I didn’t know which topic to put this under. Please move, if needed
The iNaturalist AI identifier seems to have a difficult time with dragonfly exuviae. I have uploaded many observations sometimes with more than one close-up views. I know there are web pages with explanations about how to identify them by various features. I know people in an Ohio dragonfly survey that uses iNaturalist that can identify exuviae at least to a family. I was wondering if the iNaturalist tech team could comment on how to train the AI identifier to recognize the exuviae and at least get it to say “dragonflies and damselflies”. This is not a criticism of the AI.
Exuviae observations are always going to be fewer than observations of the adults, and only a random subset of photos are chosen for training for each species. But the best way to help improve the model at this point is to make lots of observations of the exuviae that are accurately identified.
From what I’ve read on the forum about similar issues with other taxa, it seems the best way to train the AI is first, to upload lots of observations of odonate exuviae and, second, to ID those exuviae to the best of the community’s ability.
I used to know a fair amount about the exuviae of riverine dragonflies (but not damselflies) in Massachusetts. I’m way out of practice, but my memory is that to get a skin to species often involved close inspection of such characters as the number, size and shape of lateral and dorsal abdominal hooks, or the shape of the labial palps, or the relative length and width of the 9th abdominal segment. In other words, characters that need better and more photos than simply a nice shot from overhead.
I keep thinking I want to get back to being competent with exuviae; feel free to message me via iNat (I’m lynnharper there, too) if you want to form a group of people actively working to educate the AI.
I am very new to dragonflies and damselflies. I just started pursuing them during the pandemic because a buddy knew of the dragonfly survey here in Ohio that uses iNaturalist. (They just released a field guide for Ohio. I know they used photos and data from iNaturalist.) I have little expertise except for what I get from reading various field guides and websites. In my enthusiasm for finding dragonflies and damselflies, I started taking photos of the exuviae. And, someone mentioned that a good side shot isn’t the only thing needed for species ID. So, I intend to try for more detailed images. I just put my close-up lens in my camera bag. So, I will try to continue to contribute observations for the AI and the people who can identify them.
I know the Ohio Odonata Society’s website has a couple of links to exuviae identification. I will have to look them up again.
“Exuviae observations are always going to be fewer than observations of the adults,”. That’s for sure! I rarely find damselfly exuviae. Neither dragonfly or damselfly exuviae last very long. They are delicate.
Damselfly exuviae are indeed delicate and blow or wash away quickly, but I have often found dragonfly exuviae tucked up under bridges and large rocks months after they have emerged.
I suspect the problem here is not that there are too few exuviae observations for them to be incorporated into the training, but rather that they may not be eligible for being included at all.
From what I understand, the CV is not trained on higher level taxa once lower-level taxa are included (i.e., if a species in a particular genus is in the CV, it will no longer be trained on observations that are at genus level, etc.) If exuviae typically cannot be ID’d to species level, but only to family or genus, it is possible that observations of exuviae are not being included in the training set, hence the CV isn’t able to recognize them even broadly as odonates.
This is a problem, I think, for many organisms that can be identified more readily at certain life stages than others.
I have uploaded observations of live nymphs from pond dipping events at local parks. But, an exuvia is probably more difficult for the AI because there is often less detail.