We’re starting a photo-identification study on populations of wobbegong sharks – spotted wobbegongs (Orectolobus maculatus), banded wobbegongs (O. halei), and ornate wobbegongs (O. ornatus) – at Julian Rocks, Byron Bay, on the east coast of Australia, as a Sundive Research / MMF collaborative project. Geeky fun.

I’ve got prior experience with shark photo-identification (you can see my scientific work on the topic here, including a 2018 review article), but Wild Me’s computer science continues to improve and accelerate what’s possible on a week-by-week basis. With that in mind, I thought it might be useful to document the process as I go this time, in case any of you are planning something similar.

I’ll be updating this post as things progress, and please don’t hesitate to ask questions below in the interim.

Can we photo-identify these sharks?

First things first: are wobbegongs suitable for photo-ID? The technique has two fundamental requirements:

1. Individual sharks can be reliably distinguished from one another; and

2. That sharks can be re-identified over the duration of the study.

The three wobbegong species all have intricate colour patterns, so they seem like good candidates for photo-ID (hence why I’m trying it). Spotted wobbegongs are the most common of the three at Julian Rocks; here are a few different individuals…

There’s certainly some individual variation there, so yes, we can identify individual sharks. That said, looking at a bunch of photos, a lot of them are very similar. I have an advantage here in that I’ve got a pro underwater photography setup, so I can use strobes to add colour and zoom in on high-resolution files to check details.

I’m planning to look at a whole bunch of sharks to confirm that all of them are visibly different, while allowing for resightings. It’s helpful to collect other information, such as the sex, length, and any scars that might be present, to cross-check the visual matching.

What area is suitable as a ‘fingerprint’?

For the moment, I’ll photograph the upper surface of the body, from the dorsal fin forward, as my area of interest for identification (as above). That also helps to define when it’s reasonable to add a ‘new’ individual to the database – in this case, if I’ve got a good ‘overhead’ shot of the body forward of the pectoral fins. Wobbegongs often lie where they are part-obscured by rocks, so it’s useful to have a clear protocol for adding new sharks.

I’ll repeat this process with the other two species when I get more photos.

Can we re-identify sharks over time?

The second assumption, that we can reliably re-identify individuals over time, is tougher. We’ll come back to that one! It might be useful to ‘borrow’ some photos from captive sharks for that purpose.

One consideration here is that a shark doesn’t necessarily have to be identifiable across its lifetime. Leopard (zebra) sharks, for instance, do change their patterns as they grow, but have a stable colour pattern as adults. If you’re mostly encountering adult sharks then a juvenile colour or body shape change may not bias your study, but its important to be aware of the possibility.

How can we automate the photo-ID process?

Okay, now, moving on… there are a lot of wobbegongs at Julian Rocks, particularly spotted wobbies (see example below). Although some of them are likely to be resident to the area, wobbies might be more migratory than most give them credit for.

I’ll hazard a very preliminary guess that there’s likely to be tens to a few hundred individuals moving around Julian Rocks at any given time. That’s a lot of sharks. To make this a practical study (i.e. retain sanity) we’re going to need to automate the photo-matching process as much as possible.

On that topic, I’ll refer you directly to my discussion with Jason Holmberg at Wild Me: Adding Wobbegong Sharks to a Wildbook. (I’ll write it up properly soon.)

In short, the first step is to train a machine-learning algorithm to recognise a wobbegong. For this, I’ll need to annotate about 2,000 wobbegong photos from different orientations, on different backgrounds, from varying distances, etc.

Enlisting Citizen Scientists

Now, I could try to take 2,000 photos myself… which would be fun… but it would take forever. Or, I could blatantly cheat :)

So how did I cheat effectively? So far, I’ve used five data sources:

1. The awesome regular divers + staff at Sundive in Byron Bay: thanks all!

2. Asking my friends on Facebook.

3. Searching for wobbegong sightings on iNaturalist, which is an amazing resource that I’m super into now.

4. Looking for #wobbegong on Instagram.

5. Searching for ‘wobbegong’ on Flickr.

I’m just taking screenshots at this stage, rather than trying to identify and match sharks, as I’m following the automation roadmap (see Jason and my discussion linked above).

I’m up to about 1,000 photos at the moment. I’ll write more when I get to the next stage – training the algorithms!

Anything you want me to expand on at this point? Just ask below.

Simon.