6
UPSCALING WITH CONDITIONAL
COSIMULATION FOR MAPPING
ABOVE-GROUND FOREST CARBON
GUANGXING WANG AND MAOZHEN ZHANG
6.1 INTRODUCTION
Forest inventory sample plot data are often combined with remotely sensed images by
regression modeling, neural networks, and K-nearest neighbors to map forest carbon,
that is, generate spatially explicit estimates at a desirable spatial resolution (Lu et al.,
2012; Wang et al., 2009). In these methods, forest carbon observations are available
only at the sample plot locations, while remotely sensed data are available everywhere. Forest carbon at unobserved locations is interpolated by combining the sample
plot data and remotely sensed images. The image data provide the linkage of forest
carbon from the sample plot locations to the unobserved locations. Generally, the
spatial resolutions of used sample plot data and images are consistent with the sizes of
units of output maps. In practice, however, the sizes of forest inventory sample plots
vary from 10 m ´ 10 m to 50 m ´ 50 m because of limitation of high cost to collect
field data, while forest carbon maps at regional, national, and global scales are often
required to have spatial resolutions that range from 90 m ´ 90 m to 1 km ´ 1 km.
Scaling up or aggregating the sample plot data from a finer spatial resolution to a
coarser one has to be thus conducted and the existing methods lack the ability (Wang
et al., 2009, 2011).
Upscaling of spatial data has been widely studied (Marceau, 1999; Marceau and
Hay, 1999; Wang et al., 2009; Wu and Qi, 2000) and was first conducted in the field of
socioeconomy. A typical example is that Gehlke and Biehl (1934) found out that
108
Scale Issues in Remote Sensing, First Edition. Edited by Qihao Weng.
Ó 2014 John Wiley & Sons, Inc. Published 2014 by John Wiley & Sons, Inc.
UPSCALING WITH CONDITIONAL
COSIMULATION FOR MAPPING
ABOVE-GROUND FOREST CARBON
GUANGXING WANG AND MAOZHEN ZHANG
6.1 INTRODUCTION
Forest inventory sample plot data are often combined with remotely sensed images by
regression modeling, neural networks, and K-nearest neighbors to map forest carbon,
that is, generate spatially explicit estimates at a desirable spatial resolution (Lu et al.,
2012; Wang et al., 2009). In these methods, forest carbon observations are available
only at the sample plot locations, while remotely sensed data are available everywhere. Forest carbon at unobserved locations is interpolated by combining the sample
plot data and remotely sensed images. The image data provide the linkage of forest
carbon from the sample plot locations to the unobserved locations. Generally, the
spatial resolutions of used sample plot data and images are consistent with the sizes of
units of output maps. In practice, however, the sizes of forest inventory sample plots
vary from 10 m ´ 10 m to 50 m ´ 50 m because of limitation of high cost to collect
field data, while forest carbon maps at regional, national, and global scales are often
required to have spatial resolutions that range from 90 m ´ 90 m to 1 km ´ 1 km.
Scaling up or aggregating the sample plot data from a finer spatial resolution to a
coarser one has to be thus conducted and the existing methods lack the ability (Wang
et al., 2009, 2011).
Upscaling of spatial data has been widely studied (Marceau, 1999; Marceau and
Hay, 1999; Wang et al., 2009; Wu and Qi, 2000) and was first conducted in the field of
socioeconomy. A typical example is that Gehlke and Biehl (1934) found out that
108
Scale Issues in Remote Sensing, First Edition. Edited by Qihao Weng.
Ó 2014 John Wiley & Sons, Inc. Published 2014 by John Wiley & Sons, Inc.