Understanding the evolutionary processes and causes of groundwater drought using an interpretable machine learning model

• Published in: Scientific reports Vol. 15; no. 1; pp. 20981 - 14
• Main Authors: Gan, Zhiyuan, Xie, Xianjun, Su, Chunli, Ge, Weili, Pan, Hongjie, Yang, Liangping
• Format: Journal Article
• Language: English
• Published: London Nature Publishing Group UK 01.07.2025; Nature Portfolio
• Subjects: 631/158/1144; 631/158/2459; 704/172; 704/242; Climate Change; Groundwater; Humanities and Social Sciences; Machine learning; Meteorological drought; multidisciplinary; Science; Science (multidisciplinary); SHAP; XGBoost; SHAP; Groundwater; Climate Change; XGBoost; Northern China; Machine learning; Meteorological drought
• ISSN: 2045-2322; 2045-2322
• DOI: 10.1038/s41598-025-05316-2

Drought is a widespread natural disaster, and rapid assessment of groundwater drought has become a challenge due to the lack of direct spatiotemporal observation of groundwater. We employed machine learning models and the Shapley Additive Explanation (SHAP), a game theory-based interpretability method, to understand and predict the evolution of groundwater drought by evaluating eight models with SHAP analysis in the West Liao River Plain (WLRP), with a semi-arid climate. The research showed: (1) The XGBoost model, optimized by the Sparrow Search Algorithm (SSA), achieved the highest performance (AUC: 0.922, F1-score: 0.84). (2) SHAP analysis revealed that the Standardized Precipitation Evapotranspiration Index (SPEI) at 12- and 24-month scales (SPEI12 and SPEI24) were key predictors, with long-term meteorological drought causing delayed groundwater drought, exacerbated by over-extraction and urbanization. The local SHAP values confirmed the robust importance of long-term meteorological drought and precipitation, and identified the interaction between precipitation and meteorological factors responsible for groundwater drought. (3) Future projections under the SSP5-8.5 climate scenario indicated a significant increase in drought-affected areas, with earlier onset, broader extent, and greater severity. This work provides a machine learning framework for evaluating groundwater drought characteristics driven by multiple factors.