Jayaluxmi Indu , D. Nagesh Kumarcivil.iisc.ernet.in/~nagesh/pubs/Indu_AGU_Poster_Feb2012.pdf ·...
1
Poster Design & Printing by Genigraphics®-800.790.4001 Jayaluxmi Indu , D. Nagesh Kumar INTRODUCTION DATA CONCLUSIONS RESULTS REFERENCES ABSTRACT Dr. D. Nagesh Kumar Professor Water Resources & Environmental Engg. Dept. of Civil engineering Indian Institute of Science, Bangalore, India Email: [email protected] J.Indu PhD Student Water Resources & Environmental Engg. Indian Institute of Science, Bangalore, India Email: [email protected] CONTACT RAIN/NO RAIN CLASSIFICATION OVER TROPICAL REGIONS USING TRMM TMI This study focalizes on the comparative assessment of existing “rain” or “no rain” classification (RNC) methodologies over land surface. Using the data products from Tropical Rainfall Measuring Mission’s Precipitation Radar (PR) and Microwave Imager (TMI), this work reveals the shortcomings of existing overland RNC methods hinged on scattering signatures from 85GHz high frequency channel with special reference to the Indian subcontinent. Overland RNC of microwave radiometer brightness temperatures (Tb) offers a myriad of complications as land surface presents itself as a radiometrically warm and highly variable background. Hence, sensitivity analysis of Tb captured at different microwave frequencies to near surface rain (NSR) rate is of supreme importance. Variability of Tb to NSR is investigated using exploratory data analysis (EDA), together with probability density functions (pdfs). Results indicate that the inclusion of 37 GHz vertical polarization channel for the computation of estimated 85GHz Tb is avoidable, due to its prominent correlation with NSR. Also relevant, were the contributions from 19 GHz channels (both horizontal and vertical). A novel attempt has been made to assess the performance of some powerful statistical descriptors to increase the accuracy of RNC. The comparative results are presented extensively in the form of contingency tables and kappa values. The descriptor giving best results from the analysis was used to formulate a regression relation with the rain rate (RR). Furthermore, this work also conducts a detailed examination on the use of Empirical Orthogonal Functions (EOF), to improve rain retrieval accuracy. Analysis reveals that, for TMI, the first two EOF’s were deemed sufficient to fully explain the variability offered by the 9 channels. A regression based relationship was established between RR and EOF. A comparative analysis was conducted between the regression relation of RR with EOF and RR with “best RNC descriptor”. Results reveal that the use of efficient methodologies for overland RNC does improve the classification accuracy (upto less than 10%). Even though the performance was not too high, it was proven to improve the retrieval accuracy. Orbital data (version 7) from TRMM Microwave Imager (TMI) and Precipitation Radar (PR) namely 1B11, 2A25 and 2A23 were used for study. (From trmm.gsfc.nasa.gov made available since September 2011) The effective field of view (EFOV) of the different data channels along the down track (DT) and cross track (CT) directions are shown in Figures 1 & 2. A total of 9635 collocated overland samples were extracted for analysis and subjected to Rain/No Rain classification (RNC) using scattering index (SI) as performance indicator. Methods proposed by 5 different authors were applied on the data. Overall accuracy & Kappa statistic, obtained after RNC are shown in Figure 3. Defying conventional views, a combination of lower channel frequencies were found to be more responsive to near surface rain rate (NSR) than ice scattering at 85 GHz (You et al, 2011). Spearman rank correlation coefficients for the top 20 data combinations are depicted in Figure 4. Scope exists to improve RNC over land regions by checking channel combinationsfor