Graph Neural Networks for Groundwater Storage Prediction: A Comparative Analysis of Spatio-Temporal Modeling Approaches

Abstract

Predicting groundwater storage (GWS) is a major challenge due to the inherently spatio-temporal nature of hydrogeological processes. Graph Neural Networks (GNNs), particularly spatio-temporal architectures (ST-GNNs), offer a promising framework for modeling these complex dependencies by jointly capturing the spatial relationships between observation points and the temporal evolution of hydroclimatic variables. This paper presents a comparative analysis of recent work applying GNNs to groundwater prediction and related hydroclimatic variables. It examines the various graph construction strategies and the main ST-GNN architectures dedicated to jointly modeling GWS and hydroclimatic variables. A unified mathematical formulation of the modeling pipeline is provided, spanning graph construction through GWS prediction. Finally, a comparative analysis of studies applying ST-GNN models to groundwater prediction is presented, highlighting their respective performances and spatial mechanisms.