DHBlog027 · An article by Robin Walls
There are now a lot of apps that will identify plants simply from an image. Taking more than one image to improve accuracy is an option in some of these apps. Ensuring the plant is in focus and the image contains a clear view of the relevant parts of the plant - flowers, leaves, stem etc. – should also improve the ability of the app to come to the right conclusion. The better apps provide a list of species names in order of decreasing likelihood, sometimes with a numerical confidence estimate. For instance, the first name in the list might have a likeness of 30%, (conversely, that is 70% dissimilarity). A useful review in BSBI News (Jones, 2020) gives the background and an evaluation of the freely available apps. Given the rate of improvement in AI, this review was a long time ago and by now we should expect more accurate identification. Jenny Ashdown and Jon Crewe are currently evaluating a selection of apps and will report in due course.
An app may appear to be very good, but none is 100% certain and some errors are inexplicable, like identifying a Sallow tree as a Siskin (Fulford, 2023)! When first evaluating an app it is tempting to judge it on how often it finds the right name and how much more likely the first option is than the second. But of course you can only do this if you know what the plant is, which negates the need for the app in the first place. There are two types of error: accepting an incorrect name, and rejecting the correct name. The real problem is how to know when the identification is wrong and when to accept it.
Unlike the keys in floras, these apps do not work through a linear, logical, often binary, sequence to come to a decision. They use an artificial neural network (ANN). This is akin to a human identifying a plant by its jizz. You can probably identify a willow from a considerable distance even though you cannot see the details of the leaves, buds or bark. You may be able to do better and know that it is a Sallow or Crack Willow before you are close enough to see the shape of the leaves. This is because you have seen a lot of willows, as well as a lot of other trees. The ANN is programmed by presenting it with a lot of named pictures (the training set); the more it sees the more accurate it will be. If your species was not in the training set it has no chance of identifying it correctly.
Some apps ask you to enter your identification to help train the app. An excellent instance of citizen science, but there is an obvious flaw. In the example above, if too many people take pictures of Goat Willow (Salix caprea), but call it Sallow (Salix cinerea) the app will tend to incorrectly identify both of these species in the future. There may be a verification process for user entered data, but I have not seen this mentioned anywhere.
Validation
With all the caveats above you might conclude that identification apps are useless. But that would be to miss an opportunity when faced with a ‘flicking species’ i.e. one where you simply look through a picture book. The app may suggests a family or genus you have overlooked. Below is a series of check questions to detect erroneous identifications and hopefully lead you to the right name:
Does the plant look like images from a reliable source? The training set may have some misidentified images which would lead to errors not detectable from the pictures offered by the app; look elsewhere.
Is it known from the vice-county or the hectad? If not, then an expert identification is probably needed and identification by an app is unlikely to be acceptable.
Does the habitat match the species’ preferences?
Do all the critical characters of the plant, flowering time etc. conform to the description in a reliable flora?
Is there an alternative plant that seems more likely, even though it has a lower score?
The new BSBI atlas (2023) is a good source for checks (1) and (2) as well as flowering times and some information on habitat. Recommended floras for points (1) and (4) are listed in the references below.
When there is doubt, collect a specimen (assuming the population is large enough) for identification by an expert. In your efforts to identify the plant you will have a good picture and have found the essential parts to include in a specimen if it is not appropriate to pick the whole plant. A voucher is essential when you may have a new hectad record.
Summary
These apps are very useful for providing a starting point that gives you a quick access into a key, or a name to type into the internet. But they are far from infallible and the tentative identification must be confirmed by inspecting the critical features. Remember, the app is only one piece of information. Used in conjunction with all the other information it can be a great help.
References
BSBI (2023) Plant Atlas 2020 online, https://plantatlas2020.org/atlas
Jones, Hamlyn (2020) BSBI News 144, 34-40. Artificial intelligence for plant identification on smartphones and tablets.
Rose, F. (2006) The Wild Flower Key, revised by C.O’Reilly, Penguin Books Ltd
Stace, C.A. (2019) New Flora of the British Isles, 4th edition, C & M Floristics.
Streeter, D. (2018) Wild Flower Guide, 2nd edition, Collins.
Fulford, Tony (2023) pers. comm., DFG WhatsApp group.