Dual-Threshold Models for Turbidity Retrieval on Great Slave Lake using MODIS and VIIRS

Turbidity is a crucial indicator of water quality and ecosystem health in large northern lakes. However, logistical reasons, high costs and the spatial complexity of lakes result in a scarcity of observations, hindering long-term monitoring efforts. In this study, we develop a dual-threshold Random Forest (RF) model optimized for low- and high-turbidity regimes to retrieve turbidity from Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) surface reflectance and band ratios over Great Slave Lake (GSL) from 2002 to 2024. Our objectives were to create a turbidity dataset from MODIS and VIIRS, assess sensor performance against in-situ measurements and among sensors, and investigate the spatial, seasonal, and interannual dynamics of turbidity. The trained RF models performed robustly across both sensors and turbid regimes: R² = 0.80 - 0.83, r = 0.89-0.91, IA = 0.94-0.95, MBE within ±0.30 NTU, with RMSE 2.3-2.8 NTU (low-turbidity areas) and 11.4 -14.1 NTU (high-turbidity areas). VIIRS consistently slightly outperformed MODIS, with lower RMSE and higher R², and has consistently higher correlations with input ratios. We observe a significant increasing trend of turbidity across the lake up to ~2 NTU/yr. Both sensors identified persistent high-turbidity zones in GSL’s southeastern region and nearshore areas, associated with river inflow from the Slave River, shallow bathymetry, and wind-driven resuspension, with plume extent expanded during high-inflow years such as 2020 and 2022. Collectively, these results provide the first harmonized continuous MODIS-VIIRS turbidity record for GSL (https://doi.org/10.20383/103.01563), quantify sensor-specific performance, and link satellite retrievals to primary limnological drivers.