Lesson objective

State the measured species-level accuracy of identification apps against expert identification, name the plants and places those measurements came from, and explain why no such figure exists for turfgrass, ornamentals or tree pests.

Plain-language explanation

Why this matters now

Identification is the centre of this trade. What the plant is, what the insect is, what the damage is: everything downstream, the product, the rate, the timing, the price, the promise, rests on that one answer. It is also the place where an app is most tempting, because the app answers in two seconds and a qualified diagnosis can take two days.

So the honest question is not "is AI identification good?" It is "how often is it right, on what, measured by whom, compared with what?" Two peer-reviewed studies answer that for wild flora and for regulated pest plants. Nobody has answered it for a turfgrass cultivar, a landscape ornamental, a tree disease or a tree pest. Both halves of that sentence matter.

A question to sit with

If a tool named the plant correctly about four times in five, how would you build a working routine around it — and would you still call that "reliable" to a customer?

What goes wrong

Repeating a number without its subject. "Plant ID apps are 79% accurate" is not a finding. The finding attaches to specific apps, specific species, a specific place, a specific season and a specific photography method.

Carrying a figure across to a different problem. A percentage measured on wild vascular plants in Alberta plots is not a percentage for a fescue cultivar, a cultivated maple or a scale insect on a hedge.

Confusing genus with species. An app may get the genus right far more often than the species. For a treatment decision, the difference between two species in one genus can be the whole decision.

Assuming more photographs is always better. More frames is not the same as more information, and the researchers who measured this recommend something more specific.

Treating an app score as the measured accuracy. The percentage the app shows you is its internal ranking. The percentages below were measured by researchers against known identifications.

The practical answer

What was measured, on wild flora in Alberta. A study published in AoB Plants on 8 May 2026 by Wanigasinghe, Haughland, Pyle, Villeneuve and Nielsen analysed 5,291 field images of 119 vascular plant species, from 54 plots of 0.8 square metres, in grassland and forested sites in Alberta, photographed in spring and at anticipated peak bloom in 2023.

What was measured Species level Genus level
Flora Incognita 79.2% 88.4%
iNaturalist 66.5% 73.7%
iNaturalist, fusing scores from multiple images of the same individual 72.5% not stated here
Authors' estimated expert accuracy range 90-95% 90-95%

Neither app reached the authors' estimated expert accuracy range of 90-95%. That range is the authors' estimate, not a measured benchmark, and the study is two apps at one place and time, so it must never be used to rank apps for purchase.

The authors' practical recommendations are worth more to you than the percentages: capture reproductive structures, prioritise image quality, especially for graminoids, and select "a few informative perspectives" rather than maximising the number of image perspectives. That last point contradicts the instinct to take twenty pictures.

What was measured, on regulated pest plants. A study published in PLOS ONE on 2 March 2026 by Charron, Julien, Stinziano and Gagnon evaluated iNaturalist and PlantNet on two groups: 61 species established, native or naturalized, in Ontario, and 402 "outsider" plants regulated as pests and absent or barely present in Canada. Three of its findings change how you work.

  • Grasses are hard. Poaceae were identified correctly about 60% of the time, substantially below Asteraceae and other families, which the authors attribute to small, inconspicuous features requiring dissection. Read that carefully: it is Poaceae generally, in a study of wild and regulated species. It is not a turfgrass cultivar identification figure, and no such figure exists.
  • Leaves alone are not enough. "Pictures containing only leaves are less likely to be accurately identified" than inflorescence or whole-plant images, and the difference was statistically significant.
  • Long-established introductions are not harder. No meaningful accuracy difference appeared between native and long-established introduced species. The difficulty is not "foreign", it is "not in the local data", which is the subject of the next lesson.

The gap where there is no evidence at all. This is the part that gets left out of a sales conversation.

Subject Verified accuracy evidence available to this course
Wild vascular plants, Alberta plots Yes: 79.2% and 66.5% at species level, against an estimated 90-95% expert range
Regulated pest plants and established Ontario species Yes: family-level differences, about 60% for Poaceae, leaves-only worse
Crop disease detection, controlled datasets versus field images Yes: benchmark accuracy often decreases in real-field conditions
Turfgrass cultivars None
Landscape ornamentals and nursery stock None
Tree diseases None
Tree and shrub insect pests None

Four of those seven rows are empty. That is not a gap this course will fill by borrowing a number from the rows above it. When someone tells you an app is accurate on turf, ornamentals or tree pests, the correct response is to ask what that claim was measured on.

What the numbers mean for a working routine. Take the highest species-level figure in this lesson, 79.2%, on wild Alberta flora, in research conditions, with careful photography. Against an estimated expert range of 90-95%, that is a useful assistant and an unacceptable decision maker. Build the routine accordingly.

Use Appropriate? Why
Generating candidate names to check against a key or a specimen Yes A shortlist you then test is exactly what the evidence supports
Recording what was on site, with photographs, for a qualified person Yes The photograph set is the value, whatever the app says
Teaching a new crew member what to look at With care Useful for looking, never for concluding
Naming a plant in a customer report or a bid No, not on its own A named plant in a document is a claim your company makes
Selecting a product or a rate No That decision belongs to the certified applicator, from the label
Declaring a tree healthy or hazardous No That belongs to the arborist or licensed tree expert

Remember this

  • Measured species-level accuracy in the Alberta study was 79.2% for Flora Incognita and 66.5% for iNaturalist, against the authors' estimated 90-95% expert range; neither app reached it.
  • Fusing multiple images of one individual raised iNaturalist species accuracy to 72.5%, still below the expert range.
  • Poaceae were identified correctly about 60% of the time in the Ontario study; that is grasses generally, not turfgrass cultivars.
  • Pictures containing only leaves are significantly less likely to be identified correctly than inflorescence or whole-plant images.
  • There is no verified accuracy evidence for turfgrass cultivars, landscape ornamentals, tree diseases or tree pests. Never borrow one.

