Hyperspectral remote sensors

which can collect data simultaneously in dozens or hundreds of narrow and adjacent spectral bands for each pixel

have been developed. However

the bands are usually highly correlated because of the fine spectrum resolution

thereby leading to great redundancy in hyperspectral datasets. Owing to the redundancy of data

utilizing all of the bands in an algorithm does not necessarily lead to an improvement in the results. These problems result in serious difficulties in data processing. To implement data dimensionality reduction efficiently and make the data exhibit minimal redundancy and considerable information after dimensionality reduction

a band selection method based on optimal combination factors is proposed. Coarse selection is implemented to remove bands with low informative content. Shannon entropy is utilized as a measure of informative content. For a large number of bands

subspace partitioning of the entire data is conducted. An adaptive subspace partition method that can automatically complete subspace partitioning is applied. In each subspace

by calculating the errors of two bands (one is minimized and the other is small)

the reconstruction error is obtained by the linear predictive mode. Their combination factors are calculated with the product of their errors and standard deviation. The combination factors are then compared to identify which one to remove. The band that has minimized combination factors is removed. Autocorrelation matrix-based band selection

which utilizes the minimum linear prediction error as the selection criterion and searches the suboptimal subset by sequential backward selection (SBS)

is employed. The proposed method uses the same SBS strategy to remove the band one by one until the desired number of bands is obtained. Finally

the bands selected in all subspaces are merged to a new set. The same dataset is utilized for experiments on time consumption and classification accuracy. The dataset was acquired with the airborne visible infrared imaging spectrometer (AVIRIS) in 1992. The dataset has 145×145 pixels and 220 bands with a range of 400 nm to 2 500 nm. The method has high computational efficiency

as revealed by experiments. Comparison of the computation time of all the four methods showed that the proposed method has a slightly shorter computation time. Support vector machine (SVM) has elicited much attention because of its capability to handle dimensionality compared with conventional classification techniques. Therefore

SVM is used in this study to classify the band subset. The classification accuracy of this method is approximately 1.5% better than that of others. The method of combination factors considers the minimum degree of redundancy and the maximum information of the band subset. This method obtains the best band subset of the data

has minimal computational complexity compared with other methods

and is applicable to AVIRIS and other high-spectral image data.