RepViT-M1.1 Based Classification of Arabica Coffee Varieties with Blockchain-Enabled NFT Certification and On-Chain SHA-256 Anti-Duplication on Polygon Network

Abstract

Variety authentication and quality validation remain persistent bottlenecks in Indonesia’s specialty coffeechains. This study addresses both. We present an end-to-end system that integrates RepViT-M1.1, a mobile-efficient visiontransformer, with NFT certification on the Polygon Amoy blockchain. Six classes are targeted: five Arabica varieties,Peaberry, Natural Ijen, Anaerobic Carbonic, Orange Bourbon, and Blue Mountain, together with an explicit Non-Coffeeclass that allows the model to reject non-coffee inputs. The classifier was trained on 12,000 images compiled in Bondowoso,East Java, and partitioned 70/15/15. It attains 99.78% accuracy on the test set and 97.00% on validation; in the ablationstudy, it rejects 99.3% of out-of-distribution (OOD) Non-Coffee inputs. The blockchain component is equally deliberate.Our ERC-721 smart contract (V3) maintains an on-chain SHA-256 hash registry that prevents any photograph from beingminted twice, and it enforces this at the consensus layer rather than in the application, a safeguard absent from existingagricultural traceability platforms. Farmers participate directly through a permissionless, Indonesian-language ProgressiveWebApp(PWA)thatpreserves the local identity of Ijen Arabica coffee. Across 30 trials, the complete path from photographto on-chain confirmation required 14.81 ± 2.04 seconds. Benchmarked against six baseline networks, RepViT-M1.1 offersthe most favourable accuracy-latency balance, with a reference CPU latency near 8 ms (ARM Cortex-A76). The result is ascalable template for national-scale digital coffee certification.