Voxel51/PlantSeg-Test

Dataset Card for PlantSeg_Test

This is a FiftyOne dataset with 1200 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/PlantSeg-Test")

# Launch the App
session = fo.launch_app(dataset)

Dataset Card for PlantSeg

Dataset Details

Dataset Description

PlantSeg is a large-scale in-the-wild dataset for plant disease segmentation, containing 11,458 images with high-quality segmentation masks across 115 disease categories and 34 plant types. Unlike existing plant disease datasets that are collected in controlled laboratory settings, PlantSeg primarily comprises real-world field images with complex backgrounds, various viewpoints, and different lighting conditions. The dataset also includes an additional 8,000 healthy plant images categorized by plant type.

• Curated by: Tianqi Wei, Zhi Chen, Xin Yu, Scott Chapman, Paul Melloy, and Zi Huang
• Shared by: The University of Queensland; CSIRO Agriculture and Food
• Language(s) (NLP): en
• License: CC BY-NC-ND 4.0

Dataset Sources [optional]

• Repository: https://doi.org/10.5281/zenodo.13293891
• Paper [optional]: arXiv:2409.04038

Uses

Direct Use

• Training and benchmarking semantic segmentation models for plant disease detection
• Developing automated disease diagnosis systems for precision agriculture
• Image classification for plant disease identification
• Evaluating segmentation algorithms on in-the-wild agricultural imagery
• Supporting integrated disease management (IDM) decision-making tools

Dataset Structure

The dataset is organized as follows:

• images/: Plant disease images in JPEG format
• annotations/: Segmentation labels in PNG format (grayscale, where diseased pixels have class index values and background is zero)
• json/: Original LabelMe annotation files in JSON format
• PlantSeg-Meta.csv: Metadata file containing image name, plant type, disease type, resolution, label file path, mask ratio, source URL, and train/test split assignment

Statistics:

• Total images: 11,458 diseased plant images + 8,000 healthy plant images
• Disease categories: 115
• Plant types: 34
• Train/test split: 80/20 (stratified by disease type)

Plant categories are organized into four socioeconomic groups:

• Profit crops (e.g., Coffee, Tobacco): 9 diseases across 3 plants
• Staple crops (e.g., wheat, corn, potatoes)
• Fruits (e.g., apples, oranges): 39 diseases across 10 plants
• Vegetables (e.g., tomatoes): 45 diseases across 15 plants

Dataset Creation

Curation Rationale

Existing plant disease datasets are insufficient for developing robust segmentation models due to three key limitations:

1. Annotation Type: Most datasets only contain class labels or bounding boxes, lacking pixel-level segmentation masks
2. Image Source: Many datasets contain images from controlled laboratory settings with uniform backgrounds, which do not reflect real-world field conditions
3. Scale: Existing segmentation datasets are small and cover limited host-pathogen relationships

PlantSeg addresses these gaps by providing the largest in-the-wild plant disease segmentation dataset with expert-validated annotations.

Source Data

Data Collection and Processing

Images were collected using plant disease names as keywords from multiple internet sources:

• Google Images
• Bing Images
• Baidu Images

This multi-source collection strategy ensured geographic diversity, with images sourced from websites worldwide. After collection, a rigorous data cleaning process was conducted where annotators reviewed each image and removed incorrect or ambiguous images, with cross-validation by at least two annotators and expert review for discrepancies.

Who are the source data producers?

Images were sourced from websites globally, representing diverse geographic regions, environmental conditions, and imaging setups. The original photographers/sources are not individually identified, but source URLs are preserved in the metadata for reproducibility and copyright compliance.

Annotations [optional]

Annotation process

1. Standard establishment: A segmentation annotation standard was created to ensure consistent labeling of disease-affected areas
2. Annotator training: Annotators were trained on the standard and required to annotate 10 test images for evaluation before proceeding
3. Annotation tool: LabelMe (V5.5.0) was used for polygon annotation
4. Annotation guidelines:
   • Distinct lesions: annotated with individual polygons
   • Overlapping lesions: annotated as combined affected areas
   • Small clustered symptoms (rust, powdery mildew): meticulously annotated to reflect disease distribution
   • Disease-induced deformities: also annotated
5. Quality control: Each image subset was annotated by one annotator, then reviewed by another annotator, with final review by expert plant pathologists

Who are the annotators?

• 10 trained annotators who passed qualification evaluations
• Supervised by two expert plant pathologists who established standards, evaluated annotator work, and performed final reviews

Citation

BibTeX:

@article{wei2024plantseg,
  title={PlantSeg: A Large-Scale In-the-wild Dataset for Plant Disease Segmentation},
  author={Wei, Tianqi and Chen, Zhi and Yu, Xin and Chapman, Scott and Melloy, Paul and Huang, Zi},
  journal={arXiv preprint arXiv:2409.04038},
  year={2024}
}

APA: Wei, T., Chen, Z., Yu, X., Chapman, S., Melloy, P., & Huang, Z. (2024). PlantSeg: A Large-Scale In-the-wild Dataset for Plant Disease Segmentation. arXiv preprint arXiv:2409.04038.