1114
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing 2008
4. IMAGE FUSION ALGORITHM BASED ON
EXTENSIONS OF ICA AND MULTI-CLASSIFIER
SYSTEM
This part introduces the decision level remote sensing image
fusion algorithm. Existing studies have show that by fusing
multi-spectral images and panchromatic image to get the high
quality fusion images can improve the accuracy of classification
than just using single source image. This paper uses the
panchromatic image, multi-spectral images and the resulting
fused images for researching objects, extracting the spectral
features, texture features and the features in ICA and TICA
transformation domain, using different classifiers to get the
classification results corresponding to different features and
applying the method of multi-classifier system to obtain the
final classification of land cover from remote sensing images.
4.1 Image Fusion Based on Contourlet Transform
Coregister both images and resample the multi-spectral images
to make its pixel size equal to that of the panchromatic image in
order to get perfectly superposable images. Here only R/G/B
three channels are considered. The images are firstly
decomposed by contourlet transform, getting low frequency and
high frequency coefficients in different resolutions and different
directions. Then combining atrous wavelet, the fusion
procedure is choosing different rules on particular sets of
contourlet coefficients that correspond to high and low
frequency bands. The high-frequency coefficients of the
panchromatic image substitute all the high-frequency
coefficients in R/G/B three channels. The panchromatic image
is decomposed by atrous wavelet, getting a group of wavelet
plane coefficients, then adding these wavelet coefficients to the
low-frequency contourlet coefficients, that is further extracts
the detail information of panchromatic image for fusing
application. Final fused images are obtained by using reversed
contourlet transform.
The proposed method can get more information in the fused
results and the spectral reserving character is quite well. The
contourlet transform image fusion method offers a desirable
result to improve spatial resolution and information of the fused
images, which get ready for the next classification procedure.
4.2 Image Feature Extraction
4.2.1 Spectral Feature: The spectral features of multi-
spectral images are the most essential features, here the three
channels of high spatial resolution fused images are considered.
Besides the spectral features extracted in the original R/G/B
color space, other image features can be gotten by transforming
the multi-spectral fused image into different color space. How
to choose the suitable color space is an important factor for
different color space can define different useful spectral
features. So there are two key points should be considered in
choosing color space, one is that the color space can present the
irrelevant color features, the other is that the color space
remains constant in different illumination conditions. Based on
the above considerations, the paper chooses the Ohta color
space(Ohta,1985) which can be expressed as following.
I, =(R + G + B)/3
< I 2 = (R - B)/2 (14)
I 3 =(2G-R-B)/4
The components in Ohta color space are irrelevant, so it can
well apperceive the change of color in statistical view. Ohta
color space is gotten by linear transform of RGB color space,
here I] is the intensity component, I 2 and I 3 are the almost
orthogonal color components. The feature vectors obtained
from Ohta color space are denoted by T t here.
4.2.2 Texture Feature of TICA basis: Since the original
panchromatic image has high spatial resolution and low spectral
resolution, different objects have the same gray value or the
same object has different gray values in the panchromatic
image sometimes. As the fused images can have high spatial
and spectral resolution simultaneously, the fused color images
are transformed into the grayscale image named grayscale
modulation image (GMI) here by specific algorithm. GMI is
useful information source because its spatial resolution is
similar to the original panchromatic image and its gray spectral
values can reflect different objects much better. So by applying
TICA to the GMI block by block in sliding window style, the
texture features of TICA basis can be extracted.
In order to uncover the underlying structure of an image, it is
common practice in image analysis to express an image as the
synthesis of several other basis images. These bases are chosen
according to serve some specific analysis tasks. The advantage
of the TICA basis is that the estimated transform can be tailored
to the needs of the application. A set of images with similar
content to the GMI is selected for training the desired bases.
Then using the TICA basis the GMI is transformed into the
TICA domain in sliding N X N window style. All of these
extracted TICA features can well reflect the structure and
texture information of the fused images. The TICA feature
vectors are denoted by T 2 here.
4.2.3 Independent Component Feature: The existing
studies show that the correlation between the bands of multi-
spectral images sometimes brings ill effect in image
classification. ICA not only can remove the correlation in the
bands of multi-spectral images, but also can makes the resulting
components mutual independent as much as possible. The every
resulting band of independent component embodies a
concentrated reflection of certain ground objects, increasing the
degree of separation between different ground objects.
Therefore independent component analysis can effectively
remove the unfavorable influence and raise the accuracy of
classification.
Multi-spectral images can be regarded as the linear combination
of multi-source mixture signals in some sense due to their low
spatial resolution. This paper treats the panchromatic and multi-
spectral images as four dimensional mixture signals and adopts
ICA to obtain another fusion scheme of multi-spectral and
panchromatic images to get three dimensional multi-spectral
images. Though the resulting independent components can not
reserve the original spectral characteristics very well, they can
express the sharpened multi-spectral images and resolve the
problem of unmixing the mixed multi-spectral images pixels in
the hypothesis of linear spectral mixture model, which lay the
foundation for the next step of properly classification. The