Datasets:

geraldmc
/

plantdoc-full

license: cc-by-4.0
task_categories:
  - image-classification
language:
  - en
tags:
  - plant-disease
  - agriculture
  - field-conditions
  - computer-vision
size_categories:
  - 1K<n<10K
pretty_name: PlantDoc (full)

PlantDoc — full variant

A curated mirror of the PlantDoc plant disease classification dataset (Singh et al. 2020), with normalized class metadata and a stable schema, hosted as a Hugging Face Dataset for reproducible distribution. The companion plantdoc-tiny variant is a 164-image stratified subsample of this dataset for fast test-suite use.

PlantDoc was built to address a specific failure mode in earlier plant disease datasets like PlantVillage: lab-condition images don't predict field-condition performance. PlantDoc images are web-scraped photos of diseased and healthy leaves in real-world settings — heterogeneous in resolution, lighting, background, and capture device. That heterogeneity is the point.

Quick start

from datasets import load_dataset

ds = load_dataset("geraldmc/plantdoc-full", revision="v0.1.0", split="train")
print(len(ds))                # 2578
print(ds.features)            # image + 7 metadata columns
print(ds[0]["class_label"])   # 'Apple Scab Leaf'

Inside the iResearch Institute 2026 Virtual Lab, the typical entry point is:

import irilab2026 as iri

metadata_df, hf_dataset = iri.load_plantdoc()

What's in this dataset

2,578 images across 28 classes covering 13 host plant species
Train/test split shipped intact from the upstream repository — 2,342 train + 236 test
Heterogeneous resolution (189–4,000 px width, 194–4,272 px height), heterogeneous lighting, heterogeneous backgrounds — by design
Mostly RGB, some CMYK color modes — downstream loaders should .convert("RGB") defensively

Schema

Column	Type	Description	Example
`image`	image	PIL Image at original upstream resolution	(W, H) JPEG
`class_label`	string	Upstream folder name, verbatim	`"Apple Scab Leaf"`
`class_idx`	int64	0–27, case-sensitive alphabetical sort over `class_label`	`0`
`host`	string	Normalized host name	`"Apple"`
`disease`	string	Lowercased disease name, or `"healthy"` for healthy leaves	`"scab"`
`is_healthy`	bool	True iff the class is a healthy leaf	`False`
`split`	string	`"train"` or `"test"`, from the upstream partition	`"train"`
`filename`	string	Original filename, verbatim (URL-encoded chars and double extensions preserved)	`"052609%20Hartman%20Crabapple%20scab%20single%20leaf.JPG.jpg"`

About `host`

Hosts are normalized from upstream folder names with one transformation: underscores become spaces (Bell_pepper → "Bell pepper"). Capitalization and spelling are otherwise preserved as upstream had them. Notable inheritances from upstream:

Soyabean retains the upstream misspelling (canonical is "Soybean") so the normalized name links unambiguously to the original folder.
grape retains the upstream lowercase. Other hosts are Title Case.
Corn retains the upstream naming (the more standard botanical name is "Maize") to match Singh et al. 2020 and downstream benchmark papers.

If your downstream code needs cross-dataset matching (e.g., aligning PlantDoc classes to PlantVillage classes), do that normalization explicitly in your code rather than relying on these column values to be canonical — that step is meaningful research methodology and should be visible in your work.

About `disease`

Disease names are lowercased even when the upstream had Title Case (Powdery mildew → "powdery mildew", Septoria leaf spot → "septoria leaf spot"). Two non-obvious mappings worth flagging:

Tomato mold leaf → "mold" (the canonical disease is leaf mold, caused by Passalora fulva, but the upstream label format doesn't say "leaf mold" — the column stays close to upstream literal naming).
Tomato two spotted spider mites leaf → "two spotted spider mites" (a pest, not strictly a plant disease, but it's a PlantDoc class).

About `class_idx`

Class indices 0–27 are assigned by case-sensitive alphabetical sort over class_label values. With case-sensitive sort, lowercase grape leaf and grape leaf black rot land at indices 26 and 27, after the uppercase Title Case classes.

