PolSAR Terrain Classification Based on Fine-tuned Dilated Group-cross Convolution Neural Network

In the study of the terrain classification based on the Polarimetric Synthetic Aperture Radar (PolSAR), the algorithms based on general CNN do not fully utilize the phase information in different channels, and the pixel-by-pixel classification strategy with extensive redundant computation is inefficient. To mitigate these problems, a deep pixel-to-pixel mapping model in the complex domain is used for achieving a fast and accurate PolSAR terrain classification at a low sampling rate. To completely utilize the phase information, this study uses Group-Cross CNN to extend the original model to the complex domain allowing complex-number input signals and significantly improving the classification accuracy. In addition, to speed up the algorithm, a Fine-tuned Dilated Group-Cross CNN (FDGC-CNN) was adopted to directly achieve pixel-to-pixel mapping as well as improve accuracy. We verified the adopted model on two PolSAR images comprising 16 classes terrains from the AIRSAR and 4 classes terrains from the E-SAR. According to our model, the overall classification accuracies were 96.94% and 90.07% respectively while the running time was 4.22 s and 4.02 s respectively. Therefore, FDGC-CNN achieved better accuracy with higher efficiency compared to SVM and traditional CNN.