Identification of Plants using Deep learning: AReviewRakibul Sk, Ankita Wadhawan

Lovely Professional University (LPU University), Jalandhar-Delhi, G.T. Road, Phagwara,144411 Punjab

AbstractIdentification of plants is a very important field in the earth’s ecology to maintain a healthy atmosphere.Certain of these plants have significant medicinal properties. Nowadays of finding a plant is not easy bylooking at its physical properties. This paper provides an academic database of literature between theduration of 2015–2020. It has been observed that the new generation of convolutionary neural networks(CNNs) in the space area of image recognition has produced remarkable performance. In this paper,techniques are discussed the concepts of Deep learning and different leaf recognition methods.

KeywordsMachine Learning, Artificial Intelligence, Fully Connected Neurons, Convolutional Neural Network,Deep Learning, Image Processing.

1. Introduction

ISIC’21: International Semantic Intelligence Conference,February 25–27, 2021, New Delhi, India" rakibulsk.cse@gmail.com (R. Sk);ankita.23891@lpu.co.in (A. Wadhawan)

© 2021 Copyright for this paper by its authors. Use permit-ted under Creative Commons License Attribution 4.0 Inter-national (CC BY 4.0).

CEURWorkshopProceedings

http://ceur-ws.orgISSN 1613-0073

CEUR Workshop Proceedings(CEUR-WS.org)

of results; it only includes identification of thedata format and unlabelled inputs.

Deep learning offers superior results evenon big data. Deep learning, which is used inimage identification and computer vision etc,utilizes artificial neurons identical to the neu-rons of man. Deep neural networks, recurrentneural networks, and deep belief networksare used to speak identification, language pro-cessing, translation software, audio recogni-tion, bioinformatics, and drug development.Plants are indeed an important member of ev-ery natural life [1] as well as the formal nam-ing of this will ensure that every natural lifeis preserved and maintained. Plants are essen-tial for our medicinal purposes, as alternativesources of energy such as biofuels, and alsoused to fulfill our numerous domestic needssuch as wood, clothes, food, and makeup. Thepresent extinction trend is primarily the prod-uct of overt and indirect human activity. Cre-ating correct identification information andplant geography propagation is key to thesurvival of ecosystems in the future.

Many countries worldwide are now design-ing programs to create channel control sys-

Nowadays, Artificial Intelligence (AI) is themost important part of our lives, it is used inthe field of Computer Vision, Robotics, Digi-tal Marketing Transformation, Medical field,Banking, and business sectors. Artificial Intel-ligence (AI) has been mainly designed to makemachines for thinking and acting like a hu-man being and machine learning would be asub-part of Artificial Intelligence (AI), as wellas a theoretical algorithm analysis and a math-ematical model that carries out a particularcheck without explicit programming, on thebasis of the assumption and templates. Somebasic forms of artificial learning strategies areSupervised learning, unsupervised learning,and reinforcement learning. The SupervisedLearning Algorithms contain data, Unsuper-vised learning algorithm requires a collection

425

tems for national agriculture [2]. India willhave a long tradition of utilizing plants as atherapeutic source. This research is calledayurvedic [3]. Each plant on Earth has a cer-tain medicinal value according to Ayurveda.This is viewed as a worldwide type of sub-stitute to allopathic medicine. one of the bigbonuses of this is that it has no adverse effects.Taxonomists systematically classify such medic-inal plants, that are susceptible to Miscar-riages in certain situations.

Image recognition strategies that have re-cently started to appear in an attempt to sim-plify that plant inspection process. With re-spect to a plant identification research, meth-ods focused by color characteristics were of-ten used to establish a plant recognition method.Color interpretation probably depends on colordistributions in such an image, although it isnot a safe function because here are certainsituations where this feature’s temporal accu-racy is abused. The Shift of light, leaf move-ment through waves, camera jitter, changingof focus, sudden shifts of camera parametercontribute to incorrect plant category predic-tions. Giving numerous researches, the cate-gory of plants dependent on digital images isnow seen as a difficult issue. Those researcheswere focused on the study of particular plantleaves for the identification and classificationof plants [4]. In subsequent years, many stud-ies were using them to develop a model for aplant leaf recognition system.

