Detecting anisotropic scattering with hidden Markov models zyxwv L.R.Flake S.C.Ahalt A. K. Krish nam u rt hy zyxwvutsrqp Indexing terms: Anisotropic scattering, Hidden Markov models, Multiaperture synthetic aperture radar imaging zyxwvu Abstract: zyxwvutsrqp The authors propose that hidden Markov models (HMMs) and multiaperture synthetic aperture radar (MASAR) imaging can be used to construct automatic target detection algorithms. Their preliminary studies show that HMM MASAR detection: effectively exploits anisotropic scattering differences between target and multiple clutter classes; is computationally efficient; can be used with either single polarisation or multipolarisation SAR imagery; and can be used with either coherent or noncoherent subapertures. Further, the results indicate that the accuracy of HMM MASAR detection is comparable to other techniques while requiring orders of magnitude less computation. These results suggest that the HMM MASAR detection technique could be effectively deployed in fielded automatic target recognition systems. 1 Introduction Computational costs represent a fundamental problem for automatic target recognition (ATR) systems designed for use in wide-area ground surveillance. This computational burden is a result of two factors. First, broad geographical areas must be frequently imaged to detect all targets in a timely fashion. Secondly, to detect and then identify individual vehicles or other tar- gets, high-resolution sensors must be employed. Conse- quently, surveillance systems require high-throughput ATR processing of high-resolution imagery [I]. One way to reduce the required computation is to prescreen the imagery to reject portions of the images that are of no interest [l]. Ideally, a prescreener must locate all targets while simultaneously minimising the number of nontargets passed to the next stage of processing. The prescreener must accomplish this goal while minimising the amount of computation required to perform the detection. Using synthetic aperture radar (SAR) sensors, which allow all-weather and foliage-penetrating surveillance, zyxwvu 0 IEE, 1997 IEE Proceedings online no. 19971048 Paper first received 18th June and in revised form 4th December 1996 The authors are with the Ohio State University, Department of Electrical Engineering, 205 Dreese Laboratory, 2015 Neil Avenue, Columbus, Ohio 43210. USA anisotropic scattering can be exploited for discriminat- ing between targets and nontargets [2]. 2 Anisotropic scattering Anisotropic radar scattering occurs because the illumi- nated object reflects more of the incident energy back to the SAR sensor from some aspect angles rather than others. To illustrate, for a dihedral of length L illumi- nated at wavelength A, the half-power width of the backscattered field is approximately AI2L radians [2]. This phenomenon of increased reflection at some angles is sometimes referred to as glint or flush. While both natural objects and manmade objects may glint at certain angles, the magnitude of such glints is usually much less for natural objects because manmade objects are largely constructed from flat segments. Conse- quently, manmade objects exhibit more pronounced anisotropic scattering than natural objects [2, 31. By analysing the relative differences in returns from one angle to another, a prescreener can take advantage of anisotropic scattering to differentiate between man- made and natural objects. Since targets are manmade, a prescreener can significantly reduce the amount of imagery that needs to be processed by simply detecting and isolating manmade objects. Although the physics of target glint occurs for any wavelength, the use of anisotropy as a discrimination feature holds particular promise for foliage penetrating radar, which operates at L-band (1.2GHz) and lower frequencies. Foliage- penetrating SAR can image objects occluded by tree canopy, but numerous false alarms are produced by the generally isotropic scattering from tree trunks [2]. 3 Forming multiaperture images rather than high resolution images A basic assumption made in SAR image formation is that the radar return from each point in a ground patch will be isotropic [4]. SAR images are formed by coherently integrating many radar returns collected over a range of aspect angles, and this large synthetic aperture is used to increase cross-range resolution [4]. However, since glints occur for only narrow aspect- angle ranges, they are represented in only a small number of the individual radar returns. When the non- glint and glint returns are integrated together to form high-resolution images, the glints are attenuated. Thus, the goal of high-resolution SAR imagery must be bal- anced against the cost of diminishing the glints that can be used for detection. Generally speaking, as the SAR 81 IEE zyxwvutsrqponmlkj Proc.-Radar, Sonar Nuvig., Vol. 144, No. zyxwvutsrqpo 2, April 1997