DigiGreen/Crop-Disease-Image-Eval-Synthetic

Crop, Category, Disease and Pest Test Set

11,057 smallholder-farmer photographs sent to FarmerChat from Ethiopia, India, Kenya and Nigeria, each labelled with the crop, whether the problem is a disease or a pest, and which one. This is the held-out test split of a four-head classification benchmark, restricted to the rows whose labels came from an independent model council rather than from the production vendor.

Why 11,057 and not 16,275

The full held-out split is 16,275 images. Labels were built by a council of eight vision-language models with a consensus vote; where the council could not agree, the label fell back to the production vendor's own answer. Any system compared against those fallback rows is partly graded against the vendor, which makes the comparison circular.

This release keeps only rows where neither the crop nor the disease label came from that fallback: 5,189 rows are dropped for a vendor-sourced disease label and a further 27 for a vendor-sourced crop label, leaving 11,057. No vendor output is redistributed here.

Columns

Column	Meaning
image_id	stable identifier for the photograph
crop	one of 110 crops
category	disease, pest/weed or healthy
diagnosis	the disease or pest name

Composition: 7,449 disease rows and 3,608 pest rows. There are no healthy rows, because the council never returns healthy and every healthy label in the source data was vendor-sourced.

Label space

label_space/ carries the four vocabularies the benchmark is defined over (110 crops, 3 categories, 285 diseases including healthy, 92 pests) and crop_disease_mask.npy, a boolean [110, 285] table of which diseases are recorded on which crop. That mask is derived from production data and is reusable on its own: it is the part of this release most likely to be useful outside the benchmark.

What is and is not in this release

Included: the labels, the image identifiers, and the label space.

Not included: the photographs themselves. They are images of smallholder farmers' fields, and releasing them needs a review pass for faces, readable names, and the location stamps some camera apps burn into the pixels. That review is a separate piece of work. Until it is done, this repository is the label set and the benchmark definition rather than a drop-in image dataset.

How the labels were made

Eight vision-language models labelled every image independently. Five voted on the final label; three were held out of the vote so that no model being evaluated helped write the label it would be scored against. Answers were normalised onto one vocabulary, with synonym merging, before a confidence-weighted vote. Disagreements and rare classes were routed to human review.

These are consensus labels, not expert annotation. A score against them measures agreement with that council. Treat them as a benchmark for comparing systems under one rule, not as clinical ground truth.

Known limitations

• Crop coverage is uneven: a small number of crops carry most rows and many have very few.
• Photographs come from one advisory service in four countries and are not a random sample of smallholder agriculture.
• Disease and pest names follow the label list the service uses, which is more specific for some crops than others.
• The category pest/weed groups insects and weeds; the benchmark treats it as the pest branch.

Citation

Cite the paper this benchmark accompanies (Digital Green, 2026).

The model trained on this benchmark is at DigiGreen/crop-disease-pest-detection-dg.