248-4 Data-Driven Long-Term Monitoring of Groundwater Levels in Shallow, Intermediate and Deep Wells Across Different Regions of Texas by Utilizing Remote Sensing Data, GIS, and various AI Techniques

Session: Expanding Geology’s Horizons: Geoinformatics, Open Science, and Open Data

Presenting Author:

Authors:

Abstract:

In Texas, groundwater meets over 55% of the state's water needs. Texas has numerous aquifers that provide groundwater for households, municipalities, industry, farms, and ranches. The groundwater level (GWL) serves as a straightforward and clear indicator of groundwater availability and ease of access.  Accurate prediction of groundwater level (GWL) is a significant challenge in managing aquifer systems and is crucial in areas with limited surface water availability.

This study predicts groundwater level (GWL) changes in 66 monitoring wells distributed across various locations in major Texan aquifers between 2002 and 2022, using GIS, remote sensing, meteorological data, and machine learning models with explainable artificial intelligence (XAI) techniques. To improve model accuracy and unlike previous studies conducted in Texas, this research incorporates geologic properties from modeled published data and extracted at different wells locations as static inputs, and the 66 wells are classified into shallow, intermediate, and deep groups using the natural breaks method in ArcGIS Pro depend on the average of groundwater elevation measured in each well.

Two modeling approaches are evaluated using Random Forest and XGBoost algorithms. The first approach utilizes continuous inputs such as GRACE-derived terrestrial water storage (ΔTWS), precipitation, maximum and minimum temperatures, NDVI, and their time lags. The second approach builds on this by including additional static inputs such as saturated hydraulic conductivity, slope, and sand percentage. Model performance is assessed using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Nash-Sutcliffe Efficiency (NSE). Results show a significant improvement in model performance with the inclusion of static inputs. For Random Forest, NSE values increased from 0.26 to 0.95 (shallow), 0.29 to 0.94 (intermediate), and 0.60 to 0.96 (deep). Similarly, XGBoost NSE values improved from 0.11 to 0.92 (shallow), 0.17 to 0.95 (intermediate), and 0.64 to 0.95 (deep). SHAP analysis, a leading XAI technique, is used to interpret the XGBoost model by quantifying feature contributions at both global and local levels. Globally, SHAP identifies the most influential predictors across the dataset, while locally, it explains individual predictions by showing how each feature contributes to a specific output.

 This framework not only enhances the predictive performance of machine learning models for GWL changes in sparsely monitored regions but also offers a scalable and transferable approach for broader data-driven monitoring applications.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-6673

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.