Prediction of Apple Quality Indicators Under Different Bagging Treatments Using Hyperspectral Imaging Integrated With a Stacking SDAE-PLSR-RR Deep Learning Model.

Abstract

The color indices (L*, a*, and b*) and soluble solids content (SSC) serve as essential quality indicators for apples, yet conventional destructive detection methods lack the efficiency required for rapid sorting of apples with varied bagging treatments. To address this limitation, this study proposes a novel stacking model, termed SDAE-PLSR-RR, which integrates hyperspectral imaging with deep learning. Hyperspectral imaging comprehensively captured spectral-spatial features from 307 Fuji apples subjected to three bagging treatments (non-bagged, mesh-bagged, and paper-bagged), enabling systematic analysis of quality-related characteristics. The SDAE-PLSR-RR employs a stacked structure where two parallel, base-level expert models capture complementary features: one Partial Least Squares Regression (PLSR) model processes linear trends in original wavelengths data, while the other analyzes non-linear deep features from a Stacked Denoising Autoencoder (SDAE). A top-level Ridge Regression (RR) model then acts as a meta-learner to fuse the predictions from these two base models, generating a final, more robust output. The integrated SDAE-PLSR-RR model achieved enhanced prediction accuracy for all quality indicators (R 2 p > 0.84), outperforming full-spectrum (R 2 p > 0.73) and feature-wavelength-based models (R 2 p > 0.75). The experimental findings validated the applicability and efficacy of integrating hyperspectral imaging systems with neural network models for non-destructive detection of the quality indicators of apples with different bagging treatments.

Plant phenotyping relevance

リンゴの色指標と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と新規スタッキングモデルによる非破壊推定法を開発・検証しており、フェノタイピング手法が中心である。

Code and data availability

The paper's apple hyperspectral/quality datasets and SDAE-PLSR-RR model are not publicly deposited; the Data Availability Statement says they are available from the corresponding author on reasonable request. No public code, data, or model URLs are provided.

No evidence-backed public reproduction asset is currently recorded.