15-5 From Prediction to Understanding: XAI Applications for Shallow Groundwater Predictions across Texas Aquifers

Session: Integrated Digital Workflows in Geoscience: Mapping, Marine Exploration, and Machine Learning

Presenting Author:

Authors:

Abstract:

Reliable forecasts of shallow groundwater are critical for drought response and sustainable allocation, however, strong hydrogeologic heterogeneity across Texas complicate both model transferability and interpretation. This research moves from prediction to understanding by coupling tree-based machine learning with explainable AI (XAI) to not only quantify the dominant controls on monthly depth to groundwater (DTW) across multiple aquifer systems, but also to provide deeper interpretations and actionable insights to support sustainable groundwater management.

Daily DTW observations from 33 Texas Water Development Board monitoring wells (2002–2022) were aggregated to monthly means and analyzed across major aquifers (including the Ogallala, Trinity, Edwards, Edwards-Trinity, Carrizo-Wilcox, and Gulf Coast systems) spanning confined, unconfined, and mixed conditions.

Predictors were assembled from multi-source datasets representing hydroclimate (PRISM precipitation, maximum temperature, and minimum temperature), vegetation dynamics (MODIS NDVI), regional water storage variability (GRACE ΔTWS), topography (ASTER-derived elevation and slope), soil texture and hydraulic properties (gSSURGO sand percent and saturated hydraulic conductivity), land cover (NLCD), and aquifer type. Predictor values were sampled at well locations and temporally aligned to the GRACE record to avoid gap filling artifacts.

We trained Random Forest (RF) and Extreme Gradient Boosting (XGBoost) models with tuned hyperparameters and evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Nash-Sutcliffe Efficiency (NSE). RF achieved good performance (MAE: 2 m; MSE: 14.85 m²; NSE: 0.96), while XGBoost produced slightly higher errors but comparable NSE (MAE: 2.29 m; MSE: 16.34 m²; NSE: 0.96). The confined Gulf Coast well showed phase shifts and amplified variability relative to observations, consistent with pressure-affected behavior and delayed transmission beneath confining units, whereas an unconfined Trinity well tracked observed DTW more closely.

To advance interpretability, we applied SHAP analysis to the XGBoost model to derive global and local explanations. SHAP diagnostics identify elevation, land cover, hydraulic conductivity, ΔTWS, and aquifer type as dominant drivers and reveal site-specific departures from simplified elevation-DTW expectations that reflect confinement and subsurface heterogeneity. Local waterfall explanations further show that precipitation can exert strong, well-specified effects even when its global importance is muted by spatially variable and lagged recharge pathways. These results demonstrate how XAI transforms predictive models into diagnostic tools that expose aquifer-specific controls and guide hydro-stratigraphically informed forecasting and groundwater management decisions for Texas groundwater resilience planning.

Geological Society of America Abstracts with Programs. Vol. 58, No. 2, 2026

© Copyright 2026 The Geological Society of America (GSA), all rights reserved.