204
Fig3: image of HIS fusion method
3. CLASSIFICATIONS AND ANALYSIS
class
road
other
Produce
accuracy
Total
accuracy
PCA
method
road
163
8
95.3%
85%
other
37
82
68.9%
User
accuracy
81.5%
91.1%
HIS
method
road
141
18
88.7%
80%
other
42
99
70.2%
User
accuracy
77%
84.6%
TM
spectral
image
road
130
23
85%
77%
other
46
101
68.7%
User
accuracy
73.9%
81.5%
Table 1. Error matrix and accuracy appraisement table of
different fusion approaches
Because of the special geology of the study area, we adopted
computer automatic processing combined with visual
interpretation.
Comparing TM image to SAR image, it is clear for us to tell the
difference between them. Though there is plenty of color
information in TM image, it is difficult to differentiate road
from other objects in such high vegetation. While in SAR
image, with the smooth surface of the roads, the roads have
lower coarseness than the other objects. So it is easy to
recognize the roads. We respectively chose training samples for
roads and other objects, and then applied the maximum
likelihood classifier. The result images are shown in Fig4-6.
Fig6: classification result of HIS method
3.1 Visual Appraisement
The three main roads in the area are recognized continuously
and clearly from the principle component analysis fusion image,
even the narrow paths in the reed mostly interpreted.
Fragmentary tiny paths in the reed are extracted from the TM
spectral image, while the aqueduct aside the road is recognized
as road by error in the south of the area. The result of the HIS
fusion method is put into the shade of the PCA method. Part of
the paths in the reed is classified.
3.2 Accuracy Analysis
In order to compare the three classifications accurately, we
select 300 samples randomly and get error matrix just as table
1 .The overall accuracy of PCA method is 85%, being the best
among the three classifications. HIS method’s accuracy is 80%,
lower than PCA method, higher than spectral image which is
77%.
4. CONCLUSIONS
Features from individual sensor images are not only preserved
but also enhanced. With both spectral features and penetrability,
tiny objects can be extracted from the fusion image of TM and
SAR. We applied two fusion methods-PCA method and HIS
method to obtain the fusion product. Then we use the
maximum likelihood classifier to classify the images. We get
results as follows.
Instead of replacing PC-1, we replaced PC-5 by SAR image
and obtained classification with the highest accuracy in the
high vegetation region. Experimental results show that PCA
method is suitable for the study area.
The accuracy of HIS conversion method is higher than that of
TM spectral image, but lower than PCA method.
As a result, in the region with flourishing vegetation or being
difficult to acquired remote sensing data, in order to get
accurate data, it is feasible for us to apply fusion method of
radar image and multi-spectral image to extract tiny objects.
