Related Experiment Videos
Early Apple Bruise Detection via Discrete Hyperspectral Signatures with SHAP-Guided Feature Selection and a
Ying Liu1, Chen Yu1, Chaoxian Liu1
1School of Mathematics & Computer Science, Wuhan Polytechnic University, Wuhan 430023, China.
Foods (Basel, Switzerland)
|June 12, 2026
Summary
This study introduces an efficient method for detecting invisible apple bruises using optimized hyperspectral imaging (HSI) wavebands and a novel CNN-Transformer model, achieving high accuracy in non-destructive quality assessment.
Area of Science:
- Agricultural Science
- Spectroscopy
- Computer Vision
Background:
- Accurate detection of invisible apple bruises is crucial for post-harvest quality assessment.
- Hyperspectral imaging (HSI) offers rich spectral data but suffers from high dimensionality, redundancy, and signal interference.
- Existing methods struggle with the sparse, discrete spectral data characteristic of reduced HSI.
Purpose of the Study:
- To develop an integrated framework for efficient and accurate detection of invisible apple bruises using reduced hyperspectral data.
- To optimize waveband selection and develop a specialized deep learning model for sparse spectral inputs.
- To interpret the biochemical relevance of selected spectral bands for bruise detection.
Main Methods:
- Developed a Selection-Refined Improved Grey Wolf Optimization (SR-IGWO) algorithm to select 18 bruise-sensitive wavebands from 273 channels (996-2501 nm), reducing dimensionality by 93.4%.
- Employed SHAP analysis to interpret the biochemical significance of selected spectral bands.
- Designed a CNN-Transformer hybrid model (DSFormer) with pointwise convolution for band embedding and a Transformer encoder for global dependency capture, specifically for discrete spectral inputs.
Main Results:
- The SR-IGWO algorithm successfully reduced spectral dimensionality while identifying sensitive wavebands.
- The DSFormer model achieved high classification accuracy (99.11% ± 0.08%), recall (96.04% ± 1.08%), and F1-score (95.95% ± 0.39%).
- Ablation studies confirmed the effectiveness of the proposed architecture for sparse spectral data detection.
Conclusions:
- The integrated framework combining waveband optimization and the DSFormer model offers an efficient approach for non-destructive apple bruise detection.
- Reduced hyperspectral data coupled with specialized deep learning models shows significant promise for fruit quality assessment.
- Further validation across diverse cultivars and conditions is warranted to fully establish the framework's potential.