[論文レビュー] Overview of LifeCLEF Plant Identification task 2020

Hervé Goëau, Pierre Bonnet+1 more|Agritrop (Cirad)|Sep 23, 2025

Genomics and Phylogenetic Studies参考文献 2被引用数 12

ひとこと要約

この論文は、 LifeCLEF 2020 Plant Identification チャレンジを取り上げ、熱帯地域の field photo 識別を支援するための herbarium sheets を用いたクロスドメイン植物識別に焦点を当て、データセット、タスク設定、参加者の手法、結果、洞察を詳述する。

ABSTRACT

Automated identification of plants has improved considerably thanks to the recent progress in deep learning and the availability of training data with more and more photos in the field. However, this profusion of data only concerns a few tens of thousands of species, mostly located in North America and Western Europe, much less in the richest regions in terms of biodiversity such as tropical countries. On the other hand, for several centuries, botanists have collected, catalogued and systematically stored plant specimens in herbaria, particularly in tropical regions, and the recent efforts by the biodiversity informatics community made it possible to put millions of digitized sheets online. The LifeCLEF 2020 Plant Identification challenge (or "PlantCLEF 2020") was designed to evaluate to what extent automated identification on the flora of data deficient regions can be improved by the use of herbarium collections. It is based on a dataset of about 1,000 species mainly focused on the South America's Guiana Shield, an area known to have one of the greatest diversity of plants in the world. The challenge was evaluated as a cross-domain classification task where the training set consist of several hundred thousand herbarium sheets and few thousand of photos to enable learning a mapping between the two domains. The test set was exclusively composed of photos in the field. This paper presents the resources and assessments of the conducted evaluation, summarizes the approaches and systems employed by the participating research groups, and provides an analysis of the main outcomes.

研究の動機と目的

• Motivate and evaluate cross-domain plant identification bridging herbarium sheets and field photos in data-deficient tropical regions.
• Provide a large-scale dataset and task protocol to promote domain adaptation research for plant identification.
• Assess how state-of-the-art methods transfer knowledge across domains and evaluate genericity for rare species.
• Analyze submitted approaches to identify which strategies best handle few-field-photo scenarios.

提案手法

• Describe the PlantCLEF 2020 dataset with 997 species and 321,270 herbarium sheets plus 6,316 field photos for training.
• Define a cross-domain learning task where training uses herbarium sheets and limited field photos while test data are field photos.
• Evaluate submissions via Mean Reciprocal Rank (MRR) on the full test set and a difficult subset with few field photos.
• Analyze results to compare classical CNNs versus domain adaptation approaches (including adversarial and triplet/embedding-based methods).
• Discuss impact of external data and multi-task learning on performance and genericity.

実験結果

リサーチクエスチョン

• RQ1 Can herbarium-sheet data effectively transfer to field photo identification in data-scarce tropical flora?
• RQ2 What domain adaptation strategies best bridge the herbarium-field gap for plant species recognition?
• RQ3 How does external data and taxonomic information affect cross-domain plant identification performance?
• RQ4 Do multi-task and self-supervised auxiliary tasks improve identification of rare species?
• RQ5 What is the trade-off between overall performance and genericity for difficult species?

主な発見

• Best overall MRR across runs was 0.18, indicating a highly challenging task.
• Adversarial domain adaptation (FSADA) outperformed other approaches on the main MRR metric.
• Two-stream/embedding approaches using herbarium-field triplet losses achieved strong genericity across easy and difficult species.
• External data significantly boosted main MRR for some adversarial approaches, while multi-task setups leveraging taxonomy improved performance, especially for rare species.
• Explicit domain adaptation methods substantially outperformed pure CNN fine-tuning without adaptation in this cross-domain setting.
• An ensemble of FSADA variants yielded the best overall results among submissions.