The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B4. Beijing 2008
1374
Window size
Figure 2. JM-distance between residential and background
classes on different textural features with respect to window
size
3.2 Optimal parameter value of quantization
Four different quantization schemes 16, 32, 64 and 256 grey
levels were tested for GLCM texture features. Figure 3 shows
JM-distance variation with respect to the pixel grey levels. The
result shows that the quantization level dose not have
significant impact on CON, DIS and SD bands, and appears to
have negative impact on the other texture features including
ENT, HOMO, ASM and COR. To balance the accuracy and
demand on computation resources, we chose a quantization
level of 32.
Quantization level
Figure 3. JM-distance between residential and background
classes on different textural feature with respect to quantization
level
3.3 Optimal parameter value of displacements
In this study, displacements of 1,3,5 and 7 pixels were tested as
shown in Figure 4. In general there is a tendency that JM-
distance decreases with the increasing displacement. Since it
appears that low displacement yield generally better results
(greater JM-distances) for most texture bands, we therefore
selected one as displacement in texture computation.
Figure 4. JM-distance values between residential class and
background class on different textural feature image with
respect to displacement
3.4 Optimal parameter value of orientation
To assess the impact of various orientation on the derived
texture measures, we examined their separability with the
variations on orientation of 0°, 45°, 90°, 135° and the average
of four orientations. Figure 5 shows that, except for SD, JM-
distances on average orientation are significantly greater than
on the others.
Texture features
Figure 5. JM-distance values between residential and
background classes on different textural features with respect to
orientation
Table 1 summarizes the optimal parameters selected for
deriving texture measures in this study.
As the parameters of each texture feature have been selected,
nine texture images can be obtained from the original IKONOS
Panchromatic image. The nine texture features can be separated
into five categories according to their properties, namely, the
‘contrast’ group (CON, DIS and HOMO), the ‘orderliness’
group (ASM and ENT) , the ‘edge’ group (ED) , the ‘mean’
group (MEAN) and the ‘descriptive statistics’ group (SD and
COR). To understand the distribution of different kinds of land
cover types on each texture space image better, we plot the
clusters based on the training samples. Figure 6 to figure 9
shows the cluster result.