The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008 
insufficiency of spectral feature in mean shift based HR image 
segmentation, a Gabor wavelet texture descriptor (Manjunath 
and Ma,1996) is introduced and the original mean shift 
clustering method in joint range-spatial domain (Comaniciu and 
Meer, 2002)is replaced with joint spatial-texture domain. 
The remainder of this paper is organized as follows: section2 
provides a brief revision on texture feature extraction based on 
Gabor filter and the spectral feature abstraction. The proposed 
method is described in section 3, and then experimental results 
are given in section 4. Conclusions are drawn in section 5. 
2. IMAGE SPECTRAL AND SPATIAL FEATURE 
ABSTRCT TITLE 
In the HR imagery, some of the landscape elements are 
represented by a group of individual pixels, it means that the 
conventional image interpretation based on pixels will no 
longer sufficient. The combination of spectral and textural 
descriptor will provide a promising solution for remote sensing 
image interpretation (Dong-chen, 1990). The use of Gabor filter 
as spatial feature is motivated by various reasons. Firstly, The 
Gabor wavelets, whose kernels are similar to the2-D receptive 
field profiles of the mammalian cortical simple cells, exhibit 
strong characteristics of spatial locality and orientation 
selectivity, and are optimally localized in the space and 
frequency domains(Chengjun and Wechsler, 2003). Secondly, 
with a bank of Gabor filters a set of filtered images are 
produced and one can use simple statistics of gray values in the 
filtered images as texture features directly. Compared with 
texture descriptor based on statistics in the window, The Gabor 
feature may provide small error rate at region localization. 
2.1 Multiresolution Gabor Features and Dimension 
Reduction 
A two dimension Gabor mother function h(x,y) and its Fourier 
transform H(u,v) can be written as follows (LeiGuang et 
al.,2006; Manjunath and Ma,1996): 
h(x,y)=Qxp 
H(u,v)=2/zcr cr 
X 2 
y 
cos 
exp 
(u-Uo ) 2 
+- 
a 2 ., a 2 
Ti \ 
-U J 
(1) 
(2) 
where <T u = l/(2/TCT x ),<T V = \j{27T(J y j.It is obvious 
that U 0 is the frequency the sinusoidal wave along the u- 
axis .A family of Gabor filters h mn (X, y) can be obtained by 
scaling and rotation of h(x, y) : 
parameter and TYl is scale factor. 
In paper (Manjunath and Ma,1996), a filter bank design strategy 
to make the half-peak magnitude support of Gabor filter 
responses in the frequency spectrum touch each other is 
proposed and a series filtering results corresponding to the 
multi-scale frequency response and having the same spatial size 
as the original image are obtained. In our method, this strategy 
is employed. The lower and upper centre frequency of interest, 
number of stage and orientation should be predetermined in the 
algorithm. In our experiments, the number of stage and 
orientation is chosen as 4 and 6, the lower and upper centre 
frequency is 0.05 and 0.4 empirically, considering the trade off 
between computation efficiency and the overall performance. 
Even though, there are 3x4x6 dimensions Gabor features 
for a HVR image with three bands, PCA(Principle Component 
Analysis ) (Sa, 2001) is used to reduce feature further and 
retain principle components(PC) till an cumulative eigenvalue 
threshold reached. In our paper, the threshold is 0.98 and about 
half of PCs is discarded. 
2.2 Spectral Information 
In most cases, a single texture measure cannot provide enough 
information on ground object discrimination. Better 
segmentation result should be desired by considering multi 
feature fusion. Obviously, the spectral feature in HR multi 
spectral imagery is complementary with the Gabor feature 
providing texture or structure information. 
Generally speaking, Color spaces used for image processing 
purposes should have color discrimination properties which are 
comparable to those of the human visual system.But 
unfortunately, the color error in RGB color space is not 
perceptually uniform and the same Euclidean distances in 
different position of RGB gamut don’t correspond to the same 
perceived color differences. The Lab color space is another 
color space which is designed to approximate human vision. It 
aspires to perceptual uniformity and a available conversion 
formulae from RGB space to Lab space can be find in . So in 
our experiments, the spectral feature in Lab color space is 
employed. 
3. JOINT DOMAIN VARIABLE BANDWIDTH MEAN 
SHIFT SEGMENTATIN 
In this part, a segmentation frame based on adaptive bandwidth 
Mean Shift and some implementation detail is discussed. 
3.1 Adaptive Bandwidth Mean Shift 
Given Yl data points set X = {x i \ i = l,2....n} in d 
dimension space R d , let f(x) be the probability density in feature 
space, then f(x) can be estimated by multivariate kernel density 
estimator (known as the Parzen window estimator)(Comaniciu 
and Meer, 2002) with kernel K(x) and a symmetric positive 
dx-d bandwidth matrix H and given as follow: 
Kr,i. x ^y) = a ~ m s{x',y') a >\,m,n e Z (3) 
by X-x COS# + y sin 0 , y' = y COS0-X sin# 
where 0 = WT jOVieYiatiOYlS _ total is called orientation 
Y\K H (x.-x) 
f{x) = ^ l=l V - 
(4)