IMAGE-BASED CLASSIFICATION OF COCOA FRUIT DISEASES USING TRANSFER LEARNING WITH THE MOBILENETV2 MODEL

Abstract

Cocoa is one of Indonesia’s leading commodities, but its productivity often declines due to pest and disease infestations. This study develops a lightweight, Transfer Learning-based MobileNetV2 model for automatic cocoa disease identification for real-time detection. The dataset consists of 405 images across four classes: black pod rot (108), healthy (106), Helopeltis (100), and pod borer (91), divided into 70% training, 10% validation, and 20% testing. During the training phase, data augmentation techniques—including RandomFlip, Random Rotation, RandomZoom, and RandomTranslation—were applied to increase the visual variation of the samples. The training configuration utilized the Adam optimizer, a batch size of 32, a random seed of 45, and early stopping. The study tested variations in learning rate (0.001, 0.0001) and epochs (30, 50, 70, 100), yielding the best performance at a learning rate of 0.0001 with 70 epochs. The model achieved a test accuracy of 92.86%, precision of 92.78%, recall of 92.86%, and an F1-score of 92.79%. These results demonstrate that MobileNetV2 is capable of consistently identifying disease characteristics, making it highly promising as the foundation for a fast and accurate field diagnostic application for cocoa farmers.