Analisis Kinerja CNN dan Transfer Learning dalam Klasifikasi Kematangan Buah Kopi

Abstract

Manual determination of coffee cherry ripeness levels still relies on the subjective judgment of farmers, which may reduce the consistency of harvest quality. Imagebased classification using deep learning has emerged as one of the solutions developed to address this problem. This study analyzes and compares the performance of a standard Convolutional Neural Network (CNN) model and a transfer learning model based on EfficientNet B0 in classifying the ripeness level of coffee cherries from digital images. The dataset was classified into three classes, namely unripe, semi-ripe, and ripe, with a data split of 70% for training and 30% for testing, while validation data was drawn from 10% of the training set. The standard CNN model was trained using a three-layer convolutional architecture, whereas the EfficientNet B0 model applied a transfer learning approach with finetuning using pretrained weights from the ImageNet dataset. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics on each class. The result show that the standard CNN model achieved an accuracy of 78%, but exhibited a significant weakness in the semi-ripe class with a recall of 40% and an F1-score of 57%. The EfficientNet B0 model achieved an accuracy of 87% with more balanced performance across all the three classes, particularly in the semiripe class which reached a recall of 77% and an F1-score of 82%. Based on the result, the EfficientNet B0 transfer learning model demonstrated superior overall performance compared to the standard CNN model, particularly in recognizing classes with similar visual characteristics. This study contributes to the development of coffee cherry image classification methods that can support more objective and accurate harvest time determination.