Efficient hyperspectral band selection via occlusion-based neural network ranking for detecting fruit-bruise severity

Abstract

Hyperspectral imaging (HSI) provides rich spectral information across hundreds of narrow bands, making it a powerful tool for material classification. However, processing all available bands is computationally expensive and often impractical for near real-time applications. In this work, an occlusion-based band-selection method— developed earlier by the authors—is applied to a multiclass classification task to identify the most informative spectral bands for a target task while substantially reducing data dimensionality. For a given application, realistic spectral variations are first simulated through data augmentation under changing intensity and noise conditions. The augmented spectra are then used to train an artificial neural network (ANN) with the full spectral input. Band importance is subsequently evaluated by systematically occluding individual spectral bands and measuring the resulting degradation in classification performance, thereby forming a reduced candidate pool. A computationally manageable exhaustive search is then performed within this pool to identify a smaller subset of bands. As a case study, the method is applied to Honeycrisp apple bruise-severity classification using spectra in the 900–1700 nm range with 336 bands. The full-band ANN achieves 97.7% classification accuracy, while the occlusion-based 16-band and 5-band subsets achieve 89.8% and 80.7%, respectively. Under the same subset sizes, PCA-based selection achieves 84.1% and 74.4%. These results indicate that the proposed method preserves task-relevant spectral information more effectively than the PCA-based baseline, while substantial band reduction can shorten acquisition time, lower computational cost, and support on-device or edge deployment in resource-constrained platforms such as smart cameras.

Plant phenotyping relevance

ハイパースペクトル画像からリンゴ果実の bruise severity を推定するためのバンド選択法を中心的に適用・評価しており、植物器官の状態を定量化するフェノタイピング手法に該当する。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.