Cross-Temporal Knowledge Injection With Color Distribution Normalization for Remote Sensing Change Detection

Remote sensing change detection plays a critical role in monitoring and identifying alterations on the Earth's surface by comparing images captured at various time intervals. Although existing change detection methods have yielded promising results, challenges such as imaging discrepancies arising from seasonal changes, sensor variations, and numerous confounding pseudochanges in cross-temporal imaging significantly hinder their accuracy. To address these problems, we propose cross-temporal knowledge injection with color distribution normalization (CICD) method for remote sensing change detection. In CICD, color alignment for bitemporal image is designed to leverage channel statistics from bitemporal images to smoothly align the color distribution differences. Subsequently, after refining the granularity of multiscale features extracted from the backbone using learnable weights with varying preferences. The proposed bitemporal complementary knowledge cross-injection enhances semantic and temporal understanding of genuine change regions by cross-injecting spatial-layout knowledge and detailed knowledge into the refined bitemporal features. Q-mapping with fixed anchor and attention temperature scaling are employed to address query redundancy and smooth out cross-attention issues during the knowledge injection. The resulting features are processed using our proposed multidirectional change perception, which facilitates channel separation and enables independent change prediction across different channels by applying perceptual weights with varying shape preferences. This approach overcomes the limitations of fixed shape weights that lead to a singular perception of changes. Our approach outperforms nine state-of-the-art methods on three widely used high-resolution remote sensing change detection datasets.