Experimenting with Identification (1/2)

Hello all,

Today I'm presenting one of my methods I thought of for orchid identification, which are new to the best of my knowledge. I do not, unfortunately, have enough data to make them work—nor have I tested them enough to assess their accuracy—so this will be very casual and require that any reader takes innumerable pinches of salt. If these new ideas don't go anywhere, they can at least be food for thought.

See the link below for a draft of the visual method and an explanation. I'll be referring to my methods as #1 (this post) and #2 (the next).

Visual Identification: https://docs.google.com/document/d/1_uN5RaVRD5J5BOzXx2eCk12FC9ie-XQxSRjeE3TGCp8/edit?usp=sharing.

I look forward to hearing some opinions on this.

Publicado el junio 20, 2019 11:00 TARDE por arethusa arethusa


Very cool! This is definitely an interesting way to look at morphlogy! It is nice to see the range of variability in different features, something you don't get with I dichotomous key.

Publicado por afid hace alrededor de 5 años

I'm glad you like it!

Publicado por arethusa hace alrededor de 5 años

Interesting. I'd like to see if one would be able to extract the automated classification from iNaturalist itself (or other plant recognition engines) and to map it to the same formalism, so that expert judgement could evaluate those engines (i.e. in which part of the space they certainly get it right, and in which they tend to fail).

Publicado por thierrygoubier hace alrededor de 4 años

I would run this by Dr. Tony Reznicek, retired, U of M Herbarium, if I were you.

Publicado por bill_spiers1 hace casi 3 años

Really neat ideas here! Some of these figures remind me of "morphospace" (a principal components plot of shape space). From Ridley's textbook on Evolution: "..morphospace is a representation of the possible form, shape or structure of an organism. Each axis of the morphospace corresponds to a variable which describes some character of the organism. Each point in the morphospace represents an individual organism.". Only to say, should you pursue this project, it would fit really well into this framework, and be incredibly useful to amateurs like myself!


Publicado por mannfred hace casi 3 años

Thanks! I'm glad you like them. Believe it or not, I actually came across some morphospace visualizations a while after posting these, but didn't read up on them too much at that time. It was encouraging to see that visualizations similar to my own have been used, but also somewhat disappointing to realize that my work wasn't completely new. That's why I never made a second post, and also why I didn't contact Dr. Tony Reznicek as @bill_spiers1 suggested. I did look him up though, and he seems like someone certainly worth becoming acquainted with at some point. He actually has a rather rare Platanthera hybrid named in his honor (Platanthera ×reznicekii).

I've also toyed around with an identification key that generates numerical sequences that correspond to different species (for species complexes where no one characteristic is truly diagnostic) and convolutional neural nets using TensorFlow. I know that there are apps like PlantSnap out there which use similar programs, but the accurate identification of species pairs requires specially designed training sets. For Platanthera species like P. psycodes and P. grandiflora, I made hundreds of cropped photos which only showed the rostellum lobes and orifice so that TensorFlow would focus on those features and ignore characteristics which I deemed superfluous and/or unreliable. It worked well for those two, and was surprisingly good at separating P. aquilonis and P. huronensis if memory serves. For whatever reason, one of the more difficult pairs was actually P. leucophaea and P. praeclara, even though they're also distinguished via rostellum lobe positioning. The problem could've easily been solved if the model could take location into account when comparing different species, but I'm not advanced enough to tinker with it like that.

Publicado por arethusa hace casi 3 años

Awesome!! Sounds like you've put a lot of thought into this! It's impressive that you converged on the idea of morphospace without having heard of it beforehand. I haven't used it much, but I think you're on the right track to pursue a ML approach to species ID, especially in clades like Platanthera with very specific (and small) diagnostic features e.g. lobe position! Looking forward to seeing what you come up with in the future!

Publicado por mannfred hace casi 3 años

Yes, I think just being able to drag and drop an image would change everything to the extent that it's possible. It's harder to approach genera like Spiranthes because identification involves noting a combination of characteristics including the general shape of the flowers themselves. Perhaps training TensorFlow with silhouettes of the flowers and cropped images of the lateral sepals or lip would work, although it's hard to say how well that'd actually go with everyday photos.

The key I mentioned which generates numerical sequences wasn't dichotomous at all, and prompted the user with a series of about a dozen specific questions. If a person didn't know the answer or wasn't certain, they could just skip it for a slight decrease in accuracy. With the code/sequence which the key generates, a person can assess how closely their specimen matches each species. A new species (assuming its not cryptic) or a hybrid likely wouldn't match well with anything in most cases. The problem with this model is that interpreting the codes can be a bit challenging (it's just a sequence of numbers and/or letters), and that constructing something like this for a species complex takes a ton of data. Probably not practical, but certainly fun to think about.

Publicado por arethusa hace casi 3 años

Agregar un comentario

Acceder o Crear una cuenta para agregar comentarios.