AQUAVERSE-AC: a machine learning-based atmospheric correction framework for inland and coastal waters

Accurately deriving remote sensing reflectance (Rrs) from satellite imagery in optically complex inland and coastal waters is critical for water quality monitoring and management. This study presents AQUAVERSE (an aquatic inversion scheme for remote sensing of fresh and coastal waters), an innovative machine learning-based atmospheric correction (AC) framework, that leverages Mixture Density Networks (MDNs) to retrieve high-quality Rrs from multispectral data acquired by Landsat-8/9 (OLI) and Sentinel-2A/B (MSI) satellites. By coupling a well-validated ocean-atmosphere radiative transfer (RT) model with extensive simulated and in situ datasets, AQUAVERSE compensates for aerosol scattering and atmospheric absorption across diverse environmental and imaging conditions. The model’s performance was evaluated through a comprehensive matchup analysis against AERONET-OC and GLORIA datasets, wherein AQUAVERSE demonstrates substantial improvements over existing AC processors. Median residuals ranged from 15-26% for key spectral bands (coastal-blue to 705nm red-edge band), representing a 20-100% enhancement in accuracy compared to alternative methods. Notably, AQUAVERSE exhibited robust performance in the blue spectrum, where most AC methods underperform. The results, including generated Rrs maps, were benchmarked against outputs from other widely used atmospheric correction processors (ACOLITE, POLYMER, SeaDAS), demonstrating AQUAVERSE's enhanced performance across diverse aquatic environments, while also providing comprehensive uncertainty estimates. These uncertainty metrics, inherently generated by the MDN framework, provide critical confidence indicators for downstream applications. AQUAVERSE represents a significant advancement in atmospheric correction for inland and coastal waters, providing a reliable tool for monitoring water quality in diverse optically complex aquatic systems.