What is in it
Dataset · datatorch/ccmt-crop-pest · 5,325 files
CCMT Crop Pest and Disease Dataset (Tomato)
Tomato subset of the CCMT crop pest and disease dataset (Mensah et al., 2023): field photographs of tomato leaves from farms in Ghana, labelled per image with one of 5 classes: healthy, leaf blight, leaf curl, septoria leaf spot and verticulium wilt (the source's spelling of verticillium wilt). 5,325 raw 400x400 JPEG images out of 5,805 in the source's Raw Data/Tomato folders; 25 corrupt JPEGs that do not decode and 454 byte-identical duplicate copies were excluded, and one duplicate pair filed under two classes was left out entirely. Per class: septoria leaf spot 2,294, leaf blight 1,285, verticulium wilt 772, leaf curl 511, healthy 463. The augmented set is not included. Cashew, cassava and maize can follow in the same project. License: CC BY 4.0. Source: https://data.mendeley.com/datasets/bwh3zbpkpv/1. Please cite: Mensah P.K. et al., CCMT: Dataset for crop pest and disease detection, Data in Brief 49 (2023) 109306.
Open →
| Fact | Value |
|---|---|
| Source | CCMT: Dataset for crop pest and disease detection, Mendeley Data bwh3zbpkpv v1 |
| Crop | Tomato (the source also has cashew, cassava and maize) |
| Where | Field photos from local farms in Ghana, validated by expert plant virologists (source page) |
| Images in this copy | 5,325 JPEG, 400 x 400 |
| Labels | one class per image |
| Classes | septoria leaf spot 2,294, leaf blight 1,285, verticillium wilt 772, leaf curl 511, healthy 463 |
| License | CC BY 4.0 |
The counts are the hosted copy's. The source spells one class "verticulium wilt" in its folders, and the project keeps that spelling as the label name.
What was left out, and why
The source's raw tomato folders hold 5,805 files, more than the 5,435 its description gives. Of those, 25 are corrupt JPEGs that do not open, and 454 are exact duplicates of other files. One more image appears under two different classes, so both copies were left out rather than guess. The augmented images (rotations and other transforms the authors made for training) are not included. Cashew, cassava and maize may follow in the same project.
The classes are not balanced: septoria leaf spot is about five times the size of healthy. Plan for that when you split and train.
How to use it
- 1
Browse it
Open the dataset above and page through the leaves, or filter by class to see what each disease looks like in field light. No download needed.
- 2
Download it
Use Download dataset on the project page. A free account is needed. You get the images and their labels as a ZIP, or you can pull them with the DataTorch Python client.
- 3
Add your own labels
Upload the images to your own project and draw boxes or polygons on the lesions if you need detection or segmentation rather than one label per image.
Good uses
- Train or test a leaf disease classifier on real field photos, not lab shots on a plain background.
- Check class confusion. Leaf blight and septoria leaf spot can look alike at 400 x 400, which makes this a useful hard case.
- Start a detection set. The per-image classes tell you where to look. Drawing boxes on the lesions turns it into detection data.
- Compare with your own region. If you photograph tomatoes elsewhere, this is a baseline from Ghana's farms to compare against.
Source and license
Mensah Kwabena P., Akoto-Adjepong V., Adu K. et al., Dataset for Crop Pest and Disease Detection, Mendeley Data, version 1, https://doi.org/10.17632/bwh3zbpkpv.1. License: CC BY 4.0. Please cite: Mensah P.K. et al., CCMT: Dataset for crop pest and disease detection, Data in Brief 49 (2023) 109306.
Label your own field photos
Upload crop images and label them in the browser, with your team.
Frequently asked questions
What classes are in the CCMT tomato dataset?
Five, one per image. In the hosted copy, septoria leaf spot has 2,294 images, leaf blight 1,285, verticillium wilt 772, leaf curl 511 and healthy 463.
Is the CCMT dataset free to use?
Yes, it is published under CC BY 4.0 on Mendeley Data. You can reuse and adapt it, including commercially, as long as you credit the authors.
Are there bounding boxes or masks?
No. The labels are one class per image, so it suits image classification. If you need boxes or masks on the lesions, you can upload the images to your own DataTorch project and draw them there.
Why 5,325 images and not the full tomato set?
The source describes 5,435 raw tomato images. Its raw tomato folders hold 5,805 files, of which 25 do not decode and 454 are byte-identical copies. Those were left out, along with one pair filed under two different classes, which leaves 5,325.
Open the hosted copy on DataTorch.
Browse the files in your browser, then download the dataset with a free account.
