A Comparative Analysis on the Classification of Pineapple Varieties Using Thermal Imaging Coupled With Transfer Learning

Abstract

Advanced intelligent systems are becoming a significant trend, especially in the classification of tropical fruits due to their unique flavor and taste. As one of the most popular tropical fruits worldwide, pineapple (Ananas comosus) has a great chemical composition and is high in nutritional value. A non-destructive method for the determination of pineapple varieties was developed, which utilized thermal imaging and deep learning techniques. This study presents a comparative analysis of three deep learning models, including ResNet, VGG16, and InceptionV3, for the rapid classification of pineapple varieties using thermal imaging and transfer learning. The dataset comprises 3240 thermal images from three different pineapple varieties, including Moris, Josapine, and N36, under controlled temperature conditions (5°C, 10°C, and 25°C), resulting in a total of three classification classes. All convolutional neural network (CNN) architectures were fine-tuned, and data augmentation techniques were applied to improve model generalization. The efficiency of hyperparameters was evaluated to improve the model accuracy, whereas the data augmentation was carried out to avoid model overfitting. The highest classification accuracy of 99 % was achieved via InceptionV3. The precision, recall, and F1-score demonstrate promising results with the values higher than 0.85 for all pineapple varieties. This approach demonstrated that transfer learning with CNNs is significantly promising as a feature extraction method for the determination of physicochemical properties in pineapple fruit. An ablation study confirmed the added benefit of using both data augmentation and transfer learning. While model architecture innovation was not the primary goal, this work contributes by benchmarking established CNN models for agricultural thermal imaging applications.