Gaber et al. used MCA to derive visual char-acteristics from plants [5] and linear discrimi-nant analysis. Multiple researchers also haveestablished the ability of broad convolutionneural network (CNN) to outperform conven-tional object recognition or detection strate-gies centered upon ordinary working light,texture, and shape characteristics. Typically,the CNN systems are using in such big-scaleplant recognizing activities consists of a traitextractor accompanied by a classifier. Despiteof occlusions collecting plant crop does not

usually a straightforward job. In a fact, certainplants cannot even have visible parts of theleaf. On the other side, several research rele-vant to utilizing machine vision approachesto address such issues were also performed.Aerial image recognition of landscapes thatuse machine learning techniques is an illus-tration of computer recognition technologiesof agricultural computing.

In Section 2, we did a literature review ofvarious research papers related to plants de-tection using deep learning and shown a ta-ble (Table 1.) of the previous pattern method.In Section 3, we discussed the identificationmethodology. In Section 4, we discussed in de-tail about the architecture of CNN. In Section5, we discussed the process of leaf identifi-cation very preciously. At last in Section 6,we gave the conclusion about our paper anddiscussed future work.

2. Literature ReviewThe literature on Deep learning is very wide.The work done by various researchers in thefield of plant identification using Deep learn-ing is described in this section. Sapna Sharma.(2015) used principal component analysis (PCA),Hu’s moment invariant method, and morpho-logical features for classification and they haveused sixteen different classes of the leaf. ThisMatlab measures the circumference by mea-suring the gap in each connected number ofpixels along the area’s boundary [1]. T. Gaber.(2015) suggested a plant recommender methodthat uses 2D visual photographs of plants.This program used the methodology of at-tribute fusion and the process of multilabelclassification. The experimental findings re-vealed that the function fusion method’s accu-racy was much higher than other individualapplications. The tests showed their robust-ness in providing accurate recommendations[2]. T. J. Jassmann. (2015) designed a new

426

CNN they tested the usage of the newly imple-mented Exponential Linear Unit (ELU) ratherthan Rectified Linear Unit (ReLU) as CNN’snon-linearity method [3].

Hulya Yalcin (2016) suggested an architec-ture of the CNN to identify the form of plantsfrom the picture sequences obtained from smartagro-stations. the design is used as a pre-processing stage to remove the picture prop-erties. Configuration of the CNN design andbreadth are important points that should behighlighted because they impact the recog-nizing capabilities of neural network architec-ture. They used 16 kinds of plants and com-pared them with other approaches; prelimi-nary findings show that the CNN centered ap-proach’s classification performance outranksother approaches.[4]

Amala Sabu (2017) depicts that UniversalLeaf Identification is a difficult Computer Vi-sion issue. Efficient leaf recovery method forAyurvedic plant beneficial for other aspects ofsociety including Medicine, studies in Botany.Recognize the photographs of the leaf. Thestudy of the Different approaches and classi-fications for leaf identification [5]. Lee, S.H.,(2017) gathered one of the pictures of plantleaves has also been discussed based uponthe leaf characteristics use as an input andconvolution neural network is being used toidentify patterns for each plant depth infor-mation. CNN was mainly utilized here justfor the improved portrayal of the characteris-tics and for effective studies of Leaf organismsDN (Deconvolutional Network) used. It en-ables greater recognition of plant leaves andtheir populations [6]. Ghazi, M.M., (2017) im-plemented three models of transfer learningto describe the identity of the various plants.The Network was evaluated using LIFECLEF2015. These three-model used GoogleNet, VG-GNet, and AlexNet for their suggestion here[7].

Barbedo, J.G. (2018) discussed the analysisof the key factors influencing the architecture

and efficacy of deep neural networks appliedto plant pathology and the in-depth study ofthe topic, which illustrates the benefits anddisadvantages, will contribute to more con-crete findings on plant pathology [8]. Barbedo,J.G.A., (2018) explored the implementationof issues in transfer learning and the use ofdeep learning. They found that CNN is amethod used to classify plant biotechnologyissues [9]. Zhu, X., (2018) uses CNN (ComplexBackground) to recognize the small objectedplant leaves. The designed methodology im-plemented sample-normalization founding V2which enhances the accuracy of Region CNN.For processing, the quality photos sub-samplesare split into a hundred and the residual im-ages are returned to final production. The ap-proach suggested that it could be faster thanconventional region convolutional neural net-work [10]. Garcia-Garcia (2018) A writer ofthis paper used deep learning techniques tofocus on high occupancy classification. Theypresented short information on the topics ofdeep learning. Which offers the required rel-evant information on deep learning for themission ahead [11].

