In recent years, mobile applications for plant identification such as PlantNet or PictureThis have had millions of downloads worldwide. The way they work is simple: Users take a photo, the app processes it on a distant server and then returns the plant's most likely name along with other pictures.
The fact of the matter is that identifying a plant from a photograph is not easy for an experienced professional, and neither for a software based on Machine Learning techniques.
Several recent studies point out that the accuracy of these apps, especially when trying to identify endemic or rare plants, can be very low. This can also be the case when the plant being photographed lacks a flower or belongs to an ecosystem that the app doesn't cover.
In April, a group of researchers from the University of Galway in Ireland tested six of these apps on 38 species of local flora. Although their accuracy varies "considerably between plants," the authors explain, the information provided by the software "should not be considered excellent or assumed to be correct." "Particularly if the species in question may be toxic or otherwise problematic," they warn.
In August, researchers from the universities of Córdoba and Seville in Spain analyzed the performance of four of these apps on the flora of Andalusia, a region in the south of the country. "With common and ruderal plants, the ones that grow in roadside ditches, the apps get it right half the time, but with endemic [local] plants they always fail," Manuel De la Estrella, a botany professor at the University of Córdoba and one of the authors of the paper, told AlgorithmWatch. The paper's title raises a provocative question: “Could plant ID applications lead to an increase in extinction risk?”
Little impact on conservation
So, could the apps' inaccuracies with the rarest plants indeed pose a risk to diversity, given the popularity of these tools among non-experts? Probably not. All three ecosystem conservation practitioners consulted by AlgorithmWatch for this article rule this out. But they they're confronted with a different problem in research and teaching.
Sandra Saura is a lecturer in ecology at the Universitat Autónoma de Barcelona (UAB) and a researcher in conservation ecology at the Ecological and Forestry Applications Research Centre (CREAF). "Threatened species are so rare that these apps almost never get them right. Also very few people know where they are, so they're not going to be found by the average user. I don't think they have that impact right now," she says.
Emilio Valbuena, a biologist and a conservation technician in several natural parks near Barcelona, and Artur Lluent from the flora protection service of the Catalan government both see these apps being used by a growing number of visitors of natural areas. On the one hand, they highlight positive aspects such as their usefulness in guiding amateur botanists. With the help of iNaturalist, another app used for plant identification, the administration for which Valbuena works runs a citizen science project to learn more about several protected species.
But on the other hand, both stress the necessity of being critical of the results the apps provide. "Too much credibility is given to everything that comes out of a screen. It's a good tool. The problem is not the tool but the use that many people make of it," says Lluent.
Lluent works on rules for the conservation of endangered species in public spaces. He warns that in some cases it may be impossible to identify a species with a photo. "Some species are easy to identify, but others are not. Some look very similar morphologically even though they have nothing to do with each other. If it doesn't have flowers, the leaves may be the same," he points out.
Like other category-identifying AI programs, the plant ID software uses pre-defined categories to detect species according to the patterns sufficiently represented in the databases, explains data scientist and researcher Javier Sánchez-Monedero.
The apps cannot identify a species that is not recorded in their database. "This is an inability by definition of neural networks," he says.
Manuel de la Estrella criticizes that the software isn't straightforward with users about its limitations. "There is a lot of diversity that we still don't know about. If you find a new species, the app doesn't tell you that, or at most some give you a probability. But it's clear that the app doesn't tell you that it doesn't know the answer and that's a serious problem," says the botanist, pointing out that these apps give users a false sense of security.
While knowing the apps' limitations is important for a hobbyist, it is crucial for students in botany and environmental science. If students rely too much on the apps, they might fail to acquire a thorough knowledge of the traits that differentiate one plant from another, at the family and genus level.
De la Estrella is worried that such lack in expertise will extend to their professional life. “In the exercises we do, we give the students five plants. They immediately take out their cell phone and almost always get it right, because we bring them common plants. They are getting the feeling that this works... When these people graduate, when they go to do an environmental study of a highway or an airport, they are going to think: with some photos I take here I can get the species that are there."
For this reason, De la Estrella and his team put a special effort into teaching the traditional identification method with specimen, magnifying glasses, microscopes, and bibliographies. “It is a slow process and it takes time to learn. At first the terminology is complex,” he points out.
Lluent also doesn't think that the apps allow for a proper learning experience: “Plants are learned by determining them. If the effort is very small, the memory is too. If you want to dedicate yourself to botany, it is not the way to learn.”
Saura has been working in conservation for 20 years and often conducts field studies at the Mediterranean coast. She works with younger technicians, trainees and self-employed professionals. "They immediately say: ‘I'm going to look at this plant on PlantNet’", she explains.
"The problem is that if you use these apps you are not going to learn how to identify that plant, because you delegate the identification to someone else," she argues.
"When you identify a plant with a magnifying glass, a book, and a dichotomous key [a sort of decision tree], that plant is already in your memory. Because you're going to spend half an hour on it, you're going to look at all the details, the flowers, the fruit, etc.," she adds.
In her ecology classes, she has already banned plant-identifying apps: “I never use them for classes, I want to promote the skills of the scientific world, the main one being observation. If you don't know how to observe, you are not going to ask good hypotheses. And with those apps you are not observing.”