Cooperative Alignment and Boundary Refinement Network via Multisource Prior Synergy for Optical Remote Sensing Change Detection

Abstract

Optical remote sensing change detection aims to identify changed pixels between bitemporal images of the same region, holding immense application value in fields such as urban planning and disaster assessment. However, disrupted by factors including seasonal transitions, illumination variations, and sensor viewpoint discrepancies, identical ground targets frequently exhibit pronounced spectral heterogeneity across bitemporal images. This heterogeneity hinders existing methods from effectively extracting robust change features, consequently leading to detection results frequently suffering from pseudochange responses and edge blurring. To address these challenges, we propose an RSCD architecture based on multisource prior synergy, named the Cooperative Alignment and Boundary Refinement Network, or CABRNet. Specifically, during the encoding phase, the Cooperative Spatiotemporal Alignment Module is embedded after each feature extraction layer. By aggregating cooperative representations of bitemporal features across both channel and spatial dimensions and modeling interchannel correlations, this module achieves dynamic feature alignment. Consequently, it extracts robust bitemporal features. During the feature fusion phase, the Confidence-Driven Boundary Modulator deeply integrates the physical high-frequency priors extracted via a Laplacian pyramid with the deep confidence priors generated through deep supervision. This integration collaboratively optimizes the edge representations of the fused features. During the decoding phase, the Signal-Preserving Context Aggregation Module is adopted to aggregate features. This strategy effectively mitigates the information loss induced by upsampling. Simultaneously, it enhances the reconstruction integrity of change regions and provides high-quality confidence priors for boundary modulation. Experimental results demonstrate that the CABRNet achieves F1 scores of 92.03%, 97.96%, and 83.92% on the LEVIR-CD, Lebedev, and SYSU-CD datasets, respectively. These results outperform existing state-of-the-art methods.