potential location of typhoons is related to both historical and future factors Existing works donot fully consider these factors thus there is significant room for improving the accuracy of predictions To this end we presenteda novel typhoon track prediction framework comprising complex historical featuresmdashclimatic geographical and physicalfeaturesmdashas well as a deep-learning network based onmultitask learningWe implemented the framework in a distributed systemthereby improving the training efficiency of the network We verified the efficiency of the proposed framework on real datasets
1 Introduction
Typhoons are tropical cyclones that occur in the WesternPacific and adjacent waters and are common climate phe-nomena Given that typhoons have significant destructivepower and often imperil the coastal areas where they makelandfall the nature of these typhoons has long been animportant research topic [1ndash3]
Typhoon track prediction is a typical problem in ty-phoon research Traditionally typhoon paths are oftenpredicted through such methods as force analysis andmathematical statistics [4ndash7] In recent years however withthe development of artificial intelligence more researchersare using deep-learning technology to predict the movementof typhoons For example some studies have utilized cloudmaps to locate typhoons and predict their movement viaconvolutional neural networks (CNNs) and generativeadversarial networks (GANs) [89] Given that typhoon trackis a continuous process many studies also use recurrentneural networks (RNNs) and long short-term memory(LSTMs) to process the track sequence [10] e formationand movement of typhoons is a very complex process that is
affected by historical as well as future factors Although thisproblem has been widely studied some limitations remainand hinder the accurate prediction of the paths typhoonstake
Typhoons have complex historical features Existingstudies have evaluated the history of typhoons with re-spect to geopotential height wind field and atmosphericpressure however these studies did not comprehensivelyanalyse the features of previous typhoons erefore byanalysing historical data we identified additional perti-nent features and categorized them into climatic geo-graphical and physical features Further we consideredsome new featuresmdashsuch as geostrophic forcemdashfor thepurposes of this study e factors that affect typhoonmovement from many aspects were categorized undermultimodal features
Although existing works apply deep learning to evaluatetyphoons most only consider the track of a typhoon as anisolated target and ignore the multiple factors that influencethis track Likewise although a few studies have predictedtyphoon tracks via a multifaceted approach their analyses oftyphoon features are too simplistic erefore we combined
HindawiComplexityVolume 2021 Article ID 5661292 12 pageshttpsdoiorg10115520215661292
the complex features of typhoons processed the featuresthrough different learning frameworks and incorporatedmultitask learning to further improve the accuracy of ty-phoon track prediction
However the expansion of data and model parameters isaccompanied by an increase in computational power andduration of model training In this regard using distributedand parallel training methods such as SparkMLlib (httpsparkapacheorgmllib) can significantly improve the effi-ciency of model training erefore to improve the trainingefficiency of the framework proposed in this paper weimplemented it based on Ray (httpsrayio) which is anemerging distributed AI platform
e contributions of this paper are as follows
(1) We propose a typhoon track prediction frameworkthat considers both historical features and the in-teraction of multiple factors
(2) We extracted the complex featuresmdashclimatic geo-graphical and physicalmdashthat affect the movement oftyphoonsWe employed deep-learning networks anda multitask learning method to improve the accuracyof typhoon track prediction
(3) We utilized distributed implementation to improvethe training efficiency of the network
(4) We used real-life datasets to conduct the experi-ments and verify the effectiveness of the proposedframework
e remainder of this paper is organized as followsSection 2 introduces related works on typhoon track pre-diction Section 3 covers the problem definition and relatedtechnologies Section 4 introduces the proposed track pre-diction framework including feature selection and networkstructure We then verify the efficiency of the proposedframework through experiments in Section 5 and finallysummarize this paper in Section 6
2 Related Works
21 Traditional Methods Traditional methods of typhoontrack prediction include numerical statistical regressionand integrated models Weber [7] proposed a numericalmodel (STEPS) to analyse the annual performance of thenumerical orbit-prediction model the model involves a verycomplex atmospheric-dynamics formula and requiresstrong computational power to successfully predict a ty-phoonrsquos path Demaria et al [4] proposed a statistical model(SHIPS) that modifies the predictor according to the newprediction factors of every new year to make the model moresuitable for observing typhoon movement Compared withSTEPS SHIPS has a lower computational complexitynonetheless its accuracy is also relatively low Goerss andKrishnamurti et al [56] demonstrated that the integratedmodel comprising multiple models was more accurate asopposed to individual models Although traditional modelsplay a crucial role in forecasting typhoon tracks they stillhave many shortcomings With the increase in meteoro-logical detection instruments more meteorological
spatiotemporal data (big data) will be produced Howevertraditional models are inevitably becoming outmoded It isdifficult for them to capture nonlinear typhoon models fromthese huge datasets which significantly reduce the accuracyof prediction
22 Deep-Learning Methods In recent years deep learningand parameter optimization [11] have rapidly developed andprovided more powerful methods for typhoon track pre-diction Neural networks have the advantages of nonline-arity and nonlocality ey can utilize big data to train thenetwork and hence determine the mapping relationshipsbetween input and output this essentially makes the pre-dictions more accurate
CNN-based methods Wang et al [9] used 2250 in-frared satellite images to train the CNN network eaverage angular error of typhoon track prediction wasthus reduced to 278 degrees indicating the greatpotential of CNN in typhoon path prediction Giffard-Roisin et al [12] proposed a fusion neural networkcomprising a neural network using past trajectory dataand a CNN involving the reanalysis of atmosphericwind-field imagesGAN-based methods Ruttgers et al [8] used GAN inconjunction with satellite images and meteorologicaldata to forecast the central location of typhoons It hasbeen proven that GAN utilizes many features thatotherwise cannot be used by traditional models thuspreventing the otherwise inevitable errors associatedwith some traditional modelsRNN- and LSTM-based methods Moradi Kordma-halleh et al [13] used sparse RNNs with flexible to-pology in which a genetic algorithm (GA) was used tooptimize the weight connection Alemany et al [14]proposed a fully connected RNN in the grid system theproposed approach can be used to model the complexand nonlinear temporal behavior of typhoons Furtherit can accumulate the historical information of thenonlinear dynamics of the atmospheric system byupdating the weight matrix hence improving the ac-curacy of typhoon track prediction Chandra and Dayaland Chandra et al [1516] also proved that RNNs aresuitable for typhoon track prediction Lian et al [17]proposed a novel data-driven deep-learning modelcomposed of a multidimensional feature-selectionlayer a convolution layer and a gating-cycle unit layerIt uses spatial locations and a variety of meteorologicalfeatures to predict typhoon trajectories Compared withCNNs and RNNs without a feature-selection layer thenovel model has higher accuracy Using records from1949 to 2012 as the training data Gao et al [10]proposed a typhoon track prediction method based onLSTM the research shows that the model can predictthe typhoon track 6ndash24 hours in advance with betteraccuracy Kim et al [18] proposed a large number oftemporal and spatial prediction models based on theConvLSTM model
2 Complexity
Multitask learning-based methods Chandra [19]proposed a coevolutionary multitask learning algo-rithm that combines the functions of modularizationand multitask learning is approach coordinatesmultitask learning dynamic programming and co-evolution algorithms Furthermore it can trainneural networks via feature sharing and modularknowledge representation It can also be used topredict typhoon intensity with limited input [20]is shows that compared with traditional modelsthe algorithm not only solves the problem of dynamictime series but also improves the prediction accuracyMukherjee and Mitra [21] proposed a joint learningmodel that can learn the distance and direction oftyphoons simultaneously via two different structureswith multiple LSTMs and multiple fully connectedlayers initial layer parameters are shared accordingto past typhoon track data e research results showthat the model can predict direction and distance(ie displacement) simultaneously
3 Preliminaries
In this section we first introduce the relevant technologiesutilized in our framework and then proceed to define ourproblem
31CNN andResNet CNN is a type of deep-learning modelthat has been successfully implemented in image recognition[22] e convolution layer is one of the core structures ofCNNs e input of the convolution layer includes one ormore matrices of the same size each of which is called achannel Each convolution layer uses common parametersknown as convolution kernels For 2D input the function ofthe convolution layer is to weigh the corresponding sub-matrices according to the size of kernels thus the convo-lution layer output is generated
Another important structure is the pooling layer whichaims to reduce the parameters of the model and strengthenthe network while improving the computing speed estrategies of the pooling layer include the maximum andaverage pooling
ResNet is a CNN model widely used for feature ex-traction [23] To solve the migration problem in deepnetworks ResNet proposes residual learning ResNet re-places the feature H(x) obtained by convolution layers withthe residual H(x) minus x of feature and input In contrast withordinary CNN ResNet adds a shortcut mechanism betweenevery two layers to realize residual learning
