Target Detection from Limited Number of Spectral bands Using a Signature-based Machine Learning

Hyperspectral camera system captures information using large number of wavelength bands with narrow spectral width in contrast to multispectral camera with a few bands across the electromagnetic spectrum. Hyperspectral data cube can provide significant amount of information in target detection. However, such systems are bulky and generate enormous amount of data and hence the real time processing is challenging for light weight airborne platform and wearable sensor system development. With recent advancement in CMOS image sensor and colour filter technologies, multispectral camera system has become compact for the lightweight applications. This paper demonstrates the suitability of a few selected bands from the multispectral camera combined with signature based machine learning techniques can provide accurate target detection. The study has used a four-band multispectral and one hundred and thirty eight bands hyperspectral systems mounted on a drone platform to detect a camouflage sheet of size 250cm x 65cm from different heights. The results will have application in the development of compact spectral image sensor technology suitable for aerial and hand held, or helmet/body mounted applications.