3. Analysis of Polarization Signatures for the Targets
The first, complex scattering matrices were derived from PI-SAR MGP data for each pixel to make polarization signatures that indicate the polarization properties of scattering intensity of microwave for the target. Next, the complex scattering matrices were transformed Mueller matrices constructed by real number elements, and obtained matrices were averaged in homogeneous area to reduction influences of speckle noises. Generally, polarization signatures are obtained from averaged Mueller matrices. Polarization signature diagrams represent the value of polarization signatures graphically, and these shapes characterize the targets [1]~[3],[8].
In SAR images, it is well known to undergo the influences of shadows that appear scarcely in the images obtained by the sensor observing directly below. Particularly, in high resolution SAR image, including the shadows in extracted sample areas is liable to occur. In the following description, the fluctuations of polarization signatures between the area including shadows and the area excluding them are investigated [9].
Fig.2 shows the polarization signature diagrams of vegetation area and residential area including shadows, and Fig.3 shows the ones of the same areas excluding them. These diagrams were obtained from high resolution X-band data, and combination of polarization was co-polarization. Although influence of shadows for the shape of polarization signature diagrams was little, it was found that the reciprocity value of s0 for including shadows area was smaller than for excluding shadow area. Then we estimated the polarization signature ratio of including shadow to excluding one. Fig.4 shows that the dependence of the ratio of s0 was small, and it was nearly certain.
 |
| (a) Vegetation area |
(a) (b) Residential area |
Fig.2: Polarization signature diagrams for extracted area including shadows
 |
| (a) Vegetation area |
(a) (b) Residential area |
Fig.3: Polarization signature diagrams for extracted area excluding shadows
 |
| (a) Vegetation area |
(a) (b) Residential area |
Fig.4: The polarization signature ratio of including shadow to excluding one
4. Analysis of Textural Features for the Targets
In the next place, we estimated textural features from the same areas analyzed polarization signatures. For PI-SAR data, it is difficult to compare as it stands because the resolutions depend on the observation frequencies (X and L-band). Hence two kinds of sub-images were compared for X-band data. One of them was the features derived from normal pixel spacing images, another one was derived from thinned out images to match the resolution to L-band image data. Furthermore, features of thinned out X-band images are compared with the features of L-band normal pixel spacing images [10].
In this study, GLCM (Gray Level Co-occurrence Matrix) method was used to estimate textural features. In this method, textures are defined from distances and directions of neighbor two pixels. Some textural features derived by this method are proposed, Angular Second Moment (ASM), Contrast (CON), Entropy (ENT), and so on [9],[11],[12]. Each element of GLCM is expressed as P(i, j, d, q ), and it is changed into probability variable p(i, j, d, q ) by equation (1) to derive the textural features.
Where N is the size of GLCM (in this study, N is set 16), i and j are number of line and column for the remarkable element in GLCM, d is the distance of two pixels to define the texture, and q is the direction of them. ASM, CON, and ENT are derived Equation (2) to (4).
Polarization properties of textural features are expressed in the same co-ordinates as polarization signatures. Fig.5 shows the polarization properties of contrast for residential area derived from X-band data, (a) shows for normal pixel spacing image and (b) shows for thinned out image.
To analyze the difference by frequency band, the feature property diagram for the same area of L-band normal pixel spacing image data is shown Fig.6.
 |
| (a) Normal pixel spacing image data |
(b) Thinned out image data |
Fig.5: Polarization properties for X-band residential area image
Fig.6: Polarization property diagram for residential area of L-band normal pixel spacing image
