AI in Climate Risk Assessment

Abstract

Climate risk assessment -- quantifying physical climate hazards, economic exposure, and vulnerability across geographies and sectors -- has historically relied on deterministic climate models and expert-driven scenario analysis. The escalating complexity of compound climate risks, the explosion of Earth-observation data, and the urgent need for high-resolution, actionable risk estimates have created conditions where AI methods can make a decisive contribution. This study evaluates seven AI approaches for climate risk assessment across four risk dimensions: extreme event frequency prediction (floods, heatwaves, wildfires), physical asset damage estimation, agricultural yield impact forecasting, and compound risk interaction modelling. Models evaluated include gradient-boosted trees (XGBoost, LightGBM), convolutional neural networks applied to satellite imagery, graph neural networks for infrastructure network risk propagation, physics-informed neural networks (PINNs) for hydrological downscaling, and a large climate foundation model (ClimaX). Evaluation used ERA5 reanalysis data, Copernicus Sentinel-2 imagery, and asset-level exposure databases across 14 European countries (2000-2022). AI models outperform statistical baselines on extreme event frequency prediction by 12.4-18.6% in AUC; PINN-based hydrological downscaling reduces bias in 10km flood-depth estimates by 34.2% vs. statistical downscaling. Compound risk modelling using GNNs captures 28.4% more co-occurring extreme event pairs than single-hazard approaches. Critical uncertainty quantification gaps and regulatory implications for the EU Taxonomy and TCFD disclosure frameworks are discussed.