ANALYSIS OF POLARIMETRIC FEATURE COMBINATION BASED ON POLSAR IMAGE CLASSIFICATION PERFORMANCE WITH MACHINE LEARNING APPROACH
Abstract
The polarimetric features of PolSAR images includes the inherent scattering mechanisms of terrain types, which is important for classification and other earth observation applications. By the use of target decomposition methods, many polarimetric scattering components can be obtained. Besides, the elements of Coherency/Covariance Matrix, as well as polarimetric descriptors such as SPAN, SERD/DERD etc., can also provide characteristic information. However, the computation cost will be very high if all of the polarimetric features are employed as the input of the classi fi cation process. In this paper, the effective polarimetric feature combination are studied based on the classi fi cation performance of SVM (Support Vector Machine) and NRS (Nearest-Regularized Subspace) machine learning approaches. A fast strategy on basis of correlation coefficient is used to select the features for classi fi cation experiments. For the airborne PolSAR data in Flevoland, 10 features have been selected from the total 107 polarimetric features with good classi fi cation accuracy up to 93.6%. The experiments on other data sets will be shown.