Intelligent landslide detection is crucial for disaster prevention. Existing models usually target isolated visual tasks, require massive labeled data, and lack deep analytical and scene‑understanding abilities. To overcome these limits we introduce LandslideFM, the first foundation model that unifies visual perception with contextual landslide interpretation. The model adopts a Source‑Agreement Graduated Trust semi‑supervised learning framework, leveraging a small set of labeled samples together with abundant unlabeled data to build a trust gradient and adapt to multiple visual tasks. Predicted landslide regions serve as explicit spatial cues for a Vision‑Language Model that is fine‑tuned with parameter‑efficient methods, enabling comprehensive landslide reasoning. To support training and evaluation we also release LandslideVL, the first benchmark that combines remote‑sensing imagery with newly designed textual prompts. Extensive experiments on public datasets and LandslideVL show that LandslideFM outperforms prior approaches both quantitatively and qualitatively, and surpasses general‑purpose VLMs in specialized landslide reasoning and scene analysis. Code is released at https://github.com/wangshanwen001/LandslideFM.
Review