The PlantCLEF 2025 initiative aims to advance ecological research by analyzing vegetation plot inventory photographs to support standardized sampling, biodiversity monitoring, long-term ecological assessment, and large-scale remote vegetation surveys. The challenge is framed as a multi-label classification task, requiring the identification of all visible plant species in high-resolution quadrat images. In this approach, a self-supervised Vision Transformer (DINOv2) model is employed to address the task. Both the base and fine-tuned variants of DINOv2 are utilized to extract robust and generalizable feature embeddings, which serve as inputs for training classifiers capable of predicting multiple plant species present in a single image. To handle high-resolution input eficiently, the proposed processing framework divides each image into a grid of overlapping tile individually, performs classification on each tile individually and then aggregates tile-level predictions into a unified set of probabilities per image. This tiling-based strategy enhances spatial context and improves species detection, especially in complex and densely populated vegetation plots achieving a macro-averaged F1 score of 0.00106 per sample on the oficial test set.
The PlantCLEF 2025 challenge [
The PlantCLEF challenge series has played a vital role in advancing automated plant species
identification, especially under real-world ecological conditions. The 2025 edition builds upon
the foundation laid in 2024, continuing the focus on multi-label classification of high-resolution
vegetation plot images. While the 2024 task [
In prior editions of the challenge, top-performing teams explored a variety of innovative strategies to
handle the scale and complexity of the data. Leading approaches included combining multiple tiling
configurations for inference, applying false positive reduction techniques like Segment Anything (SAM),
and aggregating predictions across diferent temporal views of the same plot [
These prior works demonstrate that dense tiling, temporal aggregation, and ensemble
modeling are highly efective in managing the high variability and label density present in plot-level plant
classification. Inspired by these advancements, the present work utilizes a transfer learning approach
based on DINOv2 [
DINOv2 is a self-supervised Vision Transformer (ViT) model trained on the large-scale LVD-142M
dataset containing 142 million images. Similar to BERT [
DINOv2 is available in various sizes (from small to giant), the proposed work utilizing the ViT-B/14 (distilled) variant, which ofers a balance between performance and eficiency. This model produces a ifxed output of shape 257 × 768, comprising 256 patch embeddings and one [CLS] token embedding [10]. The challenge organizers provided two pretrained ViT-B/14 DINOv2 models: one with a frozen backbone, and another fine-tuned end-to-end [ 11]. The fine-tuned model is used to extract [CLS] embeddings from each image, leveraging its stronger feature representations for the multi-label classification task
The dataset used in the PlantCLEF 2024 challenge consists of two main components: a large single-plant training set and an unlabeled test set of vegetation plot images. The training metadata, provided in PlantCLEF2024_single_plant_training_metadata.csv, contains over 1.4 million images associated with 1, 408, 033 unique entries. Each entry has the following key attributes: • image_name: name of the image file • species_id: numeric ID representing the plant species • organ: visible plant part in the image (flower, habit, or other) • obs_id: identifier for the observation instance • Metadata: Geo-location (latitude, longitude, altitude) and contributor-related fields (author, partner, license).
A separate file provides the complete list of 7, 800 unique species IDs used for classification.The test set includes 2, 105 high-resolution images of vegetation plots. Unlike the training set, which contains single-species images, the test set presents a multi-label setting, where each image may contain multiple species. This domain shift makes the task significantly more complex, requiring robust feature extraction and efective aggregation strategies.
The overall flow diagram of the proposed system is given in Figure 1. It consists of four main stages: data downloading, preprocessing, modelling, and inference. The input images are first collected from the PlantCLEF 2025 dataset. These images then undergo preprocessing to standardize the format and ensure consistency. In the modelling phase, a pretrained DINOv2 ViT-B/14 model is used to extract features, and a classifier is trained using the [CLS] token representation. During inference, test images are passed through the trained classifier using two approaches — full image classification and grid-based classification.
Before training, the images in the single-plant dataset were preprocessed to reduce computational cost and ensure consistency in input format.
All the input images were cropped to a square and then resized to 128 × 128 pixels. This standardization helps maintain focus on the main subject while ensuring compatibility with batch processing and feature extraction. While this resizing may lead to the loss of fine details and reduce the efectiveness of advanced data augmentation techniques, it was chosen to reduce computational cost and training time for this initial exploration.
During the training phase, the preprocessed images are further used by the proposed framework. During the test phase, the vegetation plot images were divided into smaller square tiles using a gridbased approach (e.g., 3 × 3 or 8 × 8), and the same preprocessing steps were applied to each tile. Features were extracted from each tile individually, enabling the model to capture localized plant features within the high-resolution test plots. The proposed preprocessing pipeline significantly reduced the size of the dataset and made the training and inference process more eficient without compromising the quality of visual information.
The overall process of the proposed framework is illustrated in Figure 1.