sensitivity to NSR. Use of sensitive channel combinations tend to improve the accuracy of relations developed between NSR and several predictor variables when compared with the use of Empirical orthogonal functions (EOFs). Need for relation over a region, for RNC to tackle issues in basin scale hydrology. Figure 3: Results of RNC using SI Figure 4 : Channel sensitivity with NSR http://civil.iisc.ernet.in/~nagesh/ Figure 2: Anaglyph showing footprint size (not to scale) of various TMI channels {Best viewed with 3D glasses} For almost over a decade, passive sensing of upwelling microwave radiation has been recognized as a promising source for the estimation of precipitation (Ferraro et al. 1996). Microwave instruments aboard Tropical Rainfall Measuring Mission (TRMM) namely, Microwave Imager (TMI) and Precipitation Radar (PR) offer excellent opportunity to study atmospheric phenomenon over tropics. Overland rain retrieval algorithms using passive radiometers, hinge mainly on the scattering signature from high microwave frequencies. Hence, problematic compared to ocean rainfall algorithms that utilize both emission and scattering signatures from observations at multiple frequencies (You et al. 2011). Studies dealing with physically based retrieval over land, are few and have not indicated better performance relative to purely statistical algorithms. A deterministic method known as rain /no rain classification (RNC), applied before actual rain retrieval is adopted to reduce computational burden as well as for quality checks (Seto et al. 2005). 85.50GHz 37.00 GHz 21.30 GHz 19.35 GHz 10.65 GHz Seto, S., N. Takahashi, and T. Iguchi, 2005: Rain/ no-rain classification methods for microwave radiometer observations over land using statistical information for brightness temperatures under no-rain conditions, J. Appl. Meteor., 44, 1243-1259. Ferraro, R., E. A. Smith, W. Berg, and G. J. Huffman (1998), A screening methodology for passive microwave precipitation retrieval algorithms. J. Atmos. Sci., 55, 1583-1600,doi:10.1175/1520-0469. Gopalan, K., N. Wang, R. Ferraro, and C. Liu (2010), Status of the TRMM 2A12 and precipitation algorithm, J. Atmos. Oceanic Technol., 27, 1343- 1354, doi:10.1175/2010JTRCHA1454.1. You, Y., G. Liu, Y. Wang, and J. Cao (2011) On the sensitivity of Tropical Rainfall Measuring Mission (TRMM) Microwave Imager channels to overland rainfall, J. Geophysical Research 116: D12203, doi:10.1029/2010JD015345. Channels EFOV (CT x DT) TRMM Microwave Imager (TMI) 10.65 GHz (H & V) 9.1 Km x 63.2 Km 19.35 GHz (H &V) 9.1 Km x 30.1 Km 21.30 GHz (V) 9.1 Km x 22.6 Km 37.00 GHz (H & V) 9.1 Km x 16.0 Km 85.50 GHz (H &V) 4.6 Km x 7.2 Km Precipitation Radar (PR) 13.80 GHz (5.1 Km x 5.1 Km) Figure 1: EFOV of TMI & PR channel frequencies. 0 0.2 0.4 0.6 0.8 1 Adler (1994) Ferraro (1997) Grody (1991) Kummerow (2001) Kummerow & Giglio (1994) Overall Accuracy / Kappa Statistic SI method proposed by different Authors Overall Accuracy Kappa Statistic Using the best results from RNC, non linear regression relations were developed between the highly correlated channel combination and NSR. The relation is shown in Figure 5. For TMI data, first two principal components were deemed sufficient to explain the variability of the data. Hence, polynomial regression based relation, linking first two principal components with NSR is shown in Figure 6. 2 2 3 4 2 2 3 2 2 01 0 07 0 08 0 18 0 06 1 b a b a a ab b a a b . ab . a b . a . . ) b , a ( NSR 3 2 2 3 4 5 3 2 2 3 4 3 2 2 3 2 2 02 0 05 0 55 1 y x y x y x x xy y x y x x y xy y x x y xy x y . x . . ) y , x ( NSR Figure 6 : Relation of NSR with 1 st & 2 nd principal components R 2 =0.96 Figure 5 : Relation of NSR with highly correlated channel combination R 2 =0.98
Transcript of Jayaluxmi Indu , D. Nagesh Kumarcivil.iisc.ernet.in/~nagesh/pubs/Indu_AGU_Poster_Feb2012.pdf ·...
