Maximaxou2/french-trees-vit-small

French trees — ViT-S/14 (fast, mobile)

A small (22M parameters, 41 MB fp16 ONNX) image classifier for 71 common trees of France, plus an other plant class. Distilled from a fine-tuned Pl@ntNet ViT-B/14, small enough for live phone-camera use.

Sibling model: Maximaxou2/french-trees-plantnet-ft — the larger, more accurate model it was distilled from.

Usage

import json, timm, torch
from huggingface_hub import hf_hub_download
from PIL import Image

model = timm.create_model("hf-hub:Maximaxou2/french-trees-vit-small", pretrained=True).eval()
transform = timm.data.create_transform(**timm.data.resolve_data_config({}, model=model))
labels = json.load(open(hf_hub_download("Maximaxou2/french-trees-vit-small", "labels.json"), encoding="utf8"))
temperature = json.load(open(hf_hub_download("Maximaxou2/french-trees-vit-small", "config.json")))["temperature"]

x = transform(Image.open("photo.jpg").convert("RGB")).unsqueeze(0)
with torch.no_grad():
    probs = (model(x) / temperature).softmax(-1)[0]
for p, i in zip(*probs.topk(3)):
    print(f"{p:.1%}  {labels[i]['latin']} ({labels[i]['fr']})")

Divide logits by temperature (fitted on the validation split) for calibrated probabilities. Class 71 (other) means "probably not one of the 71 trees".

ONNX: model.onnx — fp16 weights, fp32 input image [1, 3, 252, 252] (RGB, center square crop, ImageNet mean/std), output logits [1, 72]. Runs with ONNX Runtime, incl. Android (CPU/XNNPACK or Qualcomm NPU via QNN).

Results

Test set: 7,090 photos of the 71 trees (100 iNaturalist observations per species, 90 for Thuja) observed in 2025–2026 by people absent from the training data, so neither the base model nor this one saw them; plus 1,000 photos of other plants, half of them close relatives of the 71 (other oaks, maples, pines, Tilia × europaea, oriental plane…).

Model	Top-1 (among the 71)	Open-set accuracy on trees	Other plants rejected
Pl@ntNet PlantCLEF 2024, zero-shot (original)	89.8	69.0	90.4
This model	91.83	89.7	89.4

• Top-3 among the 71: 97.43 %. Calibration error (ECE): 1.43 %.
• When the calibrated confidence is ≥ 90 % (84.2 % of tree photos), it is right 97.84 % of the time.
• By photo type (auto-tagged, approximate): flower 97.6 · fruit 95.4 · leaf 95.2 · needles or cones 92.4 · twig or bud 91.6 · bark 79.7 · whole tree 78.8.
• Most frequent confusions: Quercus robur → Quercus petraea (12), Tilia cordata → Tilia platyphyllos (11), Quercus pubescens → Quercus petraea (10), Pinus pinea → Pinus pinaster (10), Pinus nigra → Pinus pinaster (8), Tilia platyphyllos → Tilia cordata (8). These pairs are hard from a single photo even for botanists, and iNaturalist labels for them are themselves less reliable.

Size and speed: 22 M parameters, 252 px input; ≈ 7,000 img/s on an RTX 5090 (fp16, batch 64).

Training data

iNaturalist research-grade observations with CC-licensed photos (iNaturalist open-data bucket), France first, the rest of Europe for the less-observed species: up to 700 observations (≤ 2 photos each) per species, plus ~37k photos of ~5,000 other plant species for the other class. Observer-disjoint train / validation / test splits. Licenses of training photos: 81 % CC BY-NC, 12 % CC BY, 4 % CC0, the rest other CC variants.

Distilled from the fine-tuned Pl@ntNet model (448 px teacher → 252 px student) for 14 epochs: 70 % KL to the teacher's softened outputs (T = 2) + 30 % label cross-entropy, with colour, blur and sensor-noise augmentation to mimic phone camera frames.

Limitations

• Only 71 species. Anything else should come out as other, but ~11 % of lookalike plants are still named as one of the 71 — treat low-confidence answers as "check it".
• Bark-only and whole-tree photos are clearly harder than leaves, flowers or fruits.
• Platanus × hispanica (London plane) is not in the base Pl@ntNet model; it was learned only during fine-tuning.
• Built for learning and curiosity, not for decisions that matter (e.g. edibility or toxicity).

