Assessment of urban land cover classification using Wishart and Support Vector Machine (SVM) based on different decomposition parameters of fully-polarimetric SAR

Abstract

Urban land cover mapping is one of the most important remote sensing applications. In this research, various polarimetric SAR parameters derived from fully-polarimetric SAR were explored for urban land cover mapping. The optimization of features is an important step for improving classification accuracy. First, radiometric correction of RADARSAT-2 Single Look Complex (SLC) product data has been performed using PolSARpro5. Two speckle filters (refined Lee and sigma Lee) were selected to be tested in RADARSAT-2 for elimination of noise and smoothing of the SAR images. It was found that sigma Lee filter is better than refined Lee filter with kernel size 5*5. Secondly, geometric correction was performed. The RADARSAT-2 was primarily geometrically corrected using ASF map ready tool in PolSARpro5 software for geocoding. Then second order polynomial based on fifteen well-distributed DGPS points was performed using ENVI 5 software. After that polarimetric decomposition parameters of RADARSAT-2 fully polarimetric SAR image was extracted from polarimetric decomposition techniques like Cloude-Pottier and Yamaguchi 4 components. Preprocessing of RADARSAT-2 data was achieved using PolSARpro5. In the present paper two classification methods, namely, Wishart and Support Vector Machine (SVM), were used for classification based on Cloude-Pottier and Yamaguchi's decompositions and combination of both decompositions. Three processing schemes were proposed based on decomposition parameters and were fed to Wishart and SVM algorithms. A comparison between these three schemes has been carried out and their usefulness in classifying urban land cover type was explored. It was found that SVM is better than Wishart classifier for classification of fully polarimetric synthetic aperture radar data. When applying the classification scheme based on each theorem separately, Yamaguchi's 4 components decomposition gave higher classification accuracy than ''H/A/α' components. The 7 parameter combination gives superior results than applying each theorem separately. Results show that SVM method discriminated each class better than Wishart supervised classification method did, especially for identifying the urban area. In the Wishart supervised classification based on 7parameters, the user's accuracy of the built-up area is very poor (54.73 %). In SVM classification based on 7 parameters, the user's accuracy of the built-up area was much higher (78.03 %).