A Feature Selection Algorithm for Hyperspectual Data with SVM-RFE

Many conventional classification algorithms have difficulties to be applied to hyperspectral data directly due to huge band number and high correlation among bands.Hence,how to reduce the band number and preserve the information of original data as much as possible simultaneously is an on-going issue.An algorithm of feature selection for hyperspectral data based on one-verse-one support vector machines recursive feature eliminate(OVO SVM-RFE) is proposed.The AVIRIS hyperspectral data was utilized to summarize the principle and characteristic in feature selection.The algorithm was compared to the one-verse-all SVM-RFE(OVA SVM-RFE),MSVM-RFE,OneR,InfoGain and ReliefF approach for feature selection in classification accuracy using support vector machines classifier.Experimental results indicate that the OVO SVM-RFE is a reliable and effective approach of feature selection for hyperspectral data and is superior to other five algorithms in accuracy of classification for selected bands.