Salinity-constituent conversion in South Sacramento-San Joaquin Delta of California via machine learning.

The levels of total salinity and its ion constituents in estuarine environments are important indicators of the overall suitability of water for environmental, agricultural, and urban use. These constituents include Total Dissolved Solids (TDS), dissolved chloride (Cl⁻), dissolved sulfate (SO4²⁻), dissolved sodium (Na⁺), dissolved calcium (Ca²⁺), dissolved magnesium (Mg²⁺), dissolved nitrate (NO3), dissolved potassium (K), dissolved bromide (Br⁻), dissolved boron (B), Alkalinity, and hardness, among others. In practice, salinity is typically measured indirectly as electrical conductance (EC) via automatic sensors while the concentration of each constituent is often measured from discrete water samples (i.e., grab samples) and thus is available much less frequently than salinity. Quadratic regression equations are generally developed between salinity (as the predictor, represented by EC) and individual constituents (as the predictand) based on grab sample data. The regression models are then applied to estimate the concentrations of individual constituents given EC during the period when grab samples are not available. The current study develops four types of machine learning models: the Generalized Additive Model, Regression Trees, Random Forest, and Artificial Neural Networks, to emulate conventional regression equations in salinity-constituent conversion. A case study in the South Sacramento-San Joaquin Delta of California is presented to illustrate the development and application of these models and to compare their performance with that of the benchmark regression models. The results indicate that machine learning models can provide comparable or superior simulations to the regression models. Among the four machine models examined, the Random Forest model tends to yield the best results. The study further discusses the scientific and practical implications of the models proposed, their limitations, as well as future work to further improve their performance and applicability. [ABSTRACT FROM AUTHOR]