New data‐driven estimation of terrestrial CO₂ fluxes in Asia using a standardized database of eddy covariance measurements, remote sensing data, and support vector regression

Abstract

The lack of a standardized database of eddy covariance observations has been an obstacle for data‐driven estimation of terrestrial CO2 fluxes in Asia. In this study, we developed such a standardized database using 54 sites from various databases by applying consistent postprocessing for data‐driven estimation of gross primary productivity (GPP) and net ecosystem CO2 exchange (NEE). Data‐driven estimation was conducted by using a machine learning algorithm: support vector regression (SVR), with remote sensing data for 2000 to 2015 period. Site‐level evaluation of the estimated CO2 fluxes shows that although performance varies in different vegetation and climate classifications, GPP and NEE at 8 days are reproduced (e.g., r2 = 0.73 and 0.42 for 8 day GPP and NEE). Evaluation of spatially estimated GPP with Global Ozone Monitoring Experiment 2 sensor‐based Sun‐induced chlorophyll fluorescence shows that monthly GPP variations at subcontinental scale were reproduced by SVR (r2 = 1.00, 0.94, 0.91, and 0.89 for Siberia, East Asia, South Asia, and Southeast Asia, respectively). Evaluation of spatially estimated NEE with net atmosphere‐land CO2 fluxes of Greenhouse Gases Observing Satellite (GOSAT) Level 4A product shows that monthly variations of these data were consistent in Siberia and East Asia; meanwhile, inconsistency was found in South Asia and Southeast Asia. Furthermore, differences in the land CO2 fluxes from SVR‐NEE and GOSAT Level 4A were partially explained by accounting for the differences in the definition of land CO2 fluxes. These data‐driven estimates can provide a new opportunity to assess CO2 fluxes in Asia and evaluate and constrain terrestrial ecosystem models.

Global distribution of groundwater‐vegetation spatial covariation

“…The GPPs estimated only by a subset of algorithms used in FLUXCOM have shown better performance than the GPP from Moderate Resolution Imaging Spectroradiometer (MODIS) against site-level observations [Ichii et al, 2017;Tramontana et al, 2015;Yang et al, 2007]. The GPPs estimated only by a subset of algorithms used in FLUXCOM have shown better performance than the GPP from Moderate Resolution Imaging Spectroradiometer (MODIS) against site-level observations [Ichii et al, 2017;Tramontana et al, 2015;Yang et al, 2007].…”

“…Compared to other global GPP products, the median GPP from FLUXCOM is advantageous because it combines the strengths of multiple global satellite-based observations with site-level observations using several methods. The GPPs estimated only by a subset of algorithms used in FLUXCOM have shown better performance than the GPP from Moderate Resolution Imaging Spectroradiometer (MODIS) against site-level observations [Ichii et al, 2017;Tramontana et al, 2015;Yang et al, 2007]. At the global scale, the Model Tree Ensembles GPP product [Beer et al, 2010;Jung et al, 2011], also estimated by a subset of algorithms used in FLUXCOM, compares well with modern satellite observation of Sun-induced fluorescence [Frankenberg et al, 2013], and it has been used extensively to benchmark global land surface models [Anav et al, 2013;Bonan et al, 2012;Piao et al, 2013].…”

Trends and controls of terrestrial gross primary productivity of China during 2000–2016

“…Boosted regression tree (BRT) analysis, a nonlinear regression model, was used to evaluate the relative importance and marginal effects of individual environmental factors on GPP for China during 2000-2016. BRT analysis has strength in evaluating complex nonlinear relationship (Elith et al 2008, Ma et al 2015, 2017, which is reflected by the marginal effect and the relative influence of each independent variable on response variables. The marginal effect of an individual predictor variable is calculated based on the assumption that other independent variables are constant, and this effect will be regarded as the relative influence on the response variable.…”

Reconciliation of top-down and bottom-up CO 2 fluxes in Siberian larch forest

“…The empirical upscaling techniques provide another independent estimate of terrestrial CO 2 fluxes. We used the terrestrial net ecosystem exchange (NEE) empirically upscaled at 0.25 • spatial resolution and 8 day temporal resolution across Asia (Ichii et al 2017). The data used 54 eddy-covariance observation sites (278 site-years) and remote sensing products (e.g.…”

“…Here we present an analysis of CO 2 fluxes from the bottom-up approaches and the top-down approach. The former approach includes the eddycovariance method (Ohta et al 2014), terrestrial ecosystem models from the GRENE-TEA (GRENE Terrestrial Ecosystem of the Arctic: the terrestrial research sub-project of the Arctic Climate Change Research Project of the GRENE (Green Network of Excellence) program by the Ministry of Education Culture, Sports, Science and Technology, Japan) model inter-comparison project (Miyazaki et al 2015), and a data-driven method using a machine learning algorithm (Ichii et al 2017). The top-down approach is the inverse modeling of the atmospheric CO 2 concentrations using atmospheric transport models.…”