Siamese Meets Diffusion Network: SMDNet for Enhanced Change Detection in High-Resolution RS Imagery

Abstract

In recent years, the application of deep learning to change detection (CD) has significantly progressed in remote sensing images. CD tasks have mostly used architectures, such as CNN and Transformer to locate image changes. However, these architectures have shortcomings in representing boundary details and are prone to false alarms and missed detections under complex lighting and weather conditions. For that, we propose a new network, Siamese meets diffusion network (SMDNet), a CD model that combines discriminative and generative architecture. By leveraging the power of the Siam-U2Net feature differential encoder (SU-FDE) and denoising diffusion implicit model (DDIM), it not only improves the accuracy of object edge detection but also enhances the data through iterative denoising and thinning reconstruction detail detection accuracy. Improves the model's robustness under environmental changes. First, we propose an SU-FDE module that uses shared weight features to capture differences between time series images, refine edge detection, and combine it with the attention mechanism to identify vital coarse features, thereby improving model sensitivity and accuracy. Finally, the progressive sampling of DDIM is used to integrate further these key features, and the adaptability of the model in different environments is enhanced with the help of the denoising ability of the diffusion model and the accurate capture of the probability distribution of image data. The performance evaluation of SMDNet on LEVIR-CD, DSIFN-CD, and CDD datasets yields validated F1 scores of 89.17%, 88.48%, and 88.23%, respectively. This substantiates the advanced capabilities of our model in accurately identifying variations and intricate details.

Keywords

• Change detection (CD)
• deep learning
• diffusion model (DM)
• remote sensing (RS)
• Siamese network