Landslideagent with multimodal landslidebench: a domain-rule-augmented agent for autonomous landslide identification and analysis

Intelligent identification and analysis of landslides are critical for disaster prevention. However, current interpretation paradigms struggle to simultaneously extract visual features and geoscientific semantics, while general-purpose vision–language models (VLMs) suffer from perceptual limitations and domain hallucinations in complex geological scenarios. This paper proposes an instruction-driven agentic framework for landslide response that integrates a multimodal benchmark, a specialized VLM, and a rule-enhanced agent. First, a multimodal fine-grained landslide dataset LandslideBench with seven landslide subtype labels, high-resolution imagery, pixel-level masks, and high-quality textual descriptions is constructed through multi-VLM cross-validation and interactive annotation. Second, a landslide-oriented dual-head vision-language method, LandslideVLM, is proposed and fine-tuned on LandslideBench via low-rank adaptation, in which a discriminative classification head and a generative language head are integrated to jointly support fine-grained landslide classification and geoscientific semantic description. Finally, a domain-rule-enhanced LandslideAgent is developed by integrating structured report metadata constraints and cross-validation identification constraints. Experimental results demonstrate that LandslideBench can serve as an effective benchmark for landslide interpretation. The fine-tuned LandslideVLM achieves accuracy improvements of 12.79%, 55.71%, and 17.27% on landslide discrimination, fine-grained classification, and semantic description quality, respectively. Across 28 paired cases, rule-constrained LandslideAgent achieves an accuracy of 88.7% across six evidence-checkable fields, outperforming free-mode agent by 11.9%.