Mitigating False Positives in Complex Spectral Scenarios: A Mutual Information Guided Wavelet Disentanglement Framework for Remote Sensing Change Detection

Abstract

Remote sensing change detection (RSCD) constitutes a fundamental task for quantifying spatiotemporal variations on the Earth’s surface through bitemporal image analysis. Despite significant advances in deep learning methodologies, existing approaches often exhibit persistent limitations in complex scenarios characterized by intricate grayscale variations, particularly manifesting as elevated false positive rates in edge and shadow regions where illumination artifacts obscure genuine change signatures. To systematically address these challenges, we propose ACNet, a graph-enhanced adaptive wavelet transform network with cross-layer knowledge injection and mutual information regularization. The proposed framework integrates an adaptive wavelet decomposition with graph convolution module that combines discrete wavelet transforms with graph neural networks in non-Euclidean space to comprehensively capture multiscale frequency-domain features while modeling complex spatial relationships; a mutual information regularization mechanism that encourages latent cross-subband decoupling and reduces interfrequency redundancy, thereby minimizing interfrequency redundancy and enhancing discriminative feature representations; and a dual-temporal differential feature complementation and alignment module that performs sophisticated multiscale feature integration through bidirectional cross-attention mechanisms, ensuring optimal fusion of fine-grained spatial details with high-level semantic information. Extensive experiments on five widely-used benchmark datasets demonstrate that ACNet consistently outperforms ten state-of-the-art methods, capable of detecting both subtle and large-scale changes while substantially mitigating false positives in challenging scenarios with complex spectral variations.