Experiments with Broadband Sonar for the Detection and Identification of Endangered Shortnose Sturgeon

Abstract

AbstractExperiments were conducted in the tidal Delaware River to evaluate the use of broadband sonar to remotely detect and identify shortnose sturgeon, a federally listed endangered species. The objectives of this study were to determine if shortnose sturgeon could be detected by broadband sonar and, if so, to develop classifiers that could differentiate shortnose sturgeon from co-occurring fish species. Broadband acoustic echoes were collected from shortnose sturgeon, three non-sturgeon fish species, and the river bottom. Classifiers were developed using neural network analysis of the normalized frequency response spectra from each target class. Two independent classifiers were developed, one that distinguished “sturgeon” from the river “bottom” and a second that classified targets as “sturgeon,” “non-sturgeon fish,” and the “bottom.” The training performance of the classifiers was 100% for each class. Testing of the two-class sturgeon vs. the bottom classifier resulted in correct classification of 96.6% of the shortnose sturgeon detections and incorrect classification of 3.4% of actual sturgeon echoes as bottom. None of the bottom detections were misclassified as sturgeon. Testing of the three-class classifier resulted in the correct classification of 89.0% of the actual sturgeon detections, and incorrect classification of 8.5% of actual sturgeon detections as non-sturgeon fish and 2.5% of actual sturgeon detections as bottom. Some 27.1% of the actual non-sturgeon fish echoes and 5.9% of actual bottom echoes were incorrectly identified as sturgeon, yielding a false-positive rate of 16.5%. Given that the “non-sturgeon fish” in this study typically occur in much higher abundance than shortnose sturgeon, incorrect classification of echoes from these fish would lead to overestimation of the abundance of shortnose sturgeon. Notwithstanding this potential problem, the results of this preliminary study were promising, and further investigations to improve classifier performance are warranted.