A physically interpretable transfer learning framework for improved generalization of data-driven groundwater level...

Area of Science:

• Hydrology
• Machine Learning
• Data Science

Background:

• Accurate groundwater level (GL) prediction is crucial for water resource management.
• Data-driven models often struggle with prediction accuracy under data-scarce conditions.
• Interpreting the parameters of complex data-driven models in hydrogeology remains a challenge.

Purpose of the Study:

• To develop an interpretable transfer learning (TL) framework for groundwater level (GL) prediction.
• To address challenges of accurate prediction with limited data and hydrogeological interpretation of model parameters.
• To enable reliable prediction and interpretation using short-term GL records for newly established monitoring wells.

Main Methods:

• Integration of Gated Recurrent Unit (GRU) networks with 1D convolution and transfer learning (TL).
• A pretrained backbone model was jointly trained across 62 monitoring wells to learn generalizable precipitation-GL response patterns.
• Leave-one-well-out cross-validation (LOWOCV) was used to evaluate performance under various data-scarcity scenarios (1-3 years).

Main Results:

• The TL framework consistently outperformed conventional GRU models across all data-scarcity scenarios.
• TL model trained with 1 year of data exceeded the performance of a baseline GRU model trained with 3 years of data.
• The framework demonstrated stable predictive performance under non-stationary hydrological conditions and provided hydrogeologically meaningful interpretations.

Conclusions:

• Transfer learning effectively compensates for limited groundwater level records, enabling accurate predictions.
• The developed framework provides a practical tool for early-stage assessment of new monitoring wells.
• The interpretable nature of the framework allows for hydrogeological insights from data-driven model parameters.