32 LSTM LSTM is a special type of RNN that is delib-erately designed to avoid long-term dependence It intro-duces a gate to solve gradient disappearance or explosion[24] LSTM contains four important structures namely theforget gate the input gate the update stage and the outputgate As shown in Figure 1 this framework operates asfollows
(1) e function of the forget gate is implemented bysigmoid to determine which information needs to beforgotten according to the input xt and the outputhtminus1 of the previous cell
(2) e input gate determines the information that willbe stored in the current cell
(3) e update stage updates Ct of the current cell(4) e output gate outputs the final information to the
next cell
33 Multitask Learning In single-task learning (involved inthe previous models) the model learns only one task at atime For complex problems single-task learning decom-poses the problems into multiple independent subproblemsfor separate training and then combines them However inpractical applications these subproblems frequently containcorrelation information that is often ignored by the single-task learning method
In this regard the goal of multitask learning is to in-tegrate multiple related tasks through shared representations[25] It entails hard and soft parameter sharing Hard pa-rameter sharing shares some parameters among all tasks andonly uses the tasksrsquo unique parameters at a specific layer Insoft parameter sharing each task has unique parametersFinally the similarity is expressed by adding constraints tothe differences between parameters of different tasks
34 Problem Definition e problem of typhoon trackprediction can be expressed in terms of the features of agiven typhoon at several past instances or moments the goalis to predict the locations at certain times or instances in thefuture e past-feature sequence of the typhoon is denotedas S (s1 s2 st) where si represents the features of thetyphoon at time i and t is the length of the sequence etrack of the typhoon at the future moment isT[(xt+1 yt+1) (xt+2 yt+2) (xt+n yt+n)] where (xj yj) isthe geographical coordinate (latitude and longitude) at timetj e goal of this study is to establish the mapping modelM T⟶ S and hence calculate the future trajectory se-quence T through the historical sequence S
4 Framework
Figure 2 illustrates the structure of our proposed frameworkIt entails feature selection weighted fusion and multitaskprediction In this section we will introduce all the partsindividually
41 Features ere are three types of features in ourframework namely climatic geographical and character-istic features e climatic features include sea surfacetemperature geopotential height and specific humidityGeographical features include geostrophic forces echaracteristic features are the speed and position of thetyphoon
Complexity 3
411 Climatic Features By studying the influence of climateon typhoons we selected three main factors as climaticfeatures in this study
Sea surface temperature (SST) SST is one of the mostimportant factors in meteorological research In gen-eral SST decreases when latitude increases SST plays apivotal role in the formation and movement of ty-phoons typhoons are formed above the sea surfacewhere SST is higher than 265degC and the intensity of thetyphoon increases through continuous absorption ofenergy SST is also one of the main factors influencingthe direction of motion and landing location of ty-phoons In this study we mainly considered the regionwithin 0degN and 60degN latitudes and 100degE and 180degElongitudes e SST in this area was regularly collectedby the sensor As shown in Figure 3 we used a matrix of121 rows and 161 columns to represent SST in whichthe SST near the equator is above 30degC whereas theSST at higher latitudes is approximately 0degC We alsodistinguished land from sea the darkest shades inFigure 3 are landGeopotential height (GH) GH is an imaginary heightin meteorology expressed in terms of the work doneagainst gravity by an object of unit mass rising from sea
level to a certain height GH also plays an importantrole in maintaining the intensity and motion of ty-phoons For example the large geopotential heightgradient between the Western Pacific subtropical highand typhoon determines the direction of movement oftyphoon Ambi to a certain extent [13] We studied GHin the same region described previously In contrast to
ResNet
ResNet
ResNet
SST
GH 200 500 850 (hPa)
SH 200 500 850 (hPa)
Geographical and physical features
LSTM
LongitudeLatitude
Intensity
121 times 161 5
3 times 121 times 161
3 times 121 times 161
5
5
5
FC
Weightedfusion
FC
FC
Figure 2 e structure of our framework
σ σ Tan h σ
Tan h
x x
xt
Ctminus1 C
t
htminus1 h
t
x +
ft
it
Ct
~ otForget gate
Input gate
Update stage
Output gate
ft = σ (W
fx
t + U
fh
tminus1 + bf)
ot = σ (W
ox
t + U
oh
tminus1 + bo)
it = σ (W
ix
t + U
ih
tminus1 + bi)
ht = o
t tan h(C
t)
= tan h (WCx
t + U
Ch
tminus1 + bC)C
t
~
Ct
~= f
tminus1Ctminus1 + it
Ct
Figure 1 Structure of LSTM
100deg
E
110deg
E
120deg
E
130deg
E
140deg
E
150deg
E
160deg
E
170deg
E
180deg
W
Sea s
urfa
ce te
mpe
ratu
re
0deg0
5
10
15
20
25
30
10degN
20degN
30degN
40degN
50degN
60degN
Figure 3 An example of SST
4 Complexity
SST we choose three different GHs under different hPaFigures 4(a)ndash4(c) show examples of GH which is alsorepresented by matrices with 121 rows and 161 col-umns It is evident from these charts that GH increaseswith latitudeSpecific humidity (SH) SH refers to the ratio of the massof water vapor in the atmosphere to the total mass of airere is a strong relationship between typhoons andvertical air motion and SH is usually used when dis-cussing the vertical motion erefore we introduced SHas a distinct climatic feature Figures 4(d)ndash4(f) show 3 SHdata charts under different hPa We can observe that SHin the south is higher than that in the north
412 Geographical Feature (1) Geostrophic force (GF) GFalso known as the Coriolis force was derived to describe theforce exerted on moving objects on the surface of the Earthas a result of the Earthrsquos rotation Owing to the existence ofGF a rotating flow of air is formed and eventually a ty-phoon is formed under the combined action of variousfactors e typhoon is also affected by GF during itsmovement In the northern hemisphere the GF of the ty-phoon is to the right which determines the typhoonrsquos di-rection of movement to a certain extent GF can be expressedas
F 2mvω sin θ (1)
where m is the mass of the object v is the velocity of theobject ω is the angular velocity of the Earthrsquos rotation and θis the latitude of the object before it begins to move Giventhat the mass of typhoons is difficult to estimate we use thegeostrophic force gradient to represent the influence of GFon typhoons denoted as
zF
zm 2vω sin θ (2)
413 Physical Features We use physical characteristics todescribe the time series and tracks of typhoons
Location and direction Given that the track of a ty-phoon is a series of coordinates we used the latitudesand longitudes (lat lon) or offsets (ΔlatΔlon) to de-scribe the location and direction of motion of typhoonsSince typhoon data were collected every 6 hours wecalculated the movement and direction of the typhoonevery 6 hoursSpeed e typhoon data are coarse erefore we usedthe average of the velocities of the typhoon at twoconsecutive moments to describe the moving velocityof the typhoonIntensity e intensity of a typhoon is determined byits wind speed Existing studies have validated therelationship between the central pressure of a typhoonand the maximum wind speed [26] erefore we usedthe maximum central pressure to express the intensitycharacteristics of a typhoon
42 Network Owing to the different modes of features weused different networks to process the features and then usedfeature fusion for learning e entire network architectureis illustrated in Figure 2
421 Feature Extraction We used climatic geographicaland physical features Some of these were two-dimensionalmatrices whereas some were one-dimensional vectorsConsequently we used different networks for differentfeatures
For climatic features all inputs were two-dimensionalimages We therefore used three ResNets to process theimages e ResNets employed in our framework have 18hidden layers [23] as shown in Figure 5 GH and SH havetrichannel inputs whereas SST has single-channel inputefirst layer is a convolution layer e size of the convolutionkernel is 7 times 7 and the stride is (2 2) Based on the size of theinput we set padding as (3 3) Batch normalization (BN)and rectified linear units (ReLU) were also used in theconvolution layer After the convolution operation thenetwork performs a maximum-pooling operation ere arefour residual blocks after the first layer Each residual blockis repeated twice To simplify the representation the re-peated parts have been replaced by ellipses Each residualblock contains two convolution layers Each layer contains aconvolution kernel batch normalization and ReLUe sizeof the convolution kernel is 3 times 3 the stride is (1 1) and thepadding is (1 1) e output dimensions of each residualblock are 64 128 256 and 512 After the last residual blockthe network performs an average-pooling operatione lastlayer of the network is a fully connected network with 5-dimensional output
As for the geographical and physical features we used afully connected network and obtained a 5-dimensionalvector as the output For feature fusion we adopted a weightmodule e weight of each feature can be regarded as thecorrelation between the feature and track of the typhoonrough weighted feature fusion for each moment weobtained a 20-dimensional feature vector which then be-came the input of the predictor
422 Multitask Prediction Because LSTM has a consider-able advantage in the processing of sequence data we usedthe classic LSTM as the predictor e dimension of theinput was t times 20 where t is the length of the sequence asintroduced in Section 2 e training process is shown inFigure 6 First we used zero-state initialization to calibratethe weight h0 and C0 For each cell of the LSTM the input isthe i-th 20-dimensional feature vector It should be notedthat all LSTM cells share these parameters
