Detection, Classification, Localization and Density Estimation

Last week I attended a workshop on the detection, classification, localization, and density estimation (DCLDE) of marine mammals. This field of inquiry is maturing around the development of more productive use of “Passive Acoustic Monitoring” (PAM) – which is essentially placing hydrophones in the water and listening to what is out there. In the ocean there is often so much out there that the field of DCLDE has been cultivated to sort through all of the data.

Early iterations of the field involved recording the sounds, running them through some manner of converting the sounds into visual representations (called audiographs and spectrograms) and then combing through the visuals to detect recognizable patterns. These patterns were then sorted and classified to sources that could be identified – such as a particular species of whale or dolphin.

Over time the recording techniques have become less cumbersome and the datasets have become much larger – to a point where there were not enough undergrads in the world to comb through all of the data. So the state of the field is now developing into automated detection and classification – writing computer code that can capture and properly classify the types of sounds found in the recordings.

Oddly this process still depends mostly on “visual” pattern recognition; digitally mapping and scanning the audiographs and spectrograms to identify recognizable patterns that fit prescribed templates.

Wavelet analysis of Baird's Beaked whale clicks recorded by Ocean Networks Canada. © Agusonics

Wavelet analysis of Baird’s Beaked whale clicks recorded by Ocean Networks Canada. © Agusonics

The workshop issued a dataset for attendees to explore – in a friendly competition to see what can be pulled out of these recordings. We got to this a bit late but working with Mark Fisher of Aguasonics we plan on evaluating the data using wavelets – which can yield beautiful visual patters, but also can reveal fine details of a signal that standard spectrograms miss. We’re pretty confident that this technique will allow us to easily sort through all of the data and identify specific species and even individual animals – facilitating more accurate population estimates.

To what end does this all serve? Ultimately the sounds that can be identified and localized give a measure of the abundance and condition of the animals that produce them. So the ongoing records can be used to measure impacts of our interaction with the animal populations. Knowing this will improve noise exposure mitigation and conservation efforts – something the US Navy is supporting in a pretty big way.