Harnessing Ensemble Machine Learning Models for Improved Salinity Prediction in Large Basin Scales

This study develops a robust predictive model for average annual salinity using ensemble machine learning to combine multiple machine learning algorithms. Salt concentration is a crucial water quality indicator, and salinity issues cost $300 million annually in the U.S. Irrigated agricultural lands in the Upper Colorado River Basin contribute excessively to dissolved solid loads despite covering a small area. Baseflow accounts for most salinity loads, with higher yields from watersheds with irrigated agriculture. Controlling salinity in the Colorado River Basin is crucial for maintaining water quality and minimizing economic damage. Using twenty years of data from 150 watersheds, eleven ML algorithms were evaluated, with Extreme Gradient Boosting, Gradient Boosting, and Random Forest emerging as top performers. Bayesian Model Averaging and stacked generalization created an ensemble from these three models, demonstrating high-performance validity in both training and testing datasets. The ensemble's performance was comparable to the best solo model with slightly higher median and 10th percentile R2 values. The watershed area and river flow were the most important predictors. Overall, the ensemble model demonstrates the potential of machine learning for modeling and predicting water salinity, providing a basis for future refinements. The success highlights the validity of machine learning approaches for understanding and managing salt levels in the Colorado River Basin.