I’ve only given ArXiv 2502.17289 the most cursory scan, but I think that
- it is attempting to identify the taxonomic placement of unknown plants
- it is a proof of concept; the training data set is too sparse for general use
Looks quite a complex paper but this kind of thing is very complex from the computing and data side.
One of the issues with this kind of thing is that ecology is also complex, convergent evolution and similar may limit the possibilities of the outcomes they are attempting but no reason not to try.
You may have noticed yesterday a big report saying 88% of higher education students now use AI, a huge jump from last year. World moving way faster than educators/regulators/governments. Are these students the first of the AI generation?
The penultimate sentence of the abstract :
For unknown species the model achieved an average accuracy of 83.36%, 78.30%, 60.34% and 43.32% for predicting correct phylum, class, order and family respectively.
.
This is not a particularly spectacular result; but I’ll read on in case I’ve been too hasty in my judgement.
Is it a form of AI that I will let you read this Paper, Jo, and then hope to glean something from your feedback - or is that just old intelligence - or merely Idle Thoughts of an Idle Fellow ?
I’ve looked a smidge more carefully. Look at figures 1 and 2. The assignment to phylum and class is all over the place; in fact in the domain of interest all the plants belong to the same phylum, and the modern APG classification doesn’t use the rank of class. The older Cronquist classes are sometimes polyphyletic (Dillenidae especially). So it’s not surprising that they don’t do a good job on identifying phylum and class. One of their test cases belong to an order and family not in the training set; a failure to identify the order and family correctly is to be expected.
The authors are from a computer science department; they don’t seem to have realised the desirability of consulting a domain expert.
Indeed that is what I was implying. These days it can be very difficult to find suitable overlap as in STEM subjects everyone has to specialise so narrowly and even when you have multiple authors from different areas it can be tricky for each to understand how their parts fit with the other parts of the paper. Even having just a taxonomist may not have been enough as an evolutionary ecologist probably also needed.
I’ve been reading around….I was initially surprised at the number of articles published in the field of Deep Learning systems for medicinal plant identification, but given WHO suggests that, for millions of people around the world, traditional medicine is the first step in the journey towards health care, maybe it isn’t so surprising.
The problem:
Traditional identification ( i.e. finding a person whose identification skills can be trusted) is not easy especially in some parts of the world.
Towards a solution:
Real-time plant species detection plays an important role in fields ranging from medicine to biodiversity conservation. So, deep learning models are being trialled.
The choice & quality of database available to learn from seems to me rather important. Classical botany traditionally uses floral structures, not always readily captured digitally or even available in the living plant in the field
Identification by leaf images seems favourite, which is reasonable; however, Poland’s Vegetative key to British Flora, covering 3,000 taxa, shows it needs more than just leaf shape.
« Images captured under unconstrained environments, scale variations, different lighting conditions, leaf orientation, complicated backdrops, and leaflet structure make plant species recognition rigorous and time-consuming. » We observe that here on iSpot.
I do not doubt AI systems will become widely used; it remains to be seen what level of probability at what hierarchical level is judged adequate.
I saw, somewhere, a figure of 0.6 probability being adequate. Mmmmmm.
The issue always to bear in mind with AI systems is that they can make different errors to humans so a wrong ID may be totally wrong and not something that looks similar or is closely related.
All plants are edible, some only once.
2 Likes