e LSTM output is divided into two tasks e maintask involves locating the typhoon at the next moment andthe auxiliary task involves determining the central pressureof the typhoon (ie the intensity of the typhoon) We usedthe L2 norm as the loss function of the two tasks
For the main task the loss is the difference in distancebetween the real location and the predicted location of thetyphoon as follows
Complexity 5
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
the complex features of typhoons processed the featuresthrough different learning frameworks and incorporatedmultitask learning to further improve the accuracy of ty-phoon track prediction
However the expansion of data and model parameters isaccompanied by an increase in computational power andduration of model training In this regard using distributedand parallel training methods such as SparkMLlib (httpsparkapacheorgmllib) can significantly improve the effi-ciency of model training erefore to improve the trainingefficiency of the framework proposed in this paper weimplemented it based on Ray (httpsrayio) which is anemerging distributed AI platform
e contributions of this paper are as follows
(1) We propose a typhoon track prediction frameworkthat considers both historical features and the in-teraction of multiple factors
(2) We extracted the complex featuresmdashclimatic geo-graphical and physicalmdashthat affect the movement oftyphoonsWe employed deep-learning networks anda multitask learning method to improve the accuracyof typhoon track prediction
(3) We utilized distributed implementation to improvethe training efficiency of the network
(4) We used real-life datasets to conduct the experi-ments and verify the effectiveness of the proposedframework
e remainder of this paper is organized as followsSection 2 introduces related works on typhoon track pre-diction Section 3 covers the problem definition and relatedtechnologies Section 4 introduces the proposed track pre-diction framework including feature selection and networkstructure We then verify the efficiency of the proposedframework through experiments in Section 5 and finallysummarize this paper in Section 6
2 Related Works
21 Traditional Methods Traditional methods of typhoontrack prediction include numerical statistical regressionand integrated models Weber [7] proposed a numericalmodel (STEPS) to analyse the annual performance of thenumerical orbit-prediction model the model involves a verycomplex atmospheric-dynamics formula and requiresstrong computational power to successfully predict a ty-phoonrsquos path Demaria et al [4] proposed a statistical model(SHIPS) that modifies the predictor according to the newprediction factors of every new year to make the model moresuitable for observing typhoon movement Compared withSTEPS SHIPS has a lower computational complexitynonetheless its accuracy is also relatively low Goerss andKrishnamurti et al [56] demonstrated that the integratedmodel comprising multiple models was more accurate asopposed to individual models Although traditional modelsplay a crucial role in forecasting typhoon tracks they stillhave many shortcomings With the increase in meteoro-logical detection instruments more meteorological
spatiotemporal data (big data) will be produced Howevertraditional models are inevitably becoming outmoded It isdifficult for them to capture nonlinear typhoon models fromthese huge datasets which significantly reduce the accuracyof prediction
22 Deep-Learning Methods In recent years deep learningand parameter optimization [11] have rapidly developed andprovided more powerful methods for typhoon track pre-diction Neural networks have the advantages of nonline-arity and nonlocality ey can utilize big data to train thenetwork and hence determine the mapping relationshipsbetween input and output this essentially makes the pre-dictions more accurate
CNN-based methods Wang et al [9] used 2250 in-frared satellite images to train the CNN network eaverage angular error of typhoon track prediction wasthus reduced to 278 degrees indicating the greatpotential of CNN in typhoon path prediction Giffard-Roisin et al [12] proposed a fusion neural networkcomprising a neural network using past trajectory dataand a CNN involving the reanalysis of atmosphericwind-field imagesGAN-based methods Ruttgers et al [8] used GAN inconjunction with satellite images and meteorologicaldata to forecast the central location of typhoons It hasbeen proven that GAN utilizes many features thatotherwise cannot be used by traditional models thuspreventing the otherwise inevitable errors associatedwith some traditional modelsRNN- and LSTM-based methods Moradi Kordma-halleh et al [13] used sparse RNNs with flexible to-pology in which a genetic algorithm (GA) was used tooptimize the weight connection Alemany et al [14]proposed a fully connected RNN in the grid system theproposed approach can be used to model the complexand nonlinear temporal behavior of typhoons Furtherit can accumulate the historical information of thenonlinear dynamics of the atmospheric system byupdating the weight matrix hence improving the ac-curacy of typhoon track prediction Chandra and Dayaland Chandra et al [1516] also proved that RNNs aresuitable for typhoon track prediction Lian et al [17]proposed a novel data-driven deep-learning modelcomposed of a multidimensional feature-selectionlayer a convolution layer and a gating-cycle unit layerIt uses spatial locations and a variety of meteorologicalfeatures to predict typhoon trajectories Compared withCNNs and RNNs without a feature-selection layer thenovel model has higher accuracy Using records from1949 to 2012 as the training data Gao et al [10]proposed a typhoon track prediction method based onLSTM the research shows that the model can predictthe typhoon track 6ndash24 hours in advance with betteraccuracy Kim et al [18] proposed a large number oftemporal and spatial prediction models based on theConvLSTM model
2 Complexity
Multitask learning-based methods Chandra [19]proposed a coevolutionary multitask learning algo-rithm that combines the functions of modularizationand multitask learning is approach coordinatesmultitask learning dynamic programming and co-evolution algorithms Furthermore it can trainneural networks via feature sharing and modularknowledge representation It can also be used topredict typhoon intensity with limited input [20]is shows that compared with traditional modelsthe algorithm not only solves the problem of dynamictime series but also improves the prediction accuracyMukherjee and Mitra [21] proposed a joint learningmodel that can learn the distance and direction oftyphoons simultaneously via two different structureswith multiple LSTMs and multiple fully connectedlayers initial layer parameters are shared accordingto past typhoon track data e research results showthat the model can predict direction and distance(ie displacement) simultaneously
3 Preliminaries
In this section we first introduce the relevant technologiesutilized in our framework and then proceed to define ourproblem
31CNN andResNet CNN is a type of deep-learning modelthat has been successfully implemented in image recognition[22] e convolution layer is one of the core structures ofCNNs e input of the convolution layer includes one ormore matrices of the same size each of which is called achannel Each convolution layer uses common parametersknown as convolution kernels For 2D input the function ofthe convolution layer is to weigh the corresponding sub-matrices according to the size of kernels thus the convo-lution layer output is generated
Another important structure is the pooling layer whichaims to reduce the parameters of the model and strengthenthe network while improving the computing speed estrategies of the pooling layer include the maximum andaverage pooling
ResNet is a CNN model widely used for feature ex-traction [23] To solve the migration problem in deepnetworks ResNet proposes residual learning ResNet re-places the feature H(x) obtained by convolution layers withthe residual H(x) minus x of feature and input In contrast withordinary CNN ResNet adds a shortcut mechanism betweenevery two layers to realize residual learning
32 LSTM LSTM is a special type of RNN that is delib-erately designed to avoid long-term dependence It intro-duces a gate to solve gradient disappearance or explosion[24] LSTM contains four important structures namely theforget gate the input gate the update stage and the outputgate As shown in Figure 1 this framework operates asfollows
(1) e function of the forget gate is implemented bysigmoid to determine which information needs to beforgotten according to the input xt and the outputhtminus1 of the previous cell
(2) e input gate determines the information that willbe stored in the current cell
(3) e update stage updates Ct of the current cell(4) e output gate outputs the final information to the
next cell
33 Multitask Learning In single-task learning (involved inthe previous models) the model learns only one task at atime For complex problems single-task learning decom-poses the problems into multiple independent subproblemsfor separate training and then combines them However inpractical applications these subproblems frequently containcorrelation information that is often ignored by the single-task learning method
In this regard the goal of multitask learning is to in-tegrate multiple related tasks through shared representations[25] It entails hard and soft parameter sharing Hard pa-rameter sharing shares some parameters among all tasks andonly uses the tasksrsquo unique parameters at a specific layer Insoft parameter sharing each task has unique parametersFinally the similarity is expressed by adding constraints tothe differences between parameters of different tasks
34 Problem Definition e problem of typhoon trackprediction can be expressed in terms of the features of agiven typhoon at several past instances or moments the goalis to predict the locations at certain times or instances in thefuture e past-feature sequence of the typhoon is denotedas S (s1 s2 st) where si represents the features of thetyphoon at time i and t is the length of the sequence etrack of the typhoon at the future moment isT[(xt+1 yt+1) (xt+2 yt+2) (xt+n yt+n)] where (xj yj) isthe geographical coordinate (latitude and longitude) at timetj e goal of this study is to establish the mapping modelM T⟶ S and hence calculate the future trajectory se-quence T through the historical sequence S
4 Framework
