Evaluating lake water quality monitoring using hyperspectral remote sensing and machine learning

Abstract Enhancing lake water quality monitoring is crucial for environmental sustainability and human health. Despite recent advancements, gaps remain in accurately estimating water quality across diverse lake types using remote sensing techniques. This study leverages in situ hyperspectral data (400–800 nm) and machine learning (ML) to estimate four key water quality parameters—CHL-a, SD, TSS, and a CDOM (440)—using the Global Reflectance Community Dataset. Among six ML models evaluated, random forest (RF) performed best in CHL-a (R 2 = 0.790) and CDOM-rich (R 2 = 0.658) waters but showed lower accuracy in clear (R 2 = 0.263) and sediment-rich (R 2 = 0.408) lakes. Spatially split validation across lakes resulted in only ~12% accuracy reduction, confirming model robustness. Simulations with multispectral data revealed that Sentinel-2′s red-edge bands (705, 740, 783 nm) notably improved CHL-a (R 2 = 0.774) and TSS (R 2 = 0.697) retrieval compared to Landsat-8. Our results, based on the latest global water reflectance data compilation, support the advancement of ML-based algorithms for satellite-driven water quality monitoring.