When the original iNat app was first launched the app was unable to rule out suggested IDs by location, as a result there are a number of mass mis-IDs we are still dealing with. Examples are Rhododendron ponticum, not a species found in the wild or regularly in cultivation in North America, but the app suggested it and less knowledgeable users added it to observations which were seconded and now North America is full of incorrect R. ponticum observations, and because of all the incorrect research level observations the iNat AI keeps suggesting R. ponticum as a likely ID resulting in even more incorrect observations. With the origianl app this was eventually (somewhat) fixed as the AI was programmed to only select from likely IDs by region. Unfortunately it seems like this beneficial feature was forgotten in the iNat next release result in new mis-IDs. For some reason iNat next keeps suggesting Parthenocissus dalzielii as possible IDs for North and Central American plants so now that species is being reported in North and Central America even though it is not found there. Please can we fix the iNat Next app so that it stops doing this and is as effective as the original app at not offering up IDS that are so unlikely as to be impossible given geographic location? Thank you.
are you using iNat Next on iPhone? Did you have any screenshots you could share of this? When I open the app and try to add an ID, I can definitely choose to limit IDs based on location/expected nearby
Suggestions on my phone for using expected nearby:
and then ignoring location:
also could you clarify why you believe this is an iNat Next-specific issue, re:
I just had a look at the most recent 30 records from North/Central America where someone IDed them as Parthenocissus dalzielii (that you have since corrected), which dated back to May 2024, and only three out of 30 of them were made via iNat Next
I use iNat Next on the iPhone, and I’ve noticed that sometimes I don’t have the “Use Location”/”Ignore Location” option, and sometimes that it defaults to species that aren’t local at all. I’m not sure if this is a new issue with the app? I feel like this has only been happening recently.
This is a good point, I just assumed this was an iNat next issue as I have experienced iNat Next giving extremely unlikely/impossible IDs several times since it’s inception the way the original app used too. Here’s one from yesterday, I’ve seen bad ID’s with way higher percentages than this on iNat Next thus far however this is the most recent one I could think of. The image in question is from New York, NY yesterday and is clearly of an Eastern Grey Squirrel, while iNat Next suggested that as the most likely ID, in the list of possible IDs is also a Rail species endemic to New Zealand which shouldn’t even be close to showing up as a possible ID for Manhattan, USA.
I’ve felt like I’ve needed to crop photos more heavily to get reasonable CV suggestions than I used to, but I’m not sure if that’s a more recent CV model thing or an iNat Next thing since I haven’t systematically tested the same photos on iNat Next vs. the iNat Classic app. It might be that if you crop the photo down, it will “recognize” that this is clearly a squirrel, although I also don’t understand why the NZ endemic is even an option
Looking at the squirrel observation in question, S. carolinensis is the #1 suggestion (using the standard CV model via the browser client). I do think the large amount of asphalt and deep shadows are the reason for the disparate suggestions on this one.
Gallirallus australis is suggestion #5 (and notably is not indicated as expected nearby). If you browse through the photos from G. australis observations you’ll find quite a few with asphalt, gravel and beach shingle in the background, so that might be the reason.
For this example, I really do think that cropping the image would make a big difference. But even @edropkin’s screenshot shows that iNat Next only rated the Weka as 1.2% similar. I’m also not convinced that this has much connection with mis-IDed Parthenocissus dalzielii observations.
Users should always be cropping photos and adjusting brightness/shadows. It’s the most basic form of photo editing and can be done on the phone or on the computer.
A little-known fact to users is that iNaturalist resizes the photos to 2048px on the long edge. On a 12MP photo (3024x4032), that 300px squirrel it’s halved to 150px if the image is uploaded as-is. The closer the photo is to 2048px before submitting the observation, the less information is thrown away because the image is resized less or not at all. In other words, the squirrel will remain at 300px and that has more information for the CV to return a better guess.
20/20, the iNaturalist apps’ image editors should have a “Crop to Original Size” option in the Crop Tool.
Anyway, two things should be drilled into users’ heads with the help of the app and identifiers:
- Get closer to the subject when it’s possible and safe to do (e.g. snakes)
- ALWAYS CROP the photo to the subject when uploading.
- Adjust brightness/contrast/shadows/highlights to get some details.
