Geomodel/Computer Vision behaving incorrectly after a taxon split

[matthew_connors]

Platform: Website

Browser, if a website issue: Chrome

URLs (aka web addresses) of any relevant observations or pages: https://www.inaturalist.org/observations/upload

Description of problem:

I’ve just come across some strange behaviour from the CV suggestions after a taxon split. Recently, the widespread and common frog Litoria rubella was split up into three separate species - L. rubella, L. pyrina, and L. larisonans. As expected, L. rubella has been included in the CV and geomodel training for quite a while so it obviously comes up as an option when presented with a similar frog. After this taxon split, the CV now suggests all three species as options when uploading:

This doesn’t seem like it should happen - neither L. pyrina nor L. larisonans have been assessed by the CV, even if it has used images of them for training. It’s not ideal to have only L. rubella show up in the CV until the next training round happens so I can understand the reasoning behind why all three may show up, so perhaps this part is not a bug.

This next bit is surely a bug though - when I go to the taxon page for, for example, L. larisonans, under the ‘About’ tab it claims L. larisonans is included in the CV and geomodel. However, clicking on the link that usually takes you to the geomodel leads to a 404:

This current behaviour from both sections seems odd and not correct. I’m not entirely certain what a good solution would be though. The best option that I can suggest is that if the CV has been trained on a taxon that has since been split, instead of showing that taxon (and/or the other things it has been split into), perhaps just show the highest taxon that includes all of the output species? Keen to hear what others have to say though.

(and yes, I’m aware that my frog photo is none of those three species)

[jamie-aa]

Urtica dioica (European stinging nettle) and Urtica gracilis (American nettle) is another recent split which is acting weird too - the American taxa will be readily suggested in Europe as “expected nearby” despite not being “Expected nearby” in the geomodel. I cannot find any literature that suggests it has been found in Europe.

Here’s U. gracilis’s geomodel, expected only in North America and the tip of Greenland: https://www.inaturalist.org/geo_model/141854/explain

And here’s a U. dioica observation in Great Britain, where CV suggests the American taxa first: https://www.inaturalist.org/observations/298446969 but it’ll do this in a lot of Europe:

[DianaStuder]

With this release, we’ve updated our geomodel using a new training approach called SINR (spatially implicit neural representation).

https://www.inaturalist.org/blog/113882 trained off data exported on April 20, 2025

The next model release, 2.23, I don’t think will be substantially different from 2.22. I expect it to be much closer to 2.22 than 2.22 was to 2.21 due to its new training approach. We did say in the 2.22 release blog post that “we’ll continue to refine how the geo model supports the computer vision system”. If we expect there to be more substantial geomodel changes in future releases, we’ll likely mention that in the release blog post. - pleary

https://www.inaturalist.org/blog/99727-using-the-geomodel-to-highlight-unusual-observations

I am still working thru the huge batch of Geomodel Anomalies for Cape Peninsula. 95% Endemics - and a very few that have been worth catching.

[cthawley]

The first part at least is intended behavior, see this comment on a previous bug report for this behavior:

@pleary

Yes, this is quite similar to that reported issue. In this case though the behavior reported here was as designed. The taxon mentioned in the initial post, Unio crassus, was involved in a taxon split after the data for the currently deploy CV models were exported. As a result of the split, observations that used to be considered Unio crassus, may be considered one of several separate Unio species. As far as the currently released models are concerned though, all those observations are still considered to be the same species, and the vision model was trained to use all of their photos to learn the visual characteristics of Unio crassus.

I am not sure about the geomodel aspect though - seems like anything that trips 404 is probably a bug.

[sbrobeson]

this seems to be standard now — similar things have happened in other taxa (e.g. the split of the Erysiphe cruciferarum complex). all I would add is that this happens even when the output taxa wouldn’t normally be included in the computer vision suggestion list except under special circumstances, such as those of the rank of section, complex, etc.