Kaya, A., Keceli. (2019) suggested the con-cept of Transfer Learning for Plants Classifi-cation focused on Deep Learning. This paperindicates the impact of four separate transfer-ence training models on plant classificationdeals dependent on DNN for four availabledatabases. Finally, their theoretical researchreveals that Transfer Learning offers a ba-sis of plant classification self-estimating andanalysing. They use certain common formatsincluding End-to- End, Fine modulation, Finemodulation Cross Dataset, Deep IntegratedFinetuning, Classification by RNN-CNN [12].

427

Table 1Previous pattern recognition methods

Author Technique used Dataset Size Results

Sapna Sharma, Dr.Chitvan Gupta(2015) [1]

GLCM (grey level co-occurrence matrices: GLCMis a histogram at a given offset over an image withco-occurring greyscale values.) and Principal com-ponent analysis (PCA): the method measuring theprincipal component and to perform on the databasis of change, Super Vector Machine (SVM): usesfor classification problem methods.

16 differentclasses of theleaf.

Review

T. Gaber (2015) [2] Bagging classifier,2D based technique

using Flavia adataset which1907 coloredimages.

Accuracy: 95%

T. J. Jassmann (2015)[3]

Rectified Linear Unit (ReLU): is an activation func-tions and ReLu is the most used activation functionin the neural network, moreover in the CNNs.

Flavia datasetcontains leavesof 32 plants.

Accuracy: 60%

Hulya Yalcin (2016)[4]

Convolutional Neural Network (CNN) model.16 plantspecies.

Accuracy:97.47%

Amala Sabu (2017)[5]

K-Nearest Neighbor (KNN). Review Review

Lee, S.H (2017) [6] CNNUsing 113,205images

Accuracy:96.3%

Ghazi, M.M (2017)[7]

Transfer Learning using AlexNet GoogLeNet andVGGN: VGGN is an object-oriented Model and sup-ports 19 layers and VGGN is still the most popularused architecture for image recognition.

using 91,758images.

Accuracy:80.18%

Barbedo, J.G (2018)[8]

Using deep learning concepts

almost 50,000images usefrom plantvil-lage dataset.

Accuracy: 81%

Barbedo, J.G.A(2018) [9]

CNN

Containing12 differentplants and1383 imageswere used.

Accuracy: 84%

Zhu, X (2018) [10] RCNN

For eachspecies, 180 im-ages are taken,of which 150are taken for

Accuracy:99%

Garcia-Garcia(2018) [11]

semantic segmentation using deep learning tech-niques.

In this pa-per they use2D and 3Ddatasets.

Review

Kaya, A., Keceli(2019) [12]

Transfer learning method use on deep learning.Using totalnumber of54,306 images.

Accuracy:98.70%

Noon, S.K., Am-jad, M., Qureshi,M.A., Mannan,A(2020)[13]

deep learning-based Review Review

428

Figure 1: The architecture of our model (CNN)

3. IdentificationMethodology

For detecting the plans from the images of theleaf, we have discussed various strategies onthe CNN model utilizing various leaf datasetto show the characteristics of the visualiza-tion approaches for CNNs. CNNs are neu-ral feed forward networks that are fully con-nected. CNNs are exceptionally effective indecreasing the number of parameters with-out compromising layout efficiency. Imageshave a large size (as each pixel is regarded asa feature) that corresponds to the CNNs. ithas been developed to take account of images,but milestones were still reached in text pro-cessing also. The edges of artifacts in eachpicture are guided by CNNs.

3.1. Architecture of CNNDeep Belief Network is used in various meth-ods of Language processing, Computer Vi-sion, Speech Reorganization, and many otherapplications. Deep Neural Network has athree-layers Input, Hidden, and Output. TheDeep Neural Network [6] processes in multi-ple NN. Fig.1 illustrates how the Neural Net-work nodes and layers are connected and shareinformation.

3.1.1. CNN

CNN [7] is the part of deep neural networkclass. This is mostly used in Computer Visionto identify the given structure of the objectbeing subjected. CNN ’s primary objectiveis to identify and forecast the sequence ofthe given input datasets. It delivers enhancedperformance and accuracy.

