1. Evaluation of supervised land-cover classification by PALSAR polarimetric interferometry Masato Ohki and Masanobu Shimada Earth Observation Research Center, Japan Aerospace Exploration Agency
2. Outline Background Polarimetric interferometry (PolInSAR) PALSAR PolInSAR data Methods and data Result: Land-cover classification by PALSAR PolInSAR Discussion Advantage of PolInSAR for LC classification Comparison between classification methods Comparison with optical sensor data Conclusion and Future work
5. PALSAR Polarimetric Interferometry (PolInSAR) Issue: single satellite-> repeat-pass interferometry Various spatial distance (0.0~2.5km) Long temporal distance (≥46 days) -> Application? Master PolInSARCoherency matrix repeat pass Slave rm rs
6. The quake hit Tsukuba Space Center What can we do for disasterprevention/mitigation? 3.11 Earthquake
7. Overview of this study Feasibility study on land-cover (LC) monitoring by PALSAR 7 classes supervised LC classification by PALSAR PolInSAR data Accuracy evaluation Comparison between four cases of datasets: (1) Quad-PolInSAR (2) Dual-PolInSAR (3) Quad-PolSAR (4) Dual-PolSAR Comparison between classification methods: Wishart SVM Comparison with other LC product ALOS LC product (optical)
9. Test data PALSAR data used in this study #1 (PLR) #2 (PLR) Optical (AVNIR-2) Tsukuba city (36.05˚N,140.10˚E) HH-VVHVHH+VV(Pauli) NARITA Int’l Airport(35.77˚N,140.39˚E)
11. Class definition Water Paddy Crop Grass Forest Urban Bare Reference data(105 polygons, 8200 samples) #2 Paddy #4 Grass #3 Crop #7 Bare Ground photographs (Tsukuba city, 09 JUN 2009)
12. Processing Procedure 1. Pre-processing(imaging, pol. calibration and interferometry) Processor: SIGMA-SAR (by Dr. Shimada) 2. Classification Compared two classificationmethods: Wishart classifier and SVM Processor: developed in this study 3. Post-processing(ortho-rectification and geo-coding) Processor: SIGMA-SAR (by Dr. Shimada) Resolution of the classification map: 60m PALSARL1.0(master) PALSAR L1.0(Slave) Generate SLC Pol. CalibrationCo-registration Slope correction (option) Pol. filtering (option) Trainingdataset Classification (Wishart or SVM) DEM Ortho-rectification(geo-coding) Final classification map
13. Classifier(1) – Wishart Classifier Maximum likelihood approach assuming that the scattering matrix follows a complex Wishart distribution function (Lee et al., 1994, 1999) The pixel is assigned to the class minimizing the distance measure between the pixel and the training class Scattering matrix for the Wishart classifier (master data) (master data)
14. Classifier(2) – Support Vector Machine (SVM) Margin maximization approach discriminating a class from other classesin the higher dimensional space(Fukuda and Hirosawa, 2000 for PolSAR data; Shimoni et al., 2009 for PolInSAR data; the SVM core routine is distributed by Chen & Lin, 2005) Feature parameters for the SVM *The Cloude-Pottier decomposition (Cloude & Pottier, 1996; Pottier 1998)
24. Comparison of classification methods Some LC types (esp. urban)can have various scattering mechanism Linear classifier (e.g. Wishart) Assuming a single scattering mechanism for each class Non-linear or non-parametric classifier (e.g. SVM) More robust for LC types which have various scattering mechanisms Urban areamisclassifiedas Forest Crop fieldsmisclassifiedas Grass Grassmisclassifiedas Bare Quad-PolInSAR (SVM) Quad-PolInSAR (Wishart) Optical
25. ALOS Land-cover product (by the optical sensor) Available at http://www.eorc.jaxa.jp/ALOS/lulc/lulc_jindex.htm (free) Current version: ver. 11.02 (released on Feb 2011) Classification method: decision tree of multi-seasonal optical sensor images Coverage: Japan area No. of classes: 10 Resolution: 30m Accuracy: 87%(evaluation result) ALOS LC product (optical)
27. Comparison with ALOS LC product Advantage of PolInSAR classification: Precise detection ofForest, Urban, Bare and Water Water Paddy Crop Grass Forest Urban Bare Small urban areamisclassified as Forest PolInSAR (this study) Optical (ALOS LC) AVNIR-2 image Bare groundsmisclassifiedas Water
28. Comparison with ALOS LC product Advantage of optical classification Precise detection of low vegetation (Paddy, Crop and Grass) Water Paddy Crop Grass Forest Urban Bare Grass areamisclassifiedas Crop PolInSAR (this study) Optical (ALOS LC) AVNIR-2 image Paddymisclassifiedas Crop
29. Summary of results Comparison of datasets Accuracy: Quad-PolInSAR > Dual-PolInSAR> Quad-Pol > Dual-Pol Interferometric coherence plays important roles for discriminating LC types which have confusing scattering mechanisms Comparison of classification methods: Accuracy: SVM > Wishart Computation Speed: Wishart > SVM Non-linear classifier is more robust for LC types which have various scattering mechanisms Comparison of PolInSAR classification and ALOS (optical) LC product PolInSAR classification is good on Forest, Urban, Bare and Water classification ALOS (optical) LC product is good on Low vegetation (Paddy, Crop and Grass) classification
30. Conclusions PALSAR PolInSAR data has high capability for LC monitoring Quad-PolInSAR classification is more accurate than dual-PolInSAR and quad/dual-PolSAR The SVM is better than the Wishart classifier on classification accuracy Future Works Improvement of classification algorithm Other classification methods Other feature parameters Speckle filtering, terrain correction Extension of the test area Application for monitoring disaster, forest or agriculture PolInSAR data of ALOS-2/PALSAR-2: higher resolution, smaller and stable orbit distance...
31. Thank you for your attention… Mt. Tsukuba Tsukuba city Water Paddy Crop Grass Forest Urban Bare
34. Detail – Paddy (2) Back-scattering in paddy area changes significantly from April to May #1 02/04/2007 Before flooding #2 18/05/2007 After flooding Optical (AVNIR-2) 15/05/2007 HH-VVHVHH+VV Water surface Soil surface