ADDEDNet: Remote Sensing Image Change Detection Algorithm Based on Multiple Attention Mechanisms

Remote sensing change detection is the analysis of remote sensing images acquired at different times to identify and evaluate the dynamic changes of surface features [1]. It plays a key role in many fields such as agricultural production management, natural disaster assessment, environmental monitoring [2], urban development planning [3], and resource scheduling. At present, the dual encoder structure uses a set of Siamese networks with shared weights to extract features separately. This structure only extracts the independent features of each time phase, which lacks sufficient temporal context information and affects the detection effect. The attention mechanism is fixed for the dependence on channels and spatial dimensions, and it is forced to capture all information, which is inefficient and not suitable for remote sensing images with different sizes of ground objects, unable to extract sufficient context information, and unable to effectively fuse bi-temporal features. Based on these problems, this paper proposes a remote sensing image change detection algorithm based on hybrid attention mechanism (ADDEDNet). The channel interaction strategy is introduced into the deep dual encoder structure, multiscale strip convolution construction is introduced into each encoder block to dynamically allocate attention weights, and the adaptive attention mechanism is introduced into the decoder to enhance feature expression. And the dual-time information adaptive fusion module of feature fusion is dynamically adjusted in the fusion process. Experiments on LEVIR, SYSU and CDD datasets prove the effectiveness of the model.