Dynamic Sparse Encoding and Cross-Temporal Attention for Remote Sensing Image Change Detection

Abstract

Due to the inherent inductive bias of operations, convolutional neural networks (CNN) cannot model global information of remote sensing (RS) images. In contrast, Transformer-based methods can establish long-range dependencies of images through self-attention (SA) mechanism, but it faces the challenges of computational complexity and memory requirements, but also ignores the exploration on the feature redundancy removal of RS images. To address these two issues, we propose a network based on dynamic sparse encoding and cross-temporal collaborative attention (DSECTCA-Net) for RS image change detection (CD). First, we implement dynamic sparse encoding (DSE) by designing hierarchical sparse Transformer module (HSTM), which decreases the correlation calculation of the SA mechanism and effectively reduces the computational complexity and parameter amount of Transformer. Secondly, we propose cross-temporal collaborative attention (CTCA) to model RS images in time series and fully explore the interactivity between dual-temporal RS images, so as to better extract the global understanding of visual scenes. Extensive experiments on two large-scale public RS datasets show that the proposed method not only provides higher detection accuracy, but also achieves lower computational complexity and required storage space than most popular CD networks.