OCRT-Net: A neural network embedding ocean color radiative transfer theory for retrieving optical water constituents

Accurate retrieval of optically active components (OACs), including chlorophyll-a (Chl-a) concentration, suspended particulate matter (SPM) concentration, and the absorption coefficient of colored dissolved organic matter (CDOM) [ a g (443)], from satellite observations remains challenging in aquatic remote sensing. Although deep learning has shown promise for this goal, its black-box nature produces physically inexplicable outputs and makes it vulnerable to the ill-posed inversion problem arising from overlapping spectral information among the OACs, which is further compounded by strong variability in their specific absorption and scattering properties. To address these limitations, this study proposes the Ocean Color Radiative Transfer Network (OCRT-Net), which embeds an analytical model of ocean color in the network architecture as a fully differentiable layer. This structural design constrains the internal latent space to reconstruct physically consistent inherent optical properties (IOPs) at every forward pass, guaranteeing optical closure by architecture rather than by post-hoc regularization. The model was optimized using a two-stage transfer learning strategy: it was pre-trained on HydroLight-simulated spectra (N = 12,000) to establish optical closure, followed by fine-tuning on the global GLORIA in situ dataset (N = 2,049) to capture real-world optical variability. Independent validation across Sentinel-3 OLCI, Sentinel-2 MSI, and SNPP VIIRS configurations demonstrated robust retrieval performance for OACs spanning several orders of magnitude (uncertainty: 35%–55%), outperforming several established empirical and semi-analytical benchmark algorithms. Application to satellite matchups and full-scene OLCI imagery over ten optically diverse lakes in China confirmed that OCRT-Net produces physically consistent retrievals and spatially coherent OAC distributions, free of the noise-induced artifacts and physically implausible values commonly associated with purely data-driven approaches. By bridging radiative transfer theory with data-driven optimization, OCRT-Net moves beyond the black-box paradigm, offering an interpretable and sensor-transferable framework for operational water quality monitoring across inland and coastal waters.