Comparative evaluation of statistical and machine learning models for spatiotemporal groundwater level forecasting in a climate-sensitive coastal aquifer

The groundwater resources of the climate sensitive coastal area of Bangladesh are getting more and more sensitive to climate variability and anthropogenic stresses, and thus reliable forecasting and spatial assessment are needed for sustainable management. This study aimed to explore the spatiotemporal dynamics of the groundwater level in Barishal District through the combination of statistical and machine-learning models along with vulnerability mapping through the GIS methodology by applying monthly groundwater and hydro-climatic data for the period of 1996–2025. The predictive performance was compared by RMSE, MAE, MSE, R 2 and NSE for the models (Seasonal Autoregressive Integrated Moving Average (SARIMA), Long Short-Term Memory (LSTM), Extreme Gradient Boosting (XGBoost), and Random Forest (RF). Machine-learning models generally outperformed SARIMA, with RF generating the lowest RMSE (average = 0.123 m) and the highest NSE (average = 0.930) and XGBoost the highest accuracy at a few individual wells (lowest RMSE = 0.0337 m; highest R 2 = 0.9885). Pairwise Diebold–Mariano statistical tests further confirmed that the machine-learning models significantly outperformed the conventional SARIMA model across the monitoring wells. The temporal variability and seasonal groundwater recharge–recession cycles were successfully reproduced by LSTM. The result of the GIS-based analysis showed that Hizla, Mehendiganj, Bakerganj, Muladi, and part of Gaurnadi are persistent groundwater vulnerable areas, while Barishal Sadar, Agailjhara, and Banaripara are comparatively stable. The idea of integrating machine-learning forecasting and GIS-based vulnerability assessment is a powerful decision support tool to identify groundwater-stressed areas and to improve sustainable groundwater-resource management in climate sensitive coastal aquifers.