3.1.2. Layers In CNN

CNN is a controlled methodology in deeplearning, and has developed a ground break-ing influence on numerous applications fo-cused on machine vision and images. Thefields of which CNN is commonly employedinclude facial recognition, target identifica-tion, analysis of videos, etc. CNN platformcomponents involve convection layers, pool-ing layers, completely linked layers, activa-tion functions, etc.

• Convolution Layer: For more pro-cessing the layer provides an RGB pic-ture or an output of another layer asdata. The obtained information is re-ferred to as image pixels to producea function map reflecting characteris-tics of low levels, such as edges andcurves. Special characteristics at thehigher level can be defined via a se-quence of further convolution levels.

• Activation Layer:Nonlinearity makesa network of neurons deeper. A Non-linear activation layer shall be added di-rectly after each layer Convolution stra-tum. Specific nonlinear mechanismsare used for Add non linearity. Theyare:

Tanh: The range between [-1,1] takesthe real-valued number of this non-linearity.

Sigmoid: The range between [0,1] takesthe real-valued number of this non-linearity.

429

Rectified Linear Unit: It improves themodel’s nonlinear property by alteringthe convolution layer’s receptive fieldby altering all the lower values to 0.

• Pooling Layer: After the activationlayer a downsampling layer was addedto raising the spatial aspect without anyalteration in size. Typically, a size 2x2input filter is applied to produce an out-put based on the pooling process. Itmay be expressed either by peak pool-ing or by average pooling where thelimit or average value is calculated foreach sub-region used in the filter. Therepooling is no restrictions.

Then the pooling layer decreases thescale of the characteristic map, i.e., thewidth and length are limited but the dis-tance is not. It Decreases the numberof Weights and Parameters, lowers thepreparation period, and thus eliminatesthe computational expenses. It also re-quires overfitting controls.

• Fully Connected Layer:This layerdefines characteristics that are at a verylarge quality correlates to class or ob-ject. Entering a fully connected fieldlayer is a collection of features for pic-ture recognition without requiring totaking into consideration the spatial con-text of the pictures. Fully connectedlayer output is often a 1D vector achievedby compressing the final pooling layeroutput. This is a method of organizing3D volume in a 1D vector.

4. Process of LeafIdentification

Plants take a significant role in both humanand other life on earth. Recognition of theleaf design implements normally the phases

Figure 2: Step of leaf Identification Process

as seen in Figure 2. The demanding one thatis part of the research is to determine leaf dis-tinguishing features for the identification ofplant organisms. In this situation, a separateclassifier with high-performance statisticalmethods has been used to conduct leaf classi-fication and function extraction of the func-tionality. The improvement in image analysisand CNN significantly aided researchers inthe classification of plants by data analysis.

This is a basic image-based plant recogni-tion process is shown in fig.2 and some definesome general stages.

430

4.1. Image AcquisitionImage acquisition is the process of collectingdatasets for identification leaves. Infected leafphotographs of collected in managed settingsand are processed in JPEG format. Againsta white backdrop, the contaminated leaf isput smooth, the light source was mountedon either side of leaves at a temperature of45 degrees to remove each reflection and pro-vide even light wherever thereby increasingillumination and clarity. This crop is zoomedsuch that the photograph captured includesjust the crop and white backdrop

4.2. Image Pre-processingImage pre-processing procedures are essen-tially used to expose information that is hid-den or basically to show any features in suchan image. Such methods are largely contex-tual and are structured to alter a picture andtaking advantages of the psycho-cultural di-mensions of the human sensory system. Equal-ization of the histograms and electronic filtra-tion methods were used.

4.2.1. Histogram equalization:

it’s some of those strategies for improvingimages. A certain approach allocates imageintensities. Among this process, the contrastbetween the fields rises through local con-trast to greater intensity. The equalization ofhistograms is used to enhance computationalcomplexity, clarity, and image consistency.

4.2.2. Grayscale conversion:

Gray scale conversion is used for convertingimages into grayscale. The grayscaled conver-sion used the method of contrast feature andintensity enhancement techniques for con-verting the images and then placed them aspieces together for further processing.

4.2.3. Binary conversion:

Create binary images on a gray image scale touse the threshold method. The Binary pictureis a visual image that contains just two poten-tial meanings for each pixel. The two shadesthat a contrasting picture uses are typicallywhite and black.

