Debris flow disaster information representation and perception based on knowledge graphs and virtual geographic environments

This paper proposes a 3D visualization system for public debris flow risk communication, integrating knowledge graphs built via LLM, virtual geographic environments (VGE), and physical simulation using shallow water equations. The system is designed to adapt to different user profiles (general public, displaced persons, rescue workers, decision-makers) and heterogeneous platforms through LOD technology and quad-tree indexing within the Cesium engine.

Experimental validation conducted on 120 participants using eye-tracking demonstrates that the animated format significantly outperforms static images and text in both information recognition accuracy (79.2% vs. 56.8% vs. 31.5%) and task completion time, with highly significant statistical differences. The fine-tuned LLM (Qwen2.5-3B) achieves an F1 score of 0.79 in entity extraction, outperforming traditional BiLSTM-CRF, while the system maintains smooth frame rates (>70 FPS on smartphones, >90 on notebooks). Key limitations include validation restricted to debris flow disasters only and a homogeneous university-based sample.