|
Special Applied Math Seminar Reverse-Time Imaging and Classification for Long-Range Underwater Sensing |
|
We consider the sensing of targets in a water channel, for cases in which the target-sensor distance is large relative to the channel depth. This scenario is relevant for both shallow-water and deep-water sensing, depending on the target-sensor distance. A wideband active sonar is considered, and the transient scattered fields are imaged via a time-reversal mirror (TRM). The TRM accounts for the properties of the channel, although we demonstrate that there can be significant mismatch between the true channel properties and those used in TRM imaging. The TRM imaging algorithm yields the location of the target as well as a time-dependent signature observed at its center. We perform classification based on this time-dependent waveform, considering several different feature-extraction algorithms. The features are processed by such algorithms as a support-vector machine (SVM), relevance-vector machine (RVM) and a hidden Markov model (HMM). In the talk we will discuss the regimes over which these different classifiers are most appropriate. In addition to presenting TRM-based classification results, we will investigate several properties of the TRM itself. In particular, we will discuss the robustness of TRM to mismatch in the actual and assumed properties of the channel (e.g. channel depth, sound-speed profile, bottom characteristics, etc.). Moreover, we will investigate the relative utility of TRM based on a vertical array of receivers, vis-à-vis utilizing a single receiver. We demonstrate that the virtual (image) sources produced by the channel waveguide, for long-range propagation, imply that a single receiver |