The nonlinearity of the relationship between CO2 flux and other micrometeorological variables flux parameters limits the applicability of carbon flux models to accurately estimate the flux dynamics. However, the need for carbon dioxide (CO2) estimations covering larger areas and the limitations of the point eddy covariance technique to address this requirement necessitates the modeling of CO2 flux from other micrometeorological variables. Artificial neural networks (ANN) are used because of their power to fit highly nonlinear relations between input and output variables without explaining the nature of the phenomena. This paper applied a multilayer perception ANN technique with error back propagation algorithm to simulate CO2 flux on three different ecosystems (forest, grassland and cropland) in ChinaFLUX. Energy flux (net radiation, latent heat, sensible heat and soil heat flux) and temperature (air and soil) and soil moisture were used to train the ANN and predict the CO2 flux. Diurnal half-hourly fluxes data of observations from June to August in 2003 were divided into training, validating and testing. Results of the CO2 flux simulation show that the technique can successfully predict the observed values with R2 value between 0.75 and 0.866. It is also found that the soil moisture could not improve the simulative accuracy without water stress. The analysis of the contribution of input variables in ANN shows that the ANN is not a black box model, it can tell us about the controlling parameters of NEE in different ecosystems and micrometeorological environment. The results indicate the ANN is not only a reliable, efficient technique to estimate regional or global CO2 flux from point measurements and understand the spatiotemporal budget of the CO2 fluxes, but also can identify the relations between the CO2 flux and micrometeorological variables.
