Change-prior guided cross-scale interaction network for remote sensing image change detection

Change detection (CD) identifies differences in remote sensing imagery of a specific location across time periods, serving critical functions in environmental monitoring, disaster response, and other applications. As a core sub-task, binary change detection (BCD) labels each pixel as changed or unchanged. Currently, deep learning-based BCD methods are mainstream. However, they face a critical challenge: changes of interest (positive samples) are extremely sparse and often overwhelmed by numerous task-irrelevant variations. This leads to severe sample imbalance and noise interference. Existing approaches still have limitations in addressing this issue. On the one hand, current attention mechanism-based solution often lacks explicit prior guidance, causing them to incorporate irrelevant interference into the feature enhancement process and thus dilute the focus on target changes. On the other hand, multi-scale fusion-based solutions typically rely on simple feature concatenation or independent branches, failing to achieve deep interaction among cross-scale features at the channel level. To address these challenges, this paper proposes a change-prior guided cross-scale interaction network (CGCSNet). The network comprises two core modules. First, the change-prior guided attention module (CPAM) leverages prior relationships between changes and the scene to guide global feature aggregation, effectively suppressing irrelevant interference. Second, the cross-scale channel interaction fusion module (CIM) promotes deep interaction and fusion of features from different scales at the channel level through parallel multi-scale convolutions and a channel shuffle mechanism. Through the synergy of these two modules, CGCSNet effectively mitigates interference from high-variance irrelevant changes and addresses the problem of sparse positive samples. Comprehensive evaluations on the publicly available datasets LEVIR-CD+ and BANDON show that CGCSNet consistently surpasses state-of-the-art methods.