Multifeature-based PolSAR image classification using superpixels and DBN model

To mitigate the effects of coherent speckle noise on polarimetric SAR (PolSAR) images and obtain consistent classification results, we propose a novel supervised multifeature-based classification method. This approach integrates polarimetric signature information and texture information based on superpixels and deep belief network (DBN) model. Initially, the improved simple linear iterative clustering (SLIC) algorithm is employed to generate superpixels, which serve as the classification elements. Subsequently, a high-dimensional PolSAR feature image is constructed. The polarimetric features are obtained through polarimetric decomposition and algebraic operations, while the gray level cooccurrence matrices (GLCMs) are employed to represent the polarimetric texture information. The DBN model is then trained using the PolSAR feature image and a set of training samples.The final classification result is derived by employing superpixels as classification elements to compute the class probabilities, and the category of each superpixel is determined according to the maximum posterior criterion. Comparative experiments conducted using an airborne AIRSAR image and a spaceborne RADARSAT-2 image were performed to assess the applicability and validity of the proposed approach. The results indicate that the proposed approach achieved the highest accuracy, significantly surpassing that of other methods. The experiments demonstrate that the proposed approach effectively mitigates the interference of speckle noise, yielding consistent and accurate classification results for PolSAR images.