Research On Mutual Information Based Band Selection For Hyperspectral Data Dimensionality Reduction

With the development of remote sensing technology, hyperspectral remote sensing data have been applied more and more widely in environmental monitoring, resource exploration, feature identification and other areas of national economy and life. However, hyperspectral data have narrow bands and large quantities of channels. The signal to noise ratio for each band of hyperspectral data differs and there is a strong redundancy between the bands with the curse of dimensionality. All of these problems have brought great difficulties for the classification, matching and recognition of hyperspectral data. Therefore, it is essential to carry on dimensionality reduction for hyperspectral data. The most commonly used methods for dimensionality reduction are generally divided into two categories. O ne is to reduce the dimensionality by a series of mathematical transformation on the original data, such as PCA, CA, etc. The other is to select some typical feature bands according to certain criteria without changing the original data. This paper focuses on the latter, band selection methods.Based on the analysis of the form and characteristics of hyperspectral data, the feasibility and necessity of the band selection is summed up. The general flow of the ba nd selection is summarized after briefly introducing several classical band selection algorithms. Then this paper makes detailed research on a certain band selection algorithm which uses mutual information as the criterion. The selected bands according to the original band selection algorithm based on mutual information are often centralized in a continuous interval. Such bands are usually of much relevance with information redundancy. Thus, this paper presents a new method, MI-ASP, which combines mutual information and automatic subinterval partition. The proposed method firstly divides the whole band interval into several subintervals. Then feature bands are selected within each subinterval based on mutual information. Through this improvement measures, the problem brought by selecting too much redundant bands can be effectively avoided. Meanwhile, when applied to hyperspectral data with huge dimensionalities, such as the detection of the oil spill on sea surface and the recognition of aircraft targets, this method can reduce the computational complexity and decrease the running time. It is very favorable when processing large amount of spectral data with real-time requirements. Finally, two simulation experiments are conducted using the proposed band selection method. The validity and feasibility of MI-ASP algorithm is verified through the analysis of the experimental results.