228 
(a) ETM+Multispectral Image, (b) Panchromatic Image, 
(c) Fused Image Based on PCA, (d) Fused Image Based on 2DPCA 
Figure 1. Results of Image Fusion Experiment 
4.2 Analysis 
From images shown in Figure 1, some useful points can be 
drawn as follows. 
(1) In the fused image based on PCA-based technique, although 
the spatial resolution is greatly improved, the spectral 
information is badly lost. Furthermore, it is difficult to 
distinguish the surface features form each other. 
(2) Compared with the multispectral images, the spatial 
resolution is improved and the spectral information is well 
preserved in the fused image based on 2DPCA. 
(3) Compared with c), fused image in d) has better spatial 
resolution and enhanced spectral information; meanwhile, it is 
easy to distinguish the surface features. Therefore, 
2DPCA-based algorithm has better performance than 
PCA-based algorithm. 
Why the performance of 2DPCA-based method is apparently 
better than that of PCA-based one? We consider that there are 
two main reasons. 
(1) In PCA-based algorithm, an image must be transformed 
into a 1-D vector when PCA is applied to the image and can not 
utilize its structural information. In 2DPCA-based algorithm, 
however, 2DPCA is directly applied to the image matrices 
instead of ID vector. Therefore, the structural information of 
the image is effectively utilized. 
(2) In PCA-based algorithm, the multispetral images are 
regarded as a whole, i.e., each band of the multispetral images 
is regarded as a feature, in the analysis and reconstruction 
processes, but the case in 2DPCA-based algorithm is 
oppositional, i.e., each band of the multispetral images is 
regarded as many features whose number is equal to the height 
of the image. 
5. Conclusions 
In this paper, 2DPCA is introduced into image fusion in remote 
sensing, and a novel image fusion algorithm based on 2DPCA 
is proposed. 2DPCA-based technique has some advantages, in 
contrast to PCA-based method. 
(1) 2DPCA is directly applied on image matrices, and the 
images are regarded as 2D matrices in reconstruction instead of