plot and image data at finer spatial resolutions. The results showed both methods
accurately captured the spatial distributions and patterns of above-ground forest
carbon at a spatial resolution of 990 m ´ 990 m by combining and scaling up the forest
inventory sample plot data at a spatial resolution of 28.3 m ´ 28.3 m and TM images
at a pixel size of 30 m ´ 30 m. Moreover, both upscaling methods output the spatial
uncertainties of the block predicted values measured as estimation variances and
probabilities for the block predicted values larger than a given threshold. This implied
both PSCPS and PSCBS not only scaled up the spatial data but also modeled the
propagation of input spatial uncertainties from spatial resolutions of 28.3 m ´ 28.3 m
and 30 m ´ 30 m to a spatial resolution of 990 m ´ 990 m. These uncertainties varied
depending not only on the configuration of the sample locations but also on the sample
plot and image data values themselves. Thus, the block variances and probabilities
reflect well the spatial uncertainties of the input spatial data and their spatial
configuration.
In this study, a root mean-square error (RMSE) of 12.9 tons/ha for the estimates of
above-ground forest carbon at a spatial resolution of 30 m ´ 30 m was obtained for the
PSCPS method. But, because of a lack of field observation of above-ground forest
carbon at a spatial resolution of 990 m ´ 990 m, the RMSE of the aggregated aboveground forest carbon estimates were not obtained. That is, this study did not provide
users with guidelines on how to choose these two methods for upscaling of spatial
data based on RMSE. But the results did show that the PSCBS produced more
smoothed block estimates and their variances than the PSCPS. This finding differed
from that obtained in the Wang et al. (2004b) study in which the PSCPS led to more
smoothing of the block estimates. The reason might be that in Wang et al. (2004)
the spatial data were only scaled up from a spatial resolution of 30 m ´ 30 m to
90 m ´ 90 m, while in this study the spatial data were aggregated from a spatial
resolution of 30 m ´ 30 m to 990 m ´ 990 m. The second reason might be that in
PSCPS the conditional distribution of above-ground forest carbon was determined
and the simulation was conducted at the pixel size of 30 m ´ 30 m, while in PSCBS
these were done at the pixel size of 990 m ´ 990 m. Another reason might be that
Wang et al. (2004b) used sample plots that had much smaller sampling distances
compared to those in this study.
Instead of using field observations, an alternative to assess the accuracy of
above-ground forest carbon estimates at a coarse spatial resolution such as
1 km ´ 1 km could be the use and upscaling of the estimates as references that are
more accurate and obtained by high-spatial-resolution remotely sensed data. In
fact, comparison of the PSCPS and PSCBS methods used in this study implied
this idea. That is, PSCPS first predicted the values of above-ground forest carbon
at a spatial resolution of 30 m ´ 30 m and then scaled up the predicted values to a
spatial resolution of 990 m ´ 990 m blocks. The block estimates were finally used
to assess the accuracy of the predicted values at the pixel size of 990 m ´ 990 m
by the PSCBS. The finer spatial resolution images used were from Landsat TM.
A better choice for such higher spatial resolution remotely sensed data could be
lidar. However, the cost to acquire lidar data for a large study area is very high
122
UPSCALING WITH CONDITIONAL COSIMULATION FOR MAPPING
accurately captured the spatial distributions and patterns of above-ground forest
carbon at a spatial resolution of 990 m ´ 990 m by combining and scaling up the forest
inventory sample plot data at a spatial resolution of 28.3 m ´ 28.3 m and TM images
at a pixel size of 30 m ´ 30 m. Moreover, both upscaling methods output the spatial
uncertainties of the block predicted values measured as estimation variances and
probabilities for the block predicted values larger than a given threshold. This implied
both PSCPS and PSCBS not only scaled up the spatial data but also modeled the
propagation of input spatial uncertainties from spatial resolutions of 28.3 m ´ 28.3 m
and 30 m ´ 30 m to a spatial resolution of 990 m ´ 990 m. These uncertainties varied
depending not only on the configuration of the sample locations but also on the sample
plot and image data values themselves. Thus, the block variances and probabilities
reflect well the spatial uncertainties of the input spatial data and their spatial
configuration.
In this study, a root mean-square error (RMSE) of 12.9 tons/ha for the estimates of
above-ground forest carbon at a spatial resolution of 30 m ´ 30 m was obtained for the
PSCPS method. But, because of a lack of field observation of above-ground forest
carbon at a spatial resolution of 990 m ´ 990 m, the RMSE of the aggregated aboveground forest carbon estimates were not obtained. That is, this study did not provide
users with guidelines on how to choose these two methods for upscaling of spatial
data based on RMSE. But the results did show that the PSCBS produced more
smoothed block estimates and their variances than the PSCPS. This finding differed
from that obtained in the Wang et al. (2004b) study in which the PSCPS led to more
smoothing of the block estimates. The reason might be that in Wang et al. (2004)
the spatial data were only scaled up from a spatial resolution of 30 m ´ 30 m to
90 m ´ 90 m, while in this study the spatial data were aggregated from a spatial
resolution of 30 m ´ 30 m to 990 m ´ 990 m. The second reason might be that in
PSCPS the conditional distribution of above-ground forest carbon was determined
and the simulation was conducted at the pixel size of 30 m ´ 30 m, while in PSCBS
these were done at the pixel size of 990 m ´ 990 m. Another reason might be that
Wang et al. (2004b) used sample plots that had much smaller sampling distances
compared to those in this study.
Instead of using field observations, an alternative to assess the accuracy of
above-ground forest carbon estimates at a coarse spatial resolution such as
1 km ´ 1 km could be the use and upscaling of the estimates as references that are
more accurate and obtained by high-spatial-resolution remotely sensed data. In
fact, comparison of the PSCPS and PSCBS methods used in this study implied
this idea. That is, PSCPS first predicted the values of above-ground forest carbon
at a spatial resolution of 30 m ´ 30 m and then scaled up the predicted values to a
spatial resolution of 990 m ´ 990 m blocks. The block estimates were finally used
to assess the accuracy of the predicted values at the pixel size of 990 m ´ 990 m
by the PSCBS. The finer spatial resolution images used were from Landsat TM.
A better choice for such higher spatial resolution remotely sensed data could be
lidar. However, the cost to acquire lidar data for a large study area is very high
122
UPSCALING WITH CONDITIONAL COSIMULATION FOR MAPPING