Deep spectral–spatial learning for robust bruise detection and internal damage estimation in apples using visible/near-infrared hyperspectral imaging
4 affiliations
- 1Department of Biosystems Engineering, Seoul National University
- 2Research Institute of Agriculture and Life Sciences, Seoul National University
- 3Integrated Major in Global Smart Farm, Seoul National University
- 4Department of Smart Bio-Industrial Mechanical Engineering, Kyungpook National University
Abstract
Bruises are common postharvest defects of apples whose appearance varies with surface location, impact severity, and time after impact, complicating visual detection particularly during the early stages of bruising. To address this challenge, four three-dimensional (3D) U-Net variants of spectral–spatial encoders were compared for bruise detection from visible/near-infrared (Vis/NIR) hyperspectral images. Furthermore, features learned by the bruise detector were reused to classify impact level and estimate internal cross-sectional bruise area. Controlled bruises were induced at three impact levels and multiple surface locations on 87 Fuji apples. Hyperspectral images were acquired from six views over 18 days of storage, resulting in 870 image cubes with 1,676 annotated bruise instances. Bruise depth and internal cross-sectional area of browned tissue were measured by destructive sectioning. Among the four models, the Hybrid 3D–2D U-Net achieved the highest test bounding box F 1 score of 0.98 and a mask Dice score of 0.86. The model outperformed baseline detectors using inputs with three channels derived from principal component and band-ratio transformations, particularly for early, mild, and peripheral bruises, maintaining the F 1 score above 0.98. Half-precision inference preserved both accuracy metrics at 87.63 frames per second. For the bruise severity estimation, models incorporating pretrained encoder features of the bruise detection model achieved a test macro-averaged F 1 score of 0.83 for impact level classification and a test Pearson r of 0.82 for internal cross-sectional bruise area estimation, outperforming the baseline models based on the mean reflectance spectrum. These results indicated that deep spectral–spatial learning supported robust bruise detection while extending hyperspectral assessment from bruise localization to estimation of internal damage severity.
Related
- Hyperspectral Soil Metal Estimation Using Fractional-Order Derivatives and LSTM–KAN Modeling
- Reading the whole corpus: a reproducible cross-sample Raman workflow for materials characterisation, with a spectral 3D Gaussian splatting representation for hyperspectral maps
- Progeny Differentiation in Faba Bean Using Hyperspectral Images and Machine Learning
- A Physics-Informed TCN-LSTM Approach for Interpretable Internal Temperature Estimation of NMC811 Lithium-Ion Batteries
- A Noise-Aware Calibrated Multi-Scale Voltage-Residual Method for Internal Short-Circuit Detection in Lithium-Ion Batteries
BibTeX
@article{ryu2026deep,
title = {{Deep spectral–spatial learning for robust bruise detection and internal damage estimation in apples using visible/near-infrared hyperspectral imaging}},
author = {Jiwon Ryu and Chang-Hyup Lee and Sang-Yeon Kim and Sungjay Kim and Minhyun Kim and Dong Heun Sung and Suk-Ju Hong and Ghiseok Kim},
journal = {Frontiers in Plant Science},
year = {2026},
month = oct,
doi = {10.3389/fpls.2026.1967857},
url = {https://www.frontiersin.org/articles/10.3389/fpls.2026.1967857/full},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Deep spectral–spatial learning for robust bruise detection and internal damage estimation in apples using visible/near-infrared hyperspectral imaging
AU - Jiwon Ryu
AU - Chang-Hyup Lee
AU - Sang-Yeon Kim
AU - Sungjay Kim
AU - Minhyun Kim
AU - Dong Heun Sung
AU - Suk-Ju Hong
AU - Ghiseok Kim
PY - 2026
DA - 2026/10/27
JO - Frontiers in Plant Science
DO - 10.3389/fpls.2026.1967857
UR - https://www.frontiersin.org/articles/10.3389/fpls.2026.1967857/full
AB - Bruises are common postharvest defects of apples whose appearance varies with surface location, impact severity, and time after impact, complicating visual detection particularly during the early stages of bruising. To address this challenge, four three-dimensional (3D) U-Net variants of spectral–spatial encoders were compared for bruise detection from visible/near-infrared (Vis/NIR) hyperspectral images. Furthermore, features learned by the bruise detector were reused to classify impact level and estimate internal cross-sectional bruise area. Controlled bruises were induced at three impact levels and multiple surface locations on 87 Fuji apples. Hyperspectral images were acquired from six views over 18 days of storage, resulting in 870 image cubes with 1,676 annotated bruise instances. Bruise depth and internal cross-sectional area of browned tissue were measured by destructive sectioning. Among the four models, the Hybrid 3D–2D U-Net achieved the highest test bounding box F 1 score of 0.98 and a mask Dice score of 0.86. The model outperformed baseline detectors using inputs with three channels derived from principal component and band-ratio transformations, particularly for early, mild, and peripheral bruises, maintaining the F 1 score above 0.98. Half-precision inference preserved both accuracy metrics at 87.63 frames per second. For the bruise severity estimation, models incorporating pretrained encoder features of the bruise detection model achieved a test macro-averaged F 1 score of 0.83 for impact level classification and a test Pearson r of 0.82 for internal cross-sectional bruise area estimation, outperforming the baseline models based on the mean reflectance spectrum. These results indicated that deep spectral–spatial learning supported robust bruise detection while extending hyperspectral assessment from bruise localization to estimation of internal damage severity.
SN - 1664-462X
ER -