Implementasi Arsitektur Improved YOLOv8 untuk Deteksi Penyakit Buah Kakao Menggunakan Citra Buah Berbasis Website

Abstract

Indonesia’s cocoa plantation sector plays a strategic role in international trade, but its production stability is hampered by pest and disease outbreaks. Fluctuations in the value of cocoa bean exports during the 2022–2025 period—which briefly reached 80.6 million USD but subsequently corrected to 51.5 million USD—highlight the need for technological interventions to maintain crop quality. The main challenge is the difficulty in quickly and accurately identifying diseases such as Blackpod, Frostyrod, and the fruit-sucking pest (Mirid). Therefore, this study aims to develop a deep learning-based early detection system to minimize the risk of reduced crop yield and quality.

The method used involves implementing the Improved YOLOv8 algorithm with the integration of the GhostNet module (GhostConv and C3Ghost) to optimize computational efficiency without compromising detection accuracy. The dataset consists of 1,992 physical images covering four classes (Blackpod, Frostyrod, Mirid, and Healthy) with a 70:20:10 split. Optimization was performed using on-the-fly oversampling to address data imbalance, along with the AdamW optimizer for training stability.

The research results show that the Improved YOLOv8 model achieved an 86.1% accuracy rate (mAP@0.5) with a compact model size of 17.3 MB. The output of this study is a web-based application capable of performing detection via image uploads or real-time camera feeds. These results demonstrate that the developed system can serve as an effective technical solution to assist in the rapid and accurate identification of cocoa fruit diseases.