SEED: Spatiotemporal Explainable Geospatial Artificial Intelligence for Extreme Events and Their Drivers
SEED develops a spatiotemporal explainability framework to identify and understand the key drivers of extreme events, in particular droughts and heatwaves, as learned by existing deep learning models for extreme event detection, creating an interpretable foundation for climate geo-intelligence.
- Funding
- UC3M Programa Propio de Investigación — Ayudas para jóvenes doctores/as (2025)
- Principal Investigator
- Miguel-Ángel Fernández-Torres
- Host
- Dept. of Signal Theory and Communications, Universidad Carlos III de Madrid
- Duration
- 12 months (2025–2026)
Global warming is intensifying extreme weather and climate events, increasing the frequency and severity of heatwaves and droughts. When they co-occur, these compound events interact nonlinearly to amplify impacts such as crop failure, wildfires and water scarcity. Yet extremes still lack universal definitions and consistent characterizations across their spatial, temporal and causal dimensions.
Machine learning and computer vision models can detect and localize extremes, but they act as black boxes, which limits both their interpretability and the scientific insight they provide. SEED addresses this gap with explainable AI (xAI) tailored to the spatiotemporal nature of Earth system data, aiming to uncover what these models learn about the physical drivers of extremes and how they manifest over space and time. Grounding model explanations in physical variables aims to reveal data-driven, physically grounded definitions of extremes based on their drivers.
- OBJ1Baseline model selectionSelect and validate state-of-the-art spatiotemporal deep learning models for the detection of droughts and heatwaves, providing a stable foundation for the explainability research.
- OBJ2ExplainabilityApply and adapt explainable AI (xAI) methods to pre-trained spatiotemporal models to identify which physical drivers, such as soil moisture deficits or temperature anomalies, underlie their detections.
- OBJ3Focused pilot case studiesDemonstrate the explainability framework on regional and seasonal heatwave and drought events worldwide, showing how it supports scientific understanding and impact interpretation.
- TOBJ1DisseminationShare methods and findings through publications, conference presentations and open-source code, building a community across ML/CV and geoscience.
- TOBJ2Interdisciplinary trainingFoster collaborations that train a new generation of researchers bridging advanced ML/CV and critical geoscience challenges.
- Journal · 2026Bridging the weather and climate divide with artificial intelligenceGustau Camps-Valls, Alberto Carrassi, Francisco de Melo Viríssimo, Kevin Debeire, Miguel-Ángel Fernández-Torres, Philipp Hess, et al., Veronika EyringNature Communications, 17(1), 8578
- Journal · 2026On the Predictive Skill of Artificial Intelligence-based Weather Models for Extreme Events using Uncertainty QuantificationRodrigo Almeida, Noelia Otero, Miguel-Ángel Fernández-Torres, Jackie MaArtificial Intelligence for the Earth Systems, e250113
- Conference · 2026The Risks of Imperfect Knowledge: Reliability Trade-Offs in Physics-Aware Neural Networks for Climate ProjectionsMengxue Zhang, Kai Hendrik Cohrs, Miguel-Ángel Fernández-Torres, Esther Rodrigo-Bonet, Gustau Camps-VallsInternational Conference on Artificial Neural Networks (ICANN), pp. 729–740, Springer
- Preprint · 2026Skillful Data-Driven Subseasonal Soil Moisture Forecasting: Prospects and Limits for Flash Drought PredictionNoelia Otero, Aytaç Özer, Miguel-Ángel Fernández-Torres, Jackie MaarXiv preprint arXiv:2610.07060
- Workshop co-organizerThe 5th Explainable AI for Computer Vision (XAI4CV) WorkshopCVPR 2026 · Denver, CO, USA
- Session co-convenerOSA1.5 — Machine Learning in Weather and ClimateEMS Annual Meeting 2026 · Utrecht, Netherlands · 6–11 September 2026
- Special session co-organizerSpecial Session: Multimodal Data Fusion and Visual Artificial Intelligence for Earth ObservationEUVIP 2026 · Luxembourg
This project is funded by Universidad Carlos III de Madrid through the 2025 call for research activity grants for young doctors (Programa Propio de Investigación y Transferencia).