Classification Performance of Support Vector Machines on Topologically Filtered Synthetic Aperture Sonar Signals

Because of its inherent high-dimensionality, automatic target recognition of undersea objects using synthetic aperture sonar (SAS) is an area where support vector machines (SVM) can be a useful tool for classification. Using standard SVM algorithms and kernels, we investigated changes in classification performance from applying topological filters to SAS signatures of undersea mine-like objects (MLO). We simulated real-world data collection conditions by using systematic sampling to determine the minimum resolution and number of target looks necessary to maximize performance. Our results indicate good classification performance using the topological filters. Performance improves as the number of samples is increased and when a combination of topological structures are employed. The results also indicate that performance improves as the resolution is increased; however, beyond a certain point, performance will degrade as the structure of the samples approach the maximum possible resolution for the data set. When compared against the performance of un-filtered data, topological filtering achieved equivalent levels of accuracy but only with larger amounts of training data. The structure of the first three principle component plots on the topologically filtered data suggests that the method of vectorization may be limiting the performance of the support vector machine and that other combinations of vectorization methods and classification algorithms may provide better results.