Description of problem: (At least) Sonoma County observations of Sceloporus occidentalis uploaded and provided to Computer Vision are getting “Island Fence Lizard” S. becki as the first choice for an ID offered by CV. This has happened for my own Sonoma County observations of Sceloporus occidentalis, but also for others. In fact I provided you a URL to someone else who too that suggestion and offered it as the ID, despite the fact that we are hundreds of miles away from the Channel Islands. The screen shot I offer is from my own observation getting that same buggy suggestion from CV. My iNat handle is shxx60.
I wonder if there is an issue with the geomodel for these taxa?
Here’s the geomodel page for Sceloporus occidentalis. That does seem to cover the known distribution, even if it’s a little generous in including substantial portions of the Channel Islands, Arizona and Washington state where there are no observations.
And this should be the geomodel page for Sceloporus becki. Unfortunately, that give a 404 error. I might speculate that CV algorithm is handling the absence of geomodel data for S. becki by assuming the species is likely in any location, thereby boosting the probability of S. becki being suggested for geographically implausible locations.
The issue is that there was a taxon change that happened after the current model was trained: https://www.inaturalist.org/taxon_changes?taxon_id=540190 Island fence lizard was previously a subspecies of western fence lizard but now it’s a full species. However, the model doesn’t know that and I think it gets a bit confused.
I’ve been subscribed to Sceloporus in California for the past few weeks so I can correct these ASAP.
Thanks for identifying the cause @tiwane. It looks like the next CV model should be released in mid-May and will be based on data extracted after Sceloporus becki was raised to a full species, so we should see better suggestions for California lizards after that point.
I guess this raises a more general point about the interaction of taxonomy changes with CV models built with a previous taxonomy. It would seem that it should be possible to prevent CV from suggesting taxa it was not trained on (as appears to be the case here). Or am I missing some subtlety?
My aplogies, the more relevant taxon change is https://www.inaturalist.org/taxon_changes/170688 In a split, we map the model to both the outputs. In cases like this it’s possible the original in-model species was the more restricted one. We split the difference and map them to both, is my understanding. As to the technical details and reasoning, I don’t know for sure.
A few more follow-up after I got some more info. The current model was trained on photos that were all labeled S. occidentalis. Depending on the split, this could mess up suggestions if we just kept to the one species. So we split the suggestions evenly, and there isn’t a way to tweak that depending on the split, at least not at the moment. This case seems pretty clear cut because the populations are so clearly geopgrahically separate, but in many splits that might not be the case.
iphone