Remote Sensing Imagery and Machine Learning-Based Methods for Quantifying Total Dissolved Solids and Total Suspended Solids Concentration in River Systems: A Case Study of the Colorado River Basin

The Colorado River and its tributaries housed in the Colorado River Basin (CRB) are the primary source of water to the western United States and the Republic of Mexico. The river system is under intense stress due to prolonged drought and anthropogenic activities which have worsened its water quality. Total dissolved solids (TDS) and total suspended solids (TSS) are two water quality parameters (WQPs) that are crucial to the sustainability of the river system. These parameters are noted to have caused varied severity to the sustenance of the basin’s water. Monitoring of these WQPs has been conventionally conducted using field and laboratory analysis which are cost and labor-intensive. This study utilized a novel method to effectively develop machine learning (ML) models to estimate TDS and TSS concentrations in the CRB by utilizing the potential of optically sensitive multispectral Sentinel 2 A/B Multispectral Scanners (MSI) and Landsat 8 Operational Land Imager (OLI) remote sensing (RS) data retrieved from the Google Earth Engine (GEE) and in situ measurements collected from 2013-2022. Several standalone models such as linear regressions (LR), ridge regressions (Ridge), lasso regressions (Lasso), and k-nearest neighbor (KNN), and ensemble methods including the gradient boosting machines (GBM), random forest (RF), adaptive boosting (AdaBoost), eXtreme gradient boosting (XGBoost), and bagging were applied for the accurate estimation of TDS and TSS. Results found ensemble models like the XGBoost as the most optimal model estimating TDS using images from both Sentinel-2 MSI and Landsat 8 OLI with performance on the external validation dataset derived as 0.99, 26.52 mg/L, and 19.19 mg/L, respectively for R², RMSE, and MAE for Sentinel-2 images. The XGBoost yielded R², RMSE, and MAE of 0.97, 35.82 mg/L, and 27.90 mg/L, respectively. The AdaBoost was found to be best model for TSS estimations with values of 0.92, 29.48 mg/L, and 24.64 mg/L, respectively, for R², RMSE, and MAE for the Sentinel-2 image on the external validation dataset. The RF model was found to be the optimal model for TSS estimations with the Landsat 8 OLI with reported R², RMSE, and MAE of 0.90, 32.80 mg/L, and 22.91 mg/L, respectively on the external validation dataset. These findings show great potential of utilizing RS data to produce cost-efficient spatiotemporal changes on the WQPs which has an important implication for the continuous management of the limited water resources. Further study should consider the effect of land use land cover, runoff, and other climatic data to understand the complexity and dynamics of these parameters on TDS and TSS in the river.