Polsar Image Classification via Complex-Valued Multi-Scale Convolutional Neural Network

Convolutional neural networks (CNNs) have achieved promising results in polarimetric SAR image classification. Generally, the semantic segmentation of a large image is conducted using the image slices, for which the small slices may be in a single pure class but with insufficient information, and the large ones may contain the mixed class. Therefore, a complex-valued multi-scale CNN (CVMS-CNN) architecture is proposed to extract the hierarchical multi-scale information, i.e. local and global features, and adapt to the complex PolSAR data format, simultaneously. Moreover, the optimal feature fusion mechanism is given through comprehensive comparisons. Experiments are carried out on two benchmark datasets to verify the effectiveness. Numerical simulations show that the classification results have been significantly improved via CVMS-CNN compared with the state-of-the-arts.