Efficient multi-angle polarimetric inversion of aerosols and ocean color powered by a deep neural network forward model

Abstract

Abstract. NASA's Plankton, Aerosol, Cloud, ocean Ecosystem (PACE) mission, scheduled for launch in the timeframe of 2023, will carry a hyperspectral scanning radiometer named the Ocean Color Instrument (OCI) and two multi-angle polarimeters (MAPs): the UMBC Hyper-Angular Rainbow Polarimeter (HARP2) and the SRON Spectro-Polarimeter for Planetary EXploration one (SPEXone). The MAP measurements contain rich information on the microphysical properties of aerosols and hydrosols and therefore can be used to retrieve accurate aerosol properties for complex atmosphere and ocean systems. Most polarimetric aerosol retrieval algorithms utilize vector radiative transfer models iteratively in an optimization approach, which leads to high computational costs that limit their usage in the operational processing of large data volumes acquired by the MAP imagers. In this work, we propose a deep neural network (NN) forward model to represent the radiative transfer simulation of coupled atmosphere and ocean systems for applications to the HARP2 instrument and its predecessors. Through the evaluation of synthetic datasets for AirHARP (airborne version of HARP2), the NN model achieves a numerical accuracy smaller than the instrument uncertainties, with a running time of 0.01 s in a single CPU core or 1 ms in a GPU. Using the NN as a forward model, we built an efficient joint aerosol and ocean color retrieval algorithm called FastMAPOL, evolved from the well-validated Multi-Angular Polarimetric Ocean coLor (MAPOL) algorithm. Retrievals of aerosol properties and water-leaving signals were conducted on both the synthetic data and the AirHARP field measurements from the Aerosol Characterization from Polarimeter and Lidar (ACEPOL) campaign in 2017. From the validation with the synthetic data and the collocated High Spectral Resolution Lidar (HSRL) aerosol products, we demonstrated that the aerosol microphysical properties and water-leaving signals can be retrieved efficiently and within acceptable error. Comparing to the retrieval speed using a conventional radiative transfer forward model, the computational acceleration is 103 times faster with CPU or 104 times with GPU processors. The FastMAPOL algorithm can be used to operationally process the large volume of polarimetric data acquired by PACE and other future Earth-observing satellite missions with similar capabilities.

Long‐Term Trends in Chesapeake Bay Remote Sensing Reflectance: Implications for Water Clarity

“…These future missions will allow for improved characterization of phytoplankton community composition and better estimates of particle size distributions. In particular, polarization may allow for improved partitioning of organic matter versus inorganic matter using the real part of the refractive index and the degree of linear polarization of water leaving radiances (Gao et al, 2021;Loisel et al, 2008;Stamnes et al, 2018;Tonizzo et al, 2009). These planned developments have the potential to further our understanding of water clarity in the Chesapeake Bay using satellite remote sensing.…”

Adaptive Data Screening for Multi-Angle Polarimetric Aerosol and Ocean Color Remote Sensing Accelerated by Deep Learning

“…Accurate estimation of the water-leaving signal requires the quantification and removal of the aerosol path radiance and the ocean surface reflectance from the remote sensing measurement (Mobley et al, 2016). To advance both aerosol and ocean color characterization based on MAP measurements, simultaneous multi-parameter retrieval algorithms have been developed over both open and coastal waters (Chowdhary et al, 2005;Hasekamp et al, 2011;Xu et al, 2016;Stamnes et al, 2018;Gao et al, 2018;Fan et al, 2019;Gao et al, 2019Gao et al, , 2021.…”

An Overview of Neural Network Methods for Predicting Uncertainty in Atmospheric Remote Sensing

“…Neural networks that are used to emulate the forward model and are applied in conjunction with a Bayesian approach to solve the inverse problem [9][10][11][12]; 2.…”

“…For a data model with input and output noise, the computation of the estimate ω is not a trivial task, because the covariance matrix C δ y (z, ω), which enters in the expression of E D (ω), depends on ω. Moreover, in Equation (11), C δ y (z, ω) corresponds to a linearization of the neural network function under the assumption that the noise process in the input space is small. This problem can be solved by implicitly learning the covariance matrix C δ y (z, ω) from the loss function [29].…”

“…It should be pointed out that from Equation (25) that it is apparent that the model is stopped from predicting high uncertainty through the log σ 2 j term, but also low uncertainty for points with high residual error, as low σ 2 j will increase the contribution of the residual. On the other hand, it should also be noted that a basic assumption of the model is that the covariance matrix C δ y (z, ω) is diagonal, which unfortunately, is contradicted by Equation (11). 4.…”