|
This paper discusses work toward a fundamental, algorithm- independent view of the ATR performance that can be achieved using SAR data. Such ATR performance predictions are intended to enable evaluation of performance tradeoffs for SAR designs, including both parameter selections and added domains of SAR observation, such as 3D, full polarimetry, multiaspect and/or multifrequency. In the paper we evaluate the classification error for two tactical targets using a Monte Carlo technique. A number of approximations are made and are detailed in the paper. Data on target signatures come from pencil-beam laser data and target photographs, which determine shadowing of the ground clutter. A single aspect angle is used for each target in the initial results. A layer of radar netting is modeled on both targets. This information is used as 'ground truth', to compute the average power that would be seen in each pixel of a SAR image, for each target. SAR image trials are then generated using independent Rayleigh amplitude fades in each pixel. In an optimal Bayesian fashion, the smaller of the probabilities of target (T1) or T2 given the trial image data is the error probability for that trial. An average over the Monte Carlo image trials yields the overall classification error probability. Comments are given on reducing the number of required trials in such a Monte Carlo. THree results of the work are shown. First, a tradeoff is made of ATR performance versus SAR resolution. ATR improves as the resolution is made finer, and physical reasons for this are discussed. Second, the relative ATR utility is determined for those pixels where at least one target has scatterer as compared with those pixels where the targets differ only in the degree to which they shadow the ground clutter. Third, an early analytical result is given for interferometric SAR, showing the physical reason behind the potential of height-sensing SAR to improve ATR - the possibility of canceling the background response becomes an important factor. Finally, the ability to make absolute performance predictions versus relative predictions is discussed, with the conclusion that relative predictions are more feasible at this time.
|
