Feature Engineering and Interpretation of Active Sonar Data Using Geometric Wavelets and Support Vector Machines

This paper presents initial findings in exploring the learned weights of support vector machines trained and tested on experimental active sonar field data in both unfiltered and geometrically filtered cases. A two-dimensional Gabor wavelet is used as the filter, applied through convolution, to three different contacts or targets. After training the support vector machine, the learned weights are extracted and multiplied by the input vectors, then summed pixel-by-pixel to gain insight on the classifiers choices. Two domains, ping versus time and ping versus frequency, are considered. The largest increase in classification accuracy seen is 9.17% when training and testing on the ping versus time case due to the smoothing features ping to ping.