If it is used responsibly, then yes, it can be a good tool. But relying on it solely, without then checking range maps or some other source to determine the veracity of the ID, can make it problematic.
Just today there were two observations for plants in South Carolina, California Bay (Umbellularia californica), https://www.inaturalist.org/taxa/48807 and Ponderosa pine (Pinus ponderosa), https://www.inaturalist.org/taxa/48461. A quick look at the range map in iNaturalist would indicate it unlikely that either of these IDs were correct. And this was with what appears to be a class project; if another one of their classmates then verifies this this ID, you’ve got a “Research Grade” identification. Then others think it’s a verified species occurring in this location so a misidentification is perpetuated. If it didn’t go to species level, then it wouldn’t get “Research Grade,” solving issues like this.
For marine and non-marine mollusks worldwide (outside of the best-covered, best known areas such as New Zealand, the US etc), it would be good if the AI could offer only family-level IDs.
Even genus would be too narrow – currently for the lesser-observed parts of the world (which are vast and numerous) the species IDs from the AI are almost all coming out wrong and misleading.
I think it is a mistake that newbies make to think that every organism can, and should, easily be identified to the species level.
Since the iNat developers are actively looking at improvements to computer vision’s confidence with species rank IDs, and species rank IDs won’t be wholly removed from the suggestions, I moved this topic from Feature Requests over to General.
It might be helpful if there was a simple user selection when accepting a CV recommendation:
A. This looks good to me, I am accepting the suggestion but have no additional knowledge of the organism.
B. Based on my knowledge and experience I believe that CV is making the correct identification.
Even if this did not have any effect on determining research grade it would help to find out how people are using CV and maybe enable a flag for IDers to take a look at CV identifications.
I see no reason that the CV software can’t remember its species determination and compare it to the community ID, it does not have to tell us what it ‘thought’.
I like this approach. Also that of a little warning/info-Box “are you sure?” when choosing from the drop down menue, with short information why false observation ids are problematic, and the suggestions of the CV should be taken with care. Maybe such an info box could be shown only for, say, the first 15 uploads, like a tutorial function?
@susanhewitt. I find your comment interesting. The suggestion that some organisms can’t be easily identified to species is a good one. However, I think that there are some aspects of CV that will take this into account automatically:
Those species that can’t be accurately ID’d won’t be. There will be outliers, but for any of those organisms, they should be ID’d to genus (or higher) anyway. Because of this, they won’t be included in the prediction model as species. Presumably, any species-level IDs for these organisms (excluding misidentifications) will have been “professionally” identified, by a science organisation or similar.
When you imply that some organisms can’t or shouldn’t be identified to species-level, what you’re actually talking about is human identification. When working with these systems, I’m occasionally surprised by how consistently some predictions are correctly made between certain objects where I, as a human, can’t reliably tell the difference. This is one of the hardest things to adjust to when using AI technologies, that sometimes the system works better than you do. But I digress. The point is that we may find that with high-quality images, the automatic prediction may find something that we don’t see or look for, and make a correct ID to species (whether we trust it or not is a different story!).
I’d like to see more suggestions at genus or family level or even higher, but there are cases where species-level ID is to be preferred. For example, in the case of a family or genus that has only one species, it would be more useful to have the species suggested.
Osprey, for example, is the only species in its genus and family. So it doesn’t make much sense to add genus or family ID in that case.
There are also a few taxa (but many observations!) where I’d like to see SUB-species suggested. These would include feral forms of species which are very commonly seen, while their wild ancestors are not. Two examples from the recent Penang bioblitz are feral pigeon and domestic chicken. But that might be asking too much!
Yes, the challenge would be to “teach” the CV system about that. I suppose it could check to see if a taxon in iNaturalist has been marked as “containing all of its descendents”, as is the case with birds at the moment. But otherwise, it won’t be able to tell whether some taxa might just happen to be missing from the iNat taxonomic structure. Some groups are as yet far from complete.
Re: making sure any changes (if any are made) don’t slow down IDers:
Maybe there can be an ‘expert’ mode (needs a better name) that can be turned on/off that can bypass certain checkpoints meant to make others who may be still learning how iNat or the taxon system works. It’d be entirely user-side only, since it’s just a mode that’s there for ease of use.
