Forecasting Groundwater Level by Characterizing Multiple Spatial Dependencies of Environmental Factors Using Graph‐Based Deep Learning

Abstract

Accurate forecasting for groundwater levels is essential for water resource management and sustainable development. Regional variations in groundwater levels exhibit a complex spatial dependency structure due to the physical proximity of monitoring wells, hydrological connectivity, and shared environmental characteristics. However, existing research has mostly overlooked the multiple spatial dependencies between monitoring wells, limiting the understanding of the added value that graph‐based models bring to groundwater dynamics prediction. In this study, we characterize spatial dependencies of groundwater from multiple perspectives and investigate the impact on the forecasting results of groundwater dynamics using graph‐based deep learning. Characterizing spatial dependencies helps improve the understanding of groundwater dynamics, but its effectiveness in enhancing prediction accuracy depends on the characteristics of spatial interactions. Graph neural networks facilitate learning from diverse spatial associations, allowing for a more comprehensive representation of spatial dependencies and uncovering potential connections between monitoring wells. Results based on real data sets contribute to understanding how multiple spatial dependencies influence groundwater level forecasting. Our research provides insights into exploiting potential spatial dependencies in similarly complex and highly interconnected earth systems.The groundwater level variations between different monitoring wells within the same region exhibit correlations or interactions, which can be represented as complex spatial dependencies. These dependencies are influenced by various environmental factors, such as hydrological conditions, meteorology, and human activities. Characterizing multiple spatial dependencies provides significant value in capturing complex connections within groundwater systems. However, this advantage does not always lead to substantially improved prediction performance. This study captures multiple forms of spatial dependencies (e.g., distance and correlation) of groundwater and integrates them into a temporal graph convolutional network for regional groundwater level forecasting. The model adaptively captures the spatial dependencies between nodes during training based on a predefined graph structure and updates the graph using a weighted averaging mechanism. Comparative analysis with single spatial dependency models shows that the updated graph structure reflects more reasonable spatial dependencies, consistent with the temporal dynamics of observed data. This study contributes to understanding and forecasting complex spatial dependency structures in regional monitoring networks of geosciences. A graph neural network characterizes multiple spatial dependencies of environmental factors for groundwater level forecasting Characterizing spatial dependencies aids in understanding groundwater dynamics, but its effectiveness depends on spatial interactions The latent associations within the dynamic time‐series of groundwater levels can be uncovered