[Paper Review] Overview of PlantCLEF 2022: Image-based plant identification at global scale

Hervé Goëau, Pierre Bonnet+1 more|ArXiv.org|Sep 22, 2025

Smart Agriculture and AI23 citations

TL;DR

This paper surveys PlantCLEF 2023 (LifeCLEF 2023 Plant Identification task) and analyzes large-scale image-based plant identification using CNNs and Vision Transformers, highlighting SSL-pretrained ViTs and the benefit of web data in training. It also presents a main results table from participating runs.

ABSTRACT

It is estimated that there are more than 300,000 species of vascular plants in the world. Increasing our knowledge of these species is of paramount importance for the development of human civilization (agriculture, construction, pharmacopoeia, etc.), especially in the context of the biodiversity crisis. However, the burden of systematic plant identification by human experts strongly penalizes the aggregation of new data and knowledge. Since then, automatic identification has made considerable progress in recent years as highlighted during all previous editions of PlantCLEF. Deep learning techniques now seem mature enough to address the ultimate but realistic problem of global identification of plant biodiversity in spite of many problems that the data may present (a huge number of classes, very strongly unbalanced classes, partially erroneous identifications, duplications, variable visual quality, diversity of visual contents such as photos or herbarium sheets, etc). The PlantCLEF2022 challenge edition proposes to take a step in this direction by tackling a multi-image (and metadata) classification problem with a very large number of classes (80k plant species). This paper presents the resources and evaluations of the challenge, summarizes the approaches and systems employed by the participating research groups, and provides an analysis of key findings.

Motivation & Objective

• Motivate automated, global-scale plant species identification to support biodiversity monitoring.
• Describe the two large training datasets (trusted and web) and the test set used for evaluation.
• Summarize participating methods, architectures, and training strategies for 80,000 species.
• Analyze key findings to guide future large-scale plant identification research.

Proposed method

• Describe dataset construction: trusted GBIF-derived curated images and noisy web images totaling 4 million images across 80k species.
• Evaluate a plant identification task using multi-image observations and macro-averaged mean reciprocal rank (MA-MRR) as the metric.
• Survey participating methods, focusing on architecture choices (CNNs vs Vision Transformers) and pretraining (STL vs SSL, EVA, MAE).
• Report results from a predefined leaderboard of runs and compare performance across setups.
• Analyze the impact of web data, organ-based sub-models, and pretraining strategies on MA-MRR.

Experimental results

Research questions

• RQ1 What are the effects of large-scale, multi-image training sets on plant species retrieval performance?
• RQ2 Do Vision Transformers with self-supervised pretraining outperform CNNs with supervised transfer learning on this large-scale plant identification task?
• RQ3 How does including noisy web data influence model accuracy and generalization across many species?
• RQ4 What training configurations (taxonomy levels, organs, pretraining) yield the best MA-MRR on large-scale plant identification?

Key findings

• Vision Transformer models pretrained with self-supervised learning (EVA/MXE approaches) achieved the highest MA-MRR scores, outperforming CNN-based solutions.
• Best EVA-based approach (MingleXuRun8) achieved MA-MRR of 0.67395.
• Incorporating the Web training set improved performance substantially compared to using trusted data alone (e.g., 0.65035 to 0.67395).
• Removing fewer-populated species from training (species ablation) reduced performance, indicating the importance of including all species.
• Combining organ-specific models generally reduced performance due to shrinking species coverage.
• CNN-based methods reached up to 0.61813 MA-MRR, still below the best EVA results.