Dual Attention-Based Difference Amplification Shrinkage Network for Unsupervised Change Detection in PolSAR Images

Image change detection is crucial in polarimetric synthetic aperture radar (PolSAR) image analysis and interpretation. However, purely convolutional neural network (CNN)-based methods cannot efficiently capture highly discriminative change-aware features owing to the interference of coherent speckles. Here, an unsupervised PolSAR image change detection method based on channel-spatial dual attention is proposed in this article, combining the advantages of difference information and attention mechanisms. First, the Wishart distance-based difference image is clustered to automatically construct pseudo-labels required for unsupervised change detection. Subsequently, a dual attention-based difference amplification shrinkage network (DADASN) is developed to suppress the speckle noise and amplify the discriminability of the change-aware features. On the one hand, the channel attention (CA) mechanism uses difference information to obtain channel shrinkage thresholds and combines them with the soft threshold to suppress noise and remove change-independent features. On the other hand, the spatial attention (SA) mechanism is adopted to capture the regional features with significant changes, thereby amplifying the discriminability of the change-aware features and improving the positioning accuracy. Finally, the cross-entropy loss function is used to supervise the training process of the network for final change detection. Experimental results on two real spaceborne Gaofen-3 PolSAR datasets demonstrate the superiority of the proposed method over other state-of-the-art methods.