Locally Common but Rare on iNaturalist
The new algorithm identifying species is amazing with weeds. It seems to get the correct species in the top 3 almost all the time and usually picks it as the top species.
For arthropods, no such luck. I just had it try to identify a springtail. First option was an Oleander aphid, second a fruit fly, third a velvet mite, fourth a psyllid, and none of the suggested top ten was a springtail. This is a bit of an extreme example, it correctly identified a spider I found today, but still the system clearly needs some help.
The problem isn't the algorithm, or at least isn't only the algorithm. A quick search for springtails shows that only four species have more than twenty observations. If I limit myself to research grade submissions not a single one has the twenty observations it takes to be looked for by the algorithm. So springtails are completely invisible to it.
The problem isn't limited to springtails, 23 species of leafhopper, 18 species of isopod, 8 species of aphid, 7 species of psyllid and one species of thrip make the cut. So with taxa like this the algorithm could be dramatically better if a few species were put over the 20 submissions mark.
So I have decided to start resubmitting more often for species which are locally common but do not make the cut for the algorithm. Hopefully I can get someone to hit the agree button enough to make it more useful.
I am also going to look for species with 20+ Total observations, but less than 20 research grade observations. Adding identifications to those species could make a big difference.
Comentarios
This is a great approach, one that I think I may try to adopt as well. I wonder, for some of these very tiny insects, just how many observations it would take to properly train the algorithm to recognize certain patterns? For example, I tested it out with an image of Calophya minuta (essentially the Arizona equivalent of Calophya californica, which has 27 research grade observations). What I would expect is for the algorithm to choose C. californica (as it obviously wouldn't know what C. minuta is, and the two species are superficially very similar), but instead it picks the Red Gum Lerp Psyllid. While it's a positive that it at least picked a psyllid, it's telling that it picked the psyllid species that has the most total observations, despite it not really looking anything like the psyllid I submitted. The algorithm's second choice was a cricket, and then a fly.
Adding observations and identifications to some of these smaller insects could really help train the algorithm to at least get in the right ball park, I would hope.
When it has sufficient photos it really does quite well. I just tested and it got a Morril's lacebug, diamondback moth and redgum psyllid correct. It thinks a greenhouse thrip is a mollusk though so that one clearly needs additional photos.
In the case of aphids the 8 species it looks for are very atypical. They all produce odd galls, or are brightly colored. Only the rose aphids are typical aphids. Since they are locally quite common I am going to try and add enough Macrosiphum euphorbiae, Aphis spiraecola, Sarucallis kahawaluokalani, Acyrthosiphon lactucae, Therioaphis trifolii, Wahlgreniella nervata, Neophyllaphis varicolor, and maybe Uroleucon sonchi to hit the 20 research grade observations. With those the algorithm should at least be able to flag most aphids as aphids.
Turns out it is relatively easy to find species which need just a couple IDs to surpass the 20 research grade cutoff. Search a taxa large enough to have a few hundred species, then filter research grade submissions. Go to the species tab and scroll down until you find observations with 10-19 research grade observations. Then look for species with enough unconfirmed observations to go over 20.
Agregar un comentario