Rapid and Non-destructive Evaluation of Soybean Seed Viability Using Transmission Hyperspectral Imaging and Ensemble Learning

Abstract

Abstract Traditional soybean seed viability assessment methods are destructive, time-intensive, and incapable of rapid, non-destructive single-seed grading. To overcome these limitations, this study proposes a novel approach integrating Transmission Hyperspectral Imaging (THSI) with ensemble learning for rapid, non-destructive evaluation. Naturally aged soybean seeds were analyzed using full-spectrum (400–2500 nm) transmittance data to capture deep physiological information. Multiple datasets were constructed by comparing preprocessing techniques—including Smoothing, First Derivative (FD), Hilbert Transform (HT), Savitzky–Golay, Multiplicative Scatter Correction (MSC), and Standard Normal Variate (SNV)—with dimensionality reduction algorithms such as Principal Component Analysis (PCA), Successive Projections Algorithm (SPA), and Competitive Adaptive Reweighted Sampling (CARS). The proposed Stacking ensemble model integrates predictions from Random Forest (RF), Support Vector Machine (SVM), Naïve Bayes (NB), k-Nearest Neighbors (k-NN), and Logistic Regression (LR), achieving 98.33% accuracy and an F1-score of 98.12% on the CARS-HT dataset—significantly outperforming individual classifiers in both accuracy and robustness. Furthermore, the model identified 17 critical wavelengths that reveal physiological mechanisms ranging from chlorophyll degradation and antioxidant balance in the visible spectrum to water migration and lipid peroxidation in the infrared region. The method's high precision and reliability were validated, providing robust technical support for intelligent and precise soybean seed quality management.

Plant phenotyping relevance

大豆種子の生存性という植物状態を、透過ハイパースペクトル画像とアンサンブル学習で非破壊推定する手法を開発・検証しており、表現型取得・抽出が研究の中心である。

Code and data availability

The supplied article blocks describe soybean seed transmission hyperspectral imaging and a stacking ensemble model, but contain no data availability statement, no public dataset deposit, no author code repository, and no trained model release. No public, paper-specific phenotyping or analysis asset is identified.

No evidence-backed public reproduction asset is currently recorded.