Machine learning driven forecasts of agricultural water quality from rainfall ionic characteristics in Central Europe

Sodium hazard poses a critical threat to agricultural production globally and regionally which has been previously predicted from ground or surface water. Monitoring rainwater quality in this context is ignored but essential for agricultural water management in central Europe. Our study focused to predict sodium adsorption ratio (SAR) from 1985 to 2021 from ten ionic species of rainwater (pH, EC, Cl-, SO4−2, NO3-, NH4+, Na+, K+, Mg2+, Ca2+) employing four machine learning (random forest (RF), gaussian process regression (GU), random subspace (RSS), and artificial neural network-multilayer perceptron (ANN-MLP)) methods at three stations K-puszta (KP), Farkasfa (FAK), and Nyirjes (NYR) of Hungary, central Europe. Exploratory data analysis was performed using the Mann-Kendall test, Pearson correlation, and principal component analysis (PCA). Rainwater composition revealed the highest percentage of SO4−2 ions i.e., 21 to 31%, followed by 10 to 15% of Na+ ions. Mann-Kendall test revealed a significant (p < 0.05) increasing trend of Na+ ions and SAR portraying it a serious hazard limiting agricultural production. Machine learning results from 10 model runs of all algorithms for SAR prediction at KP station proved the efficacy of ANN-MLP as superior with RMSE range of 0.02 to 0.05, followed by RF with RMSE of 0.14 to 0.19 in scenario 2 (SC-2) (Na+, Mg2+, Ca2+). Validation of the best-selected algorithm (ANN-MLP) and scenario (SC-2) also predicted the SAR with a low RMSE of 0.08 and 0.05 at both FAK and NYR stations, respectively. Hence, the efficiency of ANN-MLP in forecasting SAR from rainwater proves it to be a meticulous tool for enhancing agricultural water management practices in Central Europe and enhancing resource efficiency and crop production in the future.