Revealing the causes of groundwater level dynamics in seasonally frozen soil zones using interpretable deep learning models

Abstract. Accurately characterizing groundwater level dynamics in seasonal frozen soil regions is of great significance for water resource management and ecosystem protection. To this end, this study proposes a new interpretable deep learning method to reveal the underlying causes of groundwater level dynamics on the basis of groundwater level simulation. Using the Songnen Plain in China as the study area and daily data from 138 monitoring wells, groundwater levels are simulated with an Long Short-Term Memory (LSTM) model, and the Expected Gradients (EG) method is employed to quantitatively identify the dominant factors and mechanisms of different groundwater level variation types.The results show that the LSTM model performs well on the test set, with the Nash-Sutcliffe Efficiency (NSE) exceeding 0.7 at 81.88 % of the monitoring sites, effectively capturing the temporal dynamics of groundwater levels. At the annual scale, three typical groundwater level variation types are identified: precipitation infiltration–evaporation type (29.0 %), precipitation infiltration–runoff type (18.1 %), and extraction type (52.9 %). Corresponding to the seasonal frozen-thaw period, groundwater level dynamics are classified into “V”-shaped (38.4 %), continuous decline (23.2 %), and continuous rise (38.4 %) types. Quantitative analysis using the EG method indicates that air temperature, precipitation, and snow thickness are the primary controlling factors of the “V”-shaped dynamics, reflecting the regulatory role of the frozen-thaw process on groundwater levels.When the initial groundwater level depth at the beginning of the freezing period is shallower than the sum of the frozen-thaw influence depth and the capillary rise height, a hydraulic connection is established between soil water and groundwater, resulting in typical “V”-shaped fluctuations. Conversely, when the depth exceeds this critical threshold, the frozen-thaw process cannot significantly influence the aquifer, and groundwater dynamics are mainly manifested as continuous rise or continuous decline, driven respectively by groundwater extraction and water level recovery following precipitation recharge. This study establishes an integrated framework of “simulation–classification–interpretation,” which not only improves the accuracy of groundwater level dynamic simulation and prediction but also provides new methods and perspectives for revealing the underlying mechanisms. The findings offer theoretical support and technical basis for regional groundwater resource management in cold regions.