PolSAR image classification using feature fusion algorithm based on feature selection and bilayer SVM

Single type of feature vector cannot fully describe objects, in order to fully use the rich object information of polarimetric SAR images and solve this problem, this paper put forward a novel feature fusion algorithm based on feature selection and bilayer SVM for polarimetric SAR image classificationthat can make full use of the completeness and dissimilarity between the features to form a more effective feature vector. Various types of feature vectors were extracted from an original image by different methods for fully describing the PolSAR data. The feature vectors were normalized to ensure each feature vector can be selected under the same standards and have the same role in classification. A spatial pyramid is introduced to get the feature vector in different size or spatial location. A mRMR feature selection method was used to obtain the optimal feature subset for given categories to avoid redundancy and overfitting phenomenon caused by the simple combination of various feature vectors. Finally, the multilayer concept was introduced and a bilayer SVM model was constructed to optimize and re-process the probabilities of the target category obtained by the first SVM. Experimental results on the two polarimetric SAR images achieved by the Jet Propulsion Laboratory show the superiority of the proposed approach.