README.md

---
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<sub>HS</sub> (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<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub>.

### Annotations
#### Annotation Process
The AGM<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub> 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<sub>HS</sub> for their work.

## Additional Information
### Dataset Curators
The AGM<sub>HS</sub> Dataset is curated by DeepPlants and AgricolaModerna. For further information, please contact us at:
- nico@deepplants.com
- etienne.david@agricolamoderna.com

### Licensing Information

### Citation Information
If you use the AGM<sub>HS</sub> 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}
}
```