DreamCD: A change-label-free framework for change detection via a weakly conditional semantic diffusion model in optical VHR imagery

• Proposes DreamCD, a change-label-free framework for VHR imagery. • Employs a weakly conditional diffusion model using pseudo-semantic masks. • Incorporates bi-temporal style differences into change image synthesis. • Introduces a large-scale semantic change detection dataset, LsSCD-Ex. • Achieves state-of-the-art unsupervised accuracy on SECOND and LsSCD-Ex datasets. The Earth is experiencing continuous anthropogenic and natural changes. Very high resolution (VHR) remote sensing imagery-based change detection provides an effective means to monitor these dynamics at fine spatial scales. Although deep learning has significantly advanced supervised change detection (CD), it heavily relies on large amounts of human-labeled samples. In real-world CD application scenarios, acquiring sufficient change samples is challenging due to the labor-intensive nature of pixel-level labeling. This challenge has motivated the development of unsupervised change detection (UCD). However, existing UCD methods still struggle in complex scenes with bi-temporal domain shifts caused by different imaging conditions. This is largely due to the absence of high-quality samples needed to guide CD-oriented optimization. To address this challenge, we propose DreamCD, a change-label-free framework that synthesizes change samples for UCD. DreamCD consists of: (1) a weakly conditional semantic diffusion model trained with pseudo-semantic masks, (2) a Content-Semantic-Style synthesis strategy that synthesizes realistic pre- and post-event image pairs of the application domain, and (3) an arbitrary contemporal deep change detector trained solely on synthetic samples. We further introduce LsSCD-Ex, a large-scale semantic change detection (SCD) dataset consistent with OpenEarthMap semantics, enabling evaluation of synthetic-sample-based SCD. Experiments on the SECOND and LsSCD-Ex datasets demonstrate that DreamCD achieves state-of-the-art (SOTA) UCD performance, improving the average F1 score by 14.01% over existing methods for binary CD and outperforming the SOTA unsupervised SCD model, Changen2, by 2.15% in F1 and 3.63% in separated kappa coefficient (SCD metric). These results suggest that DreamCD provides a promising and extensible solution for CD in real-world remote sensing applications. Code and LsSCD-Ex dataset are available at https://github.com/tangkai-RS/DreamCD .