A Comparative Analysis on the Classification of Pineapple Varieties Using Thermal Imaging Coupled With Transfer
Norhashila Hashim1,2, Maimunah Mohd Ali3,4
1Department of Biological and Agricultural Engineering, Faculty of Engineering, Universiti Putra Malaysia, Serdang, Selangor, Malaysia.
Journal of Food Science
|September 9, 2025
Summary
This study used thermal imaging and deep learning models like InceptionV3 to classify pineapple varieties non-destructively. The InceptionV3 model achieved 99% accuracy, demonstrating a promising approach for agricultural applications.
Area of Science:
- Agricultural technology
- Computer vision
- Machine learning
Background:
- Pineapple (Ananas comosus) is a globally popular tropical fruit with significant nutritional value.
- Accurate classification of pineapple varieties is crucial for quality control and market differentiation.
- Non-destructive methods are increasingly sought after for fruit quality assessment.
Purpose of the Study:
- To develop and compare deep learning models for the rapid, non-destructive classification of pineapple varieties using thermal imaging.
- To evaluate the effectiveness of transfer learning and data augmentation in improving classification accuracy.
- To benchmark established convolutional neural network (CNN) architectures for agricultural thermal imaging applications.
Main Methods:
- A dataset of 3240 thermal images from three pineapple varieties (Moris, Josapine, N36) was collected under controlled temperatures.
- Three deep learning models (ResNet, VGG16, InceptionV3) were fine-tuned using transfer learning and data augmentation.
- Hyperparameter tuning was performed to optimize model performance, with an ablation study confirming the benefits of augmentation and transfer learning.
Main Results:
- The InceptionV3 model achieved the highest classification accuracy of 99%.
- Precision, recall, and F1-scores exceeded 0.85 for all pineapple varieties, indicating robust performance.
- The study confirmed the significant potential of transfer learning with CNNs for extracting physicochemical properties from pineapple thermal images.
Conclusions:
- Deep learning models, particularly InceptionV3, combined with thermal imaging offer a highly accurate and non-destructive method for pineapple variety classification.
- Transfer learning and data augmentation are effective strategies for enhancing model generalization and preventing overfitting in agricultural image analysis.
- This research provides a valuable benchmark for applying established CNN models in agricultural thermal imaging, paving the way for advanced intelligent systems in fruit assessment.
