The distinctive characteristic of synthetic aperture radar (SAR) sensors is the ability to provide a day-or-night, all-weather means of remote sensing. Recent SAR systems can produce high-resolution images of the land under the illumination of radar beams. SAR polarimetry is a technique that employs different polarization waves during transmission toward and reception from the Earth's surface and the resultant PolSAR images can be used in identification of different classes based on analyzing different polarization backscattering coefficients; by assigning pixels into different classes using a classification technique, the information contained in the SAR/PolSAR images can be interpreted.

Classifier ensembles or multiple classifier systems (MCS) are methods in pattern recognition that are used for image classification; by combining different independent classifiers, MCS can improve classification accuracy in comparison with a single classifier. There are different methods for creating such an ensemble. These methods include modifying the training samples (e.g. bagging [1] and boosting [2]), manipulating the input features (the input feature space is divided into multiple subspaces [3]), and manipulating the output classes (multi-class problem is decomposed into two multiple class problems, e.g. the error correcting output code [3]). After creating an ensemble of classifiers, a decision fusion is used to combine the outputs of the classifiers. Several fusion algorithms have been developed and employed in the literature like majority voting, fuzzy integral, weighted summation, consensus, mixed neural network, and hierarchical classifier system [4], [5]. Class-based feature selection (CBFS) is a method that chooses features for each class separately to create a multiple classifier with manipulating input features. We used this method for pixel-based classification, and then fused single classifiers in two different ways described in the next section.

Experimental results showed that the overall accuracy of the proposed method (90.07%) has improved compared with the single SVM classifier and pixel-based multiple SVM classifiers (83.61%).