Practical aspects of providing pixel-level spectral Rrs error covariance in satellite ocean color products

Minwei Zhang^1*Amir Ibrahim²Bryan A Franz²Andrew Sayer³Jeremy Werdell²Lachlan McKinna⁴

• ¹Science Applications International Corporation (United States), McLean, United States
• ²NASA Goddard Space Flight Center, Greenbelt, United States
• ³University of Maryland Baltimore County, Baltimore, United States
• ⁴4Go2Q Pty Ltd, Sunshine Coast, QLD, Australia

We previously established a derivative-based approach to generate pixel-level spectral error covariance matrix in satellite-retrieved remote sensing reflectance, ∑Rrs. However, one practical issue is the delivery of the products without increasing file size by an order of magnitude or more, considering that for N sensor spectral bands there are N´(N+1)/2 covariance matrix elements to be specified at each pixel. The issue becomes more pertinent for hyperspectral imaging spectroradiometers such as Ocean Color Instrument (OCI) on NASA's Plankton, Aerosol, Cloud, ocean Ecosystem mission (PACE), which has 286 bands, resulting in ~40, 000 unique elements in ∑Rrs per pixel that would lead to a ~60 GB Level-2 file for one 5-minute granule. As a first step to tackle the issue, we took OCI and Moderate Resolution Imaging Spectroradiometer (MODIS) data to explore the possibility of approximating ∑Rrs using a third-degree polynomial, thereby decreasing the memory overhead to 4´N numbers. We found that ∑Rrs derived from the polynomial fitting match well with the original value, with the difference smaller than 5%. We then compared the relative uncertainty in two derived ocean color data products (chla and Kd(490)) calculated using the original fully-computed ∑Rrs and then using the polynomial model approximation for ∑Rrs, finding the absolute difference between the two approaches to be smaller than 0.5%. These evaluations suggest the polynomial approximation of ∑Rrs is suitable without degrading scientific quality. By including the coefficients derived from polynomial fitting instead of the full error covariance matrix, a typical five-minute Level-2 file for OCI decreases from ~60 GB to a more practical to handle ~1.7 GB.