28-8 Forecasting Future Shallow Groundwater Levels for Different Monitoring Wells in Texas Using LSTM Deep Learning Models

Session: Geoscience and Hydrogeology in the AI Era: From Predictive Models to Real-Time Applications (Posters)

Presenting Author:

Authors:

Abstract:

Groundwater is a vital resource for sustaining agriculture, municipal supply, and ecosystems in semi-arid regions like Texas, where shallow aquifers often serve as the primary source of water. Continuous monitoring of groundwater levels is critical for detecting seasonal trends, identifying over-extraction, and informing sustainable management strategies. However, maintaining consistent data collection across monitoring wells can be challenging, particularly in the face of extreme weather events or infrastructure limitations.

This study employs Long Short-Term Memory (LSTM) deep learning models to forecast groundwater levels one year in advance for 33 shallow monitoring wells that monitor the groundwater level below land surface in the range between 10 and 72 m and are distributed across diverse hydrogeological and climatic regions of Texas. Historical daily groundwater level data from 2002 to 2022 were aggregated into monthly averages to capture long-term trends. The model was trained on data from 2002 to 2019 and tested on data from 2020 to 2022, with forecasts generated for 12 months of 2023. 

The LSTM model demonstrated strong predictive performance across multiple wells. The Root Mean Squared Error (RMSE) values ranged from 0.05 to 0.8, while Mean Absolute Error (MAE) values ranged from 0.05 to 0.5. The lowest RMSE and MAE (~0.05) were observed at Well # 5225209, located in an unconfined igneous aquifer in Jeff Davis County, West Texas. The coefficient of determination (R²) varied from approximately 0.3 to over 0.8, with the highest R² (>0.8) recorded at Well # 6950302 in the confined Edwards (Balcones Fault Zone) aquifer in Uvalde County, South-Central Texas.

This study explores the use of Long Short-Term Memory (LSTM) deep learning models to forecast groundwater levels using only historical groundwater level data. While these models cannot replace real-time monitoring, they can serve as a complementary tool for anticipating groundwater trends at the near future, particularly in cases where temporary data gaps occur, or when long-term planning decisions must be made in the absence of full hydrologic inputs. By capturing patterns embedded in historical records, the LSTM model supports short-term forecasting and can help guide better informed data-driven decisions in regions with limited or irregularly available observations. This research contributes to the growing body of literature on deep learning applications in hydrology that support sustainable groundwater management.

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

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