Basic clic classification using click detector and ROCCA in PamGuard

Question

I would like to analyze a number of recording files containing, among other things, clicks of marine mammals. The background noise changes partly due to the strong tidal currents in the area. I was able to carry out an initial analysis with python programs, but I want to compare these results with PamGuard software. To this end, I'm using click analyses obtained with Click Detection and ROCCA for classification. Indeed, I've seen that ROCCA can analyze mammalian clicks. I use the ROCCA configuration files supplied by Pamguard (https://www.pamguard.org/downloadpage.php?id=81) in the Eastern Atlantic region (which must correspond to my geographical area of the English Channel). Rocca uses the results of mammal click pre-qualification and I'd like to know several things:

• firstly, whether my methodology was relevant to the classification of these detections (I'd at least like to know whether a click emitted could come from a dolphin, a porpoise, etc.). This approach only brings me results that are “Ambig” by ROCCA, which means that he normally can't make up his mind about classification. With other parameters, I could only find one dolphin species over the whole study period (3 months), so I think the results are wrong.
• What if there were more efficient methods for carrying out this type of analysis?
• If the simple identification of clicks by the clicks classification algorithm could provide sufficient results for an overview of mammal frequentation in the area and if certain parameters should be favored.

Thank you in advance for your answers, I'm a novice in this discipline and I'm looking to learn a little more.

Answer 1

So if I understand correctly, you're trying to:

1. Detect clicks in a three month long underwater recording made in the English Channel
2. Determine whether or not detected clicks are from marine mammals or other (echosounder, invertebrates, dumping of unsuitable product from Pop-Rocks factory, or whatever)
3. For marine mammal clicks, you then want to try to identify down to lowest taxa (e.g. species where possible)

Step 1 is straightforward and has a number of solutions, and it sounds like you don't have an issue with that.

Steps 2 and 3 are much harder -- especially if you are constrained to classifying individual clicks one-at-a-time (independently) without any additional context. One of the major problems with classification of individual clicks is that the characteristics of a detected click from an odontocetes are incredibly variable. Furthermore, much of this variability will arise from factors that are not related to the species (such as location and orientation of animal with respect to the hydrophone). Given the nature of this challenge, it might help to manage your expectations about how good the existing "solutions" to this problem actually are, and reconsider whether or not they will be useful as a "reference".

A general approach that tends to be more successful than classifying individual clicks is to try to classify bouts/sequences/series of clicks that are proximate in time to each other and likely to have come from the same group of animals (sometimes called event). Check out Rankin et al (2016) for an detailed dive (no pun intended) into click classification vs event classification using ROCCA. In their own words, they suggest the performance of the ROCCA click classifier was "mediocre", whereas their event-based classification algorithm BANTER (Bio-Acoustic eveNT classifiER) performed much better. BANTER is open source, available as an R package, and has a detailed user guide and tutorial under development by one of the authors online at: https://taikisan21.github.io/PAMpal/banterGuide.html.