Research On Band Selection Algorithms For Hyperspectral Remote Sensing Images

Hyperspectral remote sensing is the remote sensing science and techonology with high spectral resolution developed in recent years, who has strong recognition capability of ground feature. However, high spectral resolution has brought series of challenges to efficiency and precision of data processing, which is large data amount, abundant redundant information, high percentage of noise in some bands, etc. On account of these problems, this paper researched on band selection which can reduce the data source while not changing bands’ physical meaning.The band selection algorithms can be divided into two categories, which are supervised and unsupervised ones. For unsupervised algorithms with no priori knowledge, the main principles of band selection are large information amount and high degree of independence. Of multiple band selection algorithms, linear prediction algorithm (LP) is relatively efficient and effective. Through analyzing, there are three shortcomings within LP, which can affect the result and efficiency of band selection.On account of the shortcomings listed above, some improvement ideas has been presented. First, removing bands with large amount of noise according to their entropy of wavelet subband before band selection; second, measuring information amount by skewness, kurtosis, K-L divergence and mutual information while selecting the initial two bands; Third, improvement in subsequent band selection through frequently removing unnecessary bands while iteration.Experiments on classification and pixel unmixing, which are very important in hyperspectral data processing, have been implemented to confirm the effectiveness and efficiency of modified LP band selection algorithms. Support vector machine (SVM) and k-nearest neighbor (KNN) are used in classification, non-negative matrix factorization (NMF) is used in pixel unmixing. The two experiments have demonstrated that modified linear prediction band selection algorithm is both effective and efficient, and it is very significant to hypespectral data processing.