Experiments and Results
The study site, Tung-shih farm, is located at E/N (159000, 2602950) as left up corner, and (164000,2597950) as right down corner. It is an experimental farm belonging to Taiwan Sugar Corporation. The 3-band (G, R, and IR) SPOT image over the study area was received on 1994, at the receiving station of the Center for Space and remote Sensing research at the NCU campus and was further processed to produce a geocoded image on a 2°TM coordinate system. Fig. 1(a) shows the pseudo color image md¥(composed by SPOT XS3, XS2, and XS1 as R, G, and B.) of the size 400 x 400 pixels of the test site covering approximately 5km x 5km. The field is divided into 400 meters by 100 meters square blocks. As an experimental farm, it maintains complete records m for land truth reference. These two reasons make the farm an ideal place for image processing test site.
We take the three bands value of each pixel (xs1, xs2, xs3) as the spectral feature vector of each pixel, and ( ) as the spatial feature vector of pixel where is the mean, is the standard deviation, and is the difference between the maximum and minimum values of the pixels in a window of size 2d+1 and centered at (i, j).
Fig. 1(a) SPOT pseudo color image (copyright@CNES 1996), (b) MLC classified image using spectral information only, (c) MLC classified image using spatial information, (d) SOM classified image using spatial information, and (e) the color coding map for classification results.
Fig. 3(b) is the classified image of (a) by MLC method uses spectral information as feature vectors, Fig. 3(c) is the classified image of (a) by MLC method uses spatial information as feature vectors, and Fig. 3(d) is the classified image of (a) by SOFM method uses spatial information as feature vectors. Fig. 3(e) is the color coding map for classification results.
Table 1. MLC classification matrix uses spectral features only, table 2. MLC classification matrix uses spatial features, and table 3. SOM classification matrix uses spatial features.
Table 1 MLC classification matrix using spectral features.
| Class |
1
| 2
| 3
| 4
| 5
| 6
| 7
| Producer's
purity (%)
|
| 1
| 2312
| 2
| 23
| 0
| 163
| 0
| 0
| 92.5
|
| 2
| 16
| 2141
| 6
| 0
| 282
| 55
| 0
| 85.6
|
| 3
| 30
| 10
| 2116
| 0
| 309
| 35
| 0
| 84.6
|
| 4
| 50
| 0
| 59
| 2070
| 0
| 0
| 321
| 82.8
|
| 5
| 302
| 214
| 405
| 0
| 1503
| 76
| 0
| 60.1
|
| 6
| 0
| 24
| 115
| 0
| 87
| 2274
| 0
| 91
|
| 7
| 6
| 0
| 18
| 37
| 0
| 0
| 2439
| 97.6
|
| User's purity (%)
| 85.1
| 89.5
| 77.2
| 98.2
| 64.2
| 93.2
| 88.4
| |
Overall purity=84.9%, K coefficient=0.824
