geraldmc/plantvillage-full

PlantVillage (full)

A curated re-release of the PlantVillage plant-disease image dataset (Mohanty, Hughes, Salathé 2016), with structured per-image metadata and a leaf-grouped train/test split. Built for the iResearch Institute 2026 Virtual Lab mentorship's Rationale 2 track. The companion debug-grade subset is at geraldmc/plantvillage-tiny.

What's in this dataset

54,304 images of plant leaves, photographed against plain backgrounds under controlled lighting, across 38 classes — each class is a unique combination of a host plant species and a disease (or "healthy"). The image set is identical to the raw/color/ directory of the upstream PlantVillage repository at the commit this release was built from. The contribution of this release is the metadata layer (host/disease parsing, leaf-of-origin grouping, the deterministic train/test split) and the packaging as Parquet shards for fast streaming access.

How to use it

from datasets import load_dataset

ds = load_dataset("geraldmc/plantvillage-full", revision="v0.1.0", split="train")
example = ds[0]
example["image"]         # PIL Image, 256x256 RGB
example["class_label"]   # "Apple___Apple_scab"

The dataset has a single underlying "train" HF split. The "train" name is HF's default wrapping; the train-vs-held-out split for analysis lives in the split metadata column (see Data fields below).

For users of the irilab2026 library, the wrapper function iri.load_plantvillage() returns a (metadata_df, hf_dataset) tuple and provides a PyTorch Dataset wrapper. See the library's documentation.

Data fields

Field	Type	Description
image	Image	The PV image. 256×256 pixel JPEG, RGB, uniformly sized across the dataset.
class_label	string	Combined host + disease folder name from upstream, e.g. Tomato___Tomato_mosaic_virus.
class_idx	int	Integer 0–37 assigned by alphabetical sort over class_label. The integer form a classifier predicts.
host	string	Host plant species, e.g. Tomato. Parsed from the left side of class_label split on ___.
disease	string	Disease name, e.g. Tomato mosaic virus or healthy. Parsed from the right side; underscores replaced with spaces.
split	string	"train" (43,356 rows) or "test" (10,948 rows). Leaf-grouped 80/20 split — see Build provenance below.
leaf_id	string	Identifier for the physical leaf the image was taken from. Globally unique across classes.
leaf_grouped	bool	True if leaf_id reflects upstream grouping (from filtered_leafmaps/); False if synthetic (one per image).

Build provenance

Built from spMohanty/PlantVillage-Dataset at commit 7f7ecc7e1eaca78107e3affe7cb5abd9427e139a by scripts/build_pv_full_hf.py in the irilab2026 repository. The build steps:

1. Shallow-clones the upstream repository.
2. Walks raw/color/ to enumerate images and parse class labels.
3. For each of the 30 classes that have a filtered_leafmaps/ CSV, attaches the leaf-of-origin metadata as leaf_id and sets leaf_grouped = True.
4. For the remaining 8 classes, assigns each image its own synthetic leaf_id and sets leaf_grouped = False.
5. Shuffles leaf_id values with a fixed seed and produces a leaf- grouped 80/20 train/test split. Leaf-grouped means all images of a single physical leaf go to either train or test, never both — a classifier evaluated on the test set has not seen any image of any leaf it's being tested on.
6. Packages the result as Parquet shards and pushes to this repo.

The split ratio is 79.84% / 20.16% (43,356 / 10,948) rather than exactly 80/20 because leaf-grouped rounding can't always hit the nominal target.

Known caveats

Image count discrepancy. Mohanty et al. (2016) reports 54,303 images. The current upstream raw/color/ contains 54,304 images. The provenance of the one-image difference has not been traced; if it matters for your analysis, audit your loaded dataset directly rather than citing the paper's count.

Tomato___Target_Spot is effectively ungrouped despite having a leafmap CSV. The class's filtered leafmap was generated from a different upstream snapshot than the current raw/color/, so every filename lookup against it fails (1,404 entries, 1,404 misses). The build script falls through to the synthetic-leaf-ID path for this class, but leaf_grouped is set per the existence of the leafmap file, not per whether the lookups succeeded. If your analysis depends on leaf_grouped reflecting actual upstream grouping rather than "a leafmap CSV exists upstream," filter out this class explicitly or check whether the column's value matches your analytical intent.

In practice this means net leaf-grouping coverage is 29 classes (grouped, lookups work) plus 9 classes (synthetic), rather than the headline 30 + 8 from the class-name count.

License

CC0 1.0 (Public Domain Dedication), inherited from the upstream PlantVillage release.

Citation

If you use this dataset in published work, cite the original PlantVillage paper:

@article{mohanty2016using,
  title={Using deep learning for image-based plant disease detection},
  author={Mohanty, Sharada P and Hughes, David P and Salath{\'e}, Marcel},
  journal={Frontiers in Plant Science},
  volume={7},
  pages={1419},
  year={2016},
  publisher={Frontiers Media SA},
  doi={10.3389/fpls.2016.01419}
}

Curated and packaged by Gerald McCollam for the iResearch Institute 2026 Virtual Lab.