Poster Design & Printing by Genigraphics® - 800.790.4001
Jayaluxmi Indu , D. Nagesh Kumar
INTRODUCTION
DATA
CONCLUSIONS
RESULTS
REFERENCES
ABSTRACT
Dr. D. Nagesh Kumar ProfessorWater Resources & Environmental Engg.Dept. of Civil engineeringIndian Institute of Science, Bangalore, IndiaEmail: [email protected]
J.InduPhD StudentWater Resources & Environmental Engg.Indian Institute of Science, Bangalore, IndiaEmail: [email protected]
CONTACT
RAIN/NO RAIN CLASSIFICATION OVER TROPICAL REGIONS USING TRMM TMI
This study focalizes on the comparativeassessment of existing “rain” or “no rain”classification (RNC) methodologies over landsurface. Using the data products from TropicalRainfall Measuring Mission’s PrecipitationRadar (PR) and Microwave Imager (TMI), thiswork reveals the shortcomings of existingoverland RNC methods hinged on scatteringsignatures from 85GHz high frequency channelwith special reference to the Indiansubcontinent. Overland RNC of microwaveradiometer brightness temperatures (Tb) offers amyriad of complications as land surface presentsitself as a radiometrically warm and highlyvariable background. Hence, sensitivity analysisof Tb captured at different microwavefrequencies to near surface rain (NSR) rate is ofsupreme importance. Variability of Tb to NSRis investigated using exploratory data analysis(EDA), together with probability densityfunctions (pdfs). Results indicate that theinclusion of 37 GHz vertical polarizationchannel for the computation of estimated85GHz Tb is avoidable, due to its prominentcorrelation with NSR. Also relevant, were thecontributions from 19 GHz channels (bothhorizontal and vertical). A novel attempt hasbeen made to assess the performance of somepowerful statistical descriptors to increase theaccuracy of RNC. The comparative results arepresented extensively in the form of contingencytables and kappa values. The descriptor givingbest results from the analysis was used toformulate a regression relation with the rain rate(RR). Furthermore, this work also conducts adetailed examination on the use of EmpiricalOrthogonal Functions (EOF), to improve rainretrieval accuracy. Analysis reveals that, forTMI, the first two EOF’s were deemed sufficientto fully explain the variability offered by the 9channels. A regression based relationship wasestablished between RR and EOF. Acomparative analysis was conducted between theregression relation of RR with EOF and RR with“best RNC descriptor”. Results reveal that theuse of efficient methodologies for overland RNCdoes improve the classification accuracy (uptoless than 10%). Even though the performancewas not too high, it was proven to improve theretrieval accuracy.
Orbital data (version 7) from TRMM Microwave Imager (TMI)and Precipitation Radar (PR) namely 1B11, 2A25 and 2A23 wereused for study. (From trmm.gsfc.nasa.gov made available sinceSeptember 2011)
The effective field of view (EFOV) of the different data channelsalong the down track (DT) and cross track (CT) directions are shownin Figures 1 & 2.
A total of 9635 collocated overland samples were extracted for analysisand subjected to Rain/No Rain classification (RNC) using scattering index(SI) as performance indicator. Methods proposed by 5 different authorswere applied on the data.
Overall accuracy & Kappa statistic, obtained after RNC are shown inFigure 3.
Defying conventional views, a combination of lower channel frequencieswere found to be more responsive to near surface rain rate (NSR) than icescattering at 85 GHz (You et al, 2011). Spearman rank correlationcoefficients for the top 20 data combinations are depicted in Figure 4.
Scope exists to improve RNC over land regions by checkingchannel combinations for sensitivity to NSR.
Use of sensitive channel combinations tend to improve theaccuracy of relations developed between NSR and severalpredictor variables when compared with the use of Empiricalorthogonal functions (EOFs).
