Remote Sensing Change Detection With Forward–Backward Diffusion and Multidirectional Scanning

Abstract

With the development and research of remote sensing change detection methods, remote sensing change detection combined with deep learning has achieved excellent results. However, the existing techniques still struggle to achieve the accurate detection outcomes when confronted with challenges, such as low image resolution and noise. In addition, a significant issue remains in the detection of large continuous change regions, which often leads to leakage problems. In this article, we introduce a novel approach, diffusion scanning change detection, which integrates forward and backward diffusion processes with multidirectional scanning techniques. The input image is first preprocessed using a forward diffusion process. The backward diffusion process, along with a state-space model that incorporates multidirectional scanning, is subsequently employed during feature extraction to mitigate the adverse effects of low resolution and noise on detection accuracy. Finally, the multidirectional scanning strategy, which is enhanced by an attention mechanism, is applied in the decoder to address the leakage problem associated with large continuous change regions. The experimental results demonstrate that the proposed method significantly outperforms the existing change detection methods, as evidenced by improved performance metrics, including the overall accuracy, intersection over union, and F1-score.