The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
classification results are gotten, so decision level fusion of these
classification results is needed.
4.3.4 Decision Fusion Strategy: There are many strategies
for combining classification results of each individual classifier,
of which majority voting principle and Bayesian combination
strategy are the most common used fusion method.
1. Majority Voting Principle: The majority voting method
selects the relevant class by polling all the classifiers to see
which class is the most popular. Whichever class gets the
highest vote is selected. This method is particularly successful
when the classifiers involved output binary votes.
2. Bayesian Combination Strategy: Bayesian combiners are used
to carried out the classification according to the Bayes rule by
selecting the class associated with the maximum average
probability.
3. The proposed Fusion Strategy: Different from the routine
fusion strategy, this paper adopts SVMs C4 to fuse the different
classification results corresponding to different image features
to get the final fusion decision. Each classification results of
respective classifiers serve as the input feature vectors for
training SVMs, which can be regarded as stack multi-classifier
fusion style and the continuation of the aforementioned parallel
multi-classifier system. The total fusion topology is as figure 3.
Moreover the above common fusion rules are also used to get
the classification results corresponding to different features.
Figure 3. paradigm of multi-feature and multi-classifier fusion
5. EXPERIMENTAL RESULTS
In this paper, to illustrate the proposed fusion procedure with an
example, the data used for this experiment are SPOT
panchromatic and Landsat TM 5/4/3 multi-spectral images,
with the same size of 256x256 pixels. Figure. 4(a) ~ 4(b) are
the panchromatic image and the corresponding multi-spectral
images. The experimental area can be classified into water body
(including river, paddy field), naked land (including road,
residential area, bridge and other undeveloped filed) and dry
land by human visual interpretation. The fused images using
contourlet transform to fuse SPOT panchromatic and TM multi-
spectral images are shown in Figure.4(c).
The training area of water body, naked land and dry land are
selected in the images, every category has two block of 16x16
pixels training samples, theses samples of three component
values in Ohta color space correspond to the first type of input
feature vectors, i.e. Tl.
The fused color images are transformed into the GMI. Then
applying TICA to the GMI with 2x2 pixels of block in sliding
window style to get four coefficients of TICA domain,
meanwhile, the statistical parameters, such as mean, standard
deviation, average gradient are computed for the second type of
input feature vectors, i.e. T2.
Turning the original multi-spectral images into 3x65536 pixels
vectors and converting the original panchromatic image into
1x65536 pixels vector to form 4x65536 pixels input vectors,
applying ICA to the whole input vectors, three independent
component bands are shown in Figure.5 (a)~ (c). The results
indicate that the three independent components play good role
in separate the water body, naked land and dry land. Meanwhile
the resulting fused false color images in Figure.5(d) have higher
spatial resolution compare to the original multi-spectral images.
All these three independent components are chosen as feature
vectors for classification, i.e. T3.
The training areas for different classes are chosen in the images,
following the above methods to extract spectral and ICA/TICA
image features and training all the chosen classifiers, the trained
classifiers are then applying to classify the whole fused images
every pixel. Selecting suitable SVMs kernel function and
parameter to train multi-category SVMs with the input feature
vectors of every category obtaining from the afore parallel
classifiers. The trained multi-category SVMs are applying to
classify the whole fused images to gain the classification results.
This paper chooses the radial basis kernel function:
A"(a:,jc') =exp( ||jc - jc| /2a 2 )» where a =2, C=100o
The classification results corresponding to different fusion rules
are shown as follows. To test the effect of the proposed
algorithm, the common fusion rules and the proposed novel
multi-feature and multi-classifier algorithm for classification
are showed in Figure.6 (c) and (e). Besides, the traditional min-
distance method and the max-likelihood method of remote
sensing image classification results are showed in Figure. 6(a)
and (b).To evaluate the classification results objectively, the
total classification accuracy is employed to describe the
classification precision of the images by computing the degree
of confusion between the statistical samples and the actual
samples through sampling randomly in classification results.
Table 1 is the comparison of total classification accuracy
according to different classification methods.