LAUSR

Abstract Geophysical methods are laborious for monitoring the extent of the seawater intrusion (SWI) in coastal aquifers and water quality over time. However, the data-based approach can be valuable to overcome this challenge by predicting the chloride concentration and sodium adsorption ratio (SAR) using physical parameters as inputs. To this end, four ML models were developed according to four combinations of input variables using 176 and 37 samples related to the coastal aquifer of Chaouia in Morocco for the training and validation processes, respectively. Results revealed that the Stochastic Gradient Descent for linear regression (SGD), Artificial Neural Network (ANN), k Nearest Neighbors (k-NN), and Support Vector Machine (SVM) have acceptable to good performances in predicting the chloride (with r in [0.84,0.97], RMSE in [1.74, 2.67] meq L−1, and RBIAS in [ − 11, − 9]%) and SAR (with r in [0.86,0.90], RMSE in [1.02,1.1] meq0.5 L−0.5, and RBIAS in [13,18]%) during the validation phase using electrical conductivity as input. The uncertainty analysis has shown that the distribution of the model errors during the validation phase is Gaussian (small standard error), indicating the model stability. Besides, the ANN and SGD models are the most accurate and stable models of 95% confidence bands of error, 1.39 meq L−1 and 0.52 meq0.5 L−0.5 for predicting the chloride concentration and the SAR, respectively. Such results provide new insights to nowcasting the SWI trend and water quality in coastal aquifers, as the input variable is measurable in real-time by sensor technologies. Overall, the implementation of these models is useful for the prediction of the SWI trend and water quality in other coastal aquifers.