Continuous groundwater level records are essential for understanding long-term trends in aquifer dynamics and supporting sustainable water resource management. However, monitoring wells in Texas often exhibits data gaps ranging from days in some locations to months in others. These missing observations hinder accurate trend analysis and reduce the reliability of models used for forecasting and planning. Imputing historical gaps using advanced machine learning techniques not only restore data continuity but also enable robust predictions under varying hydroclimatic and anthropogenic conditions.  

This study investigated the extent to which topographic, geologic, and anthropogenic factors influence shallow groundwater level predictions across heterogeneous aquifer systems in Texas by using observation data from 33 shallow monitoring wells (depths ranging from 10 to 72 m) distributed across six major Texan aquifers. XG-Boost machine learning models were employed to predict groundwater levels from 2002 to 2022. Input variables included hydroclimatic parameters (e.g., ΔTWS from GRACE, precipitation, temperature, and NDVI), geologic properties (e.g., saturated hydraulic conductivity, sand percentage, and aquifer type), topographic features (e.g., elevation and slope), anthropogenic features (such as land use/land cover), and lagged hydroclimatic inputs (such as one-month and two-month lags). To enhance model interpretability, Shapley Additive Explanations (SHAP) were used to identify the most influential predictors. 

Model performance was robust, with a test-phase Mean Absolute Error of 2.09 m, Mean Squared Error of 13.68 m, and Nash-Sutcliffe Efficiency of 0.96. SHAP analysis revealed that ground elevation and land use/land cover were dominant predictors, however, a wide variability in SHAP values for ground elevation underscores that topography is not uniform across different Texan aquifers and reflects the heterogeneity of the aquifers. For example, in karst systems such as the Edwards aquifer, elevation strongly influences recharge through thin soils and faulted zones. But, confined aquifers such as Carrizo-Wilcox and Gulf Coast, the ground elevation effect will be the opposite as recharge depends on outcrop zones which lie in higher elevations while downdip confined sections exhibit artisan pressure reducing elevation effect. Similarly, the wide range of SHAP values for saturated hydraulic conductivity and sandy texture reflects their limitation as surface-based measurements, which cannot fully capture subsurface heterogeneity of Texan aquifers. In contrast, rapid changes in LULC correlated with lower groundwater levels, reflecting pumping demand. These findings highlight the need for stratigraphy informed models integrated with anthropogenic inputs to achieve more accurate prediction, especially in heterogeneous aquifers.