Dual baseline method enable both enlarge the width of 
Interference fringes and increase the height ambiguity so that 
precision of phase unwrapping and height inversion can be 
improved. 
5. EXPERIMENTAL RESULTS AND DISCUSSION 
5.1 Land Classification Result 
A generic object-based multi-scale idea is introduced to the land 
classification (Corr,D.G., 2003).After the polarimetric analysis 
for different lands, surface scattering fraction as feature 
parameters to distinguish surface, forest, building and cultivated 
lands via multi-threshold segmentation at the first scale analysis 
layer. As it is sensitive to texture features exhibited in different 
types orchards and crops, the two features of gray level co- 
occurrence matrix (GLCM) homogeneity and dissimilarity are 
selected to distinguish orchards and crops via region-based 
segmentation at the following scale analysis layer. Then, merge 
the two layers to generate integral segmentation image. The 
each class is finally assigned at all the objects according to 
image object features within a search domain. The classification 
result for test areal is shown in figure.6, the 8 classes are 
generated consisting of bare surface, crop, forest, building, 
lychee, mango, pawpaw and areca. By the ground data 
validation, the overall accuracy is up to 0.85. 
    
hare 
aie crop forest building areca lychee pawpaw mango 
Figure.6. Land classification result of test area 1. 
5.2 Man-made Objects Detection Result 
As the Polarimetric Analysis section described, surface 
scattering fraction, entropy, anisotropy are all proper 
polarimetric descriptors for man-made objects detection. 
Support Vector Machine (SVM) method (B. E. Boser, 1992) is 
adopted to fulfil man-made objects detection with these 
polarimetric descriptors. Since man-made targets are the main 
object concerned, it is two classes of man-made objects and 
other objects problem research of SVM method. The SVM 
construct two hyperplanes that separate respectively one class 
and the other class. The Libsvm library (C.-C.Chang and C.-J. 
Lin., 2011) is employed with the RBF kernel of 6 = 0.5 and the 
cost parameter equal to 1000. 
The detection result for test area 2 is shown in figure.7. , the 
most of man-made objects, e.g. the buildings and street lamps 
are mostly extracted. By the ground data validation, the overall 
accuracy is up to 0.89. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B7, 2012 
XXII ISPRS Congress, 25 August - 01 September 2012, Melbourne, Australia 
    
  
y 
po EU ta ES ‘a casses e 
Figure.7. Man-made objects detection result of test area 2. 
5.3 DSM Mapping 
The result of ground topography mapping for the test area 3 can 
be derived by a dual baseline method. Due to short wavelength 
of x band data, it is feasible to map the height of the top of the 
canopy rather than the height of the underlying terrain. It is 
more accurate to call the map digital surface model (DSM). As 
the Ground Control Points (GCP) for precision evaluation are 
not available in the test area, the DSMs derived from three 
different methods (MLE, ESM and the proposed method in the 
paper) are compared. 
The coherence map with a 9X9 pixel window are calculated, 
three coherence maps from different height estimation methods 
can be compared. As shown in figure.8., coherence histograms 
in green, red, black colors indicates ESM method has the 
highest coherence depending on its own characteristics of 
coherence maximum, and the dual baseline method also has 
higher coherence than MLE method because dual baseline 
method implements coherence optimization processing whereas 
MLE method not. 
  
  
  
  
  
  
Proposed i 
Sd ee € 
[ ESM i 
3x19 - MLE | 
2x10°- | 
ixurr- m" H4 
us 02 94 06 Us 
cohereace 
Figure.8.Coherence histograms of three methods 
DSMs derived from three methods can be obtained shown in 
figure.9. As shown in figure.10. , histograms in green, red, 
white colors show the height distribution of three methods. 
Although accurate GCPs are not available, the range of 
elevation of test area 3 can be searched about 20~35 meter. We 
can find in the green histogram that the height distribution of 
dual baseline is most approaching the true elevation of test area 
3 as its peak value is close to 30 meter. Contrast with the dual 
baseline method, other two methods both deviate the true 
elevation range about 10 and 30 meters. It demonstrates dual- 
baseline polarimetric interferometry method has better elevation 
precision than MLE and ESM.