Can Sentinel-2-derived spectral indices improve the accuracy of retrieving optically active water quality parameters using machine learning algorithms?

This study was conducted in the Cradle of Humankind World Heritage Site (COHWHS), South Africa, an area characterised by interconnected surface waters and sensitive dolomitic aquifers. The region is subject to increasing pressure from land use change, tourism, and nutrient enrichment, making reliable and spatially explicit water quality monitoring essential for protecting its ecological, cultural, and hydrological integrity. The study aimed to assess whether Sentinel-2-derived spectral indices improve the retrieval accuracy of optically active water quality parameters, namely Chlorophyll- a (Chl- a ) and Total Suspended Solids (TSS). Three input configurations were tested: traditional Landsat-like bands, Sentinel-2 bands, and Sentinel-2 bands combined with spectral indices. These inputs were used within Random Forest and Gaussian Process Regression models to evaluate model performance across wet (summer) and dry (winter) seasons. The results show that integrating Sentinel-2 spectral indices substantially improves Chl- a estimation during wet conditions, while TSS retrieval benefits mainly from Sentinel-2 red, red-edge, and SWIR bands. Model performance was strongly seasonal, with reduced accuracy during dry periods due to lower optical variability. The findings provide new insight into how seasonal hydrological conditions and spectral sensitivity influence water quality retrievals in optically complex inland waters of the COHWHS. This approach supports improved regional water quality monitoring and contributes to the protection of connected surface water-groundwater systems in this vulnerable heritage landscape. • Spectral indices improve Sentinel-2 retrieval of Chl- a in optically complex waters. • Gaussian Process Regression outperforms Random Forest across seasons and inputs. • Red-edge and blue bands are key drivers of Chl- a and TSS prediction accuracy. • Wet-season conditions yield stronger water quality predictions than dry seasons. • Sentinel-2 enables cost-effective monitoring of inland water quality at regional scale. • Results support improved water quality management in sensitive catchments.