Citation

Wang, Xiaoshuang and Hu, Yuhang and Ang, Chun Kit and Solihin, Mahmud Iwan and Tiang, Jun Jiat and Lim, Wei Hong (2025) Hyperspectral imaging-based deep learning benchmarks in non-destructive testing of cherry tomatoes. Applied Food Research, 5 (2). p. 101387. ISSN 2772-5022

Text Hyperspectral imaging-based deep learning benchmarks in non-destructive testing of cherry tomatoes.pdf - Published Version Restricted to Repository staff only Download (6MB)

Abstract

Non-destructive testing of fruit quality is critical for enhancing post-harvest processing efficiency and market value, particularly for delicate produce such as cherry tomatoes. Traditional methods, which are labor-intensive and destructive, limit real-time applicability. Hyperspectral imaging (HSI), by capturing spatial and spectral information, offers a promising alternative for rapid, non-invasive quality evaluation. This study evaluates the potential of combining HSI with traditional machine learning (ML) and deep learning (DL) models to predict four key physicochemical properties of cherry tomatoes: soluble solids content, acidity, sugar content, and firmness. A dataset of 310 samples was analyzed across two spectral ranges: visible/near-infrared (406–1010 nm) and near-infrared (957–1677 nm). Four preprocessing methods were applied to enhance spectral features: multiplicative scatter correction, standard normal variate, first derivative, and second derivative. To establish DL benchmarks, Bayesian optimization hyperparameter search, as neural architecture search, was employed for spectral preprocessing and DL network architecture tuning. For fair comparison, traditional ML models with preprocessing methods were optimized using grid search. Seven models, including linear regression (LR), partial least squares regression (PLSR), support vector regression (SVR), VGG, ResNet, Transformer, and DenseNet, were developed. Results indicated that DL models, particularly ResNet and Transformer, exhibit superior accuracy and robustness, achieving coefficient of determination (R2) values up to 0.96 in the test sets. Spectral analysis based on Grad-CAM confirms that these DL models consistently focus on chemically informative wavelengths. These findings highlight the advantages of optimized DL models in developing interpretable and accurate HSI for non-destructive fruit quality testing, supporting advancements in agricultural productivity, enabling manufacturing innovation in post-harvest processing, and promoting sustainable food consumption through reduced waste and improved quality control.

Item Type:	Article
Uncontrolled Keywords:	Cherry tomatoes, deep learning
Subjects:	Q Science > Q Science (General) > Q300-390 Cybernetics S Agriculture > SB Plant culture
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Nor Afiqah Mohd Adnan
Date Deposited:	07 Nov 2025 07:25
Last Modified:	09 Nov 2025 22:41
URII:	http://shdl.mmu.edu.my/id/eprint/14783

Downloads

Edit (login required)