A geoAI-driven and decision-oriented methodology for multi-hazard early warning system development

This technical paper proposes a new methodology for building smarter Multi-Hazard Early Warning Systems (MHEWS). The core argument is that existing systems suffer from a persistent "early warning without action" gap — they generate forecasts but fail to translate them into concrete decisions by the right people at the right time. To fix this, the authors propose integrating Geospatial Artificial Intelligence (GeoAI) with a decision-oriented design philosophy, moving away from broadcast-style warnings toward customised, role-specific, end-to-end decision support for multiple stakeholders across government departments and communities.

The methodology is built around four innovations: stakeholder-specific GeoAI Agents that replace generic warnings with tailored services; an evidence-based decision support mechanism structured around the "4W1H" framework (When, Where, Who, What, How); a multi-agent deep reinforcement learning (MA-DRL) algorithm enabling the system to adapt continuously to evolving risks rather than relying on static rules; and a Geographic Knowledge Graph (GeoKG) integrated with large language models (LLMs) to convert raw data into traceable, queryable decision insights. The methodology was validated through a prototype called "Smart Early Warning Town," tested at the subdistrict level in Caoyang Xincun, Putuo District, Shanghai, using real data from two 2024 typhoon events. Across seven functional modules, the system demonstrated that hourly inundation and wind speed predictions could be generated, uncertainty-quantified, and delivered as actionable, stakeholder-specific guidance — with a mean grid-level inundation prediction error of just 0.0118 metres.