Answers are grounded in this paper and cite their sources.
Within the golden 72 h after an earthquake, the rapid generation of emergency thematic maps is critical for disaster assessment and rescue decision-making, yet it faces core challenges such as the difficulty of reusing experience from historical earthquakes and the weak association among multi-source heterogeneous data. The spatial hallucination inherent in conventional spatial RAG systems can further mislead situational awareness of the affected region and increase decision risk. To address these issues, this paper proposes an adaptive dynamic thematic map generation method that fuses LLMs with Graph RAG for earthquake emergency scenarios. At the knowledge reconstruction level, a dual-channel association constraint mechanism over spatial and semantic channels is designed to link base geographic data, emergency resources, and historical cases across modalities, thereby constructing a directed heterogeneous event-region-resource spatial knowledge graph and achieving spatially grounded semantic consolidation of heterogeneous spatiotemporal multi-source data in emergency scenarios. At the retrieval and reasoning level, asymmetric query rewriting and multi-dimensional parallel recall over graph topology are introduced together with causal constraints of earthquake-triggered spatiotemporal propagation and proactive circuit breaking, which upgrades the text matching of conventional RAG to multi-hop spatiotemporal causal reasoning over the entire cross-domain chain of earthquake emergency response, suppresses spatial hallucination at the source of spatial decision making, and ensures the rigor of historical case reuse. At the collaborative mapping level, a multi-dimensional text fusion model of evolutionary evidence chains is constructed, which converts the discrete reasoning results of the LLM into standardized structured semantics and realizes mechanism-driven mapping of disaster semantics such as seismic intensity and lifeline engineering damage onto emergency thematic map layers. In addition, in-memory and distributed multi-level caching strategies are introduced to reduce the computational response latency when massive data floods in after strong earthquakes. Experiments on real earthquakes such as the Lushan and Luding events show that the proposed method raises the layer match rate to 0.791 (versus 0.295 for the strongest baseline), achieves the best layer-set F1 (0.252), reduces the in-window viewport-center displacement to 21.2 km, and attains a blind-judged semantic accuracy of 0.429 over the full 200-query set (0.881 on in-window queries) against 0.00 for all baselines, while a spatially filtered hybrid baseline reaches only 0.068. Driven by a lightweight 7B model, the cold-start generation of a single thematic map takes only 0.83 min, and the dynamic viewport refresh is kept within 12 s when the cache is activated, which meets the time window requirement for rapid post-earthquake map production. Given the moderate overall semantic accuracy and the disclosed out-of-graph coverage boundary, the method is positioned as an efficient and verifiable decision-support solution whose outputs still require human verification, rather than as a fully autonomous cartographic product.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: