Bulut et al. (2026) Toward early warning of drought impacts: a framework for predicting drought impacts in the UK

Identification

• Journal: Natural hazards and earth system sciences
• Year: 2026
• Date: 2026-03-25
• Authors: Burak Bulut, Eugene Magee, Rachael Armitage, Oyewole Adedipe, Maliko Tanguy, Lucy Barker, Jamie Hannaford
• DOI: 10.5194/nhess-26-1515-2026

Research Groups

• UK Centre for Ecology & Hydrology (UKCEH)
• School of Computing and Communications, Lancaster University
• European Centre for Medium-Range Weather Forecasts (ECMWF)
• Irish Climate Analysis and Research UnitS (ICARUS), Maynooth University

Short Summary

This study presents a data-driven framework to predict real-world drought impacts. Different modelling approaches were tested and evaluated in the United Kingdom using predictions at the time of occurrence, with the best-performing method selected for forecasting impacts months ahead. Both predictions and forecasts were validated using independent UK data and applied to Germany to test transferability, supporting early warning systems and improved drought risk planning.

Objective

• Develop multiple machine learning and linear models based on a lumped drought impact approach for the whole UK.
• Evaluate and compare the performance of these models using metrics from both training and independent “unseen” validation datasets across UK regions and selected time periods to identify the most accurate approach for near-real-time drought impact prediction (zero lead time).
• Apply the best-performing model to Germany to assess its transferability and demonstrate how models developed in data-available regions can be applied to regions with insufficient impact data.
• Evaluate the potential of lagged drought indicators for forecasting drought impacts at multiple lead times.
• Produce and assess gridded spatial drought impact predictions to characterise the spatial variability and severity of drought impacts.

Study Configuration

• Spatial Scale: United Kingdom (NUTS1 regions), Germany (NUTS1 regions), approximately 9 km grid cells for gridded predictions.
• Temporal Scale: Study period 1970–2012 (UK EDII data) and 1970–2018 (Germany EDII data); reference period for drought indices 1969–2010; forecast lead times up to 24 months.

Methodology and Data

• Models used: Random Forest (RF), Quantile Random Forest (Quantile RF), eXtreme Gradient Boost (XGBoost), Least Absolute Shrinkage and Selection Operator with Cross-Validation (LASSOCV), Linear Regression (LR).
• Data sources:
  • European Drought Impact Inventory (EDII V2.0)
  • ERA5-Land reanalysis dataset (precipitation, temperature, soil moisture at 0–7 cm, 7–28 cm, 28–100 cm, and 100–255 cm depths)
  • CORINE Land Cover dataset (for Arable Area Ratio - AAR)
  • WorldPop 2020 dataset (for Population Ratio - PopR)
  • Drought indicators: Standardized Precipitation Index (SPI), Standardized Precipitation-Evapotranspiration Index (SPEI), Standardized Soil Moisture Index (SSMI).

Main Results

• Random Forest (RF) consistently outperformed other models, achieving the lowest Root Mean Squared Error (RMSE) (3.74 for training, 4.77 for validation) and highest Area Under the Curve (AUC) values (0.94 binary for training, 0.78 binary for validation) for drought impact prediction.
• The generalized RF model successfully predicted drought impacts across UK NUTS1 regions and demonstrated promising transferability to unseen German NUTS1 regions, capturing many observed moderate and significant events.
• Predictive accuracy declined with increasing lead time, with RF providing the highest accuracy for up to 3-month forecasts, and performance significantly decreasing beyond 6 months (AUC values dropping below 0.6).
• Long-accumulation-period drought indicators, particularly SPEI24, and deep-layer soil moisture (SSMI Level 4), were identified as the most influential predictors.
• Regional variables like Population Ratio (PopR) and Arable Area Ratio (AAR) did not significantly improve model accuracy at short lead times, but PopR gained importance at longer lead times, potentially reflecting reporting behavior or exposure.
• Gridded impact predictions successfully captured the spatial distribution of observed impacts, particularly in central and southeastern England, even for unseen events like the April 2012 drought.

Contributions

• Developed a generalized and transferable machine learning framework for drought impact forecasting by aggregating impacts across all sectors and regions, addressing challenges in data-scarce areas.
• Integrated regression-based (total impact occurrences), classification-based (severity categories), and probabilistic (likelihood of occurrence) predictions into a comprehensive framework.
• Validated the model using uncensored drought impact time series, including normal and wet periods, enhancing its applicability for operational early warning systems.
• Demonstrated cross-national transferability by successfully applying a UK-trained model to Germany, highlighting its potential for regions with limited impact data.
• Produced spatially explicit, gridded drought impact predictions, offering high-resolution information for localized drought management.

Funding

• Natural Environment Research Council, Centre for Ecology and Hydrology (grant nos. NE/X012727/1 and NE/Y006208/1).

Citation

@article{Bulut2026Toward,
  author = {Bulut, Burak and Magee, Eugene and Armitage, Rachael and Adedipe, Oyewole and Tanguy, Maliko and Barker, Lucy and Hannaford, Jamie},
  title = {Toward early warning of drought impacts: a framework for predicting drought impacts in the UK},
  journal = {Natural hazards and earth system sciences},
  year = {2026},
  doi = {10.5194/nhess-26-1515-2026},
  url = {https://doi.org/10.5194/nhess-26-1515-2026}
}