Need for relation over a region, for RNC to tackle issues inbasin scale hydrology.
Figure 3: Results of RNC using SI
Figure 4 : Channel sensitivity with NSR
http://civil.iisc.ernet.in/~nagesh/
Figure 2: Anaglyph showing footprint size (not to scale) of various TMI channels {Best viewed with 3D glasses}
For almost over a decade, passive sensing of upwelling microwaveradiation has been recognized as a promising source for theestimation of precipitation (Ferraro et al. 1996).
Microwave instruments aboard Tropical Rainfall Measuring Mission(TRMM) namely, Microwave Imager (TMI) and Precipitation Radar(PR) offer excellent opportunity to study atmospheric phenomenon overtropics.
Overland rain retrieval algorithms using passive radiometers, hingemainly on the scattering signature from high microwave frequencies.Hence, problematic compared to ocean rainfall algorithms thatutilize both emission and scattering signatures from observations atmultiple frequencies (You et al. 2011).
Studies dealing with physically based retrieval over land, are few andhave not indicated better performance relative to purely statisticalalgorithms.
A deterministic method known as rain /no rain classification (RNC),applied before actual rain retrieval is adopted to reduce computationalburden as well as for quality checks (Seto et al. 2005).
85.50GHz
37.00 GHz
21.30 GHz
19.35 GHz
10.65 GHz
Seto, S., N. Takahashi, and T. Iguchi, 2005: Rain/ no-rain classification methodsfor microwave radiometer observations over land using statistical informationfor brightness temperatures under no-rain conditions, J. Appl. Meteor., 44,1243-1259.
Ferraro, R., E. A. Smith, W. Berg, and G. J. Huffman (1998), A screeningmethodology for passive microwave precipitation retrieval algorithms. J.Atmos. Sci., 55, 1583-1600, doi:10.1175/1520-0469.
Gopalan, K., N. Wang, R. Ferraro, and C. Liu (2010), Status of the TRMM2A12 and precipitation algorithm, J. Atmos. Oceanic Technol., 27, 1343-1354, doi:10.1175/2010JTRCHA1454.1.
You, Y., G. Liu, Y. Wang, and J. Cao (2011) On the sensitivity of TropicalRainfall Measuring Mission (TRMM) Microwave Imager channels tooverland rainfall, J. Geophysical Research 116: D12203,doi:10.1029/2010JD015345.
Channels EFOV (CT x DT)
TRMM Microwave Imager (TMI)
10.65 GHz (H & V) 9.1 Km x 63.2 Km19.35 GHz (H &V) 9.1 Km x 30.1 Km21.30 GHz (V) 9.1 Km x 22.6 Km
37.00 GHz (H & V) 9.1 Km x 16.0 Km85.50 GHz (H &V) 4.6 Km x 7.2 Km
Precipitation Radar (PR)
13.80 GHz(5.1 Km x 5.1 Km)
Figure 1: EFOV of TMI & PR channel frequencies.
0
0.2
0.4
0.6
0.8
1
Adler (1994) Ferraro (1997)
Grody (1991)
Kummerow (2001)
Kummerow & Giglio (1994)
Ove
rall
Acc
urac
y / K
appa
Sta
tistic
SI method proposed by different Authors
Overall Accuracy Kappa Statistic
Using the best results from RNC, non linear regressionrelations were developed between the highly correlatedchannel combination and NSR. The relation is shown inFigure 5.
For TMI data, first two principal components were deemedsufficient to explain the variability of the data. Hence,polynomial regression based relation, linking first twoprincipal components with NSR is shown in Figure 6.
223422322 010070080180061 babaaabbaab.ab.ab.a..)b,a(NSR
3223453
2234322322020050551yxyxyxxxy
yxyxxyxyyxxyxyxy.x..)y,x(NSR
Figure 6 : Relation of NSR with 1st & 2nd principal components
R2=0.96
Figure 5 : Relation of NSR with highly correlated channel combination
R2=0.98
