A Spatial–Spectral Difference Guided DDPM for Hyperspectral Image Change Detection

With the advancement of hyperspectral imaging technology, hyperspectral image (HSI) change detection (CD) has become a vital tool for monitoring the dynamics of Earth’s surface. However, existing HSI CD methods struggle to detect subtle and boundary changes, and often suffer from false positives (FPs) and missed detections in complex backgrounds. Effectively extracting spatial and spectral information, accurately identifying fine-grained changes, and generating high-precision detection results remain significant challenges. To this end, this article presents a spatial–spectral difference guided denoising diffusion probabilistic model (DDPM) for HSI CD, named SSDGD. Unlike traditional HSI CD methods that classify each pixel into either the changed or unchanged category, our SSDGD leverages the progressive denoising mechanism of DDPM to directly generate comprehensive CD results, effectively using difference features to precisely detect subtle changes and boundary variations. Specifically, we propose the spatial Convolution-Mamba (SpaCM) and spectral Convolution-Mamba (SpeCM) modules to enhance spatial and spectral feature extraction while reducing redundancy. Furthermore, we design a difference attention (DA) module, which incorporates an attention mechanism to focus on significant changes, filter out irrelevant information, and reduce information loss simultaneously. Finally, a spatial–spectral difference guidance (SSDG) module leverages spatial and spectral difference features extracted from bitemporal HSIs to guide the progressive denoising and refinement of DDPM results, ultimately reconstructing high-quality HSI CD outcomes. Experiments on multiple benchmark datasets demonstrate the superiority of our SSDGD over existing approaches and highlight its effectiveness.