A pistachio leaf in a field photograph can take a few forms before it reaches a classifier: the original image, a cut-out of the leaf, and a mask marking its outline. The distinctions are small on screen and consequential when a computer is asked to learn from labels.
A Zenodo deposit published October 21, 2026, contains 599 field photographs taken in south-eastern Türkiye. The images show leaves of Pistacia vera L., and the collection accompanies research on how observers label disease in photographs. The dataset record describes the photographs as field images.
It also includes single-leaf cut-outs and masks made by an extraction pipeline. Those files give researchers more than a folder of whole-scene photographs to work with.
The label is part of the test
The deposit accompanies a manuscript titled “Observer disagreement in photograph-based pistachio leaf disease labels and its implications for classifier evaluation.” That framing puts human judgment inside the evaluation problem: a classifier is scored against labels supplied by people, and differences among those labels can affect what its score says.
The available record does not give an agreement rate, identify how many observers labelled the images, or report classifier scores. So the release description establishes the research question and the image materials, but not the size or direction of any disagreement.
A Turkish dataset is not a California trial
For California pistachio operations, the work is relevant as a question about how image-based disease tools are judged, not as evidence that a particular tool recognizes disease reliably in local orchards. The description supplies no California images, local disease results, or recommendations for orchard management.
That leaves a useful distinction for anyone weighing a classifier: performance against a set of photographs is only as interpretable as the labels and evaluation conditions behind it. The release description does not provide enough detail to assess those conditions.