Impact‐Based Drought Detection via Interpretable Machine Learning and Causal Discovery

Abstract The intensification of drought events due to climate change presents unprecedented risks and widespread socioenvironmental impacts. Traditional drought indices, which rely on hydrometeorological anomalies, often fall short in capturing the full extent of these impacts, as they primarily focus on isolated precursors (e.g., precipitation deficit for the SPI) rather than the cumulative effects of multiple interacting variables over time, like prolonged dry periods combined with high temperatures and reduced snowpack. Here, we introduce an advanced machine learning (ML) framework named Drought Detection via Regression‐based Interpretable Extraction and Causal Relationships (DRIER), which is designed to develop interpretable, impact‐based drought indices. DRIER employs a combination of cutting‐edge ML techniques, including nonlinear feature aggregation for dimensionality reduction, conditional mutual information‐based feature selection, and multitask linear regression to model vegetation stress quantified via the Vegetation Health Index as a function of hydroclimatic drivers. A distinguishing feature of DRIER is its integration of causal discovery, which ensures that selected predictors reflect robust physical relationships rather than spurious correlations, thereby enhancing model interpretability and generalizability. Applying DRIER to the Po River Basin (Italy), we identify key hydroclimatic precursors of vegetation stress, demonstrating its potential to improve drought monitoring and inform adaptation strategies in water‐stressed regions. By offering a scalable, data‐driven approach to impact‐based drought detection, DRIER represents a significant advancement toward more effective climate resilience planning.