DEVELOPING MACHINE LEARNING MODELS TO ASSESS ANTHROPOGENIC IMPACTS ON U.S. FRESHWATER SALINIZATION SYNDROME

The salinization and alkalinization of fresh water have been occurring across the continental scale of the United States (U.S.) over the past decades. The trend of salinization and alkalinization can threaten drinking water security, agricultural production, sector water use, and ecosystem health. Previous studies have suggested that such salinization and alkalinization in the U.S. freshwater are mainly caused by human activities (i.e., road salting, wastewater, agricultural irrigation) and human-accelerated weathering. However, little is known about a mechanistic explanation for whether and to what extent human-accelerated weathering has impacted river alkalinity. In this research, dissolved sodium flux (i.e., salinity proxy) and alkalinity flux (i.e., weathering proxy) were compiled along with 32 watershed properties, including features from hydrology, climate, geomorphology, geology, soil chemistry, land use, and land cover, for 228 river monitoring sites across the contiguous U.S. After feature selection, a total of 18 features were used to build the machine learning model to predict monthly salinity and alkalinity flux at these sites. The assessment of predicted errors and residuals showed no spatiotemporal overfitting in the model development. Based on conditional permutation importance. We found that human activities (i.e., population density and impervious surface percentage) were the two most important features in sodium flux prediction. This suggests that human activities are the main driver of the variation of salts in the U.S. rivers, consistent with previous studies. However, unlike the sodium flux model, no human-related features appeared in the top five important features (the top two features were runoff and the percentage of carbonate sediment) for the alkalinity flux model. Our study indicates that natural processes were still the major drivers of U.S. river alkalinization syndrome on the continental scale.