A wavelet-assisted deep learning approach for simulating groundwater levels affected by low-frequency variability

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence in Environmental Science
• Climate Change Impact Assessment

Background:

• Accurate groundwater level (GWL) simulations are vital for understanding past variability and projecting future resources under climate change.
• Analyzing GWLs influenced by low-frequency variations is critical for reliable resource management and climate adaptation strategies.
• Deep learning (DL) models offer potential for complex time-series simulation, but their efficacy with varying GWL frequencies needs assessment.

Purpose of the Study:

• To evaluate the performance of three deep learning models (LSTM, GRU, BiLSTM) in simulating groundwater levels with distinct low-frequency behaviors.
• To investigate the impact of Maximal Overlap Discrete Wavelet Transform (MODWT) pre-processing on the accuracy of DL models for GWL simulation.
• To assess the influence of different input variables (raw precipitation, air temperature, effective precipitation) on model performance.

Main Methods:

• Three deep learning models (LSTM, GRU, BiLSTM) were employed to simulate three types of GWLs: inertial, annual, and mixed.
• Maximal Overlap Discrete Wavelet Transform (MODWT) was used as a pre-processing technique to decompose input variables (precipitation, effective precipitation).
• Shapley Additive exPlanations (SHAP) was utilized to interpret model predictions and understand feature importance, particularly frequency content.

Main Results:

• For inertial and mixed GWLs, MODWT-pre-processed DL models using raw data significantly outperformed standalone models.
• DL models performed well for annual GWLs, with effective precipitation as input; MODWT offered only minor improvements.
• MODWT-enhanced models showed better performance, particularly for low-frequency variations, with models learning from low-frequency precipitation data.

Conclusions:

• Maximal Overlap Discrete Wavelet Transform (MODWT) pre-processing is highly effective for improving deep learning simulations of low-frequency varying groundwater levels.
• The choice of input data (raw vs. effective precipitation) and pre-processing significantly impacts model accuracy depending on GWL type.
• Deep learning models, especially when combined with MODWT, demonstrate strong potential for accurate groundwater resource assessment and climate change projections.