<p>Soluble solids content (SSC) is considered one of the key internal qualities, which determines the commercial value and consumer acceptance of milk jujubes. Traditional SSC evaluation methods are destructive and time-consuming and thus are not suitable for large-scale application. Therefore, developing a rapid, nondestructive detection technology is essential for industrial development. This study aims to use visible/near-infrared (Vis/NIR) spectroscopy technology, developing a stable and accurate nondestructive detection model for milk jujube SSC. In this study, a total of 160 milk jujube samples were collected, and their reflectance spectra in the 433–1711 nm range were acquired. First, a data augmentation–based unsupervised anomaly sample elimination (DAUASE) strategy was used to identify and remove abnormal samples, and the KS algorithm was utilized for dataset partitioning. Subsequently, an adaptive Savitzky–Golay (ASG) smoothing algorithm was proposed, and it was compared with four traditional preprocessing methods and their combinations. Finally, the combination of ASG and standard normal variate (SNV) and first derivative (FD) was determined as the optimal preprocessing procedure. In order to select the most influential wavelengths, a two-stage feature selection strategy was adopted: first, four traditional algorithms were used for coarse selection, and then fine selection was conducted based on the optimal feature subset, combined with the optimized binary enzyme action optimizer with cross-validation (BiEAOCV). This strategy significantly reduced the model’s computational load and, at the same time, enhanced the predictive model’s performance. Finally, support vector regression (SVR), long short-term memory (LSTM), and partial least squares regression (PLSR) models were established based on the selected features. In the results, the SVR model obtained the best prediction accuracy, with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${R}_{p}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>R</mi> <mrow> <mi>p</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> of 0.93, and RMSEP was as low as 0.3°Brix. In conclusion, this study can provide a powerful tool for the rapid and non-destructive determination of milk jujube SSC. This also offers a valuable technical basis for developing online grading and sorting equipment, and it is expected to promote the development of the milk jujube industry.</p>

Non-destructive Detection of SSC in Milk Jujubes using Vis/NIR Spectroscopy Combined with a Coarse-to-Fine Feature Selection Strategy

期刊论文
发表日期：

2026年4月23日

查看全文

Yinhai Yang,
Shibang Ma,
Feiyang Qi,
Feiyue Wang,
Chenyu Huang,
Yujia Guo

摘要

Soluble solids content (SSC) is considered one of the key internal qualities, which determines the commercial value and consumer acceptance of milk jujubes. Traditional SSC evaluation methods are destructive and time-consuming and thus are not suitable for large-scale application. Therefore, developing a rapid, nondestructive detection technology is essential for industrial development. This study aims to use visible/near-infrared (Vis/NIR) spectroscopy technology, developing a stable and accurate nondestructive detection model for milk jujube SSC. In this study, a total of 160 milk jujube samples were collected, and their reflectance spectra in the 433–1711 nm range were acquired. First, a data augmentation–based unsupervised anomaly sample elimination (DAUASE) strategy was used to identify and remove abnormal samples, and the KS algorithm was utilized for dataset partitioning. Subsequently, an adaptive Savitzky–Golay (ASG) smoothing algorithm was proposed, and it was compared with four traditional preprocessing methods and their combinations. Finally, the combination of ASG and standard normal variate (SNV) and first derivative (FD) was determined as the optimal preprocessing procedure. In order to select the most influential wavelengths, a two-stage feature selection strategy was adopted: first, four traditional algorithms were used for coarse selection, and then fine selection was conducted based on the optimal feature subset, combined with the optimized binary enzyme action optimizer with cross-validation (BiEAOCV). This strategy significantly reduced the model’s computational load and, at the same time, enhanced the predictive model’s performance. Finally, support vector regression (SVR), long short-term memory (LSTM), and partial least squares regression (PLSR) models were established based on the selected features. In the results, the SVR model obtained the best prediction accuracy, with ${R}_{p}^{2}$ $R_{p}^{2}$ of 0.93, and RMSEP was as low as 0.3°Brix. In conclusion, this study can provide a powerful tool for the rapid and non-destructive determination of milk jujube SSC. This also offers a valuable technical basis for developing online grading and sorting equipment, and it is expected to promote the development of the milk jujube industry.