Figure 2 illustrates the structure of our proposed frameworkIt entails feature selection weighted fusion and multitaskprediction In this section we will introduce all the partsindividually
41 Features ere are three types of features in ourframework namely climatic geographical and character-istic features e climatic features include sea surfacetemperature geopotential height and specific humidityGeographical features include geostrophic forces echaracteristic features are the speed and position of thetyphoon
Complexity 3
411 Climatic Features By studying the influence of climateon typhoons we selected three main factors as climaticfeatures in this study
Sea surface temperature (SST) SST is one of the mostimportant factors in meteorological research In gen-eral SST decreases when latitude increases SST plays apivotal role in the formation and movement of ty-phoons typhoons are formed above the sea surfacewhere SST is higher than 265degC and the intensity of thetyphoon increases through continuous absorption ofenergy SST is also one of the main factors influencingthe direction of motion and landing location of ty-phoons In this study we mainly considered the regionwithin 0degN and 60degN latitudes and 100degE and 180degElongitudes e SST in this area was regularly collectedby the sensor As shown in Figure 3 we used a matrix of121 rows and 161 columns to represent SST in whichthe SST near the equator is above 30degC whereas theSST at higher latitudes is approximately 0degC We alsodistinguished land from sea the darkest shades inFigure 3 are landGeopotential height (GH) GH is an imaginary heightin meteorology expressed in terms of the work doneagainst gravity by an object of unit mass rising from sea
level to a certain height GH also plays an importantrole in maintaining the intensity and motion of ty-phoons For example the large geopotential heightgradient between the Western Pacific subtropical highand typhoon determines the direction of movement oftyphoon Ambi to a certain extent [13] We studied GHin the same region described previously In contrast to
ResNet
ResNet
ResNet
SST
GH 200 500 850 (hPa)
SH 200 500 850 (hPa)
Geographical and physical features
LSTM
LongitudeLatitude
Intensity
121 times 161 5
3 times 121 times 161
3 times 121 times 161
5
5
5
FC
Weightedfusion
FC
FC
Figure 2 e structure of our framework
σ σ Tan h σ
Tan h
x x
xt
Ctminus1 C
t
htminus1 h
t
x +
ft
it
Ct
~ otForget gate
Input gate
Update stage
Output gate
ft = σ (W
fx
t + U
fh
tminus1 + bf)
ot = σ (W
ox
t + U
oh
tminus1 + bo)
it = σ (W
ix
t + U
ih
tminus1 + bi)
ht = o
t tan h(C
t)
= tan h (WCx
t + U
Ch
tminus1 + bC)C
t
~
Ct
~= f
tminus1Ctminus1 + it
Ct
Figure 1 Structure of LSTM
100deg
E
110deg
E
120deg
E
130deg
E
140deg
E
150deg
E
160deg
E
170deg
E
180deg
W
Sea s
urfa
ce te
mpe
ratu
re
0deg0
5
10
15
20
25
30
10degN
20degN
30degN
40degN
50degN
60degN
Figure 3 An example of SST
4 Complexity
SST we choose three different GHs under different hPaFigures 4(a)ndash4(c) show examples of GH which is alsorepresented by matrices with 121 rows and 161 col-umns It is evident from these charts that GH increaseswith latitudeSpecific humidity (SH) SH refers to the ratio of the massof water vapor in the atmosphere to the total mass of airere is a strong relationship between typhoons andvertical air motion and SH is usually used when dis-cussing the vertical motion erefore we introduced SHas a distinct climatic feature Figures 4(d)ndash4(f) show 3 SHdata charts under different hPa We can observe that SHin the south is higher than that in the north
412 Geographical Feature (1) Geostrophic force (GF) GFalso known as the Coriolis force was derived to describe theforce exerted on moving objects on the surface of the Earthas a result of the Earthrsquos rotation Owing to the existence ofGF a rotating flow of air is formed and eventually a ty-phoon is formed under the combined action of variousfactors e typhoon is also affected by GF during itsmovement In the northern hemisphere the GF of the ty-phoon is to the right which determines the typhoonrsquos di-rection of movement to a certain extent GF can be expressedas
F 2mvω sin θ (1)
where m is the mass of the object v is the velocity of theobject ω is the angular velocity of the Earthrsquos rotation and θis the latitude of the object before it begins to move Giventhat the mass of typhoons is difficult to estimate we use thegeostrophic force gradient to represent the influence of GFon typhoons denoted as
zF
zm 2vω sin θ (2)
413 Physical Features We use physical characteristics todescribe the time series and tracks of typhoons
Location and direction Given that the track of a ty-phoon is a series of coordinates we used the latitudesand longitudes (lat lon) or offsets (ΔlatΔlon) to de-scribe the location and direction of motion of typhoonsSince typhoon data were collected every 6 hours wecalculated the movement and direction of the typhoonevery 6 hoursSpeed e typhoon data are coarse erefore we usedthe average of the velocities of the typhoon at twoconsecutive moments to describe the moving velocityof the typhoonIntensity e intensity of a typhoon is determined byits wind speed Existing studies have validated therelationship between the central pressure of a typhoonand the maximum wind speed [26] erefore we usedthe maximum central pressure to express the intensitycharacteristics of a typhoon
42 Network Owing to the different modes of features weused different networks to process the features and then usedfeature fusion for learning e entire network architectureis illustrated in Figure 2
421 Feature Extraction We used climatic geographicaland physical features Some of these were two-dimensionalmatrices whereas some were one-dimensional vectorsConsequently we used different networks for differentfeatures
For climatic features all inputs were two-dimensionalimages We therefore used three ResNets to process theimages e ResNets employed in our framework have 18hidden layers [23] as shown in Figure 5 GH and SH havetrichannel inputs whereas SST has single-channel inputefirst layer is a convolution layer e size of the convolutionkernel is 7 times 7 and the stride is (2 2) Based on the size of theinput we set padding as (3 3) Batch normalization (BN)and rectified linear units (ReLU) were also used in theconvolution layer After the convolution operation thenetwork performs a maximum-pooling operation ere arefour residual blocks after the first layer Each residual blockis repeated twice To simplify the representation the re-peated parts have been replaced by ellipses Each residualblock contains two convolution layers Each layer contains aconvolution kernel batch normalization and ReLUe sizeof the convolution kernel is 3 times 3 the stride is (1 1) and thepadding is (1 1) e output dimensions of each residualblock are 64 128 256 and 512 After the last residual blockthe network performs an average-pooling operatione lastlayer of the network is a fully connected network with 5-dimensional output
As for the geographical and physical features we used afully connected network and obtained a 5-dimensionalvector as the output For feature fusion we adopted a weightmodule e weight of each feature can be regarded as thecorrelation between the feature and track of the typhoonrough weighted feature fusion for each moment weobtained a 20-dimensional feature vector which then be-came the input of the predictor
422 Multitask Prediction Because LSTM has a consider-able advantage in the processing of sequence data we usedthe classic LSTM as the predictor e dimension of theinput was t times 20 where t is the length of the sequence asintroduced in Section 2 e training process is shown inFigure 6 First we used zero-state initialization to calibratethe weight h0 and C0 For each cell of the LSTM the input isthe i-th 20-dimensional feature vector It should be notedthat all LSTM cells share these parameters
e LSTM output is divided into two tasks e maintask involves locating the typhoon at the next moment andthe auxiliary task involves determining the central pressureof the typhoon (ie the intensity of the typhoon) We usedthe L2 norm as the loss function of the two tasks
For the main task the loss is the difference in distancebetween the real location and the predicted location of thetyphoon as follows
Complexity 5
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
Multitask learning-based methods Chandra [19]proposed a coevolutionary multitask learning algo-rithm that combines the functions of modularizationand multitask learning is approach coordinatesmultitask learning dynamic programming and co-evolution algorithms Furthermore it can trainneural networks via feature sharing and modularknowledge representation It can also be used topredict typhoon intensity with limited input [20]is shows that compared with traditional modelsthe algorithm not only solves the problem of dynamictime series but also improves the prediction accuracyMukherjee and Mitra [21] proposed a joint learningmodel that can learn the distance and direction oftyphoons simultaneously via two different structureswith multiple LSTMs and multiple fully connectedlayers initial layer parameters are shared accordingto past typhoon track data e research results showthat the model can predict direction and distance(ie displacement) simultaneously
3 Preliminaries
In this section we first introduce the relevant technologiesutilized in our framework and then proceed to define ourproblem
31CNN andResNet CNN is a type of deep-learning modelthat has been successfully implemented in image recognition[22] e convolution layer is one of the core structures ofCNNs e input of the convolution layer includes one ormore matrices of the same size each of which is called achannel Each convolution layer uses common parametersknown as convolution kernels For 2D input the function ofthe convolution layer is to weigh the corresponding sub-matrices according to the size of kernels thus the convo-lution layer output is generated
Another important structure is the pooling layer whichaims to reduce the parameters of the model and strengthenthe network while improving the computing speed estrategies of the pooling layer include the maximum andaverage pooling
ResNet is a CNN model widely used for feature ex-traction [23] To solve the migration problem in deepnetworks ResNet proposes residual learning ResNet re-places the feature H(x) obtained by convolution layers withthe residual H(x) minus x of feature and input In contrast withordinary CNN ResNet adds a shortcut mechanism betweenevery two layers to realize residual learning
