6.3.4 Data-Driven Approach
A data-driven approach is used to upscale fluxes from the tower footprint to the
continental scale. This approach has been fully described elsewhere (Xiao et al. 2008),
and is briefly summarized here. The data-driven method is essentially an ensemble
of regression models. This approach relies on rule-based models, each of which is
a set of conditions associated with a multivariate linear submodel. These rulebased, piecewise regression models allow both numerical (e.g., carbon fluxes,
temperature, vegetation index) and categorical variables (e.g., land cover type) as
input variables, and account for possible nonlinear relationships between predictive and target variables.
In this approach, the predictive accuracy of a rule-based model can be improved
by combining it with an instance-based/nearest-neighbor model that predicts the
target value of a new case using the average predicted values of the n most similar
cases (RuleQuest 2008). The use of the composite model can improve the predictive accuracy relative to the rule-based model alone. This approach can also
generate committee models made up of several rule-based models, and each
member of the committee model predicts the target value for a case (RuleQuest
2008). The member’s predictions are averaged to give a final prediction.
A predictive NEE model was constructed using AmeriFlux and MODIS data.
The predictive variables include a variety of MODIS data streams, such as vegetation type, EVI, LST, NDWI, and PAR. Three statistical measures are used to
evaluate the quality of the constructed predictive model, including mean absolute
error (MAE), relative error (RE), and product-moment correlation coefficient
(Yang et al. 2003; Xiao et al. 2008). MAE is calculated as:
MAE ¼
1
N
X N
i ¼ 1
y i À ^ y i
j
j
ð6:3Þ
where N is the number of samples used to establish the predictive model, and y i
and ^ y i are the actual and predicted values of the response variable, respectively. RE
is calculated as:
RE ¼
MAE T
MAE l
ð6:4Þ
where MAE T is the MAE of the constructed model, and MAE l is the MAE that
would result from always predicting the mean value.
For forest sites, the MAE and RE are 0.48 g C m
-2 day
-1 and 0.37, respectively for the predictive model. For non-forest sites, the MAE and RE are 0.73 g C
m
-2 day
-1 and 0.61, respectively. The performance of the model is slightly better
for forest sites than for non-forest sites. Given the diversity in ecosystem types, age
structures, fire and insect disturbances, and management practices, the performance of these models is encouraging (Xiao et al. 2008).
156
J. Xiao
A data-driven approach is used to upscale fluxes from the tower footprint to the
continental scale. This approach has been fully described elsewhere (Xiao et al. 2008),
and is briefly summarized here. The data-driven method is essentially an ensemble
of regression models. This approach relies on rule-based models, each of which is
a set of conditions associated with a multivariate linear submodel. These rulebased, piecewise regression models allow both numerical (e.g., carbon fluxes,
temperature, vegetation index) and categorical variables (e.g., land cover type) as
input variables, and account for possible nonlinear relationships between predictive and target variables.
In this approach, the predictive accuracy of a rule-based model can be improved
by combining it with an instance-based/nearest-neighbor model that predicts the
target value of a new case using the average predicted values of the n most similar
cases (RuleQuest 2008). The use of the composite model can improve the predictive accuracy relative to the rule-based model alone. This approach can also
generate committee models made up of several rule-based models, and each
member of the committee model predicts the target value for a case (RuleQuest
2008). The member’s predictions are averaged to give a final prediction.
A predictive NEE model was constructed using AmeriFlux and MODIS data.
The predictive variables include a variety of MODIS data streams, such as vegetation type, EVI, LST, NDWI, and PAR. Three statistical measures are used to
evaluate the quality of the constructed predictive model, including mean absolute
error (MAE), relative error (RE), and product-moment correlation coefficient
(Yang et al. 2003; Xiao et al. 2008). MAE is calculated as:
MAE ¼
1
N
X N
i ¼ 1
y i À ^ y i
j
j
ð6:3Þ
where N is the number of samples used to establish the predictive model, and y i
and ^ y i are the actual and predicted values of the response variable, respectively. RE
is calculated as:
RE ¼
MAE T
MAE l
ð6:4Þ
where MAE T is the MAE of the constructed model, and MAE l is the MAE that
would result from always predicting the mean value.
For forest sites, the MAE and RE are 0.48 g C m
-2 day
-1 and 0.37, respectively for the predictive model. For non-forest sites, the MAE and RE are 0.73 g C
m
-2 day
-1 and 0.61, respectively. The performance of the model is slightly better
for forest sites than for non-forest sites. Given the diversity in ecosystem types, age
structures, fire and insect disturbances, and management practices, the performance of these models is encouraging (Xiao et al. 2008).
156
J. Xiao