NCDP Perspectives

Source: Google Flood Hub

LEAP Wallerstein Panel Series: AI + Extreme Weather Preparedness

Based on panel presentations and discussions by Dr. Shruti Nath, Isaac Obai, Dr. Grey Nearing, and Dr. Josh DeVincenzo.

Early warning systems and anticipatory action can help to dramatically reduce the negative impacts of extreme weather events on vulnerable populations around the globe. Emerging advancements in weather and hazard forecasting enabled by machine learning (ML) and artificial intelligence (AI) present promising new solutions for adapting to increasingly frequent and severe weather events. However, several critical gaps remain that must be addressed to ensure these technologies effectively strengthen local disaster preparedness and reach the last mile.

At the LEAP Wallerstein Panel on AI + Extreme Weather Preparedness, Dr. Shruti Nath from the University of Oxford, Isaac Obai from the World Food Programme, Dr. Grey Nearing from Google Research, and Dr. Josh DeVincenzo from Columbia University, explored how AI and ML are transforming our ability to forecast, prepare for, and respond to climate disasters. The following is a synthesis of themes and ideas from their discussion.

Early Warning Systems

Early Warning Systems are “an integrated system of hazard monitoring, forecasting and prediction, disaster risk assessment, communication and preparedness activities, systems and processes that enables individuals, communities, governments, businesses and others to take timely action to reduce disaster risks in advance of hazardous events.” (United Nations Office for Disaster Risk Reduction, 2017).

Early Warning Systems (EWS) are a foundational component of disaster risk reduction (Reichstein 2025). The accuracy and impact of EWS depend not only on good data, strong hazard prediction, and an understanding of risks, but also on how quickly and clearly that information is shared and whether it leads to timely and effective decisions. In many contexts, existing systems remain reactive and fragmented. Ironically, the areas that need effective EWS the most also tend to be the most data-scarce yet disaster-prone.

AI-backed Early Warning Systems for Flood Forecasting

Many people go to Google for information about flooding. Motivated by the prevalence of flooding, the most common natural disaster affecting 19% of the world’s population, Google has invested in applied AI research to generate reliable flood predictions on a global scale. AI has helped fill gaps in areas that have historically lacked data (Matias, 2024). Flood forecasts require local data to calibrate hydrology models and stream data to calibrate riverine flood warning systems. However, the uneven distribution of open, reliable data means that places that need flood forecasts are often those where it is most difficult to build reliable predictive models (Figure 1). 

Figure 1. The inverse relationship between publicly available streamflow data in a country and national GDP (Nearing et al., 2024). 

Google Flood Hub uses AI for gap filling to create usable and reliable models for places without data (Figure 2). Google takes much of the world’s open-source stream flow data and trains one model on generalized patterns for weather forecasts and river flow. This model is applied to places that don’t have readily available data, effectively deriving regional models from global data.

AI-based forecast models have extended the reliability of global nowcasts from zero to five days. Now, Africa and Asia have improved forecasts similar to those available in Europe (Matias, 2024). Across 80 countries, Flood Hub can now provide real-time river forecasts up to seven days in advance, which can be used for anticipatory action. 

The primary value of AI in this space lies in its ability to increase forecast reliability and accuracy in data-scarce environments.

Figure 2. Google Flood Hub with a zoomed-in visualization of danger level flooding in the Peruvian Amazon River on April 22, 2025. (Google Flood Hub, 2025) 

From Early Warning Systems to Anticipatory Action

An EWS requires a robust forecasting system to enable the implementation of anticipatory action mechanisms, in which actions and finances are in place, to reduce the scale of a disaster. Traditionally, the humanitarian system has depended on a reactive approach to crises; by the time you are able to gather resources to provide aid, the situation has already compounded into a massive disaster, rendering aid largely ineffective or inadequate (Figure 3). 

Figure 3. From a reactive approach to an anticipatory humanitarian system. Adapted from Isaac Obai, World Food Programme, panel discussions slides.

Translating forecasts into action requires engaging decision-makers and clearly defining what constitutes an actionable event. The United Nations defines Anticipatory Action (AA) as “acting ahead of predicted hazards to prevent or reduce acute humanitarian impacts before they fully unfold” with the incorporation of a pre-agreed trigger, pre-agreed activities, and pre-arranged financing (UN-OCHA 2025).

