---
pretty_name: DPD-Small for Out-of-Distribution Plant Disease Identification
task_categories:
- image-classification
language:
- en
tags:
- plant-disease
- out-of-distribution
- zero-shot-learning
- compositional-generalization
- agriculture
size_categories:
- 10K<n<100K
---

# DPD-Small: Out-of-Distribution Plant Disease Identification

**语言 / Language：** **English** | [简体中文](README.zh-CN.md)

## Dataset Summary

DPD-Small is a reduced version of the Deep-Plant-Disease (DPD) dataset used for the COS30082 Applied Machine Learning project. It supports experiments in identifying crop–disease combinations, including combinations absent from the training split. The project explores compositional generalization by predicting crop and disease concepts and combining them into crop–disease pairs.

The dataset contains **11,448 images** across three provided splits. These splits are intended for training and evaluating models on seen and unseen crop–disease combinations.

## Dataset Structure

```text
DPD_Small_train.csv
DPD_Small_test_seen.csv
DPD_Small_test_unseen.csv
DPD_Small_dataset/
    <image files>
```

| Split | CSV file | Images | Description |
| --- | --- | ---: | --- |
| Train | `DPD_Small_train.csv` | 8,000 | Training images from seen crop–disease combinations |
| Seen test | `DPD_Small_test_seen.csv` | 2,000 | Test images from combinations represented in training |
| Unseen test | `DPD_Small_test_unseen.csv` | 1,448 | Test images from 10 crop–disease combinations absent from training |

The project brief describes **55 crop categories**, **175 disease categories**, **323 seen crop–disease combinations**, and **10 unseen combinations**. Category indices start at zero.

### CSV metadata

According to the assignment specification, each CSV contains seven columns in this order:

1. Image filename
2. Crop category label
3. Disease category label
4. Crop–disease composition label
5. Crop name
6. Disease name
7. Crop–disease composition name

**Note:** The exact CSV header names and the relationship between image filenames and paths should be checked against the uploaded files. This card describes the provided specification rather than asserting unverified header names.

## Intended Use

- Training crop and disease classifiers on seen combinations.
- Evaluating recognition of previously unseen crop–disease combinations.
- Comparing seen and unseen Top-1/Top-5 accuracy and their harmonic mean.
- Studying robustness to variation in real-world plant images.

The unseen test images should **not** be used for model training or hyperparameter tuning. The project protocol allows use of unseen combination names as semantic side information.

## Download

This repository is intended to be downloaded as a regular Hugging Face dataset repository, preserving the original CSV and image files:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="print1Username/COS30082-Machine-Learning-Project.dataset",
    repo_type="dataset",
    local_dir="data",
    token=False,
)
```

Anonymous downloading requires the repository to be public and not gated.

## Dataset Limitations

The dataset is designed for research and coursework, not for clinical or agricultural diagnostic decisions without independent validation. Performance on the provided splits does not by itself establish reliable performance across all crops, diseases, environments, or imaging conditions.

## Source, Licensing, and Attribution

**To be completed by the dataset uploader before public release:**

- Original DPD dataset publication, authors, and source URL.
- License or redistribution permission for the images and metadata.
- Any required attribution and usage restrictions.

This card does **not** assert that the original data may be redistributed publicly. Do not assign a license without verifying the original dataset's terms.

## Project Context

Prepared for **COS30082 Applied Machine Learning**, *Out-of-Distribution Plant Disease Identification* (2026). The course assignment specification is the source for the split counts and label descriptions in this card.