Early Apple Bruise Detection via Discrete Hyperspectral Signatures with SHAP-Guided Feature Selection and a CNN-Transformer Model.

Abstract

Accurate detection of early invisible apple bruises is important for post-harvest quality assessment. Although hyperspectral imaging (HSI) provides rich spectral information, its high dimensionality introduces substantial redundancy and weak-signal interference. This study proposes an integrated framework combining waveband optimization and discrete spectral modeling for efficient bruise detection. A Selection-Refined Improved Grey Wolf Optimization (SR-IGWO) algorithm was developed to select 18 bruise-sensitive wavebands from 273 channels (996-2501 nm), achieving a 93.4% reduction in spectral dimensionality. SHAP analysis was further used to interpret the selected bands in relation to biochemical responses associated with bruising. To address the mismatch between conventional CNNs and sparse discrete spectral inputs, a CNN-Transformer hybrid model (DSFormer) was designed using pointwise convolution for band embedding and a Transformer encoder to capture global dependencies. Experimental results across ten independent runs achieved a classification accuracy of 99.11% ± 0.08%, a recall of 96.04% ± 1.08%, and an F1-score of 95.95% ± 0.39% under the tested conditions. Ablation studies suggest that the proposed architecture supports effective detection under sparse spectral conditions. Although validation was limited to a single cultivar and controlled sampling, the proposed framework provides a promising preliminary exploration of reduced hyperspectral data for non-destructive fruit bruise detection.

Plant phenotyping relevance

リンゴ果実の打撲状態を対象に、ハイパースペクトル波長選択とCNN-Transformerによる症状検出手法を開発・検証しており、植物器官の状態取得が研究の中心である。

Code and data availability

The paper's hyperspectral apple bruise dataset (over 11,000 object-level spectral samples from ~5700 bruised and 6000 healthy ROIs) and analysis code are not publicly deposited; the Data Availability Statement states raw data are available only from the authors on request, and no public repository or code URL is given.

No evidence-backed public reproduction asset is currently recorded.