The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008 
statistic WG(i,j) of the similarity measures of spectral and 
texture distributions G s and G t . 
WG(i,j) = w t G t + w s G s 
(5) 
The weights of texture and spectral distributions W and 
W y should be adaptively determined in terms of different 
characteristics of pairs of regions. If two regions are smooth, 
the weight of spectral distribution should tend to be large. If 
two regions have obvious texture characteristic, the weight of 
texture should be larger than that of spectra. Standard deviation 
(SD) can evaluate the smoothness of a region to a certain 
extent. Smooth region produces small SD and rough region 
produces large one. So we apply SD of regions to evaluate the 
feature weights between two regions i and j, if the SD values of 
two neighboring regions are less than 40, 
w = max(SD.,SD.) 
s v ,» jj ( 6 ) 
u t = min (SD i , SD j ) 
Or else, U t is set to be the larger one. Where U t and U s are 
the weight estimation of the texture and spectral distributions; 
SD i and SD. are the SD of regions i and j. After 
normalizing the weights, we have the final result: 
w,=u,/(u, + u s ) 
w s =u s /(u,+u s ) 
3. SEGMENTATION METHODOLOGY 
Figure 1. Region-based unsupervised segmentation adaptively 
combining texture and spectral distributions. 
The whole segmentation framework in this paper includes 
three steps: hierarchical splitting, modified agglomerative 
merging and pixel-wise refinement (see Figure 1). Hierarchical 
splitting recursively split the original image into four square 
sub blocks of varying size based on a homogeneity test: 
By our experiment, a better way of calculating WG(i, j) in 
the split process is by normalizing the six G-statistics. The 
normalized G-statistics are calculated by: 
WG 
R = ^™ax > x 
WG m ;„ 
(10) 
G” r =G r lY,G r 
r* (8) 
and the weighted sum similarity between two regions 
WG(i, j) is defined as: 
WG(iJ) = w t -G: + w,-G; ... 
Where WG max and WG min represent the largest and smallest 
homogeneities among the six pairwise homogeneities of the 
four sub blocks. The initial divided window size is set to 64 and 
the smallest size is 16. The block recursively split into four sub 
blocks when R is greater than a threshold X. The value of X is 
invariant for different kind of images: X is experimentally set to 
1.3 to 1.5 for regular texture images and 1.2 for H-res satellite 
imagery. 
Once the image has been split into blocks of roughly uniform 
features, the blocks are merged through a modified merging 
procedure. At a particular stage of merging, we merge that pair 
of adjacent regions which has the smallest merger importance 
(MI) value. MI is defined as: 
MI = Jpx WG 
where p is the number of pixels in the smaller of the two 
regions and WG is the weighted sum similarity measure 
between the two regions. The reason we adopt equation (12) 
instead of MI = p X WG (Ojala and Pietikainen 1999) is