Fusion of Multiple Models for Improving Gross Primary Production Estimation With Eddy Covariance Data Based on Machine Learning

Terrestrial gross primary production (GPP) represents the magnitude of CO2 uptake through vegetation photosynthesis, and is a key variable for carbon cycles between the biosphere and atmosphere. Light use efficiency (LUE) models have been widely used to estimate GPP for its physiological mechanisms and availability of data acquisition and implementation, yet each individual GPP model has exhibited large uncertainties due to input errors and model structure, and further studies of systematic validation, comparison, and fusion of those models with eddy covariance (EC) site data across diverse ecosystem types are still needed in order to further improve GPP estimation. We here compared and fused five GPP models (VPM, EC-LUE, GOL-PEM, CHJ, and C-Fix) across eight ecosystems based on FLUXNET2015 data set using the ensemble methods of Bayesian Model Averaging (BMA), Support Vector Machine (SVM), and Random Forest (RF) separately. Our results showed that for individual models, EC-LUE gave a better performance to capture interannual variability of GPP than other models, followed by VPM and GLO-PEM, while CHJ and C-Fix were more limited in their estimation performance. We found RF and SVM were superior to BMA on merging individual models at various plant functional types (PFTs) and at the scale of individual sites. On the basis of individual models, the fusion methods of BMA, SVM, and RF were examined by a five-fold cross validation for each ecosystem type, and each method successfully improved the average accuracy of estimation by 8%, 18%, and 19%, respectively.

Automated summary

Through the comparison and fusion of eight ecosystems, we found that the EC-LUE model was more accurate in estimating CO2 uptake than other models. Additionally, random forest and support vector machine algorithms performed better in merging different models. Based on the individual models, the fusion methods of Bayesian model averaging, support vector machine, and random forest improved the average accuracy of estimation by 8%, 18%, and 19% respectively.