32 LSTM LSTM is a special type of RNN that is delib-erately designed to avoid long-term dependence It intro-duces a gate to solve gradient disappearance or explosion[24] LSTM contains four important structures namely theforget gate the input gate the update stage and the outputgate As shown in Figure 1 this framework operates asfollows
(1) e function of the forget gate is implemented bysigmoid to determine which information needs to beforgotten according to the input xt and the outputhtminus1 of the previous cell
(2) e input gate determines the information that willbe stored in the current cell
(3) e update stage updates Ct of the current cell(4) e output gate outputs the final information to the
next cell
33 Multitask Learning In single-task learning (involved inthe previous models) the model learns only one task at atime For complex problems single-task learning decom-poses the problems into multiple independent subproblemsfor separate training and then combines them However inpractical applications these subproblems frequently containcorrelation information that is often ignored by the single-task learning method
In this regard the goal of multitask learning is to in-tegrate multiple related tasks through shared representations[25] It entails hard and soft parameter sharing Hard pa-rameter sharing shares some parameters among all tasks andonly uses the tasksrsquo unique parameters at a specific layer Insoft parameter sharing each task has unique parametersFinally the similarity is expressed by adding constraints tothe differences between parameters of different tasks
34 Problem Definition e problem of typhoon trackprediction can be expressed in terms of the features of agiven typhoon at several past instances or moments the goalis to predict the locations at certain times or instances in thefuture e past-feature sequence of the typhoon is denotedas S (s1 s2 st) where si represents the features of thetyphoon at time i and t is the length of the sequence etrack of the typhoon at the future moment isT[(xt+1 yt+1) (xt+2 yt+2) (xt+n yt+n)] where (xj yj) isthe geographical coordinate (latitude and longitude) at timetj e goal of this study is to establish the mapping modelM T⟶ S and hence calculate the future trajectory se-quence T through the historical sequence S
4 Framework
Figure 2 illustrates the structure of our proposed frameworkIt entails feature selection weighted fusion and multitaskprediction In this section we will introduce all the partsindividually
41 Features ere are three types of features in ourframework namely climatic geographical and character-istic features e climatic features include sea surfacetemperature geopotential height and specific humidityGeographical features include geostrophic forces echaracteristic features are the speed and position of thetyphoon
Complexity 3
411 Climatic Features By studying the influence of climateon typhoons we selected three main factors as climaticfeatures in this study
Sea surface temperature (SST) SST is one of the mostimportant factors in meteorological research In gen-eral SST decreases when latitude increases SST plays apivotal role in the formation and movement of ty-phoons typhoons are formed above the sea surfacewhere SST is higher than 265degC and the intensity of thetyphoon increases through continuous absorption ofenergy SST is also one of the main factors influencingthe direction of motion and landing location of ty-phoons In this study we mainly considered the regionwithin 0degN and 60degN latitudes and 100degE and 180degElongitudes e SST in this area was regularly collectedby the sensor As shown in Figure 3 we used a matrix of121 rows and 161 columns to represent SST in whichthe SST near the equator is above 30degC whereas theSST at higher latitudes is approximately 0degC We alsodistinguished land from sea the darkest shades inFigure 3 are landGeopotential height (GH) GH is an imaginary heightin meteorology expressed in terms of the work doneagainst gravity by an object of unit mass rising from sea
level to a certain height GH also plays an importantrole in maintaining the intensity and motion of ty-phoons For example the large geopotential heightgradient between the Western Pacific subtropical highand typhoon determines the direction of movement oftyphoon Ambi to a certain extent [13] We studied GHin the same region described previously In contrast to
ResNet
ResNet
ResNet
SST
GH 200 500 850 (hPa)
SH 200 500 850 (hPa)
Geographical and physical features
LSTM
LongitudeLatitude
Intensity
121 times 161 5
3 times 121 times 161
3 times 121 times 161
5
5
5
FC
Weightedfusion
FC
FC
Figure 2 e structure of our framework
σ σ Tan h σ
Tan h
x x
xt
Ctminus1 C
t
htminus1 h
t
x +
ft
it
Ct
~ otForget gate
Input gate
Update stage
Output gate
ft = σ (W
fx
t + U
fh
tminus1 + bf)
ot = σ (W
ox
t + U
oh
tminus1 + bo)
it = σ (W
ix
t + U
ih
tminus1 + bi)
ht = o
t tan h(C
t)
= tan h (WCx
t + U
Ch
tminus1 + bC)C
t
~
Ct
~= f
tminus1Ctminus1 + it
Ct
Figure 1 Structure of LSTM
100deg
E
110deg
E
120deg
E
130deg
E
140deg
E
150deg
E
160deg
E
170deg
E
180deg
W
Sea s
urfa
ce te
mpe
ratu
re
0deg0
5
10
15
20
25
30
10degN
20degN
30degN
40degN
50degN
60degN
Figure 3 An example of SST
4 Complexity
SST we choose three different GHs under different hPaFigures 4(a)ndash4(c) show examples of GH which is alsorepresented by matrices with 121 rows and 161 col-umns It is evident from these charts that GH increaseswith latitudeSpecific humidity (SH) SH refers to the ratio of the massof water vapor in the atmosphere to the total mass of airere is a strong relationship between typhoons andvertical air motion and SH is usually used when dis-cussing the vertical motion erefore we introduced SHas a distinct climatic feature Figures 4(d)ndash4(f) show 3 SHdata charts under different hPa We can observe that SHin the south is higher than that in the north
412 Geographical Feature (1) Geostrophic force (GF) GFalso known as the Coriolis force was derived to describe theforce exerted on moving objects on the surface of the Earthas a result of the Earthrsquos rotation Owing to the existence ofGF a rotating flow of air is formed and eventually a ty-phoon is formed under the combined action of variousfactors e typhoon is also affected by GF during itsmovement In the northern hemisphere the GF of the ty-phoon is to the right which determines the typhoonrsquos di-rection of movement to a certain extent GF can be expressedas
F 2mvω sin θ (1)
where m is the mass of the object v is the velocity of theobject ω is the angular velocity of the Earthrsquos rotation and θis the latitude of the object before it begins to move Giventhat the mass of typhoons is difficult to estimate we use thegeostrophic force gradient to represent the influence of GFon typhoons denoted as
zF
zm 2vω sin θ (2)
413 Physical Features We use physical characteristics todescribe the time series and tracks of typhoons
Location and direction Given that the track of a ty-phoon is a series of coordinates we used the latitudesand longitudes (lat lon) or offsets (ΔlatΔlon) to de-scribe the location and direction of motion of typhoonsSince typhoon data were collected every 6 hours wecalculated the movement and direction of the typhoonevery 6 hoursSpeed e typhoon data are coarse erefore we usedthe average of the velocities of the typhoon at twoconsecutive moments to describe the moving velocityof the typhoonIntensity e intensity of a typhoon is determined byits wind speed Existing studies have validated therelationship between the central pressure of a typhoonand the maximum wind speed [26] erefore we usedthe maximum central pressure to express the intensitycharacteristics of a typhoon
42 Network Owing to the different modes of features weused different networks to process the features and then usedfeature fusion for learning e entire network architectureis illustrated in Figure 2
421 Feature Extraction We used climatic geographicaland physical features Some of these were two-dimensionalmatrices whereas some were one-dimensional vectorsConsequently we used different networks for differentfeatures
For climatic features all inputs were two-dimensionalimages We therefore used three ResNets to process theimages e ResNets employed in our framework have 18hidden layers [23] as shown in Figure 5 GH and SH havetrichannel inputs whereas SST has single-channel inputefirst layer is a convolution layer e size of the convolutionkernel is 7 times 7 and the stride is (2 2) Based on the size of theinput we set padding as (3 3) Batch normalization (BN)and rectified linear units (ReLU) were also used in theconvolution layer After the convolution operation thenetwork performs a maximum-pooling operation ere arefour residual blocks after the first layer Each residual blockis repeated twice To simplify the representation the re-peated parts have been replaced by ellipses Each residualblock contains two convolution layers Each layer contains aconvolution kernel batch normalization and ReLUe sizeof the convolution kernel is 3 times 3 the stride is (1 1) and thepadding is (1 1) e output dimensions of each residualblock are 64 128 256 and 512 After the last residual blockthe network performs an average-pooling operatione lastlayer of the network is a fully connected network with 5-dimensional output
As for the geographical and physical features we used afully connected network and obtained a 5-dimensionalvector as the output For feature fusion we adopted a weightmodule e weight of each feature can be regarded as thecorrelation between the feature and track of the typhoonrough weighted feature fusion for each moment weobtained a 20-dimensional feature vector which then be-came the input of the predictor
422 Multitask Prediction Because LSTM has a consider-able advantage in the processing of sequence data we usedthe classic LSTM as the predictor e dimension of theinput was t times 20 where t is the length of the sequence asintroduced in Section 2 e training process is shown inFigure 6 First we used zero-state initialization to calibratethe weight h0 and C0 For each cell of the LSTM the input isthe i-th 20-dimensional feature vector It should be notedthat all LSTM cells share these parameters
e LSTM output is divided into two tasks e maintask involves locating the typhoon at the next moment andthe auxiliary task involves determining the central pressureof the typhoon (ie the intensity of the typhoon) We usedthe L2 norm as the loss function of the two tasks
