Demonstrating Fresh and Coastal Water Products from PACE/OCI Proxy Observations
Ryan E. O'Shea, Science Systems and Applications, Inc. (SSAI) and NASA Goddard Space Flight Center, ryan.oshea@ssaihq.com (Presenter)
Nima Pahlevan, Science Systems and Applications, Inc. (SSAI) and NASA Goddard Space Flight Center, nima.pahlevan@nasa.gov
Brandon Smith, Science Systems and Applications, Inc. (SSAI) and NASA Goddard Space Flight Center, brandon.smith@ssaihq.com
Caren Binding, Environment and Climate Change Canada, caren.binding@canada.ca
Emmanuel Boss, School of Marine Sciences, University of Maine, emmanuel.boss@maine.edu
Ruth Briland, Division of Drinking and Ground Waters, Ohio Environmental Protection Agency, ruth.briland@epa.ohio.gov
Todd Egerton, Virginia Department of Health, todd.egerton@vdh.virginia.gov
Raphael Kudela, Ocean Sciences Department, Institute of Marine Sciences, University of California-Santa Cruz, kudela@ucsc.edu
Stephanie Schollaert Uz, NASA Goddard Space Flight Center, stephanie.uz@nasa.gov
Jennifer Wolny, Office of Regulatory Science, Center for Food Safety and Applied Nutrition, U.S. Food and Drug Administration, jennifer.wolny@fda.hhs.gov
Simultaneously estimating multiple biogeochemical parameters (BPs) and absorbing inherent optical properties (IOPs) from inland and coastal waters via remote sensing is a challenging non-unique inverse problem. Mixture density networks (MDNs), a class of machine learning models that output the probability distribution of each product, are well suited to solve this non-unique inverse problem, as the most probable of the estimated values for a given input can be selected instead of the average. We train our MDNs on a globally distributed dataset (N=8,237) of co-aligned BPs and absorbing IOPs. The BPs include total suspended sediment and the pigments chlorophyll-a and phycocyanin. The absorbing IOPs include absorption from colored dissolved organic matter, phytoplankton, and non-algal particles. We rigorously demonstrate the MDNs performance via a hold-out assessment (50/50 training/testing split), a regional cross-validation assessment (to test generalization performance), and a matchup comparison between in situ measurements and satellite-derived values. Our MDNs outperform (achieve lower uncertainty than) operational multispectral algorithms in both the hold-out assessment and regional cross-validation assessment for all products (except for phycocyanin). Product maps of Lake Erie and the Curonian Lagoon, produced by applying the MDN to imagery from the Hyperspectral Imager for the Coastal Ocean (HICO) and the PRecursore IperSpettrale della Missione Applicativa (PRISMA), have high spatial consistency, match with co-aligned in situ measurements of BPs and IOPs, and agree with the literature’s understanding of the regions. The performance of the MDNs for BP and IOP retrieval from optically complex waters via past and present hyperspectral imagers, despite associated uncertainties, demonstrates its viability for future hyperspectral missions (e.g., the Plankton, Aerosol, Cloud, ocean Ecosystem (PACE)). The BPs and IOPs produced by the MDN will serve a key role in phytoplankton community composition analysis from hyperspectral satellite imagery.
Poster: Poster_OShea_1-19_55_35.pdf
Associated Project(s):
Poster Location ID: 1-19
Presentation Type: Poster
Session: Poster Session 1
Session Date: Tue (May 9) 5:00-7:00 PM
CCE Program: OBB