8
$\begingroup$

I have tried to detect click trains of dolphins in my audio file using the click train identification, but it doesn't detect any click trains.

(The recording contains clicks of the species Tursiops truncatus dolphin)

My recording has some background noise because of sediments that hit the hydrophone and maybe this is affecting my detection. Also, my recording was done using just one hydrophone so the bearing parameter of the click train detector doesn't make any sense here.

I attached pictures where I can clearly distinguish visually the clicks and click trains after using the click detector.

These clicks are really close to each other and they are easy to distinguish but the click train identification is not capable of doing it.

I don't know how to be able to detect these click trains using the click train identification detector. I'd really appreciate if someone could help me.

Thank you in advance.

click trains

Results after using the click classifier and some pre-filtering:

Results after using the click classifier

When I use the new click train detector , the program crashes. Results:

results with the new click train detector

Click train settings:

Click train settings

Data model:

Data model

$\endgroup$
2
  • $\begingroup$ Copy and send the black window text to [email protected] when something crashes - this will work to solve why it happened in the first place $\endgroup$
    – Chloe
    Nov 9, 2022 at 15:09
  • $\begingroup$ There is an extensive PAMguard help file on the click train detector - have you looked at that? $\endgroup$
    – Chloe
    Nov 9, 2022 at 15:10

2 Answers 2

5
$\begingroup$

As I understand it, PAMGuard's automatic click train identifier uses rate of change of click bearings as one of its parameters, so will require two or more channels of data. And it doesn't tend to work very well when at low SNR - it breaks click trains into very short sequences.

Instead, can you use a click classifier to separate a high proportion of Tursiops clicks from noise, e.g., using click length and energy band comparison? If so, you could then display clicks classified as Tursiops, and perhaps include a second window to plot time vs ICI. This won't automate click train extraction, but it may help you discriminate click trains during manual review of your data in the click detector displays.

enter image description here

enter image description here

$\endgroup$
11
  • $\begingroup$ Thank you very much, I was able to use the classifier to separate some Tursiops clicks from noise by comparing click length and energy band. I have also added a second window to plot time vs ICI but it doesn't show me anything in this window, what is the point of doing the latter? $\endgroup$
    – Anaa
    Oct 29, 2022 at 19:36
  • 1
    $\begingroup$ Hi Anna - I find the time v ICI display helpful when identifying trains manually. I've added a screenshot to illustrate how the ICI trends make contiguous trains stand out, in Viewer mode. The example is a 1-ch file of harbour porpoise clicks. $\endgroup$
    – ChrisP
    Oct 31, 2022 at 10:30
  • $\begingroup$ cont... If the display is blank, it's probably a problem with display settings - are you using Viewer or 'normal' mode? In the example above, in Viewer, i linked the scrollers for the two displays, and set both to the same start time and duration. Then ticked 'display ICI for unclassified' and log scale. You should be able to display several minutes of data in the amplitude window but scroll through e.g. 1-min windows of ICI. $\endgroup$
    – ChrisP
    Oct 31, 2022 at 10:30
  • $\begingroup$ cont... If your question is about the relatively new Click Train Detector module rather than the Click Detector module..., then i'm probably not the best person to try and answer you, as i'm still struggling to understand its parameterisation myself. This maybe something tha twill help you though, as it should be able to incorporate your classifier, and colour clicks by train on a 1-ch time v amplitude display, i think. $\endgroup$
    – ChrisP
    Oct 31, 2022 at 10:30
  • 2
    $\begingroup$ After some experimentation - the Click Train Detector module (PAMGuard v2.02.x) did allow me to classify / colour click detections automatically as trains on the 1-ch click detector displays. I imagine when scrolling through longer survey data, having clusters of delphinid clicks highlighted as click trains (as well as classified by type) will make them easier to identify. They can be assigned to events directly from the ICI display (and as super-detections from a time-base FX display of ICIs). $\endgroup$
    – ChrisP
    Oct 31, 2022 at 13:21
3
$\begingroup$

Please note that there are two Click Train detectors in PAMGuard. The one you are using seems to be the one built into the click detector, which I think was written in around 2006 and hasn't been touched since.

Jamie Macaulay has written a much more sophisticated Click Train Detector module which you will find in the detectors menu and can add to your model. This should work a lot better with single channel data - though never as well as it does with multi channel data. The latest release 2.02.05 also has updated documentation for Jamies detector in the help file.

$\endgroup$
3
  • $\begingroup$ Okay, thank you so much for the information! Do I need to download again Pamguard in order to get the latest release or is there an option to update from version 2.02.03 to 2.02.05? $\endgroup$
    – Anaa
    Nov 3, 2022 at 23:32
  • $\begingroup$ Just download and install the latest release from pamguard.org. Existing configuration files won't be affected. $\endgroup$ Nov 4, 2022 at 12:44
  • $\begingroup$ I've tried the new click train detector but the program doesn't respond after running the click train detector...I've attached screenshots of the settings and the results $\endgroup$
    – Anaa
    Nov 5, 2022 at 11:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.