UNSUPERVISED POLARIZATION FEATURES SELECTION NETWORK FOR POLSAR IMAGE CHANGE DETECTION BASED ON ATTENTION MECHANISM

Polarimetric synthetic aperture radar (PolSAR) can obtain rich polarization characteristics of observed targets. Therefore, its application to land cover change detection has become a research focus. Given the scarcity of polarization SAR data, this paper proposes an improved unsupervised PolSAR image change detection network, referred to as the Polarimetric Feature Selection Network (PFS-Net). In PFS-Net, using the channel attention mechanism, a simple and flexibly pluggable PFS block is designed to augment the polarization parameters of the target. These polarization parameters are fused with the spatial location to achieve the extraction and selection of specific polarization features of different ground objects. Then, the residual backbone network captures advanced semantic features, performs unsupervised clustering on deep features to obtain pseudo labels, and dynamically adjusts the weight of PFS-Net by calculating loss backpropagation through pseudo labels. An experiment based on the fully polarized data obtained from the fully polarised data of RADARSAT-2 on 23 May and 3 August 2013, which covers the farmland area in the Inner Mongolia Autonomous Region of China is carried out. The experiment result shows that the proposed algorithm can effectively improve the accuracy of the change detection results.