Comparing four machine learning model performances in forecasting the alluvial aquifer level in a semi-arid region

Abstract

Groundwater level fluctuation is a nonlinear and non-stationary system as it depends on several factors in the time and space scales. Conceptual models require several physical parameters whose estimation is delicate in poorly monitored areas. However, data-based models may be valuable for modelling and forecasting groundwater level over short and long terms. To that end, four machine learning models, namely: Support Vector Regression, k- Nearest Neighbour (k-NN), Random Forest (RF), and Artificial Neural Network (ANN), are trained, validated, and compared for predicting groundwater level (GWL) at seven piezometers on alluvial groundwater of Tanobart aquifer in Morocco. The results revealed that the ANN models succeeded properly in simulating GWL at five piezometers out of the total seven piezometers considered in this study (NSE = 0.69 to 0.8); the RF was satisfactory at five piezometers (NSE = 0.41 to 0.72) and SVR at three piezometers (NSE = 0.57 to 0.81); the k-NN was the poorest model among all the investigated models (NSE = −1.05 to −0.15). The uncertainty analysis showed that the selected models are accurate overall; the SVR model showed the best forecasting accuracy with the smallest 95% interval prediction error (−0.25 m and 0.11 m) at one piezometer. This study provides new insight to forecast the GWL under a semi-arid context such Tanobart aquifer in Khemesset province, Morocco.