NCDP Perspectives

Image credit: Bellwether

LEAP Wallerstein Panel Series: AI + Extreme Weather Preparedness

Based on panel presentations and discussions by Ali Ahmadalipour, Elena Xoplaki, Jatan Buch, and Jorge Pérez Aracil. 

The 2025 wildfire season in the United States is forecasted to be above normal, highlighting the need to leverage emerging technologies for hazard risk mitigation. At the LEAP Wallerstein Panel on AI + Extreme Weather Preparedness, Dr. Ali Ahmadalipour from Bellwether, a Google X company, Dr. Elena Xoplaki from the Justus-Liebig-University Giessen and the MedEWSA project, Dr. Jorge Pérez Aracil from Universidad de Alcalá, and Dr. Jatan Buch from Columbia University discuss artificial intelligence and machine learning (AI/ML) models for wildfires and heatwaves. The following is a synthesis of themes and ideas from their discussion.

AI/ML in Wildfire and Heatwave Modeling: Ongoing Initiatives

Recent advances in stochastic machine learning models, such as SMLFire1.0, demonstrate the capability to systematically integrate the diverse spatiotemporal patterns of wildfire activity across the Western US (Figure 1). SMLFire1.0 models fire frequencies and sizes 12 km × 12 km grid cells across the western United States and can reproduce observed monthly and interannual variability in fire frequencies and area burned. Additionally, while property loss is a widely recognized impact of disasters, most human exposure to wildfires is through poor air quality. These impacts are computationally expensive to model via numerical weather prediction models in real-time. However, machine learning, such as graph neural networks, reduces computational costs while enabling forecasts with higher spatial and temporal resolution. LEAP researchers utilized this approach to simulate the particulate matter 2.5 concentration from prescribed fires in California, adopting their model to identify optimal burn periods by quantifying the potential air quality trade-offs involved in conducting more prescribed fires outside the peak fire season.

MedEWSa, the MEDiterranean and pan-European forecast and Early Warning System Against natural hazards project, is another example of AI/ML being leveraged for early warning and decision support systems to address a variety of hazards—including wildfires and heatwaves—across the Euro-Mediterranean and North African region. This system leverages machine learning for risk and vulnerability assessment, providing a standardized, actionable decision-support tool for users ranging from first responders to policymakers, planners, and ministries. MedEWSa covers the full early warning value chain, from data to public safety, tackling challenges of loss and damage from extreme events and natural hazards, and enhancing resilience and reducing the impact of an area.

MedEWSa’s approach addresses the complexities of multi-hazard environments and the challenges posed by data heterogeneity, varied technological capabilities, and differing regional needs. The Euro-Mediterranean and North African region MedEWSa serves is highly populated, economically vital, and rapidly warming, with a high diversity of people and hazards. In 2018, much of Europe faced unprecedented sequences of compound events and concurrent extremes, such as extreme heat and one of the most prolonged droughts since 1500. MedEWSa took advantage of existing methods and enhanced them with machine learning to improve decision-making, using AI to calculate seasonal forecasts of different extreme events across different temporal horizons in the future and their impact on the region.

CLINT, a European AI/ML framework for the “detection, causation, and attribution” of heatwaves, droughts, and tropical cyclones, aims to improve climate services by using AI-driven algorithms (Figure 2). CLINT’s interdisciplinary teams aim to improve climate services by reconstructing past events to better understand the interaction of and predict future heatwaves. Through dimensionality reduction and temporal resolution optimization, CLINT can identify anthropogenic and climatic drivers that affect the occurrence and intensity of heatwaves. Additional frameworks identify the most important drivers of different extreme events, improving the interpretability and operational usability of forecasts.

Bellwether, a Google X Moonshot company, is working on long-term wildfire risk forecasting focused on probabilistic, multi-year hazard assessment. Bellwether looks at geospatial, environmental, climatic, and other data to forecast wildfire risk one to five years into the future across the US, Canada, and Australia, with a 100-meter resolution and quarterly updates (Figure 3). This temporality best fits the insurance industry, with Swiss Re being one of Bellwether’s largest partners in the reinsurance sector. This tool has other applications, including forestry, timber growers, municipalities, and utilities.

AI for Disaster Recovery

Post-disaster recovery is another domain where AI is catalyzing operational improvements. Bellwether is working on machine learning processes that collect aerial imagery collected by civil air patrols after major disasters, such as wildfires, floods, and tornadoes, and identify and geolocate critical infrastructure, such as hospitals, schools, and transmission lines. This process accelerates both disaster response and evacuation planning. This technology is currently utilized by a few government agencies, but the implications for infrastructure resilience and recovery logistics are applicable in many different contexts. 

Bellwether is also piloting a geospatial LLM, a conversational tool that can interact with and look at geospatial data in terms of disaster recovery, weather forecasts, and more. The team is currently testing this new tool with pilot testers who can facilitate the use and application of these sorts of forecasts in the real world and will expand the user base in the coming months.

Co-design to Address Complexity

The technical and institutional challenges of integrating AI into operational climate services are considerable. For example, MedEWSa operates in a region characterized by diverse mentalities, capabilities, and approaches, complicating efforts to unify the technologies used to detect and predict extreme events and their impacts. To overcome these challenges, MedEWSa works with a large consortium of stakeholders and users to co-design and co-create its tools. 

