Hyperspectral particulate backscattering coefficient retrieval in inland and coastal waters via physics-informed mixture density networks

The absorbing and backscattering inherent optical properties (IOPs) of water and its associated constituents shape the ocean’s color. We previously developed machine-learning-based models, mixture density networks (MDNs), to retrieve the absorbing (a(λ)) components in optically complex coastal and inland waters from hyperspectral satellite imagery (O’Shea et al. 2023). Well-established algorithms, such as the Quasi-Analytical Algorithm (QAA, Lee et al. 2002), the Generalized IOP (GIOP) model (Werdell et al. 2013), and the semi-analytical Gordon model (Gordon et al. 1988) can retrieve bbp(λ) from hyperspectral imagery; however, these models were designed for open ocean conditions, not optically complex inland and coastal waters. In this work, we will create a physics-informed MDN for bbp(λ) retrieval from inland and coastal waters. The MDN will learn the relationship between a(λ), bbp(λ), and Rrs(λ) via the inclusion of the semi-analytical Gordon et al. model during model training. For training and validation of the physics-informed MDN we have a large (N~500) dataset of in situ measured multispectral bbp from multiple optically distinct regions. Additionally, we will compare bbp(λ) derived using our physics-informed MDN and operational models (e.g., QAA, GIOP) to in situ bbp(λ) from Lake Erie. Finally, we will qualitatively compare bbp(λ) retrieved using the MDN and operational models from the imagery of Lake Erie taken from the Ocean Color Instrument (OCI) aboard PACE to evaluate the impact of atmospheric correction uncertainties. Overall, we expect the physics-informed MDN to better retrieve IOPs from inland and coastal waters.