PolSAR Image Classification based on Complex-value Linear Attention Network

Advantages of active imaging and providing more diverse features with multiple polarizations have promoted the utilization of polarimetric synthetic aperture radar (PolSAR) in land cover classification. Despite the great success achieved by convolutional neural network (CNN) with attention mechanism before, there are still two major concerns. One is the real valued CNN would undoubtedly discounts the richer representations in complex-valued PolSAR and the other is the challenge of implementing attention in complex domain. In this paper, we design a complex linear attention mechanism which can be fully executed in the complex domain and then combine it with the complex-valued CNN, to extract abundant feature representations in the complex domain. Experiment on the public PolSAR dataset shows the prominent performance of the complex-valued CNN equipped with complex linear attention, which surpass the real-value CNN by 4.59%, 5.22%, 4.49% in terms of overall accuracy, average accuracy and mIoU.