Images now ship inside the dataset as images/00001.jpg upward, and image_url is replaced by image_file. This is a breaking change for anything reading image_url.
Two reasons. The dataset no longer points at a live production storage bucket, and the filenames no longer carry an internal message identifier. That identifier was the key used to attach geography to this dataset, so publishing it gave anyone with database access a route from a public row back to an internal conversation and user. Sequential filenames carry nothing.
989 images, 1026 annotation rows.
Every image was screened for personal information by two vision models plus a human pass over the disagreements. Removed: burned-in camera-app location stamps (some carrying coordinates at roughly one-metre precision), watermarks carrying the photographer's name, names and phone numbers visible in the frame, faces, and one chat screenshot.
1092 rows -> 1026 rows; 1053 -> 989 distinct images. All counts in the dataset card were regenerated from the surviving rows. The card's previous claim that no capture coordinates exist for the photos was wrong and has been replaced with a description of the screening.
Add country and state columns; document geographic coverage
Answers discussion #2. Location is the registered state of the farmer who sent the photo, not image GPS. State level only: several states hold a single image.
Add 1,092 expert crop disease annotations with CC-BY-4.0 license