First, the project “starts with the end users,” employing an extensive surveying process that engages a wide spectrum of potential users to assess their current access to information and to identify their specific needs. Given MedEWSa’s goal to support a broad range of users, the team maintains an active presence across all of its eight project sites, seeking to bridge information gaps while also introducing its aims and objectives, such as enhancing the accuracy and actionability of outputs delivered directly to first responders. This highlights the critical role of humanities and social science experts, who help translate technical outputs into actionable warnings for end users. 

Still, the team continues to navigate issues related to varying data availability, quality, and resolution, as well as the different needs of information recipients. Data sharing has presented obstacles: while climate data is generally accessible, government-held data, especially data related to national security and defense, remains difficult to obtain, particularly in the sensitive Mediterranean and North African context. Many of these issues, however, have been mitigated by leveraging global data sets to supplement local gaps and through sustained trust-building with key departments. 

MedEWSa Approach to Co-design

• Start with the end user.
• At the outset, employ an extensive survey approach to learn:
  • What information do users have?
  • What information do users need?
  • What information would users like to have?
• Engage in continuous interaction with continuous feedback.

CLINT also uses aspects of co-design through early stakeholder engagement and involving end users, such as emergency managers and European planners, from the beginning of the project, incorporating their necessities, types of indices used, and more. 

Ensuring Actionable Research 

Researchers are often siloed in the academic community and do not always have the incentive mechanisms or structures in place to align their research with the needs of end users. Furthermore, the interpretability and transparency of machine learning models remain a major challenge. Continuous feedback, in which end users interact with models starting from the prototyping phase, can help develop operational frameworks for drivers of natural hazards suitable for non-expert end users. Sharing lessons learned is also extremely important. It’s valuable to disseminate accomplishments and research in open science formats that are available to everyone, not just published in academic journals. 

Cross-regional Learning and Collaboration

One major opportunity lies in cross-regional learning and collaboration. Through four paired twin pilot sites (including wildfires in Greece and Ethiopia, heatwaves, droughts, and wildfires in Catalonia and Sweden), MedEWSa fuses experience and expertise from different areas across similar hazards to understand differentiated impacts. For example, while Sweden’s systems are highly advanced and organized, Mediterranean regions have more experience with heatwaves, droughts, and wildfires, creating a rich learning environment for countries that may start facing these challenges soon.

Advancements in Datasets, Modelling, and Analysis Tools

From a technical perspective, the growing availability and resolution of environmental datasets further expand the range of actionable opportunities. Bellwether pulls from a wealth of public datasets for vegetation indices, such as NDVI and land cover, and drought conditions, typically available at 100–200 meter resolution, as well as high-resolution terrain variables (e.g., elevation, slope, digital terrain models) down to 10 meters. Climate datasets, while coarser (generally 1–4 km for the US, with frequent updates), enable large-scale modeling and forecasting. The challenge and opportunity lie in bridging and downscaling these heterogeneous datasets, leveraging the high consistency of satellite-derived climate and vegetation data, while supplementing with site-specific human factors like proximity to firefighting infrastructure.

Another opportunity is the refinement and diversification of modeling approaches. Instead of a single unified model, practitioners are developing a suite of specialized models, such as hazard models and explainability models, which generate risk scores and identify top risk drivers and suppressors, vulnerability models, and fire spread models. This modularity enables nuanced analysis tailored to different stakeholders, such as emergency managers, insurers, and reinsurers, who may have unique requirements for risk pricing, portfolio management, and resilience planning.

Advancements in geospatial analysis tools also offer new opportunities for disaster response. While tornado-induced debris can challenge computer vision models’ geolocalization capabilities, these tools are still largely effective in identifying impacted areas, even in cases of near-total inundation. For most applications, a 1 km spatial resolution is adequate, although certain analyses, such as urban heat island analysis, may require finer detail. The ability to bridge datasets and learn from varied regional experiences and hazards represents a unique moment for advancing predictive capacity, operational response, and resilience through AI-enabled climate services.

Conclusion

“It’s the greatest time to be alive. We have tools that no other generation has had access to.
The AI revolution can enable amazing upgrades and improvements in climate forecasting.” ~ Dr. Ali Ahmadalipour

As 2025 shapes up to be another pivotal year for wildfire and heatwave risk, with Europe facing a long drought and the seasonal forecast predicting a very hot summer globally, AI-driven tools are at the forefront of forecasting and analysis efforts, providing the preconditioning needed to prepare the authorities and public for challenging times. New operational models, real-time forecasting, and the democratization of high-resolution environmental data represent unprecedented opportunities for climate resilience. However, to ensure effective risk reduction and resilience planning, the rapid expansion of AI capabilities must be matched by end-to-end user engagement, transparent methodologies, and robust communication.

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).

References

Bellwether. (2025).  Accessed 25 May 2025 https://x.company/projects/bellwether/

Buch, J., Williams, A. P., Juang, C. S., Hansen, W. D., & Gentine, P. (2023). SMLFire1. 0: A stochastic machine learning (SML) model for wildfire activity in the western United States. Geoscientific Model Development, 16(12), 3407-3433.

CLINT Logical Flow. (2025). Accessed 25 May 2025 https://climateintelligence.eu/project/

MedEWSa. (2025). Accessed 25 May 2025 https://www.medewsa.eu/