Development of a Multi-Scale Groundwater Drought Prediction Model Using Deep Learning and Hydrometeorological Data

Groundwater is an essential water resource and plays a crucial role, especially in areas with limited surface water availability. However, the exacerbation of groundwater droughts, fueled by phenomena such as climate change, urbanization, and industrialization, highlights the necessity for predictive tools to aid in sustainable groundwater management. While artificial neural networks (ANN) have been increasingly used for groundwater level prediction, most studies have focused solely on point-scale predictions from groundwater observation wells, which can be resource-intensive and time-consuming. In this study, we propose a multi-scale groundwater-based drought prediction model that can predict both zonal average values and the values at well locations for the standardized groundwater level index (SGI). Specifically, we develop a zone-scale SGI prediction model through long short-term memory (LSTM) and propose a model that can accurately predict point-scale SGI through a simple downscaling process. Our model was developed and tested for Jeju Island, a volcanic island in South Korea where groundwater serves as the primary water source. Specifically, we partitioned Jeju Island into 16 sub-watersheds, termed zones, and constructed an individual model for each zone. Forecasting the standardized groundwater level index (SGI) for each zone was based on input datasets including the daily temperature, precipitation, snowfall, vapor pressure deficit (VPD), wind speed, and preceding SGI values. Additionally, we downscaled the predicted values of each zone to the specific SGI values at groundwater monitoring wells within the zone. This was achieved by applying the spatial deviation of each well relative to the zonal mean over the preceding 4 days to the predicted zone-scale SGI value. Our findings indicate high accuracy of the model in SGI predictions across both scales, with the Nash–Sutcliffe efficiency coefficient (NSE) exceeding 0.9 and the root mean square error (RMSE) remaining less than 0.3 for both the representative zone and observation well. By leveraging the proposed model, stakeholders and policymakers can efficiently generate and utilize both zone-scale and point-scale groundwater-based drought predictions, contributing to effective groundwater management practices.