DEVELOPMENT OF AN AIoT-BASED COFFEE BEAN CLASSIFICATION AND SORTING SYSTEM USING A VISION TRANSFORMER

Dody Pintarko (Universitas Pembangunan Nasional "Veteran" Jawa Timur)
Basuki Rahmat (Universitas Pembangunan Nasional "Veteran" Jawa Timur)
Faisal Muttaqin (Universitas Pembangunan Nasional "Veteran" Jawa Timur)

Multidiciplinary Output Research for Actual and International Issue (Morfai Journal)
Vol. 6 No. 5 (2026): Multidiciplinary Output Research For Actual and International Issue

Manual coffee bean sorting is highly prone to subjectivity, inconsistency, and low operational efficiency. This study aims to develop an automated classification and sorting system based on the Artificial Intelligence of Things (AIoT). The method integrates a Vision Transformer (ViT) model, TensorFlow Lite, Firebase, and an ESP32 microcontroller within a Mobile–Cloud–Edge Computing architecture. The ViT model was trained on four coffee roast levels to perform real-time inference on Android devices linked to physical sorting actuators. Experimental results showed that the ViT model achieved a 96.87% classification accuracy, while the automated physical sorting mechanism achieved 95.83% accuracy with an average response time of 462 ms. In conclusion, the integration of Vision Transformer and AIoT provides a fast and reliable post-harvest automation solution tailored for smart agricultural applications.