A wrapper feature selection for the polarimetric SAR data classification

The main objective is to propose a wrapper feature selection algorithm for analyzing the polarimetric SAR data for forest mapping. The method is based on the concept of feature selection and classifier ensemble. Due to its ability to take numerous and heterogeneous features into account, the support vector machine (SVM) algorithm is used as the classifier. The limitation of SVM as the evaluation function for feature selection is its time-consuming optimization. To accelerate the SVM training process, a training sample reduction strategy based on the notion of support vectors is proposed. Two fine quad-polarized Radarsat-2 images, which were acquired in leaf-on and leaf-off seasons, were chosen for this study. A wide range of SAR parameters were derived from each PolSAR image. A combined dataset was also considered. The classification results (in terms of the overall accuracy) compared to the baseline classifiers demonstrate the effectiveness of the proposed wrapper scheme for forest mapping.