Known caveats

One class has 2 training images and 0 test images

Tomato two spotted spider mites leaf appears in train/ with 2 images and is entirely absent from test/ in the upstream repository. Singh et al. 2020 and downstream benchmark papers (e.g., Ahmad et al. 2023) report 27 classes — they implicitly drop this orphan. This dataset preserves all 28 classes for upstream fidelity. If you're benchmarking against the literature, drop the orphan explicitly:

df = df[df["class_label"] != "Tomato two spotted spider mites leaf"]

Class names are inconsistent in capitalization, word order, and the position of "leaf"

The upstream folder names are not internally consistent. Examples that all appear in this dataset: Apple Scab Leaf (Title Case, disease before "Leaf"), Tomato Septoria leaf spot (mixed case, "leaf spot" as a compound), grape leaf (all lowercase), Tomato mold leaf (disease before "leaf"), Tomato leaf bacterial spot (host "leaf" disease). The host and disease columns are the hand-curated normalization; the class_label column preserves the upstream string verbatim.

Filenames have quirks worth knowing about

About 6% of filenames contain URL-encoded characters (%20 for spaces, %2C for commas) — artifacts of however the upstream curators saved web-scraped images. Roughly 12% have double extensions like .JPG.jpg or .jpeg.jpg. The filename column preserves these exactly because they're identifiers — the link back to the upstream filename is exact and unambiguous.

Per-class test set sizes are small

Test split sizes range from 4 (Corn Gray leaf spot) to 12 (Corn leaf blight, grape leaf), with a median around 9. Per-class accuracy estimates on this dataset will have wide confidence intervals — frame your conclusions accordingly. Aggregate metrics (averaged across classes or grouped by host family) are more reliable than per-class numbers.

About `class_count = 28` but the README of upstream says 17

The upstream README at pratikkayal/PlantDoc-Dataset claims "13 plant species and up to 17 classes of diseases." That's 17 disease classes; the dataset includes healthy-leaf classes for many hosts, bringing the total to 28 (or 27 if the orphan is dropped — see above).

Build provenance

This dataset was built by:

Cloning pratikkayal/PlantDoc-Dataset at commit 5467f6012d78 (on a case-sensitive Linux filesystem — see note below).
Walking train/ and test/ directory trees to enumerate all image files.
Applying a hand-curated 28-row class-normalization lookup to produce the host, disease, and is_healthy columns.
Assigning class_idx by case-sensitive alphabetical sort over distinct class_label values.
Constructing an HF Dataset with the Image() feature and pushing to geraldmc/plantdoc-full with revision tag v0.1.0.

Build script: scripts/build_pd_full_hf.py in the irilab2026 repository.

A note about case-sensitive filesystems

The upstream repository contains 6 pairs of files whose names differ only in case (e.g., CAR1.jpg and car1.jpg in Apple rust leaf/). On case-insensitive filesystems (default macOS APFS/HFS+, default Windows NTFS) git clone silently drops one file per pair and produces a 2,572-image working tree instead of 2,578. This dataset was built on Colab's case-sensitive Linux filesystem to preserve all 2,578 upstream images. If you rebuild from scripts/build_pd_full_hf.py, do so on a case-sensitive filesystem.

License

This curated mirror is released under CC BY 4.0, matching the license of the upstream pratikkayal/PlantDoc-Dataset.

Attribution must include both the upstream dataset (Singh et al. 2020) and this curated mirror. See Citation below.

Citation

@inproceedings{singh2020plantdoc,
  title     = {{PlantDoc}: A Dataset for Visual Plant Disease Detection},
  author    = {Singh, Davinder and Jain, Naman and Jain, Pranjali and
               Kayal, Pratik and Kumawat, Sudhakar and Batra, Nipun},
  booktitle = {Proceedings of the 7th ACM IKDD CoDS and 25th COMAD},
  pages     = {249--253},
  year      = {2020}
}

When citing the curated mirror itself, additionally reference this Hugging Face Dataset by its repo ID and revision tag (geraldmc/plantdoc-full @ v0.1.0).

Related resources

geraldmc/plantdoc-tiny — 164-image stratified subsample for test-suite use
geraldmc/plantvillage-full — the lab-condition counterpart dataset; PlantDoc is most commonly used as a transfer test for classifiers trained on this
pratikkayal/PlantDoc-Dataset — the upstream GitHub repository