Train the vision model on cultivated plants

Perhaps the most important improvement from the perspective of encouraging teachers to use Seek with students (who will likely use iNaturalist for only a single term) would be to train the database on captive/cultivated plants. I know that almost all of my students photographed only captive/cultivated plants. I then helped with identification. I could not use Seek because Seek could not identify the captive/cultivated plants that inhabit campus. A Seek trained on captives would help instructors like me use Seek with my students. Perhaps some ability to display suggestions instead of just “cannot identify” which is what I encountered here. And a vote from here on geographically informed suggestions so my students do not identify a plant here on Pohnpei, Micronesia as one found only in northern Texas and southern Oklahoma (that was the top suggestion).

1 Like

Do you have some example species? How many observations of those species are on iNat?

https://www.inaturalist.org/observations/22401532

This is a commonly cultivated/captive plant here on island and a prime target on campus. Seek cannot identify, but I can guide my students in iNaturalist to the correct identification. Thus as a teaching tool, Seek has limits for a teacher when it cannot ID a plant. I think this connects to iNaturalist being aimed at wild organisms and being “misused” by those such as myself as a teaching tool with students, students who tend to photograph many captive/cultivated plants and who use the app for only one term. Still, iNaturalist trumps other free apps such as PlantNet which lack all capability to be used by an instructor (http://danaleeling.blogspot.com/2019/04/inaturalist-and-plantnet-reporting-back.html). As I read through the forum I see many posts that suggest a “tension” between iNaturalist being a “big tent” for all users including those who photograph captive/cultivated organisms or an app focused more exclusively on wild organisms. I gather the founding mission was wild organism, but the userbase now includes a lot of users photographing/identifying captive/cultivated plants.

Screenshot_20190413-173111|250x500

1 Like

The Seek screenshot did not upload with that post:

1 Like

Thanks. I’ve split this off into its own topic.

1 Like

Another example (at least on iNat, not sure about Seek) is cultivate pines. For instance… in Southern California coastal areas, the most common pines are Alleppo, Canary Island, and Monterey, all planted. Other than the small patch of Torrey Pines near San Diego there are no native pines in the urban areas of southern California. However the algorithm continually identifies things as montane species such as Ponderosa and Jeffrey pine. Unfortunately the proximity thing won’t help either without elevation considered because those species do occur within 100 miles,but it’s just that they are only above ~4000 feet. Though the problem is somewhat perplexing because sometime the algorithm does ID them properly and the correct id is rejected for another wrong ID further down the list. So maybe it’s user error.

Other than that, overall i find the algorithm works pretty well for cultivated plants, at least on the US east coast.

2 Likes

This post is just another example where a cultivated/captive plant can provide images of benefit to training the image model to the benefit of wild plant identifications and is provided fully aware of:

This post is also related to the discussion from What Image(s) Are Used for "Training" Computer Vision?:

I was looking through Ixora casei Research Grade and could only confirm that one of the eight is likely to be Ixora casei. Based on distribution and appearance, a couple others could potentially be Ixora javanica. The remaining observations seem more likely than not to be other Ixora. Ixora casei is a native in the islands of Micronesia, Marshalls, and Kiribati. The bulk of the images being used for training the computer vision seem to be likely to mistrain the computer vision.

I am also certain that this Ixora here on Pohnpei is Ixora casei and will be a valuable addition to proper training of the computer vision. The plant is cultivated/captive because I moved the plant from the forest nearby to a more accessible location on campus as a field specimen for my ethnobotany course over a decade ago.

I must confess that the images are of the same plant but span time. I visit the plant two or three times each academic term as part of an ethnobotany class. I could separate the images into observations for each time, but these would all be casual and would all be the exact same plant.

I would note that the lead main image on the Ixora casei taxon page is not likely to be Ixora casei. There is a beautiful type specimen for Ixora casei from Kosrae island dated September 18, 1884. Fun trivia: the location listed in the record was miscopied from the herbarium voucher. The location “Valan” should be “Ualan” which was also known as Strong’s island and is today known as Kosrae. Ualan was possibly a misunderstanding of the Kosraean word walung which refers to specific land sections at the far end of a municipality away from Lelu. The compound of the king of Kosrae was located on Lelu island which was also the harbor into which foreigners usually arrived. Pointing to the main island and asking what that place is called may have generated a response along the lines of “walunge” - away from Lelu.

Just a note if you were unaware, all users can change the photos associated with a species. Just go to its page, Click the curation dropdown and select edit photos. Remove any you wish and drag over any possible replacement remembering to save

1 Like