Modeling ion constituents in the Sacramento-San Joaquin Delta using multiple machine learning approaches

Abstract

Abstract Salinity is of paramount importance in shaping water quality, ecosystem health, and the capacity to sustain diverse human and environmental demands in estuarine environments. Electrical conductivity (EC) is commonly utilized as an indirect measure of salinity, serving as a proxy for estimating other ion constituents within the Sacramento-San Joaquin Delta (Delta) of California, United States. This study investigates and contrasts four machine learning (ML) models (Regression Trees, Random Forest, Gradient Boosting, and Artificial Neuronal Networks) for approximating ion constituent concentrations based on EC measurements, emphasizing the enhancement of conversion for constituents exhibiting pronounced non-linear relationships with EC. Among the four models, the Artificial Neuronal Networks model outshines the others in predicting ion constituents from EC, especially for those displaying strong non-linear relationships with EC. All four ML models surpass traditional parametric regression equations in terms of accuracy in estimating ion concentrations. Furthermore, an interactive web browser-based dashboard is developed, catering to users with or without programming expertise, enabling ion level simulation within the Delta. By furnishing more precise ion constituent estimations, this research enriches the understanding of salinity's effects on water quality in the Delta and fosters well-informed water management decisions.