4.2.4. Noise Removal:

Digital photographs are susceptible to a plethoraof Noise levels. Noise emerges from errors inthe virtual method of image acquisition whichresults in pixels values They can be used toeliminate linear filtering those values noiseStyles. Few filters are ideal for this purpose,such as Gaussian filters or low pass filters,averaging. An ordinary filter, for example, isuseful for having grain noise off the picture. Amedian filter and averaging filter are used forsalt and paper noise removal from an image.

4.3. Feature extractionmainly the characteristics of leaf color andform. The Specific plant leaf is generally iden-tical in color and form are considered for clas-sification and so a specific function alone can-not achieve anticipated results.

4.3.1. Color features

Dr. H.B. Kekre et al.’s suggested the approachof scanning and retrieving photographs pri-marily focused on the production of the colorfunction vector by measuring the mean. Thisthree-color Red, Green and Blue are first di-vided in the suggested algorithm. Then meansand column mean of colors are determinedfor every plane side. For each plane the sumof all means of the row and all means of thecolumns are determined. The characteristicsof all 3 planes converge to create a matrix offeatures. Until the function vectors for an im-

431

Figure 3: The basic geometric features

age is created, they are contained in a databaseof features.

4.3.2. Shape features

Based on the of Geometric features, we de-fined shape features:

a. Geometric features: We used the similar5 geometric features (DMFs), define in fig. 3,derived from the following 5 basic features:

2

1. Diameter: between any two pointsthe diameter of the leaf is the longestdistance on the closed contour of theleaf.

2. Physiological Length: It’s the lengthof the line which connects the two mainvein terminals points in the vine.

3. Physiological Width: This corre-sponds to the interval perpendicularto the physiological longitude betweenthe two endpoints of the longest linesection.

4. Leaf Area: This is the amount of bi-nary representation 1 pixels on the smoothimage of the vine.

5. Leaf Perimeter: the count of pixelsin the leaf’s closed contour.

4.4. ClassificationCommon statistical identification is the methodof defining based on the previous informationsuch as a training dataset a group of groups, orclasses to which a new phenomenon belongs.More precisely, classification in this work isthe method used to attribute a picture to acertain plant genus, based on its collection offeatures. It is a subclass of more general statis-tics and deep learning identification problems,including supervised learning.

4.5. TestingOf this phase, we test the model by givingtesting data to the model. Then check how isit identifying the object and also, we get theaccuracy.

4.6. Convolution NeuralNetwork Process

CNN is one of the types of neural networkswhich are widely used in computer visionarea. Its name stems from the form of secretlayers it consists of. Usually a CNN’s con-cealed layers comprise of convolutionary lay-

432

Figure 4: The architecture of our model (CNN)

Table 2For identification, Some Layer is used for our model

Layer No of filters Filter size Stride Value1st 16 2*2 12nd 16 2*2 13rd 32 2*2 14th 32 2*2 15th 64 2*2 17th 64 2*2 1

ers, pooling layers, completely linked layers,and layers of normalization. Here it simplymeans that convolution and pooling functionsare used as activation functions.

4.6.1. Dropout

A Dropout feature is employed in many CNNworks. That can result from the problem ofoverfitting in our model. By randomly remov-ing those connections that exist between thenodes.

Fig 4 this architecture is our work. On thevery last layer, the dense function is used.

2

4.6.2. Pooling

Pooling is a discretization method dependenton the samples. The aim is to down-sample aninput data (image, hidden-layer output matrix,

etc.), decreasing its dimensionality and mak-ing conclusions about features found in dis-carded sub-regions. There are 2 major formsof pooling generally recognized as poolingwith max and min. Like the name indicates,max pooling is focused on taking up the high-est value from the selected area, and min pool-ing is based on picking up the selected re-gion’s minimal value.

