Pl@ntNet-300K-v2 image dataset

Pl@ntNet-300K-v2 is an image dataset designed for evaluating set-valued classification methods. It is derived from the Pl@ntNet citizen observatory database and comprises 306,087 images spanning 1,000 plant species.The new version provides improved images resolution and better naming for the species.

Key Features

The dataset reflects two notable characteristics, intrinsic to both the image acquisition process and the morphological diversity of plants:

Strong class imbalance: A few species account for the majority of images.

Visual similarity: Many species are visually indistinguishable, posing challenges even for expert identification.

These attributes make Pl@ntNet-300K-v2 particularly suited for benchmarking set-valued classification approaches.

Dataset Structure

Image Organization

Images are partitioned into train, test, and val sets, stored in directories labeled 0000 to 0999.

Metadata Files

1. plantnet300K_metadata.csv

Contains 306,087 image entries, each with the following fields:

Field
Description

species_id
Numerical species index (0–999)

PN_observation_id
Unique Pl@ntNet observation identifier

organ
Plant organ in the image (leaf, flower, other, habit, fruit, etc.)

author
Photographer’s identity

license
Image license type (cc-by-sa, cc-by-nc, cc-by-nc-sa)

split
Dataset partition (train, test, val)

PN_hash
Image hash name

2. species_metadata.csv

Provides taxonomic and conservation details for each species (aligned with the World Checklist of Vascular Plants, v13):

Field
Description

species_id
Numerical species index (matches image directories)

full_species
Full species name (with author)

species
Species name (without author)

genus
Genus name

family
Family name

epithet
Species epithet

author
Species name author(s)

unmatched_terms
Unresolved terms (e.g., spp., f.)

iucn_status
IUCN conservation status (e.g., EX, CR, EN, VU, LC, DD, NE)

Resources

Scientific Publication

The dataset and baseline results are described in:

Garcin et al. (2021), NeurIPS Datasets and Benchmarks

Utilities

PyTorch tools for data loading and model training:

GitHub Repository

Citation

If you use this dataset, please cite the following publication:

@inproceedings{Garcin_Joly_Bonnet_Affouard_Lombardo_Chouet_Servajean_Lorieul_Salmon2021,
author = {Garcin, C. and Joly, A. and Bonnet, P. and Affouard, A. and Lombardo, J.-C. and Chouet, M. and Servajean, M. and Lorieul, T. and Salmon, J.},
booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},
pdf = {https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/7e7757b1e12abcb736ab9a754ffb617a-Paper-round2.pdf},
title = {Pl@ntNet-300K: a plant image dataset with high label ambiguity and a long-tailed distribution},
year = {2021},
comment = {[<a href="https://github.com/plantnet/PlantNet-300K">Code</a>]}
}

Pl@ntNet-300K-v2 是一款专为评估集合值分类（set-valued classification）方法设计的图像数据集。其源自Pl@ntNet公民观测数据库，共包含覆盖1000个植物物种的306087张图像。该新版本优化了图像分辨率，并完善了物种命名规则。

### 核心特性
该数据集具备两项源自图像采集流程与植物形态多样性的显著特征：
1. 严重的类别不平衡：仅少数物种占据了绝大多数图像样本。
2. 视觉相似性：诸多物种外观高度相似，即便专业人员进行物种识别也颇具挑战。
上述特性使得Pl@ntNet-300K-v2 特别适合作为集合值分类方法的基准测试数据集。

### 数据集结构
#### 图像组织方式
图像被划分为训练集、测试集与验证集，存储于以0000至0999编号的目录中。

#### 元数据文件
1. `plantnet300K_metadata.csv`
该文件包含306087条图像条目，每条条目包含以下字段：
- `species_id`：物种数值索引（取值范围0–999）
- `PN_observation_id`：Pl@ntNet观测记录的唯一标识符
- `organ`：图像中的植物器官类型（叶片、花朵、其他、植株形态、果实等）
- `author`：拍摄者身份信息
- `license`：图像授权类型（如cc-by-sa、cc-by-nc、cc-by-nc-sa）
- `split`：数据集分区（训练集、测试集、验证集）
- `PN_hash`：图像哈希名称

2. `species_metadata.csv`
该文件提供了每个物种的分类学与保护学详情，其数据参照《维管植物世界名录》v13版本：
- `species_id`：物种数值索引（与图像目录编号保持一致）
- `full_species`：完整物种学名（含命名人信息）
- `species`：物种学名（不含命名人）
- `genus`：属名
- `family`：科名
- `epithet`：种加词
- `author`：物种命名人
- `unmatched_terms`：未解析术语（例如`spp.`、`f.`）
- `iucn_status`：IUCN保护等级（如EX、CR、EN、VU、LC、DD、NE）

### 资源
#### 学术出版物
该数据集与基准测试结果的相关研究发表于：Garcin 等人（2021），《NeurIPS 数据集与基准赛道》（NeurIPS Datasets and Benchmarks）。

#### 实用工具
用于数据加载与模型训练的PyTorch工具：GitHub 仓库。

#### 引用说明
若使用该数据集，请引用以下文献：
bibtex
@inproceedings{Garcin_Joly_Bonnet_Affouard_Lombardo_Chouet_Servajean_Lorieul_Salmon2021,
author = {Garcin, C. and Joly, A. and Bonnet, P. and Affouard, A. and Lombardo, J.-C. and Chouet, M. and Servajean, M. and Lorieul, T. and Salmon, J.},
booktitle = {Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks},
pdf = {https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/file/7e7757b1e12abcb736ab9a754ffb617a-Paper-round2.pdf},
title = {Pl@ntNet-300K: a plant image dataset with high label ambiguity and a long-tailed distribution},
year = {2021},
comment = {[<a href="https://github.com/plantnet/PlantNet-300K">Code</a>]}
}