License and credits

Weights: CC BY-NC 4.0 — most training photos are CC BY-NC. Base models: Pl@ntNet PlantCLEF 2024 ViT-B/14 (Apache-2.0, trained on Pl@ntNet data), DINOv2 (Meta). Photos: the iNaturalist community. Thank you to every observer.

Species

Latin name	French name	Test accuracy (%)
Quercus robur	Chêne pédonculé	72.0
Quercus petraea	Chêne sessile	89.0
Quercus pubescens	Chêne pubescent	78.0
Quercus ilex	Chêne vert	91.0
Quercus suber	Chêne-liège	88.0
Quercus rubra	Chêne rouge d'Amérique	97.0
Quercus cerris	Chêne chevelu	88.0
Fagus sylvatica	Hêtre commun	93.0
Carpinus betulus	Charme commun	89.0
Castanea sativa	Châtaignier	90.0
Fraxinus excelsior	Frêne élevé	85.0
Fraxinus ornus	Frêne à fleurs	89.0
Acer pseudoplatanus	Érable sycomore	92.0
Acer platanoides	Érable plane	91.0
Acer campestre	Érable champêtre	89.0
Acer monspessulanum	Érable de Montpellier	94.0
Tilia cordata	Tilleul à petites feuilles	79.0
Tilia platyphyllos	Tilleul à grandes feuilles	89.0
Betula pendula	Bouleau verruqueux	95.0
Alnus glutinosa	Aulne glutineux	98.0
Corylus avellana	Noisetier	97.0
Juglans regia	Noyer commun	95.0
Prunus avium	Merisier	92.0
Prunus padus	Cerisier à grappes	93.0
Prunus mahaleb	Cerisier de Sainte-Lucie	95.0
Malus sylvestris	Pommier sauvage	95.0
Pyrus communis	Poirier commun	90.0
Sorbus aucuparia	Sorbier des oiseleurs	95.0
Sorbus domestica	Cormier	93.0
Sorbus torminalis	Alisier torminal	98.0
Sorbus aria	Alisier blanc	96.0
Crataegus monogyna	Aubépine monogyne	94.0
Robinia pseudoacacia	Robinier faux-acacia	96.0
Gleditsia triacanthos	Févier d'Amérique	96.0
Platanus × hispanica	Platane commun	92.0
Aesculus hippocastanum	Marronnier d'Inde	99.0
Ailanthus altissima	Ailante glanduleux	97.0
Celtis australis	Micocoulier de Provence	97.0
Populus nigra	Peuplier noir	96.0
Populus tremula	Peuplier tremble	91.0
Populus alba	Peuplier blanc	98.0
Salix alba	Saule blanc	99.0
Salix caprea	Saule marsault	88.0
Ulmus minor	Orme champêtre	89.0
Ulmus glabra	Orme de montagne	94.0
Ilex aquifolium	Houx	98.0
Buxus sempervirens	Buis	96.0
Cornus mas	Cornouiller mâle	94.0
Euonymus europaeus	Fusain d'Europe	94.0
Sambucus nigra	Sureau noir	95.0
Olea europaea	Olivier	98.0
Arbutus unedo	Arbousier	95.0
Laurus nobilis	Laurier-sauce	88.0
Prunus laurocerasus	Laurier-cerise	96.0
Ostrya carpinifolia	Charme-houblon	99.0
Pinus sylvestris	Pin sylvestre	84.0
Pinus pinaster	Pin maritime	84.0
Pinus nigra	Pin noir	82.0
Pinus halepensis	Pin d'Alep	91.0
Pinus pinea	Pin parasol	81.0
Pinus cembra	Pin cembro	82.0
Pinus uncinata	Pin à crochets	89.0
Abies alba	Sapin pectiné	92.0
Picea abies	Épicéa commun	83.0
Pseudotsuga menziesii	Douglas	94.0
Larix decidua	Mélèze d'Europe	94.0
Cedrus atlantica	Cèdre de l'Atlas	94.0
Taxus baccata	If commun	92.0
Juniperus communis	Genévrier commun	98.0
Cupressus sempervirens	Cyprès de Provence	99.0
Thuja occidentalis	Thuya du Canada	86.7
other	Autre plante	—