Landscaping example

Kestrel Tree & Shrub LLC is a fictional company that has been asked to bid on plant health care for an invented 40-unit condominium association on Fernbrook Way. Its licensed arborist, Priya Ambrose, sends two technicians to inventory the ornamental beds and hedges. One of them, Dov Halloran, uses a plant identification app throughout and comes back with a list of 34 plants, each with a name and a percentage. Priya spends an hour with him on the list rather than accepting it. She sorts it three ways. Eleven entries are shrubs photographed in flower with whole-plant frames; those she treats as a solid shortlist and confirms nine of them from her own knowledge on a site walk, correcting two. Fourteen entries are from single leaf photographs against mulch; she reminds Dov that pictures containing only leaves were significantly less likely to be identified correctly in the Ontario study, and treats all fourteen as unresolved. Nine entries are grasses and sedges in the ornamental beds, where the app gave confident species names; Priya points out that Poaceae were identified correctly only about 60% of the time in that same study, on wild and regulated species, and that no accuracy figure exists at all for a cultivated grass cultivar. She marks all nine as requiring a specimen. Dov asks whether the app is worth using. Priya says yes, for what it is: it produced a shortlist and, more usefully, a photograph set and a record of what was on site on a date. What it did not produce is an inventory. The bid that goes to the association names the eleven confirmed plants, describes the rest as "mixed ornamental shrub and grass plantings, species inventory to follow a site visit", and prices a half day for that visit. Kestrel does not put a percentage or an app name in front of the customer, and does not name a plant in a document that Priya has not confirmed.

Responsible-use warning

Do not quote an identification accuracy figure for turf, ornamentals, tree diseases or tree pests. None exists in this course's sources, and the figures that do exist were measured on wild Alberta flora and on regulated pest plants in Ontario. Borrowing one across is how a false reassurance enters a customer conversation and then a contract. Equally, do not let the honest numbers become an excuse: 79.2% at species level is genuinely useful for producing a shortlist and a photograph record. It is only unacceptable as the thing that decides.

Human-review requirement

A qualified person confirms or rejects every identification before it appears in a customer document, a bid, a work order or a treatment decision. For a pesticide decision that person is the certified applicator, or the certified applicator supervising a non-certified one, and the product label is the binding instruction. For a tree's condition it is the arborist or, in a state that licenses tree work, the licensed tree expert. The reviewer checks that leaves-only photographs and grass identifications were treated as unresolved, that the photograph set was kept with the date, and that no accuracy figure was transferred from wild flora or crops to turf, ornamentals or trees.

Practical exercise

Measure your own hit rate on material where the answer is already known. Use plants on your own premises or a fictional practice site, never a customer's identification decision.

  1. Assemble 15 plants whose identity a qualified person in your company already knows and has written down. Include at least three grasses and at least three woody ornamentals.
  2. Photograph each one twice: once as a single leaf against a busy background, and once as a set of a few informative perspectives, including reproductive structures if present, with attention to image quality.
  3. Run both through one identification app. Record the name and the number shown for each of the 30 attempts.
  4. Score each attempt against the known answer at species level and at genus level. Count them.
  5. Compute four hit rates: leaf-only at species level, informative-set at species level, and the same two at genus level.
  6. Compute the same rates for your grasses alone, and write one sentence comparing them with the rest.
  7. Write your four figures next to the published ones in this lesson, and one sentence on why yours are not comparable to them.
  8. Write the rule your company will follow: what an app result may be used for, what it may never be used for, and who confirms.
  9. Record everything on the identification verification sheet, with the date and the name of the qualified person who supplied the known answers.

Knowledge check

  1. What species-level accuracy did the Alberta study measure, and against what benchmark?

    Answer guidance: Flora Incognita achieved 79.2% for species and 88.4% for genera; iNaturalist achieved 66.5% for species and 73.7% for genera. Neither reached the authors' estimated expert accuracy range of 90-95%, which is the authors' estimate rather than a measured benchmark.

  2. What did the Ontario study find about grasses and about leaves-only photographs?

    Answer guidance: Poaceae were identified correctly about 60% of the time, substantially below Asteraceae and other families, attributed to small inconspicuous features requiring dissection. Pictures containing only leaves were significantly less likely to be identified accurately than inflorescence or whole-plant images.

  3. A vendor says their app is 90% accurate on landscape ornamentals. What do you ask?

    Answer guidance: What that was measured on, by whom, against which reference identifications, on what species, where and when. No verified accuracy evidence for landscape ornamentals, turfgrass cultivars, tree diseases or tree pests exists in this course's sources, so a figure for them cannot be taken from the published plant identification research.

  4. The Alberta authors recommend a photography approach. What is it, and what does it rule out?

    Answer guidance: Capture reproductive structures, prioritise image quality especially for graminoids, and select a few informative perspectives rather than maximising the number of perspectives. It rules out the instinct that simply taking more photographs will improve the result.

Sources and further reading