5. Conclusion and FutureWork

CNN performs so much more on pictures andvideos than traditional neural networks, sincethe convolutionary layers take advantage ofthe image’s intrinsic properties. Simple neuralfeedforward networks see little structure intheir inputs. When you combined all the im-

433

ages in the same way, the neural network willhave the same success when trained on photosthat are not shuffled. But on the other side,it optimizes local spatial picture coherence.This ensures they will significantly decreasethe number of operations needed to processan image by utilizing convolution on adjacentpixel patches as adjacent pixels are meaning-ful together. We name it central connectivitytoo. The Map is then loaded with the productof a small patch of pixels converting, slid overthe entire picture with a window. There aremany methods in the detection and classifica-tion process of automated or computer visionfor plant identification but there is still a lackof research in this field. Moreover, there arecurrently no consumer options on the market,even those that deal with the identification ofplant organisms dependent on photographs ofthe leaves. It has been concluded that a differ-ent approach using deep learning techniquesare used to automatically identify and recog-nize plants from the photographs of the leaf.The model established was able to sense theexistence of a leaf and distinguish betweenhealthy leaves.

In the future research would be to raise thesize of the dataset by raising the samples andby adding different classes of the plants leaf.

After doing the literature review of vari-ous papers, we conclude that CNN is the bestapproach for detecting the plants and leaveswith very good accuracy.

References[1] Sapna Sharma, Dr. Chitvan Gupta, “A Re-

view of Plant Recognition Methods andAlgorithms,” IJIRAE - International Jour-nal of Innovative Research in AdvancedEngineering, Vol. 2, Issue no. 6, June,2015.

[2] Hulya Yalcin, Salar Razavi, “PlantClassification using ConvolutionalNeural Networks”, IEEE International

Conference on Agro-Geoinformatics,Tianjin, China, DOI: 10.1109/ Agro-Geoinformatics.2016.7577698, Spet,2016.

[3] T. J. Jassmann, R. Tashakkori, and R. M.Parry, "Leaf classification utilizing a con-volutional neural network”, IEEE South-eastcon, Fort Lauderdale, FL, USA, DOI:10.1109/ SECON.2015.7132978, April,2015.

[4] Amala Sabu, Sreekumar K, “LiteratureReview of Image Features and ClassifiersUsed in Leaf Based Plant RecognitionThrough Image Analysis Approach”,International Conference on InventiveCommunication and ComputationalTechnologies (ICICCT), DOI: 10.1109/ICI-CCT.2017.7975176, March, 2017.

[5] 5) T. Gaber, A. Tharwat, V. Snasel, and A.E. Hassanien. "Plant Identification: TwoDimensional-Based Vs One Dimensional-Based Feature Extraction Methods”, Inter-national Conference on Soft ComputingModels in Industrial and EnvironmentalApplications, Springer, Cham, DOI:http://doi-org443.webvpn.fjmu.edu.cn/10.1007/ 978-3-319-19719-733, May, 2015.

[6] Kaya, A., Keceli, A.S., Catal, C., Yalic, H.Y.,Temucin, H. and Tekinerdogan, B, “Anal-ysis of transfer learning for deep neuralnetwork-based plant classification models.”Computers and Electronics in Agriculture,Vol. 158, March, 2019.

[7] Barbedo, J.G, “Factors influencing the useof deep learning for plant disease recog-nition”, Biosystems engineering, Vol. 172,May,2018.

[8] Ghazi, M.M., Yanikoglu, B. and Aptoula, E,“Plant identification using deep neural net-works via optimization of transfer learningparameters”, Neurocomputing, Vol. 235,April, 2017.

[9] Lee, S.H., Chan, C.S., Mayo, S.J. and Re-magnino, P., “How deep learning extractsand learns leaf features for plant classifi-cation”, Pattern Recognition, Vol. no: 71,

434

May, 2017.[10] Barbedo, J.G.A., “Impact of dataset size

and variety on the effectiveness of deeplearning and transfer learning for plant dis-ease classification”, Computers and Elec-tronics in Agriculture, Vol. 153, Oct, 2018.

[11] Zhu, X., Zhu, M. and Ren, H., “Method ofplant leaf recognition based on improveddeep convolutional neural network”, Cog-nitive Systems Research, Vol. 52, Dec, 2018.

[12] Garcia-Garcia, A., Orts-Escolano, S.,Oprea, S., Villena-Martinez, V., Martinez-Gonzalez, P. and Garcia-Rodriguez, J., “Asurvey on deep learning techniques for im-age and video semantic segmentation, Ap-plied Soft Computing”, Vol. 70, May, 2018.

[13] Noon, S.K., Amjad, M., Qureshi, M.A.,Mannan, A., “Use of deep learning tech-niques for identification of plant leafstresses: A review”, Sustainable Comput-ing: Informatics and Systems, Vol.28, Dec,2020.

435