From inundation rasters to draft action briefs: integrating spatial computing and LLM agents for urban waterlogging emergency decision support

Extreme urban waterlogging poses substantial challenges to response systems because inundation mapping, accessibility assessment, and resource instructions are often generated through separate workflows. This study presents MAEDF-UW, a decision-support framework integrating event-specific Earth observation and deterministic spatial computation with a large language model (LLM)-based multi-agent layer. For the 20 July 2021 Zhengzhou rainstorm, a 5 m GF-3 inundation footprint and a native 12.5 m ALOS PALSAR DEM were used. Water depths were estimated under a local hydrostatic assumption after resampling the DEM to an aligned 5 m grid without introducing additional topographic information. Operational depth thresholds were applied to a directed road graph to identify inaccessible road segments and modelled rescue islands. The resulting spatial facts were encoded in a structured dictionary and combined with local contingency-plan rules through retrieval-augmented agents. These agents generated schema- and rule-checked JSON action briefs for human review. The case analysis illustrated how event-specific spatial evidence and local response rules informed draft evacuation and resource-allocation instructions. The evaluation framework treats instruction consistency, spatial-input accuracy and processing latency as separate dimensions. Independent validation is still needed to establish the performance and operational effectiveness of the full workflow.