Frontiers in Remote Sensing (May 2023)
A neural network approach to the estimation of in-water attenuation to absorption ratios from PACE mission measurements
Abstract
Introduction: In preparation for the upcoming PACE mission, we explore the feasibility of a neural network-based approach for the conversion of measurements of the degree of linear polarization at the top of the atmosphere as carried out by the HARP2 instrument into estimations of the ratio of attenuation to absorption in the surface layer of the ocean. Polarization has been shown to contain information on the in-water inherent optical properties including the total attenuation coefficient, in contrast with approaches solely based on remote sensing reflectance that are limited to the backscattered fraction of the scattering. In turn, these properties may be further combined with inversion algorithms to retrieve projected values for the optical and physical properties of marine particulates.Methodology: Using bio-optical models to produce synthetic data in quantities sufficient for network training purposes, and with associated polarization values derived from vector radiative transfer modeling, we produce a two-step algorithm that retrieves surface-level polarization first and attenuation-to-absorption ratios second, with each step handled by a separate neural network. The networks use multispectral inputs in terms of the degree of linear polarization from the polarimeter and the remote sensing reflectance from the Ocean Color Instrument that are anticipated to be fully available within the PACE data environment.Result and Discussion: Produce results that compare favorably with expected values, suggesting that a neural network-mediated conversion of remotely sensed polarization into in-water IOPs is viable. A simulation of the PACE orbit and of the HARP2 field of view further shows these results to be robust even over the limited number of data points expected to be available for any given point on Earth’s surface over a single PACE transit.
Keywords
