Resources/courses for large PAM dataset- porpoises

Question

I am working on extracting porpoise clicks from a multi-year dataset from the northeast pacific using PAMGuard's click detector. I am new to the software and have come across issues in detection using settings from other studies, something I believe is due to different click peak frequencies with this subspecies of harbour porpoise and with Dall's porpoise. Though I have been reading papers and exploring tutorials, I don't currently have the skills to know exactly what settings would be best tweaked for detection.

Does anyone have any recommendations for courses or resources that would help resolve this issue and assist with development of a workflow?

Does anyone have experience (positive or negative) with the "PAM:Large Dataset Analysis" course with Ocean Science Analytics?

Thank you!

Answer 1

I won't delve much into the specifics of conducting analysis since both WMXZ and Brian answered a lot about it. However, as someone who has completed both the introductory and advanced courses by Ocean Science Analytics, I would highly recommend them. PAMGuard is indeed a powerful software, but it has a steep learning curve. It can be challenging to learn by yourself, especially when faced with a deadline and no one helping you.

I completely agree with WMXZ. If you are not familiar with analysis, or how to interpret results, running software on large datasets and hoping for the best will, in most cases, lead to a waste of time and failure. The key to good analysis is designing the right method from the start, test it and re-test it then run it on all of your data, and then understand in depth what you are looking at.

If you can attend the course, although it is impossible to cover everything in one or two courses, it will definitely help you understand the basics of PAMGuard and how to prepare for analysis. This, in turn, will help you understand the limitations of your analysis and assist you in designing the right method from the start, which will save you a lot of time.

Answer 2

Was going to write as a comment, but my response ended up WAY too long for that. I think there's actually a few questions here, and they're largely independent of each other.

WMXZ has provided a good answer to the question about workflow. If you generalise that approach till you wind up with a pithy statement, you're left with something along the lines of: break the problem up into smaller & more tractable problems that can each be verifiably solved on their own.

So with that in mind, I see separate questions about training courses, large PAM datasets, and the intricacies of the PAMGuard click detector and classifier.

Since I've never taken a PAM training course, I'll defer that question to those better able to answer (noting that one of the founders of Ocean Science Analytics does visit and post on this bioacoustics StackExchange, so could be well placed to elaborate on the content of that course either in this thread, or privately if you reached out directly).

Regarding 'analysis' of 'large' datasets: you might get a better answer from this forum if you're able to be more specific about your needs and your goals. For example, what you've described as a 'large' multi-year dataset could quite plausibly be viewed by some people as a 'typical' porpoise PAM dataset -- depending on exactly how much >1 year you mean by multi-year. You probably wouldn't need to do anything differently to analyse 2-3 years of recordings than you would for a recording of 3 months (it'd just take more computer-time). But if you've got 10 years of data with 50 hydrophones each year, you might need to take a different approach (i.e. to ensure that your analysis can be completed prior to your deadline with the compute resources/budget available). It's probably safe to assume that by posting your question here you're most likely to be in the former category than the latter, but the days of knowing what analysis you're planning prior to collecting multi-year datasets has been long gone for a while now, and these days it's all 'record first, someone will make an AI that can analyse it all later'. But I digress...

In addition to putting quotes around 'large', I also put quotes around 'analysis' too because ultimately, the type of analysis required will depend very strongly on the questions that you wish to answer. Do you just need to know if lots of porpoise clicks were present on a given day? Do you need to know the time of day that they were present? Do you need to compare across dates, times, sites, or with other activities? The first question could conceivably be answered by inspecting daily LTSAs (as WMXZ suggested) possibly without even using a click detector at all. The last question might involve detecting clicks + characterising detector bias + modelling detection range + statistical hypothesis testing. This may seem obvious, but I'm of the mindset that it's often useful to state the obvious, and also useful to start at the beginning, rather than the middle.

Regarding PAMGuard: I say the following with love and respect, and as a long-time PAMguard user, and occasional minor PAMGuard developer. As you've discovered, PAMGuard is not especially user-friendly (to put it mildly). Also, each new version comes with the potential to introduce new quirks and bugs. Getting PAMGuard to do something 'new' can very plausibly wind up being very tedious, difficult, or impossible without inside knowledge of the software, algorithms, and intended usage. PAMGuard is designed to be flexible. This flexibility makes it powerful, but also means that there are a million ways to configure things. However, among those million configurations, there's often only one (or zero) that is actually known to work --and it's not necessarily the configuration you want or need.

That said, detecting clicks, especially from porpoises, has long been part of the core functionality of PAMGuard. The PAMGuard Click Detector module has long been the most mature module, and also consistently one of the most actively developed modules. So, my recommendation regarding PAMGuard would be to reach out to:

1. the PAMGuard community
2. the people who have conducted the studies & tutorials that you've mentioned in your post

These groups will almost certainly be better placed to tell you whether PAMGuard is the right tool for the job, and if so how to configure it. In this instance, I suspect reaching out with a view towards collaboration would likely be the quickest way to obtain the specific information you're seeking, and in the long run might also be more beneficial to you than a training course.

Answer 3

When dealing with large data sets (e.g. Terrabytes of PAM data) I first try to get an overview, e.g. using compressed spectrograms, say 1 spectrogram per day, or week (keyword Log-Term-Spectral-Average LTSA).

This gives me an overview on what to expect. In paricular it gives me time and spectral information on the species I'm interested in. This results also a rough auditing of the data set.

The next step is to zoom into a particular time to have more details and to verify if this is really the species I'm looking for. At the same time I can determine the parameters I need for an automatic detector (e.g. PAMGuard).

After that I would run the detector on the selected data snippet to obtain first results. I would then analyse the detections (discuss with experts) to assess if the detector is setup properly and the results are OK. I will be spending sufficient time on this part as it will condition the overall result.

Finally, if the detector analysis part is satisfactory and I understand the consequences of the parameter settings, then I'm ready to run the detector over the whole data set.

Concluding, I would like to stress that running software (here porpoise detector) on a large data set without first testing it, re-testing it with different parameters, and verifying the results is a recipe for waste of time or even failure.

OK, that is why you ask for a course, fair enough, but I'm a fan of learning by doing.