The International Archives of the Photogrammetry, Remote Sensing and Spatial information Sciences. Vol. XXXVII. Part B7. Beijing 2008
a. b. c.
Figure 3 K-means segmentation results of two
images. The first column uses Lab color,
the second use all Gabor PCs features;
the third uses a part of PCs which is
determined by contribution threshold
0.98(described in section 2.1)
From the first column, we can find the spectral information
performed well when the great spectral difference exists in
between classes and the result has good discontinuity
preserving performance. However, the grass and tree in the first
image, farm land and tree are mixed due to their similarity in
spectral information. On the contrary, Gabor feature provide
better regional information, but weak boundary. In addition, by
comparing the second and the third column, it is obviously that
Figure 4. The k-means results using combinations of
two type of feature with different scales.
The first column, feature is combined
without any scale change. The second
column, Lab feature is scaled by
multiplying a constant 4/125; and Gabor
PCs is mapped linearly into range [0 1].
4.4 Experiment 3
In this part, our algorithm compared with a fixed bandwidth
mean shift segmentation in (Comaniciu and Meer, 2002) with
default parameters..
the performance of Gabor feature doesn’t reduce significantly
with the reduction of number of Principle Components in some
degree. From the result, we can conclude the Lab color and
Gabor feature may provide us some complementary information
about the HR imagery. In the next section the potential of
spectral and spatial feature fusion will be evaluated.
4.3 Experiment 2
In this part, the combination of spatial and spectral information
with different scales is compared (as shown in Figure 4).
From the first column, we can find that spectral feature seems
to play a dominant role in segmentation since the tree and grass
pair in the upper image, farm land and tree pair in the lower are
similar in the spectral surface and classes mixed together.
Fundamentally, it caused by the different nature of the features,
which are measured in different measurement units and
occupying quite disparate value ranges. In order to get a
balanced contribution of all features to classification and
preserve semantic information, a simple scale factor is chosen.
The segmentation result with standardized data is shown in the
second column. Compared with the first, the confusion of
different land cover classes is reduced, especially for the upper
one.
In addition, the different classes are labeled randomly since we
use K-means for final result; it means that even the same land-
cover types in different result images must be labeled
differently. So we judge the performance of results by
comparing with the original image directly.
a. synthetics texture 1 b. synthetics texture2
Figure 2. Two texture images synthesized from the same
QucikBird image of Wuhan, China. And each
raster band has been qualified into 256 gray
levels. The size of the first one is 128 X
128.the second is 256X256.
4.2 Experiment 1
To validate the differentiate power of features used in the
algorithm; the performance of individual spectral features, all
Gabor PCs features and the retained PCs are represented. In
those experiments, each feature set is put into K-means
classifier directly; classification results for both images are
shown in the Figure 3.
1219