Utilization of Machine Learning Models and Satellite Data for the Estimation of Total Dissolved Solids in the Colorado River System

Salinity or TDS-related issues are a major cause of economic loss due to their devastating impact on agricultural production. The Colorado River Basin (CRB) Salinity Control Act requires the delivery of water by the US to Mexico from the Colorado River with an average maximum TDS concentration of 115 ± 30 mg/L in addition to the annual average salinity of the river at the Imperial Dam. However, traditional TDS measurements are costly and labor-intensive, and there is a need to estimate changes in TDS along the river system in the CRB. This study developed models for the estimations of TDS using remote sensing data and Machine Learning (ML) models. The ML models used in the study are categorized as standalone and hybrid ML models. Standalone models used include linear, random forest, and extreme gradient boosting regressors. The hybrid ML models include bagging, stacking, and adaptive boosting regressors. The reflectance data from 10-m resolution Sentinel-2A/B and the ground-truth TDS data obtained from the United States Geological Survey are used for the study. The remote sensing data was processed in the Google Earth Engine (GEE) platform. The GEE platform provides an interface to use JavaScript libraries and serves as a toolbox for visualization and geospatial analysis. The model provides a cost-effective approach for the estimation of TDS. For standalone ML algorithms, the analysis produced average Wilmot’s Index of agreement (WI) and Mean Absolute Relative Error (MARE) of 0.62 and 0.36, respectively. The Ensemble ML algorithms produced average WI and MARE of 0.65 and 0.45, respectively. From the results and data processing methods attempted, the ensemble ML methods have not shown better performance compared to the standalone ML models using the model evaluation metrics.