Carbon dioxide sources for 1981–1995 are calculated using a mass-balance inversion method. Options for deriving a suitable surface constraint for the model from the available observed data are assessed. A new method, in which the longitudinal variation in the observations is accounted for, is compared to a more conventional spline fit to the data. This new constraint is applied either to the zonal mean concentration or to the concentration at all grid points. The results indicate relatively large differences between long-term mean regional sources but relatively good agreement of the interannual variations. The global and tropical sources estimated show a surprising relationship to the southern oscillation index; maximum correlation occurs for CO2 sources leading the SOI by around 9 months.