Predicting Groundwater Drought in Ireland Using a Machine Learning Ensemble

Groundwater droughts in temperate regions are typically considered rare phenomena and consequently neglected in research despite their significant socio-economic and ecological impacts. In light of increasing water demands and climate change intensity, understanding and predicting groundwater droughts are essential for sustainable water resource management.

This study aims to define, identify and predict groundwater drought events across the Irish groundwater network by integrating multiple drought identification indices with machine learning (ML) techniques. Groundwater level (GWL) time series from 100 monitoring stations, methods: (i) the Threshold Level Method (TLM), which identifies drought when GWLs fall below predefined thresholds (ii) the Percentage of Normal (PON), which quantifies deviations in mean GWL relative to a baseline reference period; and (iii) the Standardised Groundwater Index (SGI), which normalises GWLs to classify drought severity. Subsequently, these approaches were evaluated and compared based on their ability to characterise drought events, using the 2018 drought for validation. This process enabled the selection of the most suitable indicator for predictive modelling.

An ensemble of ML binary classifiers including Logistic Regression (LR), Generalized Linear Models (GLM), Decision Trees (DT), Random Forest (RF), and XGBoost (XGB) were trained using meteorological inputs such as precipitation and temperature, to predict groundwater drought occurrences. However, the imbalanced class problem (rare drought events) was found to reduce classifier accuracy therefore, datasets were resampled using the Synthetic Minority Over-sampling Technique (SMOTE) technique, using several balance conditions of 50%, 40%, 30%, 20% minority class distribution.

Analyses indicate that the TLM and PON exhibit low sensitivity for drought detection, whereas the SGI was significantly more effective in characterising drought events within the Irish hydrogeological environment. Results show that the SMOTE technique enhanced performance of LR, GLM, and DT models, demonstrated by higher area under the receiver operating characteristic curve (AUC), and area under the precision/recall curve (AUCPR) values. However, XGB showed superior stability and accuracy across all sampling conditions. Notably, with a 40% minority class, XGB achieved the highest Recall and Precision values of 91.6% and 95.2%, respectively. As expected, model interpretations highlighted precipitation as a key precursor to drought propagation, with stations showing variable vulnerability linked to cumulative precipitation lags.

Future research directions will involve developing multi-scale early-warning models for groundwater drought using machine learning and deep learning. These models will be upscaled to a national level to map spatiotemporal impacts and inform groundwater management planning under changing climatic conditions.