configuration of the used data but also the data values themselves and reveal the
spatial uncertainties of the obtained estimates. This algorithm requires normal score
transformations of all the sample plot and image data in order to meet the assumptions
of the multivariate Gaussian distribution.
The procedure of PSCBS is similar to that of PSCPS described above. The
difference lies in that PSCPS first creates estimates at the same spatial resolution as the
input data and then scales up the estimates to any coarser spatial resolution, meaning
indirect upscaling of spatial data, while PSCBS directly scales up the spatial data from
a finer spatial resolution to a coarser one. In PSCBS, the map units at the desirable and
coarser spatial resolution are considered as a block and each block consists of smaller
pixels that have the same size as the sample plots and image pixels. Following a
simulation procedure similar to the above, the predicted value and its variance of each
smaller pixel within a block are obtained using the collocated simple cokriging
estimator. Using the block mean and block variance by averaging the estimates and
cokriging variances of the pixels within the block, a conditional distribution of aboveground forest carbon for the block is determined. The block variance is calculated by
modeling the propagation of both cokriging variances of smaller pixels and covariances among them. From the distribution for the block, a value is then randomly
drawn and considered to be a realization of above-ground forest carbon at this block.
With this method, cokriging is conducted on the basis of smaller pixels and the
simulation is made on the basis of larger blocks. The spatial data and their
uncertainties are directly scaled up. The block cosimulation can also be run many
times, resulting in many predicted values for each block. From the predicted values, a
sample average and a sample variance for each block are then obtained. This method
provides the potential to improve the block conditional distribution by modeling the
propagation of uncertainty from the sample plot and image data at the finer spatial
resolution and the estimates of smaller pixels to the block and thus improve the quality
of the block estimates and their variances.
6.3 STUDY AREA AND DATA SETS
This study was conducted in Lin-An County, ZheJiang Province of East China (Wang
et al., 2011) (Figure 6.1b) . Its latitude and longitude range from 29°56ʹ N to 30°23ʹ N
and from 118°51ʹ E to 119°52ʹ E, respectively. It is characterized by a typical
subtropical climate with average annual temperature and precipitation of 16.4 °C and
1628 mm, respectively. This study area consists of 312,680 ha and has features of a
mountainous area with an elevation range of 1770 m. The main forest types include
Chinese fir and Pinus massoniana plantations and evergreen broad-leaf forests,
deciduous and evergreen broad-leaf mixed forests, bamboo, and shrubs (Wang
et al., 2011; Zhang et al., 2009).
Since the 1950s continuous forest inventory has been conducted in Lin-An County
[Chinese Ministry of Forestry (CMF), 1996; Wang et al., 2011]. A total of 50 national
permanent sample plots were established and repeatedly measured every five years.
114
UPSCALING WITH CONDITIONAL COSIMULATION FOR MAPPING
spatial uncertainties of the obtained estimates. This algorithm requires normal score
transformations of all the sample plot and image data in order to meet the assumptions
of the multivariate Gaussian distribution.
The procedure of PSCBS is similar to that of PSCPS described above. The
difference lies in that PSCPS first creates estimates at the same spatial resolution as the
input data and then scales up the estimates to any coarser spatial resolution, meaning
indirect upscaling of spatial data, while PSCBS directly scales up the spatial data from
a finer spatial resolution to a coarser one. In PSCBS, the map units at the desirable and
coarser spatial resolution are considered as a block and each block consists of smaller
pixels that have the same size as the sample plots and image pixels. Following a
simulation procedure similar to the above, the predicted value and its variance of each
smaller pixel within a block are obtained using the collocated simple cokriging
estimator. Using the block mean and block variance by averaging the estimates and
cokriging variances of the pixels within the block, a conditional distribution of aboveground forest carbon for the block is determined. The block variance is calculated by
modeling the propagation of both cokriging variances of smaller pixels and covariances among them. From the distribution for the block, a value is then randomly
drawn and considered to be a realization of above-ground forest carbon at this block.
With this method, cokriging is conducted on the basis of smaller pixels and the
simulation is made on the basis of larger blocks. The spatial data and their
uncertainties are directly scaled up. The block cosimulation can also be run many
times, resulting in many predicted values for each block. From the predicted values, a
sample average and a sample variance for each block are then obtained. This method
provides the potential to improve the block conditional distribution by modeling the
propagation of uncertainty from the sample plot and image data at the finer spatial
resolution and the estimates of smaller pixels to the block and thus improve the quality
of the block estimates and their variances.
6.3 STUDY AREA AND DATA SETS
This study was conducted in Lin-An County, ZheJiang Province of East China (Wang
et al., 2011) (Figure 6.1b) . Its latitude and longitude range from 29°56ʹ N to 30°23ʹ N
and from 118°51ʹ E to 119°52ʹ E, respectively. It is characterized by a typical
subtropical climate with average annual temperature and precipitation of 16.4 °C and
1628 mm, respectively. This study area consists of 312,680 ha and has features of a
mountainous area with an elevation range of 1770 m. The main forest types include
Chinese fir and Pinus massoniana plantations and evergreen broad-leaf forests,
deciduous and evergreen broad-leaf mixed forests, bamboo, and shrubs (Wang
et al., 2011; Zhang et al., 2009).
Since the 1950s continuous forest inventory has been conducted in Lin-An County
[Chinese Ministry of Forestry (CMF), 1996; Wang et al., 2011]. A total of 50 national
permanent sample plots were established and repeatedly measured every five years.
114
UPSCALING WITH CONDITIONAL COSIMULATION FOR MAPPING