pool to atmosphere). Second, disturbances modify soil physical and chemical
factors and microclimate, creating ecological legacies that affect carbon dynamics
over ensuing decades or even centuries. The low net carbon uptake in 2002, 2006,
and 2009 are likely mainly caused by severe extended droughts and wild fires.
6.5.4 Sources of Uncertainty
Despite the encouraging performance of the data-driven model, the resulting
gridded flux estimates for the U.S. exhibit significant uncertainties. There are
several sources of uncertainty associated with the NEE estimates, including
uncertainty in flux observations from towers, uncertainty in other input data (e.g.,
land cover), model structural uncertainty, and uncertainty resulting from the
representativeness of the AmeriFlux network (Xiao et al. 2011a).
The uncertainty of input data can propagate through model simulations and lead
to biases in the flux estimates. The NEE measurements from eddy covariance flux
towers contain significant uncertainty largely due to random measurement error
(Hollinger and Richardson 2005). The potential uncertainties associated with the
eddy covariance technique include systematic errors from insensitivity to highfrequency turbulence, random errors from inadequate sample size associated with
averaging period, vertical and horizontal advection, u
* filtering, and gap-filling
methods (e.g., Hollinger and Richardson 2005; Loescher et al. 2006). Other input
data, particularly the land cover map, also contain significant uncertainty. Landcover maps are typically derived from satellite remote sensing, and their uncertainty
is associated with the limited accuracy of image classification. The sub-grid
heterogeneity in land cover, topography, and climate can also influence the accuracy
of the gridded flux estimates.
There is also significant uncertainty associated with the algorithm of the datadriven approach. As mentioned earlier, a variety of variables derived from satellite
remote sensing including EVI, LST, NDWI, and LAI are used as explanatory
variables for the prediction of NEE. Although these variables can partly account
for the climatic, physiological, and hydrological factors controlling NEE, some
important factors influencing NEE such as soil organic carbon pools and disturbances are not represented. In addition, it is debatable whether NDWI provides a
sufficient measure of ecosystem water stress, although it is strongly related to leaf
water content (Jackson et al. 2004) and soil moisture (Fensholt and Sandholt
2003). Microwave sensors including AMSR-E provide global estimates of soil
moisture that can be potentially used in upscaling efforts. These estimates, however, are not available for densely vegetated areas that are important for terrestrial
carbon cycling. For a given cell, the data-driven approach can also introduce
biases to the flux estimates if the values of the explanatory variables are beyond
the range of the training data.
The representativeness of the AmeriFlux network also leads to uncertainty in
the gridded flux estimates. Although the AmeriFlux sites are fairly representative
6 Assessing Net Ecosystem Exchange
163
factors and microclimate, creating ecological legacies that affect carbon dynamics
over ensuing decades or even centuries. The low net carbon uptake in 2002, 2006,
and 2009 are likely mainly caused by severe extended droughts and wild fires.
6.5.4 Sources of Uncertainty
Despite the encouraging performance of the data-driven model, the resulting
gridded flux estimates for the U.S. exhibit significant uncertainties. There are
several sources of uncertainty associated with the NEE estimates, including
uncertainty in flux observations from towers, uncertainty in other input data (e.g.,
land cover), model structural uncertainty, and uncertainty resulting from the
representativeness of the AmeriFlux network (Xiao et al. 2011a).
The uncertainty of input data can propagate through model simulations and lead
to biases in the flux estimates. The NEE measurements from eddy covariance flux
towers contain significant uncertainty largely due to random measurement error
(Hollinger and Richardson 2005). The potential uncertainties associated with the
eddy covariance technique include systematic errors from insensitivity to highfrequency turbulence, random errors from inadequate sample size associated with
averaging period, vertical and horizontal advection, u
* filtering, and gap-filling
methods (e.g., Hollinger and Richardson 2005; Loescher et al. 2006). Other input
data, particularly the land cover map, also contain significant uncertainty. Landcover maps are typically derived from satellite remote sensing, and their uncertainty
is associated with the limited accuracy of image classification. The sub-grid
heterogeneity in land cover, topography, and climate can also influence the accuracy
of the gridded flux estimates.
There is also significant uncertainty associated with the algorithm of the datadriven approach. As mentioned earlier, a variety of variables derived from satellite
remote sensing including EVI, LST, NDWI, and LAI are used as explanatory
variables for the prediction of NEE. Although these variables can partly account
for the climatic, physiological, and hydrological factors controlling NEE, some
important factors influencing NEE such as soil organic carbon pools and disturbances are not represented. In addition, it is debatable whether NDWI provides a
sufficient measure of ecosystem water stress, although it is strongly related to leaf
water content (Jackson et al. 2004) and soil moisture (Fensholt and Sandholt
2003). Microwave sensors including AMSR-E provide global estimates of soil
moisture that can be potentially used in upscaling efforts. These estimates, however, are not available for densely vegetated areas that are important for terrestrial
carbon cycling. For a given cell, the data-driven approach can also introduce
biases to the flux estimates if the values of the explanatory variables are beyond
the range of the training data.
The representativeness of the AmeriFlux network also leads to uncertainty in
the gridded flux estimates. Although the AmeriFlux sites are fairly representative
6 Assessing Net Ecosystem Exchange
163