Exploring the Generalisation Ability of Deep Learning Models for Large-Sample Groundwater Level Predictions across Space and Time

In hydrology, deep learning (DL) models have already achieved remarkable breakthroughs in predicting streamflow. These DL models are fed with meteorological time-series and static catchment attributes across large samples of catchments, and predict streamflow remarkably well in both gauged and ungauged situations. In recent years, some studies have transferred the idea of constructing multi-basin/station DL models – particularly Long-Short Term Memory (LSTM) neural networks – to large-sample groundwater level modelling to explore their potential for temporal and spatial extrapolation. To the best of authors’ knowledge, existing multi-station LSTM applications are limited to three, covering 76 climate-sensitive stations in Northern France, 108 nationwide stations in Germany, and 1,800 coastal stations across nine countries/regions. Notably, spatial generalisation was investigated solely in the German study, which suggests that the model utilised static features primarily as ‘unique identifiers’ to memorise local behaviour rather than deriving the generalisable hydrological insights required for spatial extrapolation. Given the limited number of studies and the potentially biased datasets, the generalisation ability of multi-station DL models for groundwater level modelling is still under exploration.

A newly released large-sample groundwater dataset by the Environment Agency of England, comprising more than 200,000 daily and 200 million sub-daily sampling observations for over 3,400 wells, offers a unique opportunity to test the generalisation ability of multi-station DL models in time and space, and whether these models can yield process-relevant insights on groundwater dynamic mechanisms. In this study, we want to investigate the following questions:

• 1) How well can multi-station DL models simulate the groundwater variability across England?
• 2) Which input features does the DL model use to make its predictions (especially in places where it does well)?