iForum LIVE!: AI and iSpot: a spotlight on the PlantNet API

Do you have many in iSpot that could be used to train the AI ?

Another question from before:
Example of AI getting it wrong?: https://www.ispotnature.org/communities/uk-and-ireland/view/observation/849378/hornbeam-two-for-price-of-one#new

Trunks are the most difficult type of view for identification. It would have been necessary to take a picture of the leaf. The trunk should be used only if there is really no other choice.

2 Likes

The visual aspect of the trunk depends on many other factors than the species (age, condition, etc.)

1 Like

ANother older question: ā€œWill Experts stop interacting because they cannot discuss anything with AI?ā€

That one is sort of understandable - itā€™s not a typical hornbeam (a typical hornbeam in Britain, discounting hedges, is a youngish ā€˜Fastigataā€™ planted as an amenity tree), and the AI might have seized on the split trunk as a prominent feature on which to base an identification.
Ā 
The web-based interface only gives a percentage of 20.83% for the top choice, and the others are in the low single figures.

1 Like

It is the experts who train the Pl@ntNet AI through the photos they share. It is fundamental that the experts continue to interact and maintain the knowledge. The pictures that the AI sees are exactly the ones available in the photo galleries of Pl@ntNet. So that experts can ā€œdiscussā€ with the AI by exploring and revising the images.

How frequently is the AI updated?

Approximately every month currently. We try to keep a good balance between frequency and consumption (carbon impactā€¦)

I would think not. Pl@ntNetā€™s stats have an average between 500 and 1000 images per species, though I expect that there is a wide variation in the actual numbers. I think iSpotā€™s number run lower, and are lower still for algae and non-vascular plants.
Ā 
If Pl@ntNet was to work to multiple taxonomic ranks, there might be enough for it to be able to identify them as non-vascular plants, and refrain from trying to identify them as vascular plants.

1 Like

Yes, or maybe at family level ?

The number of images needed to train an AI presumably differs depending on the group of plants - for critical groups like Taraxacum or Rubus or Hieracium Iā€™d guess that you need lots of images, while for groups with few well separated species youā€™d get away with fewer.
Ā 
Do you have any rules of thumb on how many images are needed for training?

1 Like

The number of images really depends on how difficult it is to distinguish the species from others. A species with very remarkable morphological features can be very well identified with a few images. But for species belonging to very difficult/ambiguous genera and having high intra-specific variability, several hundreds or thousands will be needed.

1 Like

But the rule of thumb is 100 pictures :wink:

It is interesting having Alexis and others online in a video conference at the same time as this forum is going on so sometimes the replies are much more extensive there but can only type a few words, in this case he explained about tails of distributions and other more complexity

On a broader view, what are Pl@ntNet and iSpot trying to achieve with this integration of the two systems?
Ā 
A radically different way of integrating them would be to have an ā€œask Pl@ntNetā€ button in the Add Observation process, rather than adding some suggestions from Pl@ntNet after the event. Pl@ntNetā€™s API, as used by iSpot, is rather erratic; sometimes getting fairly difficult plants (e.g. a Cotoneaster divaricatus - recent confirmed by Dirk Derdeyn) correct, and sometimes getting relatively simple plants badly wrong (I blame that on iSpot not providing locational data to the API). But Iā€™d be hard-pressed to find a case where the Pl@ntNet suggestions have contributed to reaching an identification.

If you have examples of erratic results when using Pl@ntNet in iSpot, we could check that there is no technical issue (e.g. related to the way the images are submitted to the API compared to Pl@ntNet classical photos).

lavateraguy you could put examples in forum or send direct to me via inbox

We would like to organise a Zoom session for further discussion related to AI and iSpot. Would you be interested in joining us?
  • Yes
  • No
  • Maybe

0 voters

This is neither category, but for an unidentifiable shrub Pl@ntNet offered 3 forest cacti. But it did cause me to start considering whether there was a technical issue. (I abandoned a comment on the forum to this effect pending some experimental investigation, which has now been performed - see below.)
Ā 

Ā 
If one goes to the web interface one gets the same (so not technical issue) but finds that all three are in the low single figure percentages. (As I said at the start of the chat passing back low probability results without any indication that is what they are makes the AI look bad.)
Ā 
The AI has failed to identify Lotus corniculatus correctly several times, preferring Lotus ucrainicus and other Lotus species. Thatā€™s understandable to a degree - Lotus ucrainicus is very like Lotus corniculatus, but itā€™s not found in Britain. Itā€™s also failed on Urtica dioica a few times, preferring other Urtica species, and other nettle-like plants. I donā€™t recall specific examples, but Iā€™m fairly sure that thereā€™s been cases where itā€™s picked up 3 non-British plants rather than the correct ID.