Band selection and performance analysis for multispectral target detectors using truthed Bomem spectrometer data

This paper investigates the problem of spectral band selection for multispectral target detection techniques. Based on truthed high-resolution spectrometer data, an empirical study is performed. Both detector performance and optimal band selection are analysed for a wide range of targets and background pairs under a variety of conditions. In particular, the authors investigate multispectral detector performance as a function of spectral band selection and other factors such as noise and spectral bandwidth. Bomem spectrometer data is used exclusively in this study. The Bomem spectrometer collects data in the range of 2.85 μm to 14.32 μm distributed over 728 spectral bands. The multispectral target detector considered here is based on a Bayes classifier. The selection of optimal spectral bands is investigated using the Mahalanobis distance criteria. The authors believe that this is an appropriate quantitative measure of target and background class separability. Given the truthed Bomem data, this metric provides a systematic method for performing spectral band selection and provides a quantitative measure of detector performance