Interpretable hyperspectral analysis of soluble solids content in apples: spectral attribution and mechanistic insights from linear and deep learning models.

Abstract

Soluble solids content (SSC) is a key determinant of apple sweetness and market quality, and its rapid, nondestructive assessment is essential for postharvest grading. Hyperspectral imaging (HSI) provides rich spectral information for SSC prediction; however, conventional wavelength selection strategies are largely data-driven and lack interpretability, while the underlying mechanisms of spectral information utilization across different modeling approaches remain insufficiently understood. In this study, an interpretable hyperspectral analysis framework was developed to investigate both effective wavelength selection and model-dependent spectral response mechanisms. Average spectra extracted from the pulp region of interest (ROI) were used for analysis. A linear model (PLSR) combined with SHAP (SHapley additive exPlanations) was employed to quantify global feature contributions, while a one-dimensional convolutional neural network with dual-attention mechanisms (BrixCNN) was constructed and interpreted using integrated gradients (IG) to capture nonlinear spectral dependencies. The consistency and divergence between the two attribution strategies were further quantitatively analyzed. Results showed that both SHAP-PLSR and IG-BrixCNN identified informative wavelength subsets that significantly reduced spectral dimensionality while maintaining comparable predictive performance (R 2 ≈ 0.83-0.84, RPD > 2.4). Despite similar predictive accuracy, the two models exhibited distinct spectral utilization patterns: The linear model primarily relied on dominant, high-variance spectral variations, whereas the deep learning model captured weaker, more distributed, and nonlinear spectral patterns. Meanwhile, partial overlap in the 1100-1300 nm region suggested that both models may utilize correlated spectral variations within similar wavelength domains for SSC prediction. These findings indicate that comparable predictive performance can arise from distinct yet complementary spectral utilization patterns, reflecting model-dependent information extraction mechanisms rather than direct chemical specificity. This study provides new insights into wavelength selection strategies and enhances the interpretability and reliability of hyperspectral analysis for fruit quality assessment.

Plant phenotyping relevance

リンゴ果実のSSCという植物器官形質を対象に、ハイパースペクトル画像、波長選択、SHAP/IG解釈を統合した予測・解析フレームワークを開発しており、形質取得手法が研究の中心である。

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.