"Our model identifies plants with 96.8% accuracy." It is a sentence that appears, in one form or another, in many AI announcements. It is also a sentence that says very little about whether a model can name the trees in a smallholder plot in Sumatra.
Species matters for carbon work. It chooses the allometric equation, affects expected growth and permanence, and in some supply chains is part of the evidence a buyer needs. So it is worth reading what published studies actually measured.
Same model family, very different scores
96.8
BioCLIP 2 zero-shot top-1 on a plant benchmark (PlantNet subset)
25.1
the same model on PlantVillage, a leaf-disease benchmark
77.9%
accuracy for 84 tropical tree species from close-up images in a drone study
One model, two benchmarks, two stories
BioCLIP 2 is an open model trained on a large collection of biological images. In its published results, its zero-shot top-1 accuracy is 96.8 on the PlantNet benchmark and 25.1 on PlantVillage. Across ten benchmarks, its mean is 55.6. The scores depend entirely on what each benchmark contains. The PlantNet column is a subset from the Meta-Album collection, not the 1,000-class Pl@ntNet-300K test set.
That bigger Pl@ntNet-300K-v2 dataset has 306,087 images of 1,000 plant species, and the best published top-1 test accuracy in its repository is 0.8457. BioCLIP 2's model card, for its part, warns that predictions may be biased toward well-represented species, because the training data is long-tailed.
Older work shows the same dependence on conditions. The LeafSnap study reported about 73% top-1 and 96.8% top-5 accuracy for 184 North American tree species, using leaf photos against plain backgrounds. Top-5 means the right answer appears in the model's five best guesses. A field worker standing under a canopy usually needs top-1.
None of these benchmarks is a collection of tropical trees photographed in the field. In a literature search for this article we did not find a published Indonesian field-photo benchmark either, so treat any such number you see with care.
What drone studies of tropical species report
Closer to the problem are studies that classify tree crowns from the air.
A 2026 arXiv study from a tropical forest labelled 5,302 tree crowns, of which 1,999 had species labels from an expert. The experiments were limited to 84 species that met a minimum label count. Accuracy was 77.9% using close-up images from a single date, and 74.3% using crown-view aerial imagery aggregated over 16 dates. The authors note that the gap is most pronounced in the long tail, where few labelled examples exist and the differences between species are subtle.
The long tail is the heart of the matter. Tropical forests contain many species and a few of them dominate. A model can score well overall by getting the common species right and failing on the rare ones.
What satellite studies report
Satellites are cheaper per hectare than drones, but they see less.
- A WorldView-3 object-based study at La Selva Biological Station, Costa Rica, classified six tropical tree species with 85.37% accuracy, 7.89% commission error and 14.63% omission error. The authors found that shadowing significantly affected segmentation and classification. - A review of crown delineation in tropical forests quotes work in the Brazilian Amazon where the best individual-crown method reached 65%, and notes that raster canopy-height methods were ineffective for lower-storey trees. - The same review cites automatic delineation from high-resolution multispectral satellite images at 79%, effective for large crowns but not for understorey and shadowed trees. It also records that star-shaped palm crowns needed manual correction in most cases, and calls perfect crown delineation in tropical forests unrealistic. - For contrast, a temperate benchmark, CentralBark, covers 25 hardwood species and over 19,000 bark images in three US states. A model reached 83.21% overall accuracy on it.
Six species, 25 species and 84 species are different problems. A score from one does not predict the others.
Reading a species-ID claim
Before relying on a number, ask:
1. How many species, and which? Six common species is not 84 species with a long tail. 2. What kind of images? Plain leaf scans, close-up photos, drone crowns and satellite pixels are different inputs. 3. Top-1 or top-5? Verification needs top-1. 4. Where was it tested? Temperate and urban benchmarks say little about tropical agroforests. 5. What happens with a species the model has never seen? Research on open-set recognition notes that conventional classifiers trained on known classes tend to specialise in them and struggle with unknowns. 6. Who checks the result? For carbon or supply-chain evidence, a human check on a sample is part of a defensible design.
For why a single number per tree is fragile in the first place, see why single-tree monitoring fails. For what species changes downstream, see species-specific allometry.
Frequently Asked Questions
Published results vary widely with the number of species, the image type and the location. In a tropical drone study covering 84 species, accuracy was 77.9% from close-up images. We found no published Indonesian field-photo benchmark in our search.
Evaluating species ID for tree data?
Species matters for allometry, permanence and EUDR evidence. If you are weighing AI against field identification, we are happy to compare notes.