The proposed framework utilizes the preprocessed images which is of 128×128 pixel size. This helps standardize the input size for our model and reduces the overall storage and processing time. The ifne-tuned DINOv2 ViT-B/14 model is used to extract image embeddings. This model produces a 768-dimensional embedding vector from the special [CLS] token, which summarizes the overall content of the image. Although the DINOv2 model was trained on 518×518 images, compatibility with our 128×128 inputs was ensured by enabling the dynamic_img_size option from the timm library.
All 1.4 million image embeddings from the training dataset were used without applying any indexing optimizations. Although this increased memory usage and inference time, it remained feasible for our experimentation. A simple k-NN classifier [ 12] is used by storing the extracted features to recognize the plant species. The k-NN approach was chosen because it can efectively identify multiple possible species in each tile based on similarity to training images. This also improves the chances of detecting smaller or less dominant plants present in localized regions of the image. A value of = 5 was used as per the default setting of the scikit-learn k-NN classifier.
During test phase, each test image is divided into a grid of smaller parts. Each part is processed through the DINOv2 model to get predictions for the top most likely species in that part. Finally, combine all the probability values from the tiles by filtering the duplicates of the species ID and the ifnal list of species for each image have been generated. This final output is submitted for evaluation.
Two distinct approaches such as full-image and grid-based image prediction were employed to evaluate the performance of the proposed framework as shown in Figure 3. The test dataset is not cropped and resized to preserve the high quality of multi-label images.
Full-Image Prediction: In this approach, the entire test image is processed in its original dimension. The fine-tuned ViT model evaluates the image and outputs probabilities for each of the 7806 plant species classes. Top 20 probabilities was then taken out, representing the most likely species present, and map these probabilities to their corresponding species IDs.
Grid-based Image Prediction: The test image was divided into an × grid of tiles, resulting in tiles. Each tile is independently processed using the fine-tuned ViT model, which outputs probabilities for each species class. The top probabilities (default = 10) are mapped to their respective species classes for each tile. For example, a 3 × 3 grid yields nine tiles, each with ten top probabilities, selecting the top five species in each tile, totaling 45 species IDs and probability mappings.
The predictions obtained using the oficial pretrained DINOv2 model resulted in a macro-averaged F1 score of only 0.00012 per plot, which is relatively low. In contrast, the multi-class full-image inference method, using a linear classifier trained on fine-tuned DINOv2 embeddings, achieved a macro-F1 score of 0.00098, representing an 8× improvement over the base pretrained model. For a more advanced approach, a grid-based inference method was adopted to enhance multi-label classification performance. Among the grid-based methods, using the top species per tile achieved the highest scores across all metrics.
For inference with the fine-tuned model, a 3 × 3 grid size was found to strike a good balance between computational eficiency and spatial coverage of species distribution. We also experimented with 2 × 2 and 5 × 5 grid sizes but observed no substantial performance improvements, as shown in Table 1. A total of 5 species per tile were selected to maximize the macro-averaged F1 score per sample, focusing on capturing the most represented species in the dataset. Although this approach predicts up to 45 species per image, it was chosen to prioritize recall, even at the cost of lower precision.
Our final submission achieved a macro-averaged F1 score of 0.00106 on 89% of the test dataset, and 0.00293 on the remaining 11% of the dataset.
This work presents a lightweight multi-label plant species classification approach based on selfsupervised Vision Transformer (DINOv2) models. A scalable solution is proposed for multi-label image classification utilizing only single-label training data. Future improvements may involve enhancing model training and inference by incorporating additional data augmentation techniques, experimenting with various grid sizes, exploring alternative dimensionality reduction methods, and employing loss functions such as binary cross-entropy and asymmetric loss [13]. Furthermore, the development of more sophisticated aggregation strategies for multi-label prediction holds potential to further improve classification performance.
We extend our gratitude to the developers of the CLEF2025 team for their pre-trained DINOv2 model and the organizers of PlantCLEF and LifeCLEF for hosting the competition.
During the preparation of this work, the authors used Grammarly in order to: check grammar and spelling, paraphrase, and reword. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [9] L. Wu, W. Zhang, T. Jiang, W. Yang, X. Jin, W. Zeng, [cls] token is all you need for zero-shot semantic segmentation, arXiv preprint arXiv:2304.06212 (2023). [10] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, et al., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020). [11] H. Goëau, J.-C. Lombardo, A. Afouard, V. Espitalier, P. Bonnet, A. Joly, 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, https://doi.org/10.5281/zenodo.10848263, 2024. doi:10. 5281/zenodo.10848263. [12] G. Guo, H. Wang, D. Bell, Y. Bi, K. Greer, Knn model-based approach in classification, in: R. Meersman, Z. Tari (Eds.), On The Move to Meaningful Internet Systems 2003: CoopIS, DOA, and ODBASE (OTM 2003), volume 2888 of Lecture Notes in Computer Science, Springer, 2003, pp. 986–996. doi:10.1007/978-3-540-39964-3_64. [13] T. Ridnik, E. Ben-Baruch, N. Zamir, A. Noy, I. Friedman, M. Protter, L. Zelnik-Manor, Asymmetric loss for multi-label classification, in: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 82–91.