Abstract
Despite considerable progress in scaling carbon fluxes from eddy covariance sites to globe, significant uncertainties still exist when estimating the global net ecosystem exchange (NEE). In this study, the site-level NEE was estimated from FLUXNET, a global network of eddy covariance towers, using a random forest (RF) model based on remote sensing products and precipitation data. The plant function type (PFT) had the highest relative explanatory power in predicting the global site-level NEE. However, within PFTs, water-related variables (i.e., the total precipitation, remotely sensed evapotranspiration, land surface water index, and the difference between daytime and nighttime land surface temperature) and soil respiration ( R s) were strong predictors of NEE variability. Cross-validation analyses revealed the good performance of RF in predicting the spatiotemporal variability of monthly NEE at 168 global FLUXNET sites, with R 2 of 0.72 and RMSE of 0.96 g·C·m−2·day−1. The performance was also good when predicting across-site ( R 2 = 0.75) and seasonal patterns ( R 2 = 0.92) over the 58 sites with available data being longer than two years and the 12-month value being present for each year. The RF-estimated NEE showed better relationships with the tower-measured NEE than a global NEE product from FLUXCOM across all PFTs. The difference between the RF-estimated NEE and FLUXCOM NEE was likely linked to the different predictor sets, such as those with more water-related variables and R s. This study indicates the importance of considering the influence of water-related variables and R s in the estimation of NEE at the global scale.
Highlights
TERRESTRIAL ecosystems play an important role in the global carbon cycle [1]
These models contain a number of complex processes and parameters, which leads to considerable differences in the simulated net ecosystem CO2 exchange (NEE) among different terrestrial ecosystem models [7]
The objectives of this study were: (1) to identify the most important variables in predicting global site-level NEE; (2) to evaluate the performance of the random forest (RF) model in characterizing temporal and spatial variations in NEE at global FLUXNET sites by utilizing remote sensing, climate, and eddy-covariance flux datasets; (3) to analyze how the RF model performance varies among different plant functional types (PFTs); (4) to compare the consistency between the RF-estimated NEE and FLUXCOM NEE [3] in describing the monthly variations in NEE at global FLUXNET sites, and across different plant function type (PFT)
Summary
TERRESTRIAL ecosystems play an important role in the global carbon cycle [1]. modeling and mechanistically understanding the net ecosystem CO2 exchange (NEE) between terrestrial ecosystems and the atmosphere is crucial for the accurate estimation of the global carbon budget [2]. The global estimation of NEE can be obtained using process-based terrestrial ecosystem models, atmospheric CO2 inversion, and data-driven machine learning approaches [4, 5]. Process-based terrestrial ecosystem models consider the mechanistic behaviors of photosynthesis and respiration responding to environmental conditions (e.g., climate and atmospheric CO2 concentration) [6]. These models contain a number of complex processes and parameters, which leads to considerable differences in the simulated NEE among different terrestrial ecosystem models [7]. Unlike the above two methods, data-driven machine-learning approaches are simple and effective for estimating NEE because they are fully data-adaptive and do not require initial assumptions regarding functional relationships [3, 11]
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
