PolSAR Image Classification Via a Novel Selective Kernel Encoder-Decoder Fully Convolutional Network

In recent years, convolutional neural networks (CNN) have made significant advances in polarimetric synthetic aperture radar (PolSAR) image classification. However, the current CNN-based PolSAR classifiers are patch-wise frameworks. The images need to be divided into overlapping patches, which have the disadvantage of repeated calculation. Additionally, the receptive field in CNN-based methods has fixed sizes, which restricts feature extraction capability. In this paper, an attention-based selective kernel encoder-decoder fully convolutional network (SK-EDNet) is introduced to PolSAR classification. Compared to CNN-based methods, the encoder-decoder fully convolutional network (EDNet) can use arbitrary-size images as input and output the whole image classification results. Meanwhile, it is utilized to extract global spatial information, which can improve the performance of classification. Besides, a selective kernel module is added to the EDNet, in which softmax attention is used to fuse several branches with different receptive field sizes. Hence, the selective kernel module can capture features with different scales and improve classification performance. The experiment results demonstrate that the SK-EDNet can obtain superior performance compared with CNN-based and traditional fully convolutional network methods.