A Diffusion Model-Based Landslide Detection Method for Optical Remote Sensing Images

As a frequent geological hazard, landslides pose significant challenges to deep learning-based detection methods due to their complex morphology and background blending. Most current object detection methods rely on empirical object priors or learnable object queries, which are difficult to fully adapt to the complexity and diversity of the landslide. To address this, we propose a diffusion model-based landslide detection method (DMLD) for optical remote sensing images. This method maps the object detection task to a denoising process. It starts from randomly generated noisy boxes that contain no learnable parameters. Then, it gradually refines these boxes to obtain the bounding boxes. To suppress background interference, we propose a gated feature fusion module (GFFM), which dynamically generates adaptive weights for fusing multi-stage features. Considering that diffusion models are fundamentally iterative processes, we develop a corrected sampling module (CSM) to improve the fast sampling precision of diffusion probabilistic models (DPMs). Experimental results demonstrate that DMLD outperforms other detection methods, confirming its effectiveness for landslide detection. The code is available at https://github.com/yuanchaosu/DMLD-GRSL.