A neural network for classifying clicks of Blainville's beaked whales (Mesoplodon Densirostris)

David K. Mellinger


Beaked whales are difficult to detect visually, and researchers have thus proposed using acoustic detection and classification. Because of the large data volumes often involved in acoustic detection and classification, automatic methods are often used. Here a neural network classification method is investigated. Using backpropagation, a feedforward neural network with one hidden layer was trained to classify clicks of Blainville's beaked whales and other odontocetes recorded in the Bahamas. Training and testing data consisted of approximately 1600 Blainville's beaked whale clicks and 3100 clicks from other odontocetes. Networks with 2-10 hidden units were trained and tested, with performance curves (ROC curves) calculated at several levels of signal-to-noise ratio. Results for most networks were quite good when compared with previous classification efforts, with less than 3% errors in both the wrong-classification and missed-call categories. Future work includes testing the network on sounds recorded in different noise backgrounds and from other populations of Blainville's beaked whales, and combining it with a detector and evaluating the joint performance.


Acoustic variables measurement; Acoustic waves; Acoustics; Computer networks; Feedforward neural networks; Image classification; Population statistics; Signal to noise ratio; Acoustic detection; Odontocetes

Full Text:



  • There are currently no refbacks.