Band Selection Using Support Vector Machines for Improving Target Detection in Hyperspectral Images

This paper examines the use of Support Vector Machines (SVMs) in the context of Hyperspectral Remote Sensing, an imaging technique where hundreds of contiguous energy-bands are used to identify ground materials. The purpose of the study is to select a reduced set of features using an SVM-based algorithm whilst maintaining or improving the target detection accuracy. We use an existing algorithm - the SVM- Confident Margin (SVM-CM), to identify only the necessary spectral bands (features) to discriminate between military targets and backgrounds. A limited selection of bands not only improved computational performance but also sub-pixel detection accuracy. The results were evaluated through a multiple regression framework used for sub-pixel detection. An optimal 59 bands out of 128 was selected from SVM- CM for which all 12 targets were detected at a false- detection cost that was 270 times less than the all-band case. All testing were carried out on Multi-Sensor Trial data (MUST 2000) involving military targets.