deep-plants/AGM_HS

---
license: cc
dataset_info:
features:
- name: image
dtype: image
- name: mask
dtype: image
- name: crop_type
dtype: string
- name: label
dtype: string
splits:
- name: train
num_bytes: 22900031.321
num_examples: 6127
download_size: 22010079
dataset_size: 22900031.321
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
---
# Dataset Card for AGM_HS Dataset

## Dataset Summary
The AGM_HS (AGricolaModerna Healthy-Stress) Dataset is an extension of the AGM Dataset, specifically curated to address the challenge of detecting and localizing plant stress in top-view images of harvested crops. This dataset comprises 6,127 high-resolution RGB images, each with a resolution of 120x120 pixels, selected from the original AGM Dataset. The images in AGM_HS are divided into two categories: healthy samples (3,798 images) and stressed samples (2,329 images) representing 14 of the 18 classes present in AGM. Alongside the healthy/stressed classification labels, the dataset also provides segmentation masks for the stressed areas.

## Supported Tasks
Image classification: Healthy-stressed classification
Image segmentation: detection and localization of plant stress in top-view images.

## Languages
The dataset primarily consists of image data and does not involve language content. Therefore, the primary language is English, but it is not relevant to the dataset's core content.

## Dataset Structure
### Data Instances
A typical data instance from the AGM_HS Dataset consists of the following:
```
{
'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=120x120 at 0x29CEAD71780>,
'labels': 'stressed',
'crop_type': 'by'
'mask': <PIL.PngImagePlugin.PngImageFile image mode=L size=120x120 at 0x29CEAD71780>
}
```

### Data Fields
The dataset's data instances have the following fields:

- `image`: A PIL.Image.Image object representing the image.
- `labels`: A string representation indicating whether the image is "healthy" or "stressed."
- `crop_type`: An string representation of the crop type in the image
- `mask`: A PIL.Image.Image object representing the segmentation mask of stressed areas in the image, stored as a PNG image.

### Data Splits
- **Training Set**:
- Number of Examples: 6,127
- Healthy Samples: 3,798
- Stressed Samples: 2,329

## Dataset Creation
### Curation Rationale
The AGM_HS Dataset was created as an extension of the AGM Dataset to specifically address the challenge of detecting and localizing plant stress in top-view images of harvested crops. This dataset is essential for the development and evaluation of advanced segmentation models tailored for this task.

### Source Data
#### Initial Data Collection and Normalization
The images in AGM_HS were extracted from the original AGM Dataset. During the extraction process, labelers selected images showing clear signs of either good health or high stress. These sub-images were resized to 120x120 pixels to create AGM_HS.

### Annotations
#### Annotation Process
The AGM_HS Dataset underwent a secondary stage of annotation. Labelers manually collected images by clicking on points corresponding to stressed areas on the leaves. These clicked points served as prompts for the semi-automatic generation of segmentation masks using the "Segment Anything" technique \cite{kirillov2023segment}. Each image is annotated with a classification label ("healthy" or "stressed") and a corresponding segmentation mask.

### Who Are the Annotators?
The annotators for AGM_HS are domain experts with knowledge of plant health and stress detection.

## Personal and Sensitive Information
The dataset does not contain personal or sensitive information about individuals. It exclusively consists of images of plants.

## Considerations for Using the Data
### Social Impact of Dataset
The AGM_HS Dataset plays a crucial role in advancing research and technologies for plant stress detection and localization in the context of modern agriculture. By providing a diverse set of top-view crop images with associated segmentation masks, this dataset can facilitate the development of innovative solutions for sustainable agriculture, contributing to increased crop health, yield prediction, and overall food security.

### Discussion of Biases and Known Limitations
While AGM_HS is a valuable dataset, it inherits some limitations from the original AGM Dataset. It primarily involves images from a single vertical farm setting, potentially limiting the representativeness of broader agricultural scenarios. Additionally, the dataset's composition may reflect regional agricultural practices and business-driven crop preferences specific to vertical farming. Researchers should be aware of these potential biases when utilizing AGM_HS for their work.

