The former likely reflect animal movements, which is important information for managers in an area where interactions with shipping is a major conservation concern, while the diel variation likely represents varying amounts of foraging effort over the 24hr cycle.
Finally, we did also do some biology – both detectors showed strong seasonal and diel patterns in the presence of sperm whale clicks in all our sites in the waters between Ibiza and Mallorca.
Problem is, models trained on one site don’t transfer well to unseen sites, so the next challenge to tackle is how to optimise transfer learning – minimising the effort needed to get a pre-trained model up to speed on a new unseen set of recordings… watch this space!
So if you have a trained NN model, the speeds are game-changing – but if you also have to train the model in the first place (which means lots of manual auditing to generate training data) then you haven’t gained much.
While performance metrics in terms of detection were similar, the neural network approach, once trained, took just *12hrs* on a desktop PC to analyse over 50k 4 min recordings (~3500hrs of data) while the ‘traditional’ detector took 352hrs.
... bear that it mind next time you read a paper where a single analyser did the auditing 🤔 The two detectors agreed with each other more than the humans. What even is truth?! We didn’t mean to prompt an existential crisis, yet here we are…
How do you test detectors? By comparing them to truth – but who determines truth? Humans. One striking result though was that even for very stereotyped species like sperm whales, humans agreed only around 70% of the time as to whether their clicks were present or not 😬...
... so we did using long term moored PAM data from our ongoing sperm whale project around the Balearic Islands
Deep learning is being increasingly used in bioacoustics with the promise of automatically analysing large datasets very rapidly. But whether its detection performance significantly improves existing techniques is less often asked...
New paper: Comparing neural networks against click train detectors to reveal temporal trends in passive acoustic sperm whale detections (DM lgf3@st-andrews.ac.uk for a copy) 🧵 pubs.aip.org/asa/jasa/art...
@pamguard.org @lrendell.bsky.social @seamammalresearch.bsky.social
