Point Target Detection in Hyper-Spectral Images

The analysis of hyperspectral imagery promises to provide technical solutions to problems in many areas of research; this is particularly true of target acquisition. Exploiting high spectral resolution data contributes greatly to the discrimination power of standard image processing techniques. This additional dimension of information is based on the physical characteristics of the target material under consideration. The present research addresses the problem of the detection of a point target, moving with sub-pixel velocity, from a time sequence of hyperspectral data cubes. The emphasis in this paper will be on the degree of improvement in target detection algorithms that can be expected as a function of the degree of difference between the target and background signatures. Differences obtained between the use of real spectral signatures, compared to synthetic ones, for the noise, background and target end-members, and their implication on the detection results will be discussed. The standard matched filter for target detection is broadened and improved by advanced non-data dependent techniques. In order to estimate algorithm performance, five different tests (detection methods of varying sophistication) were applied to the real hyper-spectral data. The results were compared to the synthetic data outcome; conclusions regarding the threshold needed for spectral differences for the target detection to be notably improved are reached. The major focus of the research is a comparative understanding of the target detection results in different scenarios: strongly, partially and lightly cluttered sequences.