Machine learning and remote sensing-based hierarchical framework for assessing non-optically observable riverine water quality parameters

Optical Remote Sensing (RS)-based satellite imagery have been increasingly used for monitoring water quality in large riverine ecosystems, especially for water quality parameters with high optical sensitivity, such as suspended solids, turbidity and chlorophyll content. However, for non-optically observable components (NOOCs) of water quality, such as dissolved oxygen (DO), electrical conductivity (EC), biochemical oxygen demand (BOD), calcium ( C a 2 + ), sulphate ( S O 4 2 − ), chloride ( C l − ), pH, and fecal coliform, their no direct correlation with reflectance of water necessitates the development of improved techniques. To address this, the present study demonstrated four hierarchical Machine Learning (ML) tool-integrated RS-based methods for assessing these NOOCs using Sentinel-2 (S2) imagery and reanalysis meteorological datasets as the proxy inputs. These ML tools included—Partial Least Squares Regression (PLSR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR), wherein the input datasets were the S2 data (bands, spectral indices), RS-derived total suspended solids (TSS) and turbidity, air temperature from the European Centre for Medium-Range Weather Forecasts Reanalysis version 5 (ERA5). To enhance the radiometric quality of Sentinel-2 imagery affected by clouds and cloud shadows, a Gaussian Mixture Model (GMM) was applied to mask the impacted pixels. These hierarchical methods were assessed in the urban-waste-dominated, ungauged lower reaches of the Ganges (Hooghly River) as a typical case. The results demonstrated that ML models using only remotely sensed and reanalysis temperature data can outperform models that incorporate ground-based measurements, underscoring the potential of remotely sensed and climate reanalysis inputs to assess the NOOCs. Examination of predictor variables using SHapely Additive exPlanations (SHAP) and Variance Importance in Projection (VIP) scores identified the influential predictors in estimating NOOCs. The proposed approach offers a replicable, scalable solution for assessing the non-optically observable water quality parameters in data-scarce and urban-impacted river systems. • Demonstrated satellite and reanalysis climate data as surrogates for water quality prediction. • Estimated non-optically observable water quality by substituting in-situ data with remote inputs. • Partial least squares regression delivered the highest predictive performance. • Identified key predictors influencing water quality using SHAP and VIP analysis.