Six LULC types were derived from each ASTER image: urban, forest, grassland,
agriculture, water, and barren land based on unsupervised classification. Postclassification smoothing process was executed to improve the accuracy of image classification. Image refinement was possessed to manually correct certain confusing pixels
to further improve the classification accuracy. The overall accuracy of each classified
image was above 85%. The spatial resolution of each LULC map is 15 m.
Existing ASTER land surface kinetic temperature images (90 m spatial resolution)
were purchased in order to investigate the scaling effects in studying the relationship
between landscape pattern and land surface temperature. Each LST map was divided
into six temperature zones by using a standard deviation method of data classification
(Smith, 1986). Please refer to Liu and Weng (2008) for statistic details of LST
classification. Figure 11.2 shows the LULC and LST maps for each image date.
In order to examine the scaling-up effects in studying the relationship between
LULC and LSTs in the landscape, we resampled each image (both LULC and LST) to
different spatial resolutions: 30 30, 60 60, 120 120, 250 250, 500 500, and
1000 1000 m. Each resampled LST map was divided into six temperature zones by
using the same method as the one used for classifying the LST image with original
90 m resolution (standard deviation method). We intended to better understand their
possible relationships at different scales by examining the LULC and LST at various
spatial resolutions. As a result, 32 LULC and LST images in all were created for
further analysis.
Four class-level landscape metrics (patch percentage, patch density, landscape
shape index, and perimeter–area fractal dimension) and four landscape-level metrics
(patch density, landscape shape index, perimeter–area fractal dimension, and mean
perimeter–area ratio) were derived from all 32 maps (see Table 11.2 for metrics
details). According to McGarigal et al. (2002), patch percentage index is the
proportion of each patch type within the study area; patch density index implies
the number of patches per 100 hectares; landscape shape index measures the
aggregation of landscape and its value increases as the patch type becomes more
disaggregated; perimeter–area fractal dimension index reflects the shape complexity
of patch, class, or the whole landscape. Mean perimeter–area ratio is a simplified
method for evaluating the shape complexity. All the indices were computed by the use
of FRAGSTATS (McGarigal et al., 2002) that were next used to analyze the scalingup effect on the analysis of landscape and LST patterns.
11.2.3 Results
11.2.3.1 Scaling-Up Effect on Class-Level Landscape Metrics Figure 11.3
shows the area percentages of LULC and LST patches across eight spatial scales
for two image dates, June 16, 2001 and October 3, 2000. According to the
calculations, LULC patch percentages did not possess obvious variations across
the spatial scales, although slight changes were captured when the pixel size changed
from 500 to 1000 m for urban, forest, and grassland for both image dates. When the
scale changed from 500 to 1000 m, an increase was observed in urban patch
percentage; however; decreases were identified in both forest and grassland patch
220
SCALING ISSUES IN STUDYING THE RELATIONSHIP
agriculture, water, and barren land based on unsupervised classification. Postclassification smoothing process was executed to improve the accuracy of image classification. Image refinement was possessed to manually correct certain confusing pixels
to further improve the classification accuracy. The overall accuracy of each classified
image was above 85%. The spatial resolution of each LULC map is 15 m.
Existing ASTER land surface kinetic temperature images (90 m spatial resolution)
were purchased in order to investigate the scaling effects in studying the relationship
between landscape pattern and land surface temperature. Each LST map was divided
into six temperature zones by using a standard deviation method of data classification
(Smith, 1986). Please refer to Liu and Weng (2008) for statistic details of LST
classification. Figure 11.2 shows the LULC and LST maps for each image date.
In order to examine the scaling-up effects in studying the relationship between
LULC and LSTs in the landscape, we resampled each image (both LULC and LST) to
different spatial resolutions: 30 30, 60 60, 120 120, 250 250, 500 500, and
1000 1000 m. Each resampled LST map was divided into six temperature zones by
using the same method as the one used for classifying the LST image with original
90 m resolution (standard deviation method). We intended to better understand their
possible relationships at different scales by examining the LULC and LST at various
spatial resolutions. As a result, 32 LULC and LST images in all were created for
further analysis.
Four class-level landscape metrics (patch percentage, patch density, landscape
shape index, and perimeter–area fractal dimension) and four landscape-level metrics
(patch density, landscape shape index, perimeter–area fractal dimension, and mean
perimeter–area ratio) were derived from all 32 maps (see Table 11.2 for metrics
details). According to McGarigal et al. (2002), patch percentage index is the
proportion of each patch type within the study area; patch density index implies
the number of patches per 100 hectares; landscape shape index measures the
aggregation of landscape and its value increases as the patch type becomes more
disaggregated; perimeter–area fractal dimension index reflects the shape complexity
of patch, class, or the whole landscape. Mean perimeter–area ratio is a simplified
method for evaluating the shape complexity. All the indices were computed by the use
of FRAGSTATS (McGarigal et al., 2002) that were next used to analyze the scalingup effect on the analysis of landscape and LST patterns.
11.2.3 Results
11.2.3.1 Scaling-Up Effect on Class-Level Landscape Metrics Figure 11.3
shows the area percentages of LULC and LST patches across eight spatial scales
for two image dates, June 16, 2001 and October 3, 2000. According to the
calculations, LULC patch percentages did not possess obvious variations across
the spatial scales, although slight changes were captured when the pixel size changed
from 500 to 1000 m for urban, forest, and grassland for both image dates. When the
scale changed from 500 to 1000 m, an increase was observed in urban patch
percentage; however; decreases were identified in both forest and grassland patch
220
SCALING ISSUES IN STUDYING THE RELATIONSHIP