Blockchain-Anchored Provenance Verification of Crop-Disease Images Using Perceptual Hashing and Metadata Consistency Checks

Photographs of diseased leaves now serve as evidence: extension staff log them to follow outbreaks, and crop insurers accept them as proof of loss. Once such an image leaves the field, an ordinary file store cannot show whether its label, plot or date was changed, a lesion pasted in or painted out, or an old picture resubmitted. This paper describes AgriProv, a provenance scheme that binds content-aware image fingerprints to a permissioned ledger run by a small set of known validators. Each image is segmented, described by colour and texture statistics, and fingerprinted with a cryptographic hash plus global and tile-level perceptual hashes. Four consistency checks covering plot boundary, capture and upload time, label plausibility and near-duplicate search screen each submission before its digests, metadata hash and a hash-chained processing history are batched under a Merkle root. A Python implementation, mirrored in MATLAB, was tested on synthetic leaf images from 48 smallholder plots under nineteen scenarios. The scheme accepted 99.2 percent of benignly re-encoded images and every untouched one, rejected all metadata edits and image substitutions, and caught 95.3 percent of lesion splices covering two percent of the leaf, compared with 26.0 percent for a global perceptual hash alone. Registration checks flagged 95.3 percent of mislabelled and 98.7 percent of neighbouring-plot submissions, at a 5.4 percent false-flag rate driven mainly by late uploads. Each on-chain record needs 359 bytes, about 2.2 percent of the mean image. Three limitations are reported: a three percent crop defeats every perceptual test, the descriptor distance added no measurable detection, and clock rollback shorter than the upload window goes unnoticed.