FEATURE EXTRACTION USING MULTI-TEMPORAL FULLY POLARIMETRIC SAR DATA

摘要

The main objective of this study was to explore the potential of the multi-temporal PolSAR data in LULC mapping and to evaluate the accuracy of classification using single date and multi-temporal data. Multi-temporal data acquired on three different dates were used. Advanced classification techniques Support Vector Machine and Rule Based Hierarchical approaches were performed on multitemporal ALOS PALSAR data to classify features at different temporal combinations, In this study, SVM classification was applied on the derived output of Yamaguchi decomposition model, for which kernel approach of second order polynomial was used, In Rule Based Hierarchical approach, Backscattering coefficients, Yamaguchi and H/A/Alpha decomposition statistics are computed and analyzed to estimate the decision boundaries of the features to separate feature at different hierarchical levels. SVM classified the PolSAR data efficiently of single data, highest overall accuracy and kappa statistics achieved was 67.65% and 0.61 from the individual image. Rule based classified map of single date, highest overall accuracy and kappa statistics achieved was 68% and 0.67. Based on the accuracy assessment, SVM and Rule Based classification both are approximately of same accuracy but comparatively Rule Based classification was accurate temporally. Rule Based classification was further considered for multi-temporal classification and achieved high overall accuracy and kappa statistics of 80% and 0.76. This proves that multi-temporal PolSAR data helps to increase the accuracy of classification in LULC mapping.