Combining flux variance similarity partitioning with artificial neural networks to gap-fill measurements of net ecosystem production of a Pacific Northwest Douglas-fir stand

We propose a gap-filling model for carbon dioxide fluxes measured by eddy covariance (EC) that combines the flux variance similarity (FVS) partitioning approach with the artificial neural network (ANN) technique (FVS–ANN). 18 years of EC-measured net ecosystem production (NEP) of a Douglas-fir (Pseudotsuga menziesii) stand in British Columbia, Canada were used. FVS-partitioned fluxes were used as a training dataset for ANNs based on the environmental variables to model missing respiration and photosynthesis values. Results from the FVS–ANN model were compared with three established models using different partitioning and gap-filling approaches, namely the nighttime relationship, the daytime intercept, and an ANN-based model informed by the nighttime relationship (ANN–nighttime). In the gap-filled long-term NEP record, the 18-year mean NEP values were similar (between 225 and 239 gC m−2 year−1) and individual years showed small differences in annual NEP among the four models, demonstrating the reliability of estimated NEP when any of the models were used for gap-filling. However, the largest differences in annual gap-filled photosynthesis and respiration estimates among the four models in a given year were as large as 813 and 835 gC m−2 year−1, respectively. The FVS–ANN model consistently resulted in the smallest annual photosynthesis and respiration estimates, while the nighttime-based models resulted in the highest photosynthesis and respiration. The differences between the FVS–ANN and any of the other three models are explained by the different definitions of partitioned fluxes (i.e., values from the FVS–ANN are smaller than the other three models by leaf respiration). However, the differences may also be partly due to light inhibition of daytime respiration, which is known as the Kok effect that is believed to be adequately considered in the new FVS–ANN combination and partially in the daytime intercept model, but not in the other two models. The results suggest that previous EC-based estimates of global photosynthesis and respiration obtained by using nighttime relationships are potentially biased high.