Target detection in non-gaussian clutter noise

Automatic target detection algorithms depend upon the characterization of several key sensor parameters. The most important of these, perhaps, involve the optics prescription (i.e., aperture size, focal length, obscuration, etc), electronics, thermal environment, and the line of sight (LOS) stability of the telescope (“jitter”). Combined, these factors determine the detector’s sensitivity against various backgrounds. The magnitude of the contribution from each of these various sources of noise is highly dependent upon the spectral region in which the measurement is made. For example, in the visible region, the day-lit earth background not only provides a large amount of “shot noise,” but also is associated with potentially insurmountable “clutter” noise arising from the telescope jitter. In an infra-red (IR) absorption band, on the other hand, the clutter noise is negligible and the sensor noise (thermal, electronics, quantum,. . .) dominates. Although the clutter-free background of the absorption bands is highly desirable, it is, of course, only useful for targets that are above the molecular absorbers in the earth’s atmosphere. In order to detect missiles at launch (from the earth’s surface), it is necessary to employ a sensor that is active in a “see-to-theground” spectral region, despite the associated clutter noise. Although sensor noise distributions can often be approximated as Gaussian, this approximation is usually quite poor for earth background clutter noise. Typically, clutter noise distributions have long tails, which make setting target detection thresholds difficult: If the threshold is set too high, the probability of detecting the target will be unacceptably low. On the other hand, if the threshold is set too low, the associated false alarm rate will be unacceptably high. In this paper, we present clutter noise distributions calculated for scenes with a variety of spatial gradients and under several lighting conditions. We show that these distributions can be parametrized based upon spectral band, observer geometry, solar geometry, spatial gradient, image quality, and jitter magnitude, and may used to predict probability of detect as a function of false alarm probability (i.e., Receiver Operator Characteristic “ROC” curves are calculated). It is expected that this analysis may be extended to allow recommendations for improved detection algorithms. TABLE OF CONTENTS