Scholarlink

<p>Visible–near-infrared (Vis–NIR) reflectance spectroscopy (500–1050&#xa0;nm) is a proven non-destructive tool for fruit quality assessment, yet calibration robustness can be limited by spectral collinearity, scatter effects, and date-to-date drift. Here, Harum Manis mangoes were analysed using chemometric pre-processing, principal component analysis (PCA), and modern machine-learning regressors. Four datasets (two acquisition dates × two targets: soluble solids content (SSC) and pH) were acquired from both the front and back sides of each mