Comparative Evaluation of Statistical and Machine Learning Models for Spatiotemporal Groundwater Level Forecasting in a Climate-Sensitive Coastal Aquifer

Abstract Coastal groundwater systems and deltaic systems portray a high level of spatiotemporal variability due to the seasonality in climate and human-induced stressors, and the assessment of the groundwater level (GWL) remains a consistent challenge. This paper gives a comparative data-driven modeling model to predict the level of groundwater variations with long-term records of the monitoring wells of Barishal, Bangladesh. The framework combines both a classical seasonal time-series model (SARIMA), deep learning (Long Short-Term Memory; LSTM) and ensemble machine learning (XGBoost and Random Forest) in order to model both linear and nonlinear groundwater dynamics. These methods were used in forecasting the changes in groundwater level (GWL) in susceptible wells in Barishal. RMSE, MAE, R 2, and Nash Sutcliffe Efficiency (NSE) were used as a measure of model performance. Findings show that LSTM and Random Forests models are better than traditional statistical models at modeling complex groundwater swings, especially in longer predictions. GIS-based spatial analysis also indicates how significant are regional heterogeneity in groundwater responses and the impact of climatic drivers in local groundwater variation. The results indicate that multi-model data-driven models can be useful in the evaluation of groundwater in hydro-climatically complex and data-constrained coastal areas that can be valuable information in the area of groundwater monitoring and management.