Uncertainty analysis of gap filled satellite-based primary production in the Ross Sea

For estimations of the ocean primary production (PP), there are three representive algorithms such as the Vertically Generalized Production Model (VGPM), Carbon-based Production Model (CbPM), and the algorithm of Arrigo et al. (1998). These algorithms are mainly based on chlorophyll-a concentration (Chl-a) data. However, the satellite-based Chl-a data have limitations due to heavy clouds in the high latitudes. To resolve such limitations, most of researchers have widely used time-averaging data (monthly or yearly composites) or spatially averaged data. Therefore, the objective of this study is to estimate the PP in the Ross Sea with gap filled CHL data. In order to do that, we employed a machine learning technique. Among machine learning algorithms, the Random Forest was used with several input data based on microwave measurements and reanalysis data and the Chl-a as a target data derived from Park et al. (2019). Then, the reconstructed PP was compared and verified with in-situ Chl-a data. In order to check the uncertainties of spatiotemporally averaged Chl-a data, we compared the reconstruded PP with standard PP product and analyzed the differences in seasonal and interannual variation.