A comparison of multiple regression and neural network techniques for mapping in situ pCO2 data

Using about 138 000 measurements of surface pCO(2) in the Atlantic subpolar gyre (50-70 degrees N, 60-10 degrees W) during 19951997, we compare two methods of interpolation in space and time: a monthly distribution of surface pCO(2) constructed using multiple linear regressions on position and temperature, and a self-organizing neural network approach. Both methods confirm characteristics of the region found in previous work, i.e. the subpolar gyre is a sink for atmospheric CO2 throughout the year, and exhibits a strong seasonal variability with the highest undersaturations occur-ring in spring and summer due to biological activity. As an annual average the surface pCO(2) is higher than estimates based on available syntheses of surface pCO(2). This supports earlier suggestions that the sink of CO2 in the Atlantic subpolar gyre has decreased over the last decade instead of increasing as previously assumed. The neural network is able to capture a more complex distribution than can be well represented by linear regressions, but both techniques agree relatively well on the average values of pCO(2) and derived fluxes. However, when both techniques are used with a subset of the data, the neural network predicts the remaining data to a much better accuracy than the regressions, with a residual standard deviation ranging from 3 to 11 mu atm. The subpolar gyre is a net sink of CO2 of 0.13 Gt-C yr(-1) using the multiple linear regressions and 0.15 Gt-C yr(-1) using the neural network, on average between 1995 and 1997. Both calculations were made with the NCEP monthly wind speeds converted to 10 m height and averaged between 1995 and 1997, and using the gas exchange coefficient of Wanninkhof.