Globally validated non-unique inversion framework to estimate optically active water quality indicators using in situ and space-borne hyperspectral data sets

Abstract

As water quality is a combination of multiple optically active parameters, there is a growing interest in probabilistic models to predict water quality. This study aims to add to the water quality prediction studies by introducing ensemble learning with deep learning-based mixture density networks with multiple probabilistic Gaussian distributions. We named the approach as Ensembled Gaussian Mixture Density Network (GMDN). Many existing water quality algorithms rely on localized data sets, which limits their applicability. This research addresses this by developing and evaluating the proposed model using the global in situ water quality data set GLORIA (Global Reflectance community data set for Imaging and optical sensing of Aquatic environments). We focused on estimating two key biogeochemical components (BPs): Total Suspended Solids (TSS) and Chlorophyll-a(Chla), along with one inherent optical property (IoP), the absorption coefficient of colored dissolved organic matter (α<inf>CDOM</inf>). The proposed approach performs quite reliably when evaluated on the data samples of individual countries. The GMDN algorithm has been fine-tuned on the satellite-matchup for the river Ganga near Varanasi city. The fine-tuning was implemented using the remote sensing reflectance (R<inf>rs</inf>) of the spaceborne hyperspectral data set PRISMA (PRecursore IperSpettrale della Missione Applicativa). The contribution of the riverbed floor to the Rrs of PRISMA has been computed using physics-based simulations in the Water Color Simulator (WASI). Overall, the simultaneous use of multiple probabilistic distributions and ensembled architectures improves the predictive accuracy of WQ parameters compared to the existing operational algorithms. © Higher Education Press 2025.