Bridging Drought Indices and Impacts: Forecasting Future Outcomes

Drought events significantly challenge communities and ecosystems worldwide, emphasising the urgent need for effective predictive methods to facilitate proactive management and to mitigate their impacts. A clear gap exists between theoretical drought indices, such as SPI, SPEI, and SSMI, and the real-world impacts of droughts. This study aims to address this disparity by leveraging machine learning (ML) techniques to predict reported drought impacts, using data from the European Drought Impact Database (EDID). A variety of ML algorithms, including Random Forest, Quantile Random Forest, Least Absolute Shrinkage and Selection Operator, XGBoost and Linear Regression were assessed. The study also uses likelihood forecasting to quantify the probability of drought impacts. This probabilistic approach and use of lagged indices allows for a deeper understanding of the range of possible outcomes, enabling decision-makers to plan and prepare for varying levels of drought severity.
 
Unlike location-specific modelling approaches, this study proposes a generalized ML model applicable across the UK. The model's robustness was validated using independent datasets from different regions and periods. The findings indicated that categorising impacts into severity levels, rather than predicting the exact number of impacts and improved the model's accuracy and interpretability. Additionally, the model was applied at a grid scale to generate impact-based drought maps, providing a valuable tool for decision-making in drought risk management. This methodological approach enhances decision-making processes for drought risk management, demonstrating the practical utility of ML techniques that can be applied globally, beyond the UK.