Predicting soil salinity in wastewater land application systems under field conditions with tree-based machine learning approaches

Wastewater land application is widely recognized as a solution to global water scarcity. However, concerns about its environmental sustainability arise due to potential soil salinization. Three tree-based machine learning models were devised to predict soil salinity using data from previous field studies, evaluated using R-2 and RMSE, and interpreted using permutation importance, partial dependence plots, Shapley value plots, and Break Down plots. Data was preprocessed and split into 80% training and 20% test sets prior to being subjected to 10-fold cross-validation and hyperparameter tuning via grid search. Random forest (R-2 = 0.920, RMSE = 0.374 dS/m) performed better than decision tree (R-2 = 0.832, RMSE = 0.614 dS/m) and gradient boosting decision trees (R-2 = 0.918, RMSE = 0.396 dS/m) on test data. The permutation importance order was: initial soil EC > initial soil pH > wastewater pH > total irrigation days > total precipitation > total irrigated wastewater amount > soil depth > wastewater EC. This study provides new insights into how different input features impact soil salinity and has important implications for sustainable management of wastewater land application to ensure that this solution to water scarcity is both effective and environmentally sustainable.