Subspace/Discriminate Ensemble-based Machine Learning on Visible/Near-infrared Spectra as an Effective Procedure for Non-destructive Safety Assessment of Spinach

In this study, an orthogonal signal correction (OSC)-based partial least squares (PLS) model and ensemble-based machine learning classifiers, combined with visible/near-infrared (Vis/NIR) spectroscopy, were proposed for non-destructive nitrate prediction in spinach leaves and sample safety evaluation. The OSC method was applied before developing the PLS model to enhance prediction accuracy. Spinach safety assessment was based on the maximum permissible nitrate accumulation level. Various ensemble classifiers, including subspace/discriminate, subspace/k-nearest neighbor, boosted trees, bagged trees, and random under-sampling boosted trees, were evaluated for distinguishing safe and unsafe samples. The best classification results were obtained using the subspace/discriminate ensemble classifier, achieving sensitivity, specificity, and accuracy of 95.24%, 98.73%, and 98.45% for the calibration dataset and 100%, 91.8%, and 92.31% for external validation. The receiver operating characteristic (ROC) curve indicated superior discrimination ability, with an area under the curve (AUC) of 0.95. Additionally, the best model demonstrated a high prediction speed of approximately 280 observations per second. These findings highlight that combining Vis/NIR spectroscopy with the subspace/discriminate ensemble classifier provides an effective, rapid, and non-invasive method for detecting nitrate contamination in spinach leaves, making it a promising approach for food safety monitoring.