[論文レビュー] Overview of PlantCLEF 2021: cross-domain plant identification

Hervé Goëau, Pierre Bonnet+1 more|ArXiv.org|Sep 23, 2025

Genomics and Phylogenetic Studies参考文献 1被引用数 50

ひとこと要約

LifeCLEF 2021 PlantCLEF チャレンジは、熱帯植物の標本写真から野外写真への横断的な植物識別を評価し、rare speciesの識別においてドメイン適応とメトリック学習を重視します。

ABSTRACT

Automated plant identification has improved considerably thanks to recent advances in deep learning and the availability of training data with more and more field photos. However, this profusion of data concerns only a few tens of thousands of species, mainly 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 systematically collected, catalogued and stored plant specimens in herbaria, especially in tropical regions, and recent efforts by the biodiversity informatics community have made it possible to put millions of digitised records online. The LifeCLEF 2021 plant identification challenge (or "PlantCLEF 2021") was designed to assess the extent to which automated identification of flora in data-poor regions can be improved by using herbarium collections. It is based on a dataset of about 1,000 species mainly focused on the Guiana Shield of South America, a region known to have one of the highest plant diversities in the world. The challenge was evaluated as a cross-domain classification task where the training set consisted of several hundred thousand herbarium sheets and a few thousand photos to allow learning a correspondence between the two domains. In addition to the usual metadata (location, date, author, taxonomy), the training data also includes the values of 5 morphological and functional traits for each species. The test set consisted exclusively of photos taken in the field. This article presents the resources and evaluations of the assessment carried out, summarises the approaches and systems used by the participating research groups and provides an analysis of the main results.

研究の動機と目的

• Assess cross-domain plant identification performance between herbarium sheets (source) and field photos (target) for tropical flora.
• Evaluate the impact of adding plant trait metadata to improve domain adaptation.
• Provide a benchmark and analysis of methods and results for data-poor species in tropical regions.

提案手法

• Use a large herbarium sheet dataset (321,270 sheets) and a smaller field photo set (6,316 train; 3,186 test) to learn domain mapping between two image domains.
• Introduce five functional traits as species-level metadata from the Encyclopedia of Life to guide training.
• Evaluate submissions with Mean Reciprocal Rank (MRR) on whole test set and on a difficult subset with few field photos.
• Participants employed domain adaptation and metric-learning approaches, including two-stream herbarium-field triplet loss networks (HTFL) and one-stream mixed networks (OSM).
• Organizer and participants augmented models with external data (e.g., PlantCLEF2019, GBIF) and multi-task cues (genus/family traits).
• Provide detailed working notes per submission to ensure reproducibility.

実験結果

リサーチクエスチョン

• RQ1 Can herbarium-based training transfer to field photos for data-poor tropical species?
• RQ2 Do trait-based auxiliary tasks improve cross-domain plant identification performance?
• RQ3 What is the impact of external training data on domain adaptation performance for rare species?

主な発見

Run	MRR (whole test set)	MRR (difficult species)
Neuon AI Run 7	0.181	0.158
Neuon AI Run 10	0.176	0.153
Neuon AI Run 8	0.169	0.150
Neuon AI Run 2	0.152	0.117
Neuon AI Run 9	0.147	0.129
Neuon AI Run 6	0.143	0.126
Neuon AI Run 5	0.137	0.116
Neuon AI Run 4	0.088	0.073
Neuon AI Run 1	0.071	0.066
LU Run 9	0.065	0.037
Domain Run 8	0.065	0.037
LU Run 8	0.065	0.037
LU Run 7	0.063	0.040
Domain Run 4	0.063	0.046
Neuon AI Run 3	0.060	0.056
Domain Run 6	0.060	0.039
LU Run 6	0.057	0.042
To Be Run 2	0.056	0.038
LU Run 4	0.051	0.031
LU Run 3	0.050	0.026
To Be Run 1	0.045	0.022
Domain Run 5	0.038	0.011
LU Run 5	0.037	0.022
LU Run 1	0.036	0.009
LU Run 2	0.034	0.013
Domain Run 3	0.031	0.015
Domain Run 1	0.019	0.009
Domain Run 2	0.018	0.009
Domain Run 7	0.006	0.003
Organizer’s submission 7	0.198	0.093
Organizer’s submission 6	0.192	0.091
Organizer’s submission 4	0.188	0.073
Organizer’s submission 5	0.184	0.077
Organizer’s submission 3	0.183	0.073
Organizer’s submission 2	0.153	0.056
Organizer’s submission 1	0.052	0.042

• Cross-domain identification remains very challenging; the best single model reaches MRR around 0.18–0.20 on the full test set, with lower performance on difficult species.
• Domain-adaptation CNN-based approaches outperform traditional CNNs but still struggle, especially for rare field-photo species.
• External data substantially improves performance (e.g., organizer’s runs rise from 0.052 to 0.153 when adding external data).
• Two-stream herbarium-field triplet loss networks with ensemble strategies achieve strong generalization across species with varying field-photo availability.
• Multi-task extensions using genus, family, and trait information provide measurable gains in MRR.
• The organizer’s multi-discriminator/sub-task submissions (traits and taxonomic levels) yield the top primary MRR among organizer runs.