Determining the Groundwater Level in Hilly and Plain Areas From Multisource Observation Data Combined With a Machine Learning Approach

Developing an accurate model that can effectively simulate groundwater levels is of immense significance for water resource management and aquifer protection. To achieve this, it is crucial to identify key factors in remote sensing, topography, and meteorology, and to improve hydrological models to enhance prediction accuracy. This study proposes a multistep modelling framework, the RF-PSO-GRNN algorithm model, to improve the accuracy of groundwater level simulations in data-scarce hilly regions. The framework combines the random forest (RF) model with the particle swarm optimization (PSO) algorithm and the generalised regression neural network (GRNN). First, the study area was divided into hilly and plain regions, decreasing mean absolute error (MAE) by 0.2 m in plain areas and 0.1 m in hilly areas. The RF-Gini index combination was then used to calculate the contributing factors for each region, facilitating the determination of an optimal balancing strategy, which reduced RMSE by 4.35 m in hilly areas and 3.82 m in plain areas. Subsequently, the PSO algorithm was employed to compute the optimal smoothing factor for GRNN, further reducing RMSE by approximately 10 m. Additionally, MAE decreased by 11 m in hilly areas and 7.5 m in plain areas. Finally, the RF-PSO-GRNN model was applied to simulate the spatiotemporal evolution of groundwater levels in three counties within the Fu River Basin of Jiangxi Province, China. The findings confirm the effectiveness of GRNN in simulating groundwater levels with limited data samples. This study provides a practical solution for hydrological modelling and groundwater management under data-scarce conditions, contributing to the understanding and predicting groundwater dynamics.