Strengthening Early Warning Systems for Anticipatory Actions (SEWAA), funded by Google.org, is a collaboration between the World Food Programme, Oxford University, the Intergovernmental Authority on Development Climate Prediction and Applications Centre (ICPAC) and the European Centre for Medium-Range Weather Forecasts (ECMWF) along with the Kenya Meteorological Department and the Ethiopia Meteorological Institute. The project integrates AI into regional and national weather models to improve rainfall forecast accuracy, leading to high-resolution predictions and reducing the need for costly supercomputers (Oxford University, 2024). These improved, low-resource forecasts are being integrated into an anticipatory action framework to help prepare and mitigate the risks of extreme weather events. According to a representative at the Kenya Meteorological Department, the forecasts demonstrate significant improvements in accuracy and are being compared with current operational methods (Oxford, 2024).

Even though EWS development is expected to cost only one-tenth of the losses and damages they offset (Global Commision on Adaptation, 2019), economic value must be communicated in terms end users understand: by weighing the cost of anticipatory action versus the estimated loss caused by inaction, based on defined climate or weather thresholds. By analyzing historical forecast performance against various cost-loss scenarios, stakeholders can determine when forecasts offer the greatest value and guide decisions accordingly (Figure 4).

Figure 4. The Anticipatory Action (AA) matrix was adapted from MacLeod et al. (2021) with the relative economic value framework integrated based on Shruti Nath’s panel discussion.

Why does anticipatory action look different in different parts of the world?

Although the Sendai Framework emphasizes the need to reduce fatalities and loss of life, federal disaster response in the U.S. is primarily governed by the Stafford Act. This framework places the burden of decision-making and responsibility on a single individual or organization, typically an emergency or disaster manager, who must determine when and how to act. As more decision-making authority is shifting to localities, it is concentrating greater responsibility in the hands of fewer individuals who have more demanding work but fewer resources. AI-powered EWS can supplement needed information and bridge the resource gaps emergency managers face.

AI Value-Add: Low Computational Needs and Addressing Data Gaps

AI models are much lighter weight than traditional models and can fill in for areas where in situ hydrometeorological data is scarce or unevenly distributed. Limited area models are regional models that can post-process finer-scale regional features to explore more extreme localized impacts. However, limited area models are computationally heavy: they require a lot of computational infrastructure and time to run. This limits the ability to generate multiple forecasts to explore uncertainty and to generate reforecasts to test the accuracy of forecasts based on historical data. AI can post-process global models to provide localized regional extreme features and to explore the uncertainty, which can ultimately better inform regional stakeholders (Palmer 2020). SEWAA uses a generative AI-based post-processing system that ingests the outputs of the Integrated Forecast System global model from the ECMWF to produce localized rainfall forecasts specialized onto regionally accurate observational data. This allows for a much more robust exploration of the uncertainty space with minimal computational cost. 

What’s needed for AI to effectively inform EWS and Anticipatory Action in practice?

Collaboration and trust building

National agencies are mandated to produce reliable, historically continuous operational forecasts, and they can be hesitant to embrace new ML research methods. It’s critical to build trust in the models, demonstrating that they can reliably predict extreme events at a localized level with sufficient lead time (World Food Programme, 2023). There is also a need to demonstrate the value-added benefits of ML models, showing improved accuracy, reliability, and cost-effectiveness. If this cannot be demonstrated, there may be hesitation or reluctance to fully adopt the new technology, despite its benefits. Here, partners in academia and the private sector can undertake some of this higher-risk research that can then be transferred to the national scale. Lingering challenges that still need to be addressed include communication delays, technology gaps, and reluctance to share data.     

Local ownership and integration

“If you don’t work with the National Meteorological and Hydrological Services (NHMS) then you cannot operationalize anything you are doing with Artificial Intelligence” – Isaac Obai, World Food Programme

While improved forecasts are being developed, it’s critical to consider integration: Where will the AI-enhanced forecasts reside in relation to decision makers?

The needs of every government partner are different, and it can be a challenge, especially when building global flood forecasting models and alerts, to create a bespoke solution for each partner. While some governments may welcome Google issuing alerts directly to the public, others prefer to integrate Google’s models into their own forecasting workflows or maintain full control over alert dissemination. Here, providers can define the range of options and design a system that tries, as efficiently as possible, to accommodate the varied needs. The end goal is for local partners to take full ownership of the forecasts, leveraging technological advancements offered by cloud computing. In the meantime, real-time communication with local actors continues to pose a challenge to operationalizing these forecasts and alerts.

