Performances of human listeners and an automatic aural classifier in discriminating between sonar target echoes and clutter

Human listening tests were conducted to investigate if participants could distinguish between samples of target echoes and clutter obtained from a broadband active sonar experiment. For each echo, the listeners assigned a rating based on how confident they were that it was a target echo or clutter. The measure of performance was the area under the binormal receiver-operating-characteristic (ROC) curve, A(z). The mean performance was A(z)=0.95 ± 0.04 when signals were presented with their full available acoustic bandwidth of approximately 0-2 kHz. It was A(z)=0.77 ± 0.08 when the bandwidth was reduced to 0.5-2 kHz. The error bounds are stated as 95% confidence intervals. These results show that the listeners could definitely hear differences, but their performance was significantly degraded when the low-frequency signal information was removed. The performance of an automatic aural classifier was compared against this human-performance baseline. Results of statistical tests showed that it outperformed 2 of 13 listeners and 5 of 9 human listeners in the full-bandwidth and reduced-bandwidth tests, respectively, and performed similarly to the other listeners. Given its performance, the automatic aural classifier may prove beneficial to Navy sonar systems.