Towards the Operational Implementation of Seasonal Drought Impact-based Forecasting with Explainable Machine Learning

Accurate drought impact forecasting is fundamental for effective decision-making, yet forecasting drought impacts rather than hazards remain difficult due to the complex, non-linear relationship in which they can materialise. Impact-based drought forecasting at seasonal timescales is particularly challenging, thereby benefitting from methods that ensure reliability and transparency. Here, we present a novel machine-learning (ML) framework for drought impact-based forecasting that explicitly evaluates model performance, actionability, and explainability. 

The study is structured as a sequence of experiments. First, an autoencoder and several ML models—Gradient Boosting (XGB), Random Forest (RF), and Support Vector Machine (SVM) are trained with observed drought hazard indicators at multiple aggregations (e.g. SPI-12, SPI-24, SPEI-1, SPEI-3, SPEI-6, FAPAR-1, FAPAR-3, SMA-1, SMA-3) up to the current date to understand the data and how ML models perform to predict water scarcity levels in Italy, chosen as the drought impact indicator. The U-Net and ConvLSTM models were chosen as baseline models, as they directly predict gridded water scarcity levels. The framework is then extended by incorporating seasonal climate forecasts (precipitation and temperature) up to six months ahead to enable real-time impact prediction. Model sensitivity to spatial resolution is evaluated by testing inputs at 1 km and 25 km scales. To ensure that results are physically meaningful, explainable AI (xAI) techniques are applied to quantify predictor importance using SHAP, identify spatial hotspots using Integrated Gradients, and determine the most informative periods of the year using Partial Dependence Plots. 

Results show clear performance differences among models. Tree-based approaches, particularly Gradient Boosting and Random Forest, consistently outperformed deep learning baselines at both spatial resolutions. At 1 km resolution, xAI identifies SPEI-6 and SMA-1 as the most influential predictors, while at 25 km resolution SPEI-6 and FAPAR-3 emerge as the dominant drivers. Model performance improves at coarser resolution, with tree-based models providing the most accurate and robust predictions. Overall, the study (i) presents a workflow for assessing the effectiveness of ML in enhancing the seasonal prediction of drought impacts, (ii) leverages xAI to evaluate the relationship between the drought hazard indicators and drought impact data, including the most informative periods of the year and the spatial hotspots; and (iii) enabling real-time drought impact-based forecasting at seasonal scale.