DWFDA-CD:A Diffusion Model-Based Wavelet Frequency Domain Attention Network for Remote Sensing Change Detection

In Remote Sensing Change Detection (RSCD), noise is a key factor affecting detection accuracy. Remote sensing images are often affected by global noise such as environmental interference, sensor noise, and changes in lighting conditions, which can obscure subtle changes and lead to misjudgments or missed detections. At the same time, local noise such as object edges and building shadows may also interfere with the detection process. In recent years, diffusion models have made significant progress in data distribution modeling and denoising, with their gradual generation and denoising mechanisms providing new solutions for change detection. To address these challenges, this paper proposes an end-to-end framework based on diffusion models-A Diffusion Model-based Wavelet Frequency Domain Attention Network for Remote Sensing Change Detection (DWFDA-CD). This framework integrates a weight-sharing feature encoder to extract shallow features from dual-time images, reducing redundant information and noise propagation. To enhance edge features in complex backgrounds and reduce local noise caused by shadows, this paper introduces a Wavelet Detail Enhancement Module (WDEM), which is embedded into the denoising process to improve edge sensitivity. In addition, we design a Multilevel Differential Feature Extraction Module (MDFEM) to precisely focus on change areas and suppress irrelevant background noise interference. The experimental results show that DWFDA-CD significantly improves performance on four public datasets (including building and land use change detection). Specifically, on the four datasets, the F1 and IoU metrics are improved over state-of-the-art methods as follows: On LEVIR, we achieved an F1 score of 91.72% and an IoU score of 84.69%, which is an improvement of 0.45% and 0.75% over Changer; on WHU, we achieved an F1 score of 93.64% and an IoU score of 88.03%, which is an improvement of 1.10% and 1.51% over GCD-DDPM; on GZ, we achieved an F1 score of 87.96% and an IoU score of 78.51%, which is an improvement of 1.76% and 2.76% over GCD-DDPM; on CDD, we achieved an F1 score of 96.06% and an IoU score of 92.42%, which is an improvement of 1.13% and 1.86% over GCD-DDPM. This method has high demands on computational and storage resources, and is especially suitable for application scenarios with strict detection accuracy requirements, such as high-resolution remote sensing image processing and large-scale dataset analysis. To further improve practicality, future research will focus on optimizing the channel pruning strategy and adaptive diffusion sampling mechanism, significantly enhancing computational efficiency while ensuring detection accuracy.