For the main task the loss is the difference in distancebetween the real location and the predicted location of thetyphoon as follows
Complexity 5
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
411 Climatic Features By studying the influence of climateon typhoons we selected three main factors as climaticfeatures in this study
Sea surface temperature (SST) SST is one of the mostimportant factors in meteorological research In gen-eral SST decreases when latitude increases SST plays apivotal role in the formation and movement of ty-phoons typhoons are formed above the sea surfacewhere SST is higher than 265degC and the intensity of thetyphoon increases through continuous absorption ofenergy SST is also one of the main factors influencingthe direction of motion and landing location of ty-phoons In this study we mainly considered the regionwithin 0degN and 60degN latitudes and 100degE and 180degElongitudes e SST in this area was regularly collectedby the sensor As shown in Figure 3 we used a matrix of121 rows and 161 columns to represent SST in whichthe SST near the equator is above 30degC whereas theSST at higher latitudes is approximately 0degC We alsodistinguished land from sea the darkest shades inFigure 3 are landGeopotential height (GH) GH is an imaginary heightin meteorology expressed in terms of the work doneagainst gravity by an object of unit mass rising from sea
level to a certain height GH also plays an importantrole in maintaining the intensity and motion of ty-phoons For example the large geopotential heightgradient between the Western Pacific subtropical highand typhoon determines the direction of movement oftyphoon Ambi to a certain extent [13] We studied GHin the same region described previously In contrast to
ResNet
ResNet
ResNet
SST
GH 200 500 850 (hPa)
SH 200 500 850 (hPa)
Geographical and physical features
LSTM
LongitudeLatitude
Intensity
121 times 161 5
3 times 121 times 161
3 times 121 times 161
5
5
5
FC
Weightedfusion
FC
FC
Figure 2 e structure of our framework
σ σ Tan h σ
Tan h
x x
xt
Ctminus1 C
t
htminus1 h
t
x +
ft
it
Ct
~ otForget gate
Input gate
Update stage
Output gate
ft = σ (W
fx
t + U
fh
tminus1 + bf)
ot = σ (W
ox
t + U
oh
tminus1 + bo)
it = σ (W
ix
t + U
ih
tminus1 + bi)
ht = o
t tan h(C
t)
= tan h (WCx
t + U
Ch
tminus1 + bC)C
t
~
Ct
~= f
tminus1Ctminus1 + it
Ct
Figure 1 Structure of LSTM
100deg
E
110deg
E
120deg
E
130deg
E
140deg
E
150deg
E
160deg
E
170deg
E
180deg
W
Sea s
urfa
ce te
mpe
ratu
re
0deg0
5
10
15
20
25
30
10degN
20degN
30degN
40degN
50degN
60degN
Figure 3 An example of SST
4 Complexity
SST we choose three different GHs under different hPaFigures 4(a)ndash4(c) show examples of GH which is alsorepresented by matrices with 121 rows and 161 col-umns It is evident from these charts that GH increaseswith latitudeSpecific humidity (SH) SH refers to the ratio of the massof water vapor in the atmosphere to the total mass of airere is a strong relationship between typhoons andvertical air motion and SH is usually used when dis-cussing the vertical motion erefore we introduced SHas a distinct climatic feature Figures 4(d)ndash4(f) show 3 SHdata charts under different hPa We can observe that SHin the south is higher than that in the north
412 Geographical Feature (1) Geostrophic force (GF) GFalso known as the Coriolis force was derived to describe theforce exerted on moving objects on the surface of the Earthas a result of the Earthrsquos rotation Owing to the existence ofGF a rotating flow of air is formed and eventually a ty-phoon is formed under the combined action of variousfactors e typhoon is also affected by GF during itsmovement In the northern hemisphere the GF of the ty-phoon is to the right which determines the typhoonrsquos di-rection of movement to a certain extent GF can be expressedas
F 2mvω sin θ (1)
where m is the mass of the object v is the velocity of theobject ω is the angular velocity of the Earthrsquos rotation and θis the latitude of the object before it begins to move Giventhat the mass of typhoons is difficult to estimate we use thegeostrophic force gradient to represent the influence of GFon typhoons denoted as
zF
zm 2vω sin θ (2)
413 Physical Features We use physical characteristics todescribe the time series and tracks of typhoons
Location and direction Given that the track of a ty-phoon is a series of coordinates we used the latitudesand longitudes (lat lon) or offsets (ΔlatΔlon) to de-scribe the location and direction of motion of typhoonsSince typhoon data were collected every 6 hours wecalculated the movement and direction of the typhoonevery 6 hoursSpeed e typhoon data are coarse erefore we usedthe average of the velocities of the typhoon at twoconsecutive moments to describe the moving velocityof the typhoonIntensity e intensity of a typhoon is determined byits wind speed Existing studies have validated therelationship between the central pressure of a typhoonand the maximum wind speed [26] erefore we usedthe maximum central pressure to express the intensitycharacteristics of a typhoon
42 Network Owing to the different modes of features weused different networks to process the features and then usedfeature fusion for learning e entire network architectureis illustrated in Figure 2
421 Feature Extraction We used climatic geographicaland physical features Some of these were two-dimensionalmatrices whereas some were one-dimensional vectorsConsequently we used different networks for differentfeatures
For climatic features all inputs were two-dimensionalimages We therefore used three ResNets to process theimages e ResNets employed in our framework have 18hidden layers [23] as shown in Figure 5 GH and SH havetrichannel inputs whereas SST has single-channel inputefirst layer is a convolution layer e size of the convolutionkernel is 7 times 7 and the stride is (2 2) Based on the size of theinput we set padding as (3 3) Batch normalization (BN)and rectified linear units (ReLU) were also used in theconvolution layer After the convolution operation thenetwork performs a maximum-pooling operation ere arefour residual blocks after the first layer Each residual blockis repeated twice To simplify the representation the re-peated parts have been replaced by ellipses Each residualblock contains two convolution layers Each layer contains aconvolution kernel batch normalization and ReLUe sizeof the convolution kernel is 3 times 3 the stride is (1 1) and thepadding is (1 1) e output dimensions of each residualblock are 64 128 256 and 512 After the last residual blockthe network performs an average-pooling operatione lastlayer of the network is a fully connected network with 5-dimensional output
As for the geographical and physical features we used afully connected network and obtained a 5-dimensionalvector as the output For feature fusion we adopted a weightmodule e weight of each feature can be regarded as thecorrelation between the feature and track of the typhoonrough weighted feature fusion for each moment weobtained a 20-dimensional feature vector which then be-came the input of the predictor
422 Multitask Prediction Because LSTM has a consider-able advantage in the processing of sequence data we usedthe classic LSTM as the predictor e dimension of theinput was t times 20 where t is the length of the sequence asintroduced in Section 2 e training process is shown inFigure 6 First we used zero-state initialization to calibratethe weight h0 and C0 For each cell of the LSTM the input isthe i-th 20-dimensional feature vector It should be notedthat all LSTM cells share these parameters
e LSTM output is divided into two tasks e maintask involves locating the typhoon at the next moment andthe auxiliary task involves determining the central pressureof the typhoon (ie the intensity of the typhoon) We usedthe L2 norm as the loss function of the two tasks
For the main task the loss is the difference in distancebetween the real location and the predicted location of thetyphoon as follows
Complexity 5
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
SST we choose three different GHs under different hPaFigures 4(a)ndash4(c) show examples of GH which is alsorepresented by matrices with 121 rows and 161 col-umns It is evident from these charts that GH increaseswith latitudeSpecific humidity (SH) SH refers to the ratio of the massof water vapor in the atmosphere to the total mass of airere is a strong relationship between typhoons andvertical air motion and SH is usually used when dis-cussing the vertical motion erefore we introduced SHas a distinct climatic feature Figures 4(d)ndash4(f) show 3 SHdata charts under different hPa We can observe that SHin the south is higher than that in the north
412 Geographical Feature (1) Geostrophic force (GF) GFalso known as the Coriolis force was derived to describe theforce exerted on moving objects on the surface of the Earthas a result of the Earthrsquos rotation Owing to the existence ofGF a rotating flow of air is formed and eventually a ty-phoon is formed under the combined action of variousfactors e typhoon is also affected by GF during itsmovement In the northern hemisphere the GF of the ty-phoon is to the right which determines the typhoonrsquos di-rection of movement to a certain extent GF can be expressedas
F 2mvω sin θ (1)
where m is the mass of the object v is the velocity of theobject ω is the angular velocity of the Earthrsquos rotation and θis the latitude of the object before it begins to move Giventhat the mass of typhoons is difficult to estimate we use thegeostrophic force gradient to represent the influence of GFon typhoons denoted as
zF
zm 2vω sin θ (2)
413 Physical Features We use physical characteristics todescribe the time series and tracks of typhoons
Location and direction Given that the track of a ty-phoon is a series of coordinates we used the latitudesand longitudes (lat lon) or offsets (ΔlatΔlon) to de-scribe the location and direction of motion of typhoonsSince typhoon data were collected every 6 hours wecalculated the movement and direction of the typhoonevery 6 hoursSpeed e typhoon data are coarse erefore we usedthe average of the velocities of the typhoon at twoconsecutive moments to describe the moving velocityof the typhoonIntensity e intensity of a typhoon is determined byits wind speed Existing studies have validated therelationship between the central pressure of a typhoonand the maximum wind speed [26] erefore we usedthe maximum central pressure to express the intensitycharacteristics of a typhoon
