Predicting multi-sectoral drought impacts in the Mediterranean with spatio-temporal deep learning

Droughts are hydroclimatic anomalies driven by precipitation deficits and increased evapotranspiration, posing an escalating threat under a warming Mediterranean climate. Assessing drought risk remains challenging due to the complex interactions between biophysical conditions and human systems, as well as limitations in impact reporting. Moreover, drought impacts are highly heterogeneous across sectors, as different types of drought affect socio-environmental systems differently. In this context, we develop a spatio-temporal deep learning framework to predict sector-specific drought impacts and identify the environmental and climatic drivers of these impacts.

We combine two primary data sources: the European Drought Impact Database (EDID), which contains above 13,000 georeferenced drought impact reports spanning 1970 to 2023 and aggregated into four sectors (agriculture, ecosystem, energy, and socio-economic); and a set of physical drivers, including precipitation, temperature, drought indices, vegetation indices, and population density, derived from various sources for the period 2001–2021.

The prediction task is formulated as a spatio-temporal segmentation problem using a 3D U-Net architecture to capture dependencies in climate and environmental conditions over a one-year period. The preprocessing workflow harmonizes all variables to a spatial resolution of 0.25° and an 8-day time step. Seasonally varying predictors are transformed into anomalies, and all variables are normalized. Input samples are arranged as tensors with shape 36×48×16×16 (C×T×H×W), representing one year of conditions, while the target consists of a binary impact map (1×1×16×16) corresponding to the subsequent 8-day period. The training dataset is balanced through equal sampling of impact and no-impact cases. Consequently, the model learns to use one year of spatio-temporal context to predict drought-affected areas at the next time step.

Initial results for the agricultural sector indicate that traditional drought indices have limited predictive skill for drought impacts. A baseline evaluation of the Standardized Precipitation-Evapotranspiration Index (SPEI) across multiple thresholds shows that the 1-month SPEI achieves a PR-AUC of 0.13 and an ROC-AUC of 0.32 for the impact class over the 2018-2020 test period, with similarly low performance for the 3-, 6-, and 12-month variants. In contrast, preliminary model experiments demonstrate a substantial improvement over the baseline, achieving an F1 score of 0.43, a PR-AUC of 0.41, and a ROC-AUC of 0.71, despite remaining limitations in predictive performance.

These limitations are primarily attributed to noise and spatial uncertainty in the ground-truth labels, as EDID impacts are reported at coarse administrative units (NUTS3) and uniformly assigned to all grid cells within each region, constraining pixel-level learning. In addition, drought impacts are influenced by large-scale atmospheric circulation patterns and remote climate teleconnections (e.g., ENSO and NAO) that are not explicitly represented in the current feature set. Future work will address these limitations by incorporating large-scale circulation and teleconnection indicators, developing strategies to mitigate label noise, and extending the modelling framework to additional sectors. Once predictive performance is optimized, explainable AI methods based on Integrated Gradients will be applied to identify the most influential climatic and environmental drivers, enabling sector-specific interpretation of drought impact mechanisms and their temporal dynamics.