Scholarlink

<p>Rapid, non-destructive evaluation of internal fruit quality is critical for optimizing harvest timing and ensuring consumer satisfaction. Previous hyperspectral imaging (HSI) studies on fruit quality mostly relied on single wavelength selection or full-spectrum chemometric models combined with linear regression or shallow machine learning, which may not fully capture informative spectral features or complex non-linear relationships. Here, we present a framework integrating near-infrared HSI (450-1000nm) with chemometric and deep learning approaches to predict soluble solids content (SSC) in loquat. Our method advances prior work by combining multiple wavelength selection algorithms, including competitive adaptive reweighted sampling (CARS), iteratively retained informative variables (IRIV), and a genetic algorithm (GA), with a one dimensional convolutional neural network (1D-CNN) for more effective feature selection and robust modeling. After harvesting 240 uniformly ripe fruits, hyperspectral cubes were acquired and corrected via dark/white calibration. Spectra from manually defined fruit regions were preprocessed using savitzky golay smoothing (SGS), multiplicative scatter correction (MSC), and standard normal variate (SNV), individually or combined. Quantitative SSC models built using partial least squares regression (PLSR), least squares support vector machine (LS-SVM), and 1D-CNN showed that the CNN consistently outperformed traditional methods, with the best model achieving <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R_p^2 = 0.945\)</EquationSource> </InlineEquation>, RMSEP = 0.365&#xa0;<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\({}^{\circ }\text {Brix}\)</EquationSource> </InlineEquation> and RPD = 3.45. Pixel-wise application of the optimized CNN enabled high-resolution SSC mapping, revealing within-fruit variability. These results demonstrate that integrating multi-algorithm wavelength selection with deep convolutional modeling provides an accurate and efficient approach for non-destructive SSC assessment in loquat and other fruit crops.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feature-optimized convolutional modeling for predicting loquat soluble solids content from hyperspectral imaging with multi-algorithm wavelength selection

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Hailiang Zhang,
  • Jingru Wei,
  • Hanxu Xu,
  • Wei Luo,
  • Xuemei Liu

摘要

Rapid, non-destructive evaluation of internal fruit quality is critical for optimizing harvest timing and ensuring consumer satisfaction. Previous hyperspectral imaging (HSI) studies on fruit quality mostly relied on single wavelength selection or full-spectrum chemometric models combined with linear regression or shallow machine learning, which may not fully capture informative spectral features or complex non-linear relationships. Here, we present a framework integrating near-infrared HSI (450-1000nm) with chemometric and deep learning approaches to predict soluble solids content (SSC) in loquat. Our method advances prior work by combining multiple wavelength selection algorithms, including competitive adaptive reweighted sampling (CARS), iteratively retained informative variables (IRIV), and a genetic algorithm (GA), with a one dimensional convolutional neural network (1D-CNN) for more effective feature selection and robust modeling. After harvesting 240 uniformly ripe fruits, hyperspectral cubes were acquired and corrected via dark/white calibration. Spectra from manually defined fruit regions were preprocessed using savitzky golay smoothing (SGS), multiplicative scatter correction (MSC), and standard normal variate (SNV), individually or combined. Quantitative SSC models built using partial least squares regression (PLSR), least squares support vector machine (LS-SVM), and 1D-CNN showed that the CNN consistently outperformed traditional methods, with the best model achieving \(R_p^2 = 0.945\) , RMSEP = 0.365  \({}^{\circ }\text {Brix}\) and RPD = 3.45. Pixel-wise application of the optimized CNN enabled high-resolution SSC mapping, revealing within-fruit variability. These results demonstrate that integrating multi-algorithm wavelength selection with deep convolutional modeling provides an accurate and efficient approach for non-destructive SSC assessment in loquat and other fruit crops.