Estimation of phytoplankton size fractions based on spectral features of remote sensing ocean color data

Abstract

AbstractThrough its influence on the structure of pelagic ecosystems, phytoplankton size distribution (pico‐, nano‐, and micro‐plankton) is believed to play a key role in “the biological pump.” In this paper, an algorithm is proposed to estimate phytoplankton size fractions (PSF) for micro‐, nano‐, and pico‐plankton (fm,fn, andfp, respectively) from the spectral features of remote‐sensing data. From remote‐sensing reflectance spectrum (Rrs(λ)), the algorithm constructs four types of spectral features: a normalizedRrs(λ), band ratios, continuum‐removed spectra, and spectral curvatures. Using support vector machine recursive feature elimination, the algorithm ranks the constructed spectral features andRrs(λ) according to their sensitivities to PSF which is then regressed against the sensitive spectral features through support vector regression. The algorithm is validated with (1) simulatedRrs(λ) and PSF, and (2)Rrs(λ) obtained by Sea‐viewing Wide Field‐of‐view Sensor (SeaWiFS) and PSF determined from High‐Performance Liquid Chromatography (HPLC) pigments. The validation results show the overall effectiveness of the algorithm in estimating PSF, withR2of (1) 0.938 (fm) for the simulated SeaWiFS data set; and (2) 0.617 (fm), 0.475 (fn), and 0.587 (fp) for the SeaWiFS satellite data set. The validation results also indicate that continuum‐removed spectra and spectral curvatures are the dominant spectral features sensitive to PSF with their wavelengths mainly centered on the pigment‐absorption domain. Global spatial distributions offm,fn, andfpwere mapped with monthly SeaWiFS images. Overall, their biogeographical distributions are consistent with our current understanding that pico‐plankton account for a large proportion of total phytoplankton biomass in oligotrophic regions, nano‐plankton in transitional areas, and micro‐plankton in high‐productivity regions.