@InProceedings{PahlevanAWSOSKMZ:2024:MaLeCe,
author = "Pahlevan, Nima and Ashapure, Akash and Wainwright, William and
Smith, Brandon and O'Shea, Ryan E. and Saranathan, Arun and Kabir,
Sakib and Maciel, Daniel Andrade and Zhai, Pengwang",
affiliation = "{Science Systems and Applications} and {Science Systems and
Applications} and {Science Systems and Applications} and {Science
Systems and Applications} and {Science Systems and Applications}
and {University of Massachusetts Amherst} and {Science Systems and
Applications} and {Instituto Nacional de Pesquisas Espaciais
(INPE)} and {University of Maryland Baltimore County}",
title = "A machine learning centered atmospheric correction model for multi
and hyperspectral satellite observations",
booktitle = "Proceedings...",
year = "2024",
organization = "Ocean Sciences Meeting",
publisher = "AGU",
abstract = "A primary challenge in aquatic remote sensing is the development
of a robust atmospheric correction (AC) method to estimate remote
sensing reflectance (Rrs) products, defined as the ratio of
water-leaving radiance to downwelling irradiance just above the
water. Our AC method, based on a machine learning model referred
to as Mixture Density Networks (MDNs), has been implemented and
widely tested with 2000+ co-located Rrs matchups of Landsat-8/-9
(Operational Land Imager; OLI) and Sentinel-2 (Multispectral
Instrument; MSI) images over inland and nearshore coastal waters.
Here, we describe an extension of this approach to the
hyperspectral domain by generating products for the Hyperspectral
Imager for the Coastal Ocean (HICO). Our processing system, termed
Aquaverse, begins by leveraging a coupled ocean-atmosphere
radiative transfer model to simulate Rayleigh-corrected
reflectance ( ) for various imaging geometries and atmospheric
conditions void of absorbing gases and Rayleigh effects, via in
situ hyperspectral Rrs measurements. The in situ Rrs and simulated
spectra are then resampled with HICO spectral response functions
and subsequently used to train an ensemble of MDNs to retrieve Rrs
from . Our trained ensemble is then applied to Rayleigh-corrected
HICO imagery, which is validated with available in situ data in
select lakes and coastal estuaries across the U.S. The performance
of our models trained with hyperspectral data is expected to match
that of Landsats OLI and Sentinel-2s MSI with median uncertainties
ranging from ~ 16% (green bands) to ~ 35% (blue bands), indicating
major improvements (i.e., ~2x in the blue bands) with respect to
the performance of existing AC models for inland and coastal
waters. Aquaverse-generated maps (Rrs and downstream products like
chlorophyll-a, total suspended solids, and the absorption by
colored dissolved organic matter) for several HICO images will be
demonstrated and compared with those produced via other
processors. Validation uncertainties of Aquaverse in fresh and
coastal waters on HICO imagery will serve as a proxy for future
global space-borne spectrometers.",
conference-location = "New Orleans",
conference-year = "18-23 Feb. 2024",
urlaccessdate = "13 maio 2024"
}