Enhancing Groundwater Level Prediction in Data-Scarce Regions: An Integrated Framework Coupling SWAT-MODFLOW with Cluster-Based Deep Learning

Precise forecasting of groundwater levels is fundamental to the sustainable management of water resources; however, this task frequently impeded by sparse and discontinuous observational data. This study proposes an integrated modeling framework that couples physically based hydrological simulations with cluster-based deep learning to overcome these limitations. Focusing on the Yom and Nan River Basins in Thailand, outputs from the coupled SWAT-MODFLOW model were utilized to provide a continuous and physically consistent dataset. To further handle spatial heterogeneity, K-Means clustering was applied to segment 34 observation wells based on intrinsic hydrogeological parameters (hydraulic conductivity, specific storage, and specific yield). Four deep learning architectures including RNN, LSTM, GRU, and CNN were subsequently evaluated using both generalized and cluster-specific training strategies. Results reveal that the basin-wide approach is inherently suboptimal for geologically diverse basins. Conversely, the cluster-based strategy can resolve complex localized dynamics, with improvements being most pronounced in hydrogeologically distinct zones, such as high-storage aquifers and complex transition zones. In these areas, the proposed framework can effectively correct amplitude biases found in the general models. Furthermore, regarding model architecture, GRU and CNN demonstrated predictive accuracy comparable to, and occasionally superior to, LSTM networks while offering greater computational efficiency. These findings demonstrate that integrating physical simulation with hydrogeological clustering offers a robust solution for enhancing groundwater forecasting in data-scarce regions, bridging the gap between physical hydrology and data-driven artificial intelligence. Consequently, it establishes a scalable methodological blueprint for generating reliable groundwater inventories in ungauged basins, supporting long-term climate resilience planning.