A knowledge-guided machine learning framework for reconstructing climate-driven subsurface water dynamics from time-lapse electrical resistivity tomography

Abstract

Understanding and anticipating climate-driven subsurface water dynamics is essential for assessing water availability and ecosystem function, yet it remains challenging in heterogeneous mountain environments. Time-lapse electrical resistivity tomography (TL-ERT) provides spatially distributed observations of subsurface change but has been used mainly to monitor subsurface change rather than to model its evolution. Here, we develop a knowledge-guided machine learning framework to reconstruct the spatiotemporal evolution of subsurface resistivity from daily meteorological forcing and continuous TL-ERT observations on a monitored hillslope. Before training, the TL-ERT data were corrected to a common reference temperature to remove the direct temperature dependence of resistivity, allowing temporal resistivity changes to be interpreted as a proxy for subsurface water-content change. Three neural architectures were evaluated within a shared recursive framework using structured loss terms that preserve reconstruction accuracy, local spatial structure, and coherent temporal variation. All models reproduced the dominant seasonal patterns, and the PredRNN-based architecture performed best, with an RMSE of 9.5 Ω·m over held-out test windows. Under rolling recursive reconstruction without subsequent ground-truth resistivity observations, the framework maintained seasonally coherent trajectories within the observed record. Under a + 4 °C warming scenario, the simulated temperature-corrected resistivity showed earlier shallow spring decreases about 12% and slight shallow summer increases, with weaker responses at depth. Through a water-content conversion, these changes are consistent with earlier spring wetting and shallow summer drying. These results provide a site-specific proof of concept that continuous TL-ERT observations combined with knowledge-guided machine learning can reconstruct the seasonal evolution of subsurface resistivity within the observed record and support scenario-based hydrologic interpretation.