Neural network-based estimates of North Atlantic surface pCO2 from satellite data: A methodological study

A new method is proposed to estimate ocean surface pCO(2) from remotely sensed surface temperature and chlorophyll data. The method is applied to synthetic observations provided by an eddy-resolving biogeochemical model of the North Atlantic. The same model also provides a perfectly known simulated pCO(2) "ground truth'' used to quantitatively assess the success of the estimation method. Model output is first sampled according to realistic voluntary observing ship (VOS) and satellite coverage. The model-generated VOS "observations'' are then used to train a self-organizing neural network that is subsequently applied to model-generated "satellite data'' of surface temperature and surface chlorophyll in order to derive basin-wide monthly maps of surface pCO(2). The accuracy of the estimated pCO(2) maps is analyzed with respect to the "true'' surface pCO(2) fields simulated by the biogeochemical circulation model. We also investigate the accuracy of the estimated pCO(2) maps as a function of VOS line coverage, remote sensing errors, and the interpolation of missing remote sensing data due to cloud cover and low solar irradiation in winter. For a simulated "sampling'' corresponding to VOS lines and patterns of optical satellite coverage of the year 2005, the neural net can successfully reproduce pCO(2) from model-generated "remote sensing data'' of SST and Chl. Basin-wide RMS errors amount to 19.0 mu atm for a hypothetical perfect interpolation scheme for remote sensing data gaps and 21.1 mu atm when climatological surface temperature and chlorophyll values are used to fill in areas lacking optical satellite coverage.