Fig.2. Comparison of cumulative GPP estimates from the flux towers and the models. The color lines represent the GPP value of cumulative comparison between the EC-tower and model for each site. The red dashed line is the 1:1 reference to the differences of modeled GPP and EC-tower. Inset histogram shows the frequency distribution of the percentage biases (PB). The two shadowed plots are two new models developed in this study.

3.2 Biases in remote sensing data products and consequences on global GPP estimation

Evaluating the quality of input data and understanding the impact of data biases on GPP simulation are prerequisites for improving GPP simulation accuracy. First, various biases were found when the spatial datasets that feed the models for global GPP simulations were evaluated at the site scale (Fig.3; Fig.S3). For example, the spatial PAR dataset only explained 57% of the observed PAR variation at the EC-towers, and the slope and intercept were 1.2 and 0.57, respectively, indicating that the PAR data fields overestimated PAR as a whole and slightly underestimated PAR at the low value. The determination coefficients of the global datasets of CO2, LE, and H at the EC-towers were less than 20% (R2<0.2), only that of the temperature data was efficient in representing site-scale variation (R2=0.89).
Second, the biases in the spatial datasets had a significant impact on GPP simulations. Before correcting these biases, the simulated GPP by the LUE-EF and LUE-NDWI models explained only 49% and 61% of the EC-tower GPP variation, and the slopes of the linear regression between simulated and tower-estimated GPP were 1.54 and 1.31, respectively, and the corresponding intercepts were -2.09 and -1.23. These results indicate that both models overestimated GPP as a whole, but underestimated low GPP values (Fig.4b and f). After correcting the biases in the spatial datasets, the R2 of LUE-EF and LUE-NDWI models improved to 0.80 and 0.79, with the slopes closer to 1 (1.20 and 1.18 values, respectively) and the intercepts closer to 0 (-0.59 and -0.91, respectively) (Fig.3c and g). The results also indicated that the LUE-NDWI model was less sensitive to the biases in the spatial data fields than the LUE-EF model, as shown by the smaller differences in R2 before and after data correction, probably attributed to the fact that the LUE-NDWI relies on NDWI, a factor that can be derived directly from remote sensing data and thereby less prone to error propagation than the LUE-EF model.
<Fig 3 roughly here>