My Soapbox Tax: this is one of mine and below are the source images (resized from 16MP to save iNat some $$$): https://www.inaturalist.org/observations/312298363
Yes, I’m aware of this and crop my images. But my workflow for phone photos is usually to take the initial photos in the app so I can have an observation to make field notes in and have the uncropped photos for context, save the observation, and then crop the photos and add them when I’m not in the field. And like I said, my recent impression is that uncropped photos that would have previously been recognized as “Lepidoptera” are now getting suggestions for “Insecta” or “Arthropoda”.
Your comment about photo pixels reminds me that I also got a new phone between last season and this one and it has a fancier camera. So it may also be that my photos are now starting off with a lot more pixels than they used to.
the iNaturalist apps’ image editors should have a “Crop to Original Size” option in the Crop Tool.
As far as I’m aware, the iNat Next and Classic iPhone apps don’t have any image editors, let alone a Crop Tool. At least in iNat Next, all I see are the options to add more photos or delete the existing ones. Is this something that the Android app has??
The “asphalt issue” occasionally creaps into the ID misidentifications of many urban/suburban species. In Texas, we’ve seen this arise in photos of squirrels, white-tailed deer, distant coyotes, and even some plants like Ashe Juniper. CV is “learning” that a lot of pavement, a curb, or even parked cars should be part of the criteria for IDing such species, even when the subject is something else entirely. When the subject animal is small in relation to the non-biological components in a picture, CV displays it’s non-human failings!
7 posts were split to a new topic: How to retain location metadata when cropping/editing photos
Yep. When the photo is mostly background, the CV is mostly looking at the background.
Just another of the many issues with CV that have been brought up on this forum.
We can say whatever we want about what users should be doing, but most users will never look at the forum, so it is preaching to the clergy.
I rarely crop photos as well, although I definitely try to get as close as possible to the subjects. But I don’t photograph birds; I could see how cropping would be useful for many bird photos. For the plants and inverts I most commonly photograph, bending down or holding a branch that’s waving in the wind is all that’s needed. At any rate, I don’t remember anyone suggesting I ought to crop my photos, and I’ve added over 67,000 observations to iNat.
My cropping practice is if I’m not sure of an ID, I crop each photo down to various “closer” versions and test the CV on each one of them. When I go to upload, I’ll switch which cropped photo is at the front of the observation, ask the CV what it thinks that one is, and then repeat. If it gives me the same result at all the “closeness” levels, I’ll go with the CV ID. If it repeatedly changes its mind when I adjust how close to the organism I am, I lose all confidence and just use a generic high-level ID. Here’s an example from last week. I’m bad at ferns, but CV called all 6 of the photos Thelypteris palustris, whether I was cropped to a partial leaf, a whole leaf, or a cluster of several leaves, all from slightly different angles, so I went with that ID:
https://www.inaturalist.org/observations/310673875
If CV had suggested a different species for any one of the photos, I’d have just called it “ferns”. I still sometimes take a CV ID that turns out to be wrong, but I’ve avoided a lot of would-be mistakes by noticing that my “white turtlehead” turned into a “purple loosestrife” when I took one step closer to it.
Lynn, I just checked out a sample of your contributions and they are beautiful! Your macro images of moths and other insects are just fine–even exemplary.
The magnitude of the issue of small subjects in wide backgrounds–particularly from phone camera images–is pervasive on iNat. As others might remember, I used to be pretty vocal about suggesting to OPs that cropping would be useful. But it takes up too much time and energy now. At present, I confine such comments to such occasions when someone gets an erroneous ID from CV (wrong genus or family) for the handful of genera I’m subscribed to. If the subject in a picture is so small that CV places it in the wrong genus or family (especially when it is otherwise quite recognizable), that’s when I offer a comment.
Thank you! One of the reasons I stopped taking photos with my phone and swiched to a camera is that I couldn’t get good enough photos with my phone. Maybe I just didn’t know how, but I also found my phone harder to handle, compared to a camera (an Olympus Tough, which is pretty small).
Human failing needs to be prompted - picture of pavement / sidewalk.
WHAT are we looking at? (Caps are mine)
Also plants. But I always crop, with a wide view for growth habit / scale / context. But flower, leaf, field mark - I crop.
This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.