Feature Extraction and Selection Based on Automatic Subspace Partition and Genetic Algorithm for Hyperspectral Images

摘要

Classification and pattern recognition of high dimensional remote sensing data are distinctly different from traditional multi-channel remote sensing classification techniques. In this paper, a newly integrated feature extraction and selection algorithm based on Automatic Subspace Partition (ASP) and Genetic Algorithm (GA) is proposed for high dimensional data reduction and classification. This method involves two steps: subspace partition of whole data space, feature extraction and features election based on Genetic Algorithm in subspace. Hyperspectral data bands are firstly partitioned different subspaces base on neighboring correlation of bands and extracted spectral feature of different subspace. Followed by the decomposition phase is optimal feature subset selection, in which the optimal feature subset acquired the best divergence is obtained according to interclass/infraclass distance of the training samples. A Genetic Algorithm implements this procedure, with each possible feature subset encoded as chromosome. Fitness scores in GA are calculated and evaluated based on Jeffries- Maturity distance of the selected training samples. In order to testify the effectiveness of the proposed method, classification experiments of hyperspectral images are conducted on Hyperion data. The experiment investigation shows that the classification result in our new method is improved compared with segmented principal component transformation (SPCT).