42 Network Owing to the different modes of features weused different networks to process the features and then usedfeature fusion for learning e entire network architectureis illustrated in Figure 2
421 Feature Extraction We used climatic geographicaland physical features Some of these were two-dimensionalmatrices whereas some were one-dimensional vectorsConsequently we used different networks for differentfeatures
For climatic features all inputs were two-dimensionalimages We therefore used three ResNets to process theimages e ResNets employed in our framework have 18hidden layers [23] as shown in Figure 5 GH and SH havetrichannel inputs whereas SST has single-channel inputefirst layer is a convolution layer e size of the convolutionkernel is 7 times 7 and the stride is (2 2) Based on the size of theinput we set padding as (3 3) Batch normalization (BN)and rectified linear units (ReLU) were also used in theconvolution layer After the convolution operation thenetwork performs a maximum-pooling operation ere arefour residual blocks after the first layer Each residual blockis repeated twice To simplify the representation the re-peated parts have been replaced by ellipses Each residualblock contains two convolution layers Each layer contains aconvolution kernel batch normalization and ReLUe sizeof the convolution kernel is 3 times 3 the stride is (1 1) and thepadding is (1 1) e output dimensions of each residualblock are 64 128 256 and 512 After the last residual blockthe network performs an average-pooling operatione lastlayer of the network is a fully connected network with 5-dimensional output
As for the geographical and physical features we used afully connected network and obtained a 5-dimensionalvector as the output For feature fusion we adopted a weightmodule e weight of each feature can be regarded as thecorrelation between the feature and track of the typhoonrough weighted feature fusion for each moment weobtained a 20-dimensional feature vector which then be-came the input of the predictor
422 Multitask Prediction Because LSTM has a consider-able advantage in the processing of sequence data we usedthe classic LSTM as the predictor e dimension of theinput was t times 20 where t is the length of the sequence asintroduced in Section 2 e training process is shown inFigure 6 First we used zero-state initialization to calibratethe weight h0 and C0 For each cell of the LSTM the input isthe i-th 20-dimensional feature vector It should be notedthat all LSTM cells share these parameters
e LSTM output is divided into two tasks e maintask involves locating the typhoon at the next moment andthe auxiliary task involves determining the central pressureof the typhoon (ie the intensity of the typhoon) We usedthe L2 norm as the loss function of the two tasks
For the main task the loss is the difference in distancebetween the real location and the predicted location of thetyphoon as follows
Complexity 5
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
Lmain
(x minus 1113954x)2
+(y minus 1113954y)2
1113969
(3)
where (x y) is the location of the typhoon at the nextmoment and (1113954x 1113954y) is the output of the predictor elongitude and latitude offset can also be used as input andthe corresponding loss will become the difference in theoffset For the auxiliary task the loss function is denoted as
Lauxiliary
(p minus 1113954p)2
1113969
(4)
where p is the central pressure of the typhoon and 1113954p is theprediction result
erefore the total loss of our framework is as follows
Ltotal αLmain +(1 minus α)Lauxiliary (5)
In this loss function α is a hyperparameter
43 Distributed Implementation To ensure that the pro-posed framework can handle big data and consequentlyimprove the efficiency of training we implemented a dis-tributed framework based on Ray Ray is a very populardistributed AI platform implemented via Python is fa-cilitates the rapidly distributed computing of the Python
Avg pooling
FC 5
3 times 3 conv 64
BNBN
ReLu
3 times 3 conv 128
BN
ReLu
BNBN
ReLu
3 times 3 conv 256
BN
ReLu
3 times 3 conv 256
BN
ReLu
3x3 conv 512
ReLu
3 times 3 conv 512
BN
ReLu
3 times 3 conv 512
BN
ReLuMax pooling
Input3 times 121 times 161
Block 1 Block 2 Block 3 Block 4
BN
7 times 7 conv 64 2padding = (3 3)
BN
ReLu
Shortcut Shortcut Shortcut Shortcut
3 times 3 conv 64 3 times 3 conv 128
hellip hellip hellip hellip
BN
ReLu
Figure 5 Details of ResNet in our framework
0deg
10degN
20degN
30degN
40degN
50degN
60degN
11600
11800
12000
12200
12400
Geo
pote
ntia
l hei
ght (
200
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(a)
0deg
10degN
20degN
30degN
40degN
50degN
60degN
5400
5500
5600
5700
5800
Geo
pote
ntia
l hei
ght (
500
hPa)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(b)
0deg1350
Geo
pote
ntia
l hei
ght (
850
hPa)
1400
1450
1500
1550
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(c)
0deg
10degN
20degN
30degN
40degN
50degN
60degNSp
ecifi
c hum
idity
(200
hPa)
00000
00002
00004
00006
00008
00010
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
(d)
0deg 00000 Spec
ific h
umid
ity (5
00 h
Pa)
00005000100001500020000250003000035
10degN
20degN
30degN
40degN
50degN
60degN
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W(e)
100deg
E11
0degE
120deg
E13
0degE
140deg
E15
0degE
160deg
E17
0degE
180deg
W
0000 Spec
ific h
umid
ity (8
50hP
a)
000200040006000800100012
0deg
10degN
20degN
30degN
40degN
50degN
60degN
(f )
Figure 4 Example of GH and SH (a) GH at 200 hPa (b) GH at 500 hPa (c) GH at 850 hPa (d) SH at 200 hPa (e) SH at 500 hPa (f ) SH at850 hPa
6 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
code In the implementation each network structure (suchas convolution layer pooling layer and FC layer) isimplemented as a class also known as an actor in RayMultiple actors construct the entire network through thedata flow In the calculation process each calculation nodestarts multiple workers as the basis of calculation Each actoris assigned to the corresponding worker for execution In thetraining process the data flows through gRPC and sharedmemory to the corresponding worker for calculation Forexample in each ResNet after the calculation of the currentlayer is completed the data will flow to the worker of thenext layer ere is no data dependence between multipleResNets therefore parallel training can be realized
5 Experiments
51 Setup We use a real dataset to verify the effectiveness ofour framework e dataset is the Western Pacific Typhoontrack data from the JTWC (be TyphoonWarning Center theJoint Typhoon Warning Center) e dataset contains ty-phoon tracks from January 1 2001 to December 31 2005e attitude is from 0degN to 60degN and the longitude is from100degE to 180degE Statistics of the experimental setup areshown in Table 1
We use the metric of distance error (same as Lmain) toverify the effectiveness of our framework We first verify thebenefits of multitask learning technology to this frameworkNext we use different weights to discuss the relationshipbetween features and resultse framework is implementedby Python 3 and the experiments are conducted on a clusterin which each node has Intel Purley 4110 CPUs and TeslaP100 GPUs
52 Results In this section we will introduce the experi-mental results in the real-life dataset We report and analysethe results by changing the parameters en we choosesome real typhoon tracks to show our prediction results
Distance error with respect to multitask and single-tasklearning firstly we compare the results of multitasklearning (MTL) and single-task learning (STL) asshown in Figure 7 We can obverse that MTL can getbetter results than STL in most cases In the 6 h pre-diction results MTL is similar to STL However inother cases MTL can achieve about 20 performanceimprovement It proves that it is feasible to improve the
effect of track prediction by auxiliary tasks What ismore the best results in 6 h 24 h 48 h and 72 h areabout 40 km 70 km 220 km and 380 km which arebetter than most existing models It also proves theeffectiveness of our frameworkDistance error with respect to |T||T| secondly wereport the distance error with different size of input |T|e results are also shown in Figure 7 We find that |T|
has a great influence on our framework in differentcases e optimal value is 3 7 4 and 5 in 6 h 24 h48 h and 72 h As |T| becomes larger or smaller thedistance error gradually increases In the later experi-ments we selected the best value of |T| in each case toverify the effect of feature weight on the distance error
en we study the relationship between features andprediction results
Distance error with respect to wSSTwSST to study theeffect of SST we keep wGH and wSH unchanged andthen adjust the value of wSST from 01 to 10 e resultsare shown in Figure 8 We can obverse that SST willgreatly affect the results e best choice is to reducewSST as small as possibleDistance error with respect to wGHwGH to study therelationship between GH and prediction results wekeep wSST and wSH unchanged and then adjust thevalue of wGH from 01 to 10 As shown in Figure 9 wecan get the best results when wGH is set as 08 edifference between the best result and the worst resultin 6 h 24 h and 48 h is about 30 km to 100 km In 72 hthe difference could be more than 300 km An ap-propriate wGH can improve the results by 30 to 40e experimental results show that there is a strongcorrelation between GH and prediction resultsDistance error with respect to wSHwSH to study therelationship between SH and prediction results wekeep wSST and wGH unchanged and then adjust thevalue of wSH from 01 to 10 e results are shown inFigure 10 To get better resultswSH is smaller thanwGHIn 6 h and 24 h cases we can get the best results whenwSH is set as 01 In 48 h and 72 h cases it is better to setwSH as 03 An appropriate wSH can improve the resultby 40 to 50 e experimental results show that SHis also related to the prediction results but the cor-relation is less than GH
Zero
-sta
te in
itial
izat
ion
LSTM cell LSTM cell LSTM cellhellip
Task 1
Task 2
hellip
f1 f2 ft
hellip hellip
hellip
helliphellip
Figure 6 e details of multitask prediction
Complexity 7
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
Table 1 Statistics of the experimental setup
Region Date range Dimension of featuresAttitude Longitude January 1 2001 to December 31 2005 SST GHSH Others0degN to 60degN 100degE to 180degE 121 times 161 3 times 121 times 161 20 times 1
