It is well known that hyperspectral remote sensing technique is one of the breakthroughs in the earth observation. Hyperspectral images have the characteristics of contiguous spectral range and narrow spectrum interval. Typically

each image pixel is represented by hundreds of values

corresponding to various wavelengths. Today

with the wide application of hyperspectral images in many fields

such as surveillance

geology

environmental monitoring

and meteorology

the high dimensionality and huge amount data has become a key problem. Feature selection

especially band selection

plays an important role in hyperspectral image processing. In order to achieve the reduction of inter-spectral dimensionality effectively

a segmented two-dimensional principal component analysis (2DPCA) algorithm for band selection of hyperspectral image is proposed. The proposed algorithm is based on the traditional 2DPCA feature extraction method. It combines the advantages of 2DPCA and band selection. The whole band selection process can be divided into two steps. First

the spectral bands in a hyperspectral image are grouped into different clusters based on the correlation coefficient between them. Then

a band selection based on 2DPCA is taken for the bands in each group separately. In the second step

the image data dimensionality is converted in order to get a covariance matrix corresponding to the number of bands in each group. The number of selected bands is determined by the cumulative contribution rate of principal components. The bands are selected according to the amount of information of each band mapped into the principal components. This method has many advantages. It only needs to calculate all of the eigenvalues and eigenvectors of the covariance matrix

while it does not have to transform the original image matrix. Therefore it has a small amount of calculations and will not change the physical characteristics of the original image. According to the inter-spectral correlation coefficients

the original image data can be divided into four groups. The proposed algorithm extracts a total of 18 bands from the whole 189 in 34.937 seconds

effectively reduces the spectral dimension of hyperspectral image. While PCA

segmented PCA and 2DPCA extracts 13

16

11 bands respectively in 192.375

50.829

92.453 seconds. The bands selected by 2DPCA algorithm

which are very typical and contain a wealth of information

can reflect the difference between the pixel vectors corresponding to different surface features clearly. The new image that has fewer bands can be used in subsequent processing such as classification and recognition

the results of them will not be adversely affected. On the basis of 2DPCA feature extraction algorithm

in this paper

we have realized the band selection algorithm based on the 2DPCA and the segmented 2DPCA. Furthermore

they are compared with the existing PCA-based and segmented PCA-based algorithm from the aspects of selected bands

time-consumation

and pixel vector curving. The experiment results show that the proposed algorithm can work faster than the traditional algorithm and the selected bands can preferably retain the local important information of the original image. It can be seen that the algorithm proposed is feasible

efficient and has a high practical value.