The World Food Programme serves as a convener, facilitating conversations among academia, regional climate centers, and National Meteorological and Hydrological Services to enhance climate forecasting. However, the sustained use of forecasting systems depends on national governments assuming ownership and stewardship to ensure their long-term viability. Strengthening Regional Climate Centers, such as ICPAC, which can cascade capacity strengthening to their member countries’ National Meteorological and Hydrological Services, also offers a sustainable way to build and scale AI-enhanced forecasting capacity in lower-resource settings, while still providing local ownership and capacity strengthening to the regions.

Closing takeaways

AI-driven forecasts have enhanced accuracy and precision, addressed data gaps, and reduced computational time and costs. By leveraging diverse data sources, these models generate more holistic and nuanced insights than traditional approaches. When paired with strong communication and collaboration mechanisms, these advancements create opportunities for bottom-up adoption by local governments, ultimately supporting more informed policymaking for early warning systems, disaster management, and anticipatory action.

This was produced by the National Science Foundation’s Center for Learning the Earth with Artificial Intelligence and Physics (LEAP) (Award #2019625)  in collaboration with the National Center for Disaster Preparedness (NCDP).

Leap Flood Research Highlight

Dr. Candace Agonafir, an associate research scientist at LEAP, studies improving urban flood forecasting using AI and data science. Her work aims to help cities like New York better predict and manage street-level flooding. Dr. Agonafir’s work has analyzed New York City resident reports of street flooding, sewer basins, and clogged catch basins to identify areas prone to flooding, and she has used machine learning models to determine which factors (e.g., rainfall intensity, infrastructure issues, and neighborhood characteristics) most significantly contribute to flooding. Recently, Dr. Agonafir has explored advanced models that can capture spatial and temporal aspects of flooding, which show how water moves through urban landscapes over time, leading to more accurate and timely flood predictions.

References

Global Commission on Adaptation. (2019). Adapt Now: A Global Call for Leadership on Climate Resilience. Access 22 April 2025 https://gca.org/wp-content/uploads/2019/09/GlobalCommission_Report_FINAL.pdf

Google Flood Hub. (2025). Accessed 22 April 2025. https://sites.research.google/floods

MacLeod, D., Kniveton, D. R., & Todd, M. C. (2021). Playing the long game: Anticipatory action based on seasonal forecasts. Climate Risk Management, 34, 100375.

Matias, Y. (2024). Using AI to expand global access to reliable flood forecasts. Accessed 22 April 2025. https://research.google/blog/using-ai-to-expand-global-access-to-reliable-flood-forecasts/

Nearing, G., Cohen, D., Dube, V., Gauch, M., Gilon, O., Harrigan, S., & Matias, Y. (2024). Global prediction of extreme floods in ungauged watersheds. Nature, 627(8004), 559-563.

Oxford University. (27 June 2024). New AI-led science initiative will help protect communities hit by climate change in East Africa. https://www.ox.ac.uk/news/2024-06-27-new-ai-led-science-initiative-will-help-protect-communities-hit-climate-change-east

Palmer, T. (2020). A Vision for Numerical Weather Prediction in 2030. Arxiv Physics. https://doi.org/10.48550/arXiv.2007.04830 

Reichstein, M., Benson, V., Blunk, J. et al. Early warning of complex climate risk with integrated artificial intelligence. Nature Communications, 16, 2564 (2025). https://doi.org/10.1038/s41467-025-57640-w

United Nations. (2025, March 26). Anticipatory action. Office for the Coordination of Humanitarian Affairs (OCHA). https://www.unocha.org/anticipatory-action 

United Nations Office for Disaster Risk Reduction (UNDRR). 2017. The Sendai Framework Terminology on Disaster Risk Reduction. “Early warning system”. Accessed 22 April 2025.  https://www.undrr.org/terminology/early-warning-system.

World Food Programme. November 2023. Machine Learning for Early Warning Systems. Accessed 22 April 2025. https://docs.wfp.org/api/documents/WFP-0000154485/download/?_ga=2.259653258.1231888055.1741730087-1905278763.1741730087