40
80
120
160
200
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
6 h STL6 h MTL
24 h STL24 h MTL
(a)
200
300
400
500
600
700
800
900
2 3 4 5 6 7 8
Dist
ance
erro
r (km
)
|T|
48 h STL48 h MTL
72 h STL72 h MTL
(b)
Figure 7 Results of varying |T| (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
250
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSST
(a)
0
200
400
600
800
1000
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSST
48h prediction72h prediction
(b)
Figure 8 Results of varying weight of wSST (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
8 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
Case study We use some real typhoons to compare thereal tracks and the prediction results We select SaolaDamrey and Longwang that are formed in 2005 thereal tracks and 6 h prediction results are shown inFigures 11ndash13 Typhoon Saola was formed on Sep-tember 20th the average distance error of 6 h
prediction results is 4033 km Typhoon Damrey wasformed on September 21 the average distance error is4059 km the minimum error is 89 km and themaximum error is 6033 km Typhoon Longwang wasformed on September 26 the average distance error is4651 km
20
40
60
80
100
120
140
160
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WGH
(a)
100
200
300
400
500
600
700
800
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WGH
48h prediction72h prediction
(b)
Figure 9 Results of varying weight of wGH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
0
50
100
150
200
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
6h prediction24h prediction
WSH
(a)
150
250
350
450
550
650
01 02 03 04 05 06 07 08 09 1
Dist
ance
erro
r (km
)
WSH
48h prediction72h prediction
(b)
Figure 10 Results of varying weight of wSH (a) Results of 6 h and 24 h (b) Results of 48 and 72 h
Complexity 9
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
Comparison with existing works Finally we compareour framework with several existing works [8101227]According to the previous introduction Ruttgers et al[8] introduced a GAN-based model used satelliteimages as the input and predicted locations after 6hours Gao et al [10] introduced an LSTM-basedmodel e work by Giffard-Roisin et al [12] was based
on CNN and feature fusion Lv et al [27] used the leastsquare method and FC network to predict the locationsWe still use distance error to verify the effectiveness andthe results are shown in Table 2 Compared with theseworks our framework can achieve high predictionresults especially in 48 h and 72 h cases In 72 h resultsour framework improves the accuracy by 60
20
24
28
32
36
40
136 140 144 148 152La
titud
eLongitude
Real track6 h prediction
Figure 11 6 h prediction results of Saola
16
17
18
19
20
21
112 114 116 118 120 122 124
Latit
ude
Longitude
Real track6h prediction
Figure 12 6 h prediction results of Damrey
19
20
21
22
23
24
25
26
115 120 125 130 135 140 145
Latit
ude
Longitude
Real track6h prediction
Figure 13 6 h prediction results of Longwang
10 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
53 Summary In this section we verify the effect of differentparameters on the performance of our framework in the realdataset In general our framework can achieve good resultsbased on multitask and feature weighting We find that GHhas a strong correlation with the movement of typhoonsfollowed by SH and SST has the weakest correlationrough the training results the optimal prediction resultscan be obtained by selecting the appropriate parameters fordifferent scenes
6 Conclusion
In this paper we proposed a typhoon track predictionframework based on multitask learning and featureweightingWe analysed the correlation between the climaticgeographical and physical features and typhoon movementthrough the method of feature weighting We designed anetwork based on ResNet and LSTM and used a multitasklearning method to improve the prediction accuracy Weimplemented the network in a distributed platform Finallywe conducted experiments on real datasets to prove theeffectiveness of the framework In future works we willanalyse more features and use the attention mechanism toautomatically process the weight of features
Data Availability
e data are available from the corresponding author uponrequest
Conflicts of Interest
e authors declare that they have no conflicts of interest tothis work
Acknowledgments
e work was supported by the National Key RampD Programof China (Grant no 2016YFC1401902) the National NaturalScience Foundation of China (Grant no 61972077) and theLiaoNing Revitalization Talents Program (Grant noXLYC2007079)
References
[1] W Liu K Fujii Y Maruyama and F Yamazaki ldquoInundationassessment of the 2019 typhoon hagibis in Japan using multi-temporal sentinel-1 intensity imagesrdquo Remote Sensing vol 13no 4 p 639 2021
[2] J Cai Y Zhang R J Doviak Y Shrestha and P W ChanldquoDiagnosis and classification of typhoon-associated low-al-titude turbulence using HKO-TDWR radar observations and
machine learningrdquo IEEE Transactions on Geoscience andRemote Sensing vol 57 no 6 pp 3633ndash3648 2019
[3] J Li Q Zheng M Li Q Li and L Xie ldquoSpatiotemporaldistributions of ocean color elements in response to tropicalcyclone a case study of typhoon mangkhut (2018) past overthe northern south China seardquo Remote Sensing vol 13 no 4p 687 2021
[4] M Demaria MMainelli L K Shay J A Knaff and J KaplanldquoFurther improvements to the statistical hurricane intensityprediction scheme (SHIPS)rdquo Weather and Forecastingvol 20 no 4 pp 531ndash543 2005
[5] J S Goerss ldquoTropical cyclone track forecasts using an en-semble of dynamical modelsrdquo Monthly Weather Reviewvol 128 no 4 pp 1187ndash1193 2000
[6] T N Krishnamurti C M Kishtawal Z Zhang et alldquoMultimodel ensemble forecasts for weather and seasonalclimaterdquo Journal of Climate vol 13 no 23 pp 4196ndash42162000
[7] H C Weber ldquoHurricane track prediction using a statisticalensemble of numerical modelsrdquo Monthly Weather Reviewvol 131 no 5 pp 749ndash770 2003
[8] M Ruttgers S Lee S Jeon and D You ldquoPrediction of atyphoon track using a generative adversarial network andsatellite imagesrdquo Scientific Reports vol 9 no 1pp 6057ndash6115 2019
[9] C Wang Q Xu X Li et al ldquoCNN-based tropical cyclonetrack forecasting from satellite infrared imagesrdquo in Pro-ceedings of the IEEE International Geoscience and RemoteSensing Symposium pp 5811ndash5814 Waikoloa HI USASeptember 2020
[10] S Gao P Zhao B Pan et al ldquoA nowcasting model for theprediction of typhoon tracks based on a long short termmemory neural networkrdquo Acta Oceanologica Sinica vol 37no 5 pp 8ndash12 2018
[11] J Chen M Zhong J Li D Wang T Qian and H TuldquoEffective deep attributed network representation learningwith topology adapted smoothingrdquo IEEE Transactions onCybernetics 2021
[12] S Giffard-Roisin M Yang G Charpiat C Kumler BonfantiB Kegl and C Monteleoni ldquoTropical cyclone track fore-casting using fused deep learning from aligned reanalysisdatardquo Frontiers in Big Data vol 3 p 1 2020
[13] M Moradi Kordmahalleh M Gorji Sefidmazgi andA Homaifar ldquoA sparse recurrent neural network for tra-jectory prediction of atlantic hurricanesrdquo in Proceedings of theGenetic and Evolutionary Computation Conference pp 957ndash964 Lille France July 2016
[14] S Alemany J Beltran A Perez et al ldquoPredicting hurricanetrajectories using a recurrent neural networkrdquo in Proceedingsof the irty-ird AAAI Conference on Artificial Intelligencepp 468ndash475 Honolulu HI USA January 2019
[15] R Chandra and K Dayal ldquoCooperative neuro-evolution ofElman recurrent networks for tropical cyclone wind-intensityprediction in the south pacific regionrdquo in Proceedings of the
Table 2 Results compared with the existing works
6 h 24 h 48 h 72 hOur framework 3875 6954 19661 3681Gao et al [10] 4595 10568 33254 97450Giffard-Roisin et al [12] mdash 1361 mdash mdashRuttgers et al [8] 956 mdash mdash mdashLv et al [27] mdash 15834 36176 mdash
Complexity 11
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity
IEEE Congress on Evolutionary Computation (CEC)pp 1784ndash1791 Sendai Japan May 2015
[16] R Chandra K Dayal and N Rollings ldquoApplication of co-operative neuro-evolution of Elman recurrent networks for atwo-dimensional cyclone track prediction for the South Pa-cific regionrdquo in Proceedings of the International Joint Con-ference on Neural Networks (IJCNN) pp 1ndash8 KillarneyIreland July 2015
[17] J Lian P Dong Y Zhang J Pan and K Liu ldquoA novel data-driven tropical cyclone track prediction model based on CNNand GRU with multi-dimensional feature selectionrdquo IEEEAccess vol 8 pp 97114ndash97128 2020
[18] S Kim H Kim J Lee et al ldquoDeep-hurricane-tracker trackingand forecasting extreme climate eventsrdquo in Proceedings of theWinter Conference on Applications of Computer Vision(WACV) pp 1761ndash1769 Waikoloa HI USA January 2019
[19] R Chandra ldquoDynamic cyclone wind-intensity predictionusing co-evolutionary multi-task learningrdquo in Proceedings ofthe International Conference on Neural Information Process-ing pp 618ndash627 Guangzhou China November 2017
[20] R Chandra Y-S Ong and C-K Goh ldquoCo-evolutionarymulti-task learning for dynamic time series predictionrdquoApplied Soft Computing vol 70 pp 576ndash589 2018
[21] A Mukherjee and P Mitra ldquoJoint learning for cyclone tracknowcastingrdquo in Proceedings of the ECMLPKDD CEURWorkshop Ghent Belgium September 2020
[22] A Krizhevsky I Sutskever and G E Hinton ldquoImagenetclassification with deep convolutional neural networksrdquoAdvances in Neural Information Processing Systems vol 25pp 1097ndash1105 2012
[23] K He X Zhang S Ren et al ldquoDeep residual learning forimage recognitionrdquo in Proceedings of the IEEE Conference onComputer Vision and Pattern Recognition (CVPR) pp 770ndash778 Las Vegas NV USA June 2016
[24] S Hochreiter and J Schmidhuber ldquoLong short-term mem-oryrdquo Neural Computation vol 9 no 8 pp 1735ndash1780 1997
[25] K-H ung and C-Y Wee ldquoA brief review on multi-tasklearningrdquo Multimedia Tools and Applications vol 77 no 22pp 29705ndash29725 2018
[26] K Chen ldquoCalculation of the maximum wind speed of ty-phoon in the western pacificrdquo Marine Science Bulletin 1985
[27] Q P Lv J Luo K Zhu et al ldquoExperiments on predictingtracks of tropical cyclones based on artificial neural networkrdquoGuangdong Meteorology pp 19ndash22 2009
12 Complexity