## Additional Information
### Dataset Curators
The AGM_HS Dataset is curated by DeepPlants and AgricolaModerna. For further information, please contact us at:
- [e-mail]
- [e-mail]

### Licensing Information

### Citation Information
If you use the AGM_HS dataset in your work, please consider citing the following publication:

```bibtex
@InProceedings{Sama_2023_ICCV,
author = {Sama, Nico and David, Etienne and Rossetti, Simone and Antona, Alessandro and Franchetti, Benjamin and Pirri, Fiora},
title = {A new Large Dataset and a Transfer Learning Methodology for Plant Phenotyping in Vertical Farms},
booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
month = {October},
year = {2023},
pages = {540-551}
}
```

原始信息汇总

数据集卡片 for AGM_HS 数据集

数据集概述

AGM_HS（AGricolaModerna Healthy-Stress）数据集是AGM数据集的扩展，专门用于解决在收获作物的顶视图图像中检测和定位植物应激的挑战。该数据集包含6,127张高分辨率的RGB图像，每张图像的分辨率为120x120像素，从原始AGM数据集中精选而来。AGM_HS中的图像分为两类：健康样本（3,798张图像）和应激样本（2,329张图像），代表了AGM中的14个类别。除了健康/应激分类标签外，数据集还提供了应激区域的分割掩码。

支持的任务

• 图像分类：健康-应激分类
• 图像分割：在顶视图图像中检测和定位植物应激。

数据集结构

数据实例

AGM_HS数据集的一个典型数据实例包括以下内容： json { image: <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=120x120 at 0x29CEAD71780>, labels: stressed, crop_type: by, mask: <PIL.PngImagePlugin.PngImageFile image mode=L size=120x120 at 0x29CEAD71780> }

数据字段

数据集的数据实例具有以下字段：

• image：表示图像的PIL.Image.Image对象。
• labels：表示图像是否为“健康”或“应激”的字符串。
• crop_type：图像中作物类型的字符串表示。
• mask：表示图像中应激区域分割掩码的PIL.Image.Image对象，存储为PNG图像。

数据分割

• 训练集：
  • 样本数量：6,127
  • 健康样本：3,798
  • 应激样本：2,329

数据集创建

创建理由

AGM_HS数据集作为AGM数据集的扩展，专门用于解决在顶视图图像中检测和定位植物应激的挑战。该数据集对于开发和评估针对此任务的高级分割模型至关重要。

源数据

初始数据收集和规范化

AGM_HS中的图像从原始AGM数据集中提取。在提取过程中，标注者选择了显示良好健康或高应激迹象的图像。这些子图像被调整为120x120像素以创建AGM_HS。

标注

标注过程

AGM_HS数据集经历了第二阶段的标注。标注者通过点击叶子上对应应激区域的点来手动收集图像。这些点击点作为半自动生成分割掩码的提示，使用“Segment Anything”技术。每张图像都标注了分类标签（“健康”或“应激”）和相应的分割掩码。

标注者

AGM_HS的标注者是具有植物健康和应激检测知识的领域专家。

个人和敏感信息

该数据集不包含有关个人的个人或敏感信息。它仅包含植物的图像。

使用数据集的注意事项

数据集的社会影响

AGM_HS数据集在推进现代农业背景下植物应激检测和定位的研究和技术方面发挥着关键作用。通过提供多样化的顶视图作物图像和相关的分割掩码，该数据集可以促进可持续农业创新解决方案的发展，有助于提高作物健康、产量预测和整体粮食安全。

偏见和已知限制的讨论

虽然AGM_HS是一个有价值的数据集，但它继承了原始AGM数据集的一些限制。它主要涉及来自单一垂直农场设置的图像，可能限制了更广泛农业场景的代表性。此外，数据集的组成可能反映了特定于垂直农业的区域农业实践和商业驱动的作物偏好。研究人员在使用AGM_HS时应意识到这些潜在的偏见。