A Siamese network-based large-size remote sensing change detection network based on differential enhancement

Abstract

Existing change detection algorithms often face challenges in large-size remote sensing images, such as boundary discontinuity, insufficient correlation between semantic and change information, and inadequate extraction of differential information from dual-temporal images. To address these issues, this paper proposes a large-size remote sensing change detection network based on the design concept of differential enhancement, named DECD. By integrating attention mechanisms and staged difference extraction techniques, we have designed a large-scale dual-temporal difference enhancement module to accurately capture and enhance change features. Additionally, by leveraging the synergistic effect of change loss and segmentation loss, we have developed a segmentation-enhanced loss function, significantly improving the model’s segmentation performance. Compared with nine advanced algorithms on the WHU-CD, LEVIR-CD and MSRS-CD datasets, the F1 score of DECD was the best, reaching 90.98%, 91.75% and 76.66% respectively. In addition, the DECD inference speed was 11.78 ms, which is faster than FCCDN (15.29 ms) and Changeformer (28.78 ms).