Connecting the world and beyond

Leveraging AI to Enhance Multi-Hazard Early Warning Systems


various hazards - drought, storm, cyclone





​​​​​​​​​​​​​







Leveraging AI to Enhance Multi-Hazard Early Warning Systems ​discusses​ the use of artificial intelligence for early warning systems (EWS). The report, produced by the Early Warnings for All​ (EW4All) initiative, demonstrates AI's potential to address existing gaps in EWS and to make the systems more effective, resilient, and inclusive. 

The report builds on the work of the AI for EW4All Group​ and allow​s policymakers, humanitarian partners, and local leaders to leverage emerging technology for protecting lives and livelihoods.​​​​​​​​​​​ Besides highlighting concrete solutions and pilots, the report emphasizes that AI is not a solution in isolation but requires strong infrastructure, partnerships, and a human-centered and responsible approach.

How AI supports early warning systems

AI is strengthening the entire early warning value cycle by (1) enhancing disaster risk knowledge, (2) improving hazard detection and forecasting, (3) enabling more targeted and inclusive warning dissemination, and (4) supporting preparedness and anticipatory action. 

Across these stages, AI processes large and diverse datasets to improve risk assessments, delivers faster and more accurate forecasts, tailors warning messages to different populations, and optimizes decision-making for early action. Realizing these benefits depends on high-quality data, interoperable systems, robust governance, transparency, and human oversight to address biases, data gaps, limited connectivity, and uncertainty. 

Future investment should therefore prioritize end-to-end or integrated, community-centred, and ethically deployed AI solutions that connect the entire early warning value cycle.​

Featured case studies

Strengthening disaster risk knowledge in Tonga's Tongatapu Island: Limited​ data on buildings, mangroves, and other critical assets have constrained risk assessment and preparedness​. The use of AI-driven tools, including a pre-trained AI model to map and count critical assets, inc​luding buildings and coconut trees, has produced a scenario analysis and ​realistic inundation simulation to produce information on vulnerable assets, allowing ​authorities to ​significantly improve local evacuation planning.​ 
Zero min simulation of Tonga floods
Four min simulation of Tonga floods











Images above show the AI-simulated Digital Twin of Tongatapu, allowing scenarios for sea level rise and flooding within a four minute gap.

Advancing forecasting and hazard detection during Hurricane Melissa: ​To improve tropical cyclone track and intensity predictions, the National Hurricane Center of the United States has examined experimental machine learning (ML)–based guidance alongside conventional dynamical models. An experimental ML tropical cyclone ensemble forecasting system developed by Google​ generates probabilistic predictions of storm track and intensity by producing up to 50 ensemble forecast scenarios. Forecast discussions and subsequent analyses indicated that a large portion of the Google ensemble members projected that Hurricane Melissa in October 2025 could intensify to Category 5 strength, highlighting the possibility of extreme intensification and providing additional probabilistic guidance on storm evolution.

Improving warning dissemination and communication in Liberia: Under the AI for EW4All Group​, ITU, in collaboration with Microsoft, IHME, and Planet Labs, developed the Early Warning Connectivity Map​ (EWCM​) to identify connectivity coldspots where populations lack the network access needed to receive emergency notifications. ​Integrating a 100-metre resolution population dataset generated using AI, the tool was deployed in Liberia. By combining ITU data with cell site details provided by the Liberia Telecommunications Authority and mobile operators, the system generated precise maps to show where people were covered by cellular networks. This information was combined with maps on flood-prone areas to show where people at risk were not reachable via mobile alert systems. 
EWCM in Liberia



















Supporting preparedness and anticipatory action through Sketch Map: ​Despite legal recognition granting community personhood to protect Colombia's Río Atrato Basin, data sources remain scarce and fail to incorporate ancestral knowledge and traditional conservation practices. This participatory mapping methodology digitizes community data through maps that overlay with authoritative sources for integrated conservation analysis. Using the AI-embedded Sketch Map Tool, communities lead workshops where participants draw conservation project locations on paper maps, with AI automatically detecting sketches to generate digital datasets integral to consultation processes.

Enabling an integrated end-to-end approach​​ to EWS: Some AI-based solutions apply across the entire early warning value chain. China's Multi-hazard, Alert, Zero-gap, Universal (MAZU) AI Agent for early warnings, developed by the Shanghai Meteorological Service and the China Meteorological Administration, is one such example, integrating advanced algorithms and multi-source data — including satellites, radar, and localized AI models — to support every stage of the early warning process. 

China MAZU system






















Aligned with the four EW4All pillars, seen in image above, MAZU provides online access to multi-hazard risk assessment tools and databases for disaster risk knowledge; flexible monitoring and forecast analysis tools for detection and observation; AI-generated disaster condition reports for warning dissemination; and role-based, disaster-specific emergency guidelines for preparedness and response. A "three-terminal" architecture tailors this to different users — an all-in-one machine for professional departments, a tablet for sectors like ports and shipping, and a mobile app that delivers location-based alerts and evacuation guidance directly to the public.​

Recommendations

​​​To support responsible, transparent, and deeper integration of AI into EWS​, the report identifies priority actions:
Icon for robust infrastructure









Robust observational infrastructure — ground networks, satellites, and in-situ sensors — must underpin AI performance, particularly in SIDS, LDCs, and LLDCs.​

Icon for AI governance









AI integration needs strong governance: national focal points, human oversight for life-safety decisions, and clear accountability frameworks.

Icon for people centric









AI design must be human-centred and equity-driven, embedding humanitaria​n principles, ensuring multilingual and low-connectivity compatibility, and co-designing tools with affected communities

Icon for integration









EWS should be built for interpillar integration from the outset, using modular, interoperable architectures and feedback loops that let data from one pillar strengthen the others — risk models, forecasts, communication, and preparedness all reinforcing each other.
Icon for investment









Investments in AI have enormous potential and sustained funding is essential to move AI from pilot projects to operational scale, supported by institutional capacity-building and partnerships across governments, the private sector, and research institutions.​
Icon for targetted approach









Efforts to integrate AI into EWS must focus on addressing existing gaps.

Background

Climate change is escalating the frequency and complexity of extreme weather events, making multi-hazard early warning systems (MHEWS) more critical than ever. While 128 countries now report some capacity for MHEWS, many communities remain unprotected or insufficiently reached by life-saving alerts, especially in many of the most vulnerable countries. 

Barriers include infrastructure and data limitations, and gaps in governance, human capacity, and financing.