The Animas watershed, located in Colorado and New Mexico, US Accurate prediction of river salinity is challenging due to complex nonlinear relationships between salinity and various environmental factors. Machine learning models have demonstrated remarkable predictive power for critical environmental factors. However, insufficient input data required by the model has challenged salinity prediction. Thus, this study proposed an integrated approach to predict river salinity by combining a Random Forest (RF) model and outputs from the APEX-MODFLOW model. Simulated streamflow, precipitation, air temperature, solar radiation, soil moisture, soil temperature, and groundwater discharge from a calibrated Agricultural Policy / Environmental eXtender integrated with MODFLOW (APEX-MODFLOW) model were used as inputs for the RF model to predict monthly river salinity. The predicted salinity was compared with the observed salinity at six monitoring stations. The prediction results showed a strong predictive performance, with an R 2 value of 0.81 and an RMSE of 79.88 µS/cm. The predicted river salinity increased gradually from upstream to downstream and showed seasonal variation, with lower salinity in warm seasons (May to July) and higher salinity in cold seasons (October to March). This study provides a novel integrated approach for predicting river salinity, particularly in ungauged basins. • The RF model coupled with APEX-MODFLOW was used to predict salinity. • The RF successfully assimilated discontinuous salinity monitoring data. • The integrated APEX-MODFLOW-RF model was used to predict the salinity of 75 ungauged streams. • The combination of RF and APEX-MODFLOW improved predictive performance.
Han et al. (Fri,) studied this question.