Application of ResNet50 for Oil Palm Fruit Image Classification Based on Ripeness Level

Abstract

Manual ripeness assessment still has limitations as it is subjective and highly dependent on human expertise. Therefore, this study aims to apply a deep learning approach based on the ResNet50 architecture to classify oil palm fruit ripeness into three categories, namely unripe, ripe, and overripe. The dataset used in this study consists of 1,350 RGB images of oil palm fruits, which are divided into training, validation, and testing sets with a ratio of 70:10:20. All images are preprocessed by resizing them to 224 × 224 pixels and normalizing pixel values, while data augmentation is applied to the training set to improve model generalization. A pre-trained ResNet50 model on the ImageNet dataset is employed as a feature extractor and trained using the Adam optimizer with a learning rate of 1 × 10⁻⁴ for 50 epochs. Experimental results show that the model achieves an accuracy of 89.7% on the training data and 84.1% on the validation data. Evaluation on the testing data yields an accuracy of 84.07%, with average precision, recall, and F1-score values of 84.71%, 84.07%, and 84.32%, respectively. These results indicate that the proposed ResNet50-based model demonstrates good and stable performance in classifying oil palm fruit ripeness levels.