Hyperspectral Band Selection Based On Image Feature Distribution

Hyperspectral image often includes hundreds of bands and a lot of pixels,though these abundant spectral information and local spatial information can make hyperspectral image accurately characterize different land cover types,it also brings difficulties in practical applications,such as the data storage,transmission and processing.The high redundancy between neighboring bands and a lot of pixels make the data compression possible.Data compression mainly consists of feature extraction and feature selection.If the feature selection is used to select bands of hyperspectral image,then it is known as band selection.Depending on the availability of labeled samples,band selection can be divided into supervised,semi-supervised and unsupervised methods.But these unsupervised methods often cannot obtain satisfactory results without any useful prior information,given that they cannot determine the quality of each band,and these noised bands will be selected because they have low correlation with other bands.In order to solve the problem that these unsupervised methods cannot determine the quality of each band,this paper introduces the objective image quality assessment to determine the quality of each band,and a new unsupervised band selection algorithm is proposed to select bands that are high quality and diverse with each other.Finally,the performance of the proposed method is tested with real hyperspectral images.For conventional hyperspectral image processing,such as segmentation or classification,the samples are image pixels,the dimensionality for each sample is usually several hundreds.However,for band selection,the samples are spectral bands.With the rapid development of remote sensing technology,the spatial resolution of hyperspectral images is also very high,there are more and more pixels,so the dimension of each band image is very high,this may cause dimensionality of the disaster,which in turn affects the performance of the selected subset of bands in the subsequent band selection process.Though some researchers have noticed this problem and reduce the dimensionality of local spatial information by randomly selecting a small portion of the pixels,the random selection will discard some important local spatial information and deteriorate the performance of band selection.In this paper,the feature extraction is introduced to reduce the dimensionality of local spatial information.Then,the extracted features are used for band selection.Finally,the effectiveness of the proposed method is tested with real hyperspectral images.