Sometimes photographic evidence is insuffient principally

In general, ID’s are based on photos in iNat. But there are genera (or other
orders above species level) in which photographic evidence is insufficient.
Still iNat sometimes suggests species names without mentioning the above
problem. See e.g. https://www.inaturalist.org/observations/65040043
Here the species had been suggested by iNat and I had accepted it, being a
layman. In this respect iNat could help educating people and avoiding false accuracy.

1 Like

What you have selected is a “suggested species” and not the one marked as “We are pretty sure this is…”. Suggested species should be investigated. If you do not investigate them, then you should select the “pretty sure” one. And sometimes it’s wrong too.

In this case, you have been educated by the expert who gave the genus ID so iNat does educate.

“without mentioning the above problem”

There’re many pitfalls in identification and it’s not practical to list them on every ID suggestion. The suggestions are described correctly as “pretty sure” and other suggestions.

9 Likes

There’s no false accuracy, it’s a machine that is stupid, as it can’t analyze much more than some parameters, you as human is who decided to follow the lead and agree, it’s a normal human mistake and it’s not on iNat side as it’s stated everywhere that those are suggestions, they’re often wrong, you have to id only taxons you know and can id. I guess staff answered to very similar posts before you may want to check them!.)

7 Likes

Thank you for posting this example. exonie gave a great reply but let me add a little more.

The leading ID (usually made by the observer) is a critical ID. Fortunately in your case, others became aware of your ID (perhaps because they subscribe to genus Sarcophaga) and were able to correct the problem relatively early on. Other genera are not so closely monitored, I’m afraid.

A basic rule of thumb for observers is: don’t lead with a species-level ID unless you are “pretty sure” that’s what it is. Think of it like this: if you lead with a species-level ID, you are basically saying “I’m pretty sure this ID is correct, so much so that a second ID is not really needed in this case.” If you are unable to make that claim, then lead with a higher-level ID instead.

I agree with you, but it’s not easy for the system to know when the observer is over-specifying the ID. If it knew that, it might avoid making unlikely suggestions in the first place.

5 Likes

Keep in mind that it is the users that determine how iNat provides ID suggestions. People provide ID’s for observations, and the research grade ones are taken into account. The computer vision itself does not know that some taxa are unidentifiable to a certain taxon level through photographs alone, that error comes from the fact that people who are not experts provides an ID more specific than what is possible, and that causes a chain effect in which others will follow suit.

I suggest taking a peek at these threads:
https://forum.inaturalist.org/t/dont-use-computer-vision/13327
https://forum.inaturalist.org/t/false-research-grade-observations/14193

2 Likes

This is probably a thing for curators, but is it possible to add to the taxon page or in the ID suggestion drop down list something like “this taxon cannot be identified to species level through photographs in the wild alone”? That seems like a lot of work, but it could tidy things up a bit?

1 Like

To me, pretty sure sounds determined. The wording needs to better reflect this could be Genus species. Could be.

1 Like

Something weird seems to be going on with this species. Looking through a few of the Research Grade records for Sarcophaga carnaria, I see several where people have added disagreeing IDs at the genus level or higher, yet the disagreements are not being considered when calculating the community taxon. As a result these records are all stuck at Research Grade, and presumably this is why the AI keeps suggesting the species.

Here are a few examples, there are probably many more, I just checked a handful:
https://www.inaturalist.org/observations/27994782
https://www.inaturalist.org/observations/32157101
https://www.inaturalist.org/observations/12142554

1 Like

The taxon appended to the phrase “We are pretty sure this is…” is thought to be accurate with around 95% probability. See: https://forum.inaturalist.org/t/identification-quality-on-inaturalist/7507

If you are saying that only Research Grade observations feed back into the computer vision model, that is not true. The photos of an observation with a single ID are eligible to be included in the training set provided the observation meets some basic data quality standards. All else being equal, the photos of an observation with a single ID have as much chance of being in the training set as the photos of a Research Grade observation.

See: https://www.inaturalist.org/blog/31806-a-new-vision-model

That’s why the leading ID (usually made by the observer) is so important. Most of the non-RG observations out there are actively contributing to the computer vision model. Even not-wild observations contribute to the model.

There was a feature request and discussion this summer about indicating whether a suggested taxon is “visually identifiable”. It would definitely be valuable if it could be implemented, but the prospect is more complicated than it initially seems at first glance.

1 Like

They do change community taxon for me.

While 95% might be accurate for plants and animals it certainly isn’t accurate regarding fungi. For instance the Russula genus contains a few dozen North American red capped species that are macroscopically identical and can only be properly identified through microscopy or gene sequencing however iNat will still throw a suggestion at species level even though they cannot be identified by sight. As a result there are tons of inaccurate RG observations in the genus Russula.

1 Like

iNat never suggests a species in the “pretty sure” claim, at most a genus level. It’s the users’ fault that they select uncritically some of the other suggestions. I’ve posted and used the computer vision on many Russulas and the computer vision does very well suggesting the genus Russula as “pretty sure”. Note that whatever a user selects from the suggestions, not just the “pretty sure” one - the identification will still be marked with the computer vision icon.

No matter how you word it, many people will still be selecting things that resemble their observation thinking “oh, that’s it”. The only way to stop that would be to remove the suggestions altogether.

2 Likes

That’s a fair point after back checking a few recent RG observations, the AI indeed only makes a pretty sure claim at the genus level for Russula as far as I can see. Disregard my previous statement.

2 Likes

Ah right I see that now. Upon initial reading that seemed to be a counter-intuitive method, but I guess statistically for a very large sample that method may seem to provide more reliable suggestions (and the graphs seem to corroborate that). Regardless of the method, a big part of the computer model working reliably depends on the ID’s us users provides, so yes I agree.

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.