MobileNetV4-Conv-S for southwestern European flora

This is a compact image-classification model for identifying 7,806 plant species from the flora of southwestern Europe. It was fine-tuned on the Pl@ntCLEF2024 training set using knowledge distillation from the Pl@ntCLEF ViT-B teacher PlantNet/vit_base_patch14_reg4_dinov2_lvd142m.pc24_all.

This MobileNetV4-Conv-S model makes plant recognition more accessible on mobile and edge hardware, where the ViT-B model can be too computationally expensive.

Performance

Evaluation on the Pl@ntCLEF2026 mono-specie test split.

The student model was trained for 50 epochs on the Pl@ntCLEF2024 training set using the images resized to a maximum side of 800 pixels provided for PlantCLEF2026, by distilling the knowledge from the teacher model following the method Energy-Efficient Plant Monitoring via Knowledge Distillation, cited below.

Load with timm

pip install "timm>=1.0.12" huggingface_hub safetensors pillow

import timm
import torch
import json
from PIL import Image
from huggingface_hub import hf_hub_download
from timm.data import create_transform, resolve_data_config

model_id = "PlantNet/mobilenetv4-conv-s-plantclef2024"
model = timm.create_model(f"hf-hub:{model_id}", pretrained=True).eval()
transform = create_transform(**resolve_data_config(model.pretrained_cfg, model=model))
labels_path = hf_hub_download(model_id, "labels.json")
image_path = hf_hub_download(model_id, "example.jpg")
with open(labels_path, encoding="utf-8") as handle:
    labels = json.load(handle)

image = Image.open(image_path).convert("RGB")
with torch.inference_mode():
    probabilities = model(transform(image).unsqueeze(0)).softmax(dim=-1)[0]

for probability, class_id in zip(*probabilities.topk(5)):
    taxon = labels[class_id.item()]
    print(taxon["species"], probability.item())

labels.json contains the ordered species ID, scientific name, genus, and family. Entry i describes output logit i.

Model details

• Model: MobileNetV4-Conv-S fine-tuned for Pl@ntCLEF2024
• Initialization: ImageNet-1K weights from timm/mobilenetv4_conv_small.e2400_r224_in1k
• Distillation teacher: PlantNet/vit_base_patch14_reg4_dinov2_lvd142m.pc24_all
• Training data: Pl@ntCLEF2024 training split
• Training duration: 50 epochs
• Geographic and taxonomic scope: 7,806 species from the flora of southwestern Europe
• Input: 512 × 512 RGB
• Normalization: ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225)

Intended use

The model expects a photograph of a single plant observation and returns a probability distribution over its 7,806 southwestern European training species. Images containing several species, including vegetation plots, require a separate multi-label method such as tiled inference, score aggregation, and calibrated thresholds.

Citations

Method

@misc{moummad2026,
  title         = {Energy-Efficient Plant Monitoring via Knowledge Distillation},
  author        = {Ilyass Moummad and Reda Bensaid and Kawtar Zaher and Hervé Goëau and Jean-Christophe Lombardo and Joseph Salmon and Pierre Bonnet and Alexis Joly},
  year          = {2026},
  eprint        = {2604.27178},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2604.27178}
}

Teacher Model

@misc{goeau2024plantclef,
  author       = {Goëau, Hervé and Lombardo, Jean-Christophe and Affouard, Antoine and Espitalier, Vincent and Bonnet, Pierre and Joly, Alexis},
  title        = {PlantCLEF 2024 pretrained models on the flora of the south western Europe based on a subset of Pl@ntNet collaborative images and a ViT base patch 14 dinoV2},
  year         = {2024},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.10848263},
  url          = {https://doi.org/10.5281/zenodo.10848263},
  note         = {Figure}
}