Improving the accuracy of remotely sensed TSS and turbidity using quality enhanced water reflectance by a statistical resampling technique

Satellite-based optical Remote Sensing (RS) presents a unique opportunity for monitoring the water quality of complex riverine systems stretching over large regions. However, due to the weak reflectance and resulting low signal-to-noise ratio of water bodies, interference of clouds and cloud shadows significantly impacts the accuracy of remotely sensed water quality parameters. This study presents a scalable, innovative cloud and cloud-shadow masking using the Gaussian Mixture Model (GMM) combined with a spatially aggregated reflectance sampling approach that can robustly monitor the Total Suspended Solids (TSS) and turbidity over selected sections of the study site, the Hooghly River in West Bengal, India. The statistical resampling approach based on GMM was applied to Sentinel-2 (S2) imagery to produce input to Machine Learning (ML) algorithms to retrieve the TSS and turbidity for target river sections. The resampled reflectance data was spatially aggregated over the selected regions of interest to further improve the input quality. Out of 80 cloud-contaminated images, we were able to use 70 images with 40%–50% clouds/cloud shadows for TSS and turbidity retrievals after applying the GMM-based masking and spatial aggregation. The resampled spectral data and in-situ TSS and turbidity measurements were used to train four ML models: Partial Least Squares Regression (PLSR), Random Forest Regression (RFR), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR). Our approach improved the estimation accuracy of the TSS by 33% and turbidity by 28% compared to the input processed using the built-in Sentinel-2 cloud and cloud shadow masks. This work emphasizes the importance of careful satellite image preparation for cloud and cloud-shadow screening and the utility of spatially aggregated reflectance samples over homogeneous regions of interest to enable robust water quality assessment under variable atmospheric conditions. • Sentinel-2 pixels aggregated from homogeneous ROIs for water quality sensing. • A statistical technique significantly improved Sentinel-2 based water quality sensing. • Robust performance of Machine Learning models with enhanced spectral data. • Key Sentinel-2 data for TSS/turbidity sensing identified by SHAP and VIP analysis.