Comparison between multivariate regression and neural network
Answers
Answered by
3
Using about 138 000 measurements of surface pCO2 in the Atlantic subpolar gyre (50–70◦N, 60–10◦W) during 1995–
1997, we compare two methods of interpolation in space and time: a monthly distribution of surface pCO2 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 occurring in spring
and summer due to biological activity. As an annual average the surface pCO2 is higher than estimates based on available
syntheses of surface pCO2. 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 pCO2 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 µatm. 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.
1997, we compare two methods of interpolation in space and time: a monthly distribution of surface pCO2 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 occurring in spring
and summer due to biological activity. As an annual average the surface pCO2 is higher than estimates based on available
syntheses of surface pCO2. 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 pCO2 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 µatm. 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.
Similar questions