kunley2/plant_image_degradation

Plant Disease Degradation 224

Dataset Summary

This dataset contains 224×224 plant disease images and their synthetically degraded variants for image-classification and robustness experiments. It is designed for transfer-learning workflows that require a fixed input size of 224×224 while still exposing models to realistic quality issues that occur in field capture.

The degraded images are generated from clean plant-leaf images using a field-oriented degradation pipeline. Instead of applying arbitrary corruption, the pipeline focuses on distortions that are plausible for smartphone and in-field agricultural imaging, such as:

• illumination changes
• mild blur
• sensor-like noise
• JPEG compression
• resolution loss simulated by downsample-then-restore
• mild occlusion or shadow
• small color-temperature shifts

All final saved images remain 224×224, which makes the dataset suitable for common transfer-learning backbones such as ResNet, VGG, DenseNet, MobileNet, and EfficientNet.

This repository is intended for:

• plant disease classification
• robustness benchmarking
• cross-domain generalization experiments
• degradation-aware training
• ablation studies on image quality sensitivity

Supported Tasks and Leaderboards

• Image classification
• Robustness evaluation under image degradation
• Domain generalization / deployment-readiness testing

Languages

This dataset contains images and metadata only. The metadata and card are written in English.

Dataset Structure

A practical repository layout is:

PlantDiseaseDegradation224/
├── README.md
├── metadata/
│   ├── degradations.csv
│   └── degradations.json
├── train/
│   ├── Apple___Black_rot/
│   │   ├── img_001.png
│   │   └── ...
│   ├── Apple___healthy/
│   └── ...
├── val/
│   └── ...
└── test/
    └── ...

If you are uploading only degraded images, use a class-folder structure so the dataset can be loaded with an ImageFolder-style workflow. Hugging Face documents ImageFolder as a no-code option for image datasets, including class-folder layouts and optional metadata files. citeturn860043search2turn860043search8

If you want to keep both clean and degraded versions, another practical layout is:

PlantDiseaseDegradation224/
├── README.md
├── clean/
│   ├── train/
│   ├── val/
│   └── test/
├── degraded/
│   ├── train/
│   ├── val/
│   └── test/
└── metadata/
    ├── degradations.csv
    └── degradations.json

Data Instances

Each image is a color image with shape 224×224×3.

Typical metadata fields in degradations.csv or degradations.json include:

• source_image: relative path to the original source image
• class_name: class label
• degraded_image: relative path to the degraded image
• severity: mild, moderate, severe, or mixed policy result
• output_height: expected to be 224
• output_width: expected to be 224
• num_steps: number of applied degradation steps
• steps_json: serialized list of applied operations and parameters

Example row:

{
  "source_image": "train/Tomato___Late_blight/img_0123.jpg",
  "class_name": "Tomato___Late_blight",
  "degraded_image": "images/Tomato___Late_blight/img_0123__deg_moderate_00.png",
  "severity": "moderate",
  "output_height": 224,
  "output_width": 224,
  "num_steps": 5,
  "steps_json": [
    {"op": "resize_with_padding", "target_size": [224, 224]},
    {"op": "resize_restore", "scale": 0.67},
    {"op": "gaussian_blur", "kernel": 3},
    {"op": "brightness_contrast", "alpha": 0.94, "beta": -9},
    {"op": "jpeg", "quality": 68}
  ]
}

Data Fields

Image fields

• image: RGB image at 224×224
• label: disease or healthy class label

Metadata fields

• source_image: source filename or relative path
• degraded_image: saved degraded image path
• severity: degradation severity tier
• steps_json: exact ordered degradation recipe used for that image

Class Labels

Replace this section with your final class list.

Example placeholder format:

0  Apple___Black_rot
1  Apple___healthy
2  Tomato___Early_blight
3  Tomato___Late_blight
4  Tomato___healthy
...

Dataset Creation

Source Data

This dataset is derived from an existing plant disease image collection. You should name the original dataset here, describe how it was obtained, and confirm that redistribution is permitted under the source license.

Suggested fields to add before publication:

• source dataset name
• source citation
• source URL
• collection conditions
• camera/device information if known
• class distribution
• train/validation/test split policy

Degradation Pipeline

The degraded variants were produced with a plant-disease-specific synthetic degradation pipeline that preserves final image size at 224×224 while introducing realistic capture defects.

Core operations may include:

• resize with aspect-ratio preservation and padding to 224×224
• downsample then restore to simulate resolution loss
• Gaussian blur
• mild motion blur
• Gaussian noise
• Poisson-like noise
• brightness and contrast shifts
• gamma change
• slight color-temperature shift
• shadow simulation
• JPEG compression
• mild occlusion

The design goal is to mimic field conditions without destroying disease evidence such as lesion boundaries, chlorotic regions, vein patterns, or texture cues.

Why 224×224?

The dataset is standardized to 224×224 so it can plug directly into common transfer-learning pipelines. This avoids shape mismatches during training and keeps experiments comparable across architectures.

Recommended Uses

This dataset is appropriate for:

• training plant disease classifiers with degradation-aware augmentation
• benchmarking robustness to real-world image quality issues
• evaluating performance drop under mild, moderate, and severe degradation
• comparing clean-only training versus mixed clean+degraded training
• studying which disease classes are most sensitive to image quality loss

Out-of-Scope Uses

This dataset is not intended to:

• diagnose plant disease in production without expert validation
• replace agronomist or plant pathologist review
• support safety-critical crop treatment decisions without additional validation
• claim universal generalization across crops, regions, devices, or diseases

Bias, Risks, and Limitations

• Synthetic degradations are approximations of field capture and may not cover every real-world failure mode.
• If the source dataset is lab-curated, the degraded variants may still differ from true field imagery.
• Some diseases depend on very small lesions or texture patterns; aggressive blur or compression may erase class-defining evidence.
• Class imbalance in the original dataset may remain or worsen after degradation if not managed carefully.
• The dataset may inherit geographic, seasonal, crop-variety, or acquisition-device bias from the original source data.

Preprocessing and Loading

Hugging Face documents image-dataset creation through ImageFolder layouts and local Python workflows, including Dataset.push_to_hub() for Python-created datasets. citeturn860043search2

Example local loading with datasets:

from datasets import load_dataset

dataset = load_dataset("imagefolder", data_dir="./PlantDiseaseDegradation224/train")

If your repository includes splits such as train/, val/, and test/, you can point data_files or separate data_dir values accordingly.

Hub Upload Notes

On the Hugging Face Hub, dataset cards live in README.md and use YAML front matter for metadata. Hugging Face’s Hub documentation also recommends upload_folder() in huggingface_hub as a standard way to upload a local folder to the Hub. citeturn860043search1turn860043search2

Citation

Replace this with the citation you want users to use.

@dataset{yourname_plant_disease_degradation_224,
  title={Plant Disease Degradation 224},
  author={Your Name},
  year={2026},
  publisher={Hugging Face}
}

Acknowledgements

Acknowledge the original dataset creators here if this dataset is derived from an existing public collection.

License

Set this section to match the actual redistribution terms of your source dataset and your own added annotations or transformations.