Global retrieval of phytoplankton functional types based on empirical orthogonal functions using CMEMS GlobColour merged products and further extension to OLCI data

High Resolution Water Column Phytoplankton Composition Across the Atlantic Ocean From Ship-Towed Vertical Undulating Radiometry

“…Applying the Akaike information criterion (i.e., the significance of an EOF mode in terms of each term's removal for each specific PG or TChla model, see Table 2), proved that for our models also higher EOFs still were significant for the predictions. All our models follow the EOF mode selections found for case-1 waters in Bracher et al (2015) and Xi et al (2020) where most of the variation in the spectral shape was caused by phytoplankton pigments (groups) absorption in addition to water absorption itself: E.g., the EOF-2 mode is the most important term for TChla and nearly all PG Chla models, except for Prochlorococcus and cyanobacteria where the other EOF modes take over and more EOF modes are included in these PG Chla models. This is because for these two groups their Chla does not co-vary with TChla.…”

“…Step 2: For each PG, we subsequently developed the corresponding multiple linear regression model using the collocated PG Chla data and the EOF modes extracted from the AOP data, in which the log-transformed PG Chla data derived from the HPLC measured pigment concentrations, ln(C train o ), are expressed as a function of a subset of the EOF scores (U). As in Xi et al (2020), the EOF modes with standard deviations (singular values from ) that are less than 0.0001 times the standard deviation of the first EOF mode were considered insignificant and thus omitted. Following Bracher et al (2015) a stepwise routine was applied to search for smaller regression models (for each PG model) based on fewer prediction terms through minimization of the Akaike information criterion.…”

“…The first four modes of EOF already explain 99.29, 99.45, and 99.77% of the total variance for the T(z, λ), K d (z 1 → z 2 , λ), andǨ d (λ) based PG prediction models (Table 1), with the first mode explaining 68.83, 71.77, and 84.39%, respectively. Other studies (e.g., Craig et al, 2012;Taylor et al, 2013;Bracher et al, 2015;Xi et al, 2020) have detailed the EOF modes selected for predictions of water constituents. They have investigated the underlying bio-optical signature that several EOF modes may carry.…”

“…The distinct linkage between the EOF modes and the specific pigments or PGs was not identified, as the PG information cannot, to first-order, be reflected by these EOF modes (Craig et al, 2012). We followed Bracher et al (2015) and Xi et al (2020) where not only TChla but also pigment concentrations and PG Chla were predicted, respectively, by including in the prediction models higher EOF modes. Though these contributed only a minute portion to the total AOP variance, they still might inherit the optical signatures of phytoplankton (partly group specific) pigments.…”

“…Previous studies by Taylor et al (2013) and Bracher et al (2015) have retrieved from underwater radiometric measurements, either upwelling radiance or remote sensing reflectance, concentrations of various phytoplankton pigments using empirical orthogonal function (EOF) analysis on the spectral data. Xi et al (2020) have further utilized the method to directly retrieve the Chla of six PGs which explain most of the TChla. In the present study we further adapted this EOF based retrievals by using AOPs derived from downwelling underwater irradiance profile measurements.…”