Effective groundwater management requires tools capable of distinguishing and attributing hydrogeologic changes to natural and anthropogenic stressors. Such tools are especially vital when working with the several overlapping stressors that impact the High Plains Aquifer, a critical water source for U.S. agriculture that faces significant depletion in Texas, Oklahoma, and Kansas. Despite recent advances in artificial intelligence, existing methods remain limited in their ability to isolate individual stressors.

This study presents a novel approach that integrates the Hilbert-Huang Transform (HHT), an empirically-based adaptive signal decomposition method, with artificial neural networks (ANNs), a computational model for optimizing data categorization. The integrated model decomposed time series data from over 6,300 wells across the High Plains Aquifer into distinct intrinsic mode functions and labeled each as correlated with a regional groundwater level stressor. Stressor identifications were informed by historical records and previous hydrogeologic studies. Key stressors included pumping for irrigation, underground injection, climate oscillations, seasonal precipitation fluctuations, and extended recharge events. Model performance was assessed using classification accuracy, cross-entropy log loss, and the number of stressors categorized. 

The HHT-ANN model demonstrated strong performance in isolating and classifying natural and anthropogenic groundwater level stressors, achieving up to 86.7% accuracy and 0.462 cross-entropy log loss. Across multiple trials, the model mitigated common issues related to overfitting and data imbalance. Notably, the model performed better at classifying higher-frequency groundwater stressors, challenging previous literature on the frequency principle of ANNs. These results suggest that integrating signal processing with ANNs enables them to overcome bias towards low-frequency data, with significant implications for modeling across various domains.

Combining the HHT with ANNs offers a promising technique for disentangling groundwater level stressors in data from the High Plains Aquifer and other large, complex regional aquifers. These findings underscore the value of computational methods in advancing hydrogeologic analyses and resource management.