Feature selection for hyperspectral data based on modified recursive support vector machines

In this paper, an improved support vector machines recursive feature elimination (SVM-RFE) approach for feature selection of hyperspectral data is proposed. An automatic model selection (AMS) algorithm using radius margin bound is integrated into the process of feature selection before feature ranking, and the ranking criterion used by standard SVM-RFE is replaced with a new criterion derived from recursive support vector machines (RSVM). To evaluate the effectiveness and efficiency, we apply the new approach to a benchmark Airborne Visible/ Infrared Imaging Spectrometer (AVIRIS) dataset. Experimental results indicate that the new approach improves the SVM-RFE in terms of classification accuracy and computational efficiency; moreover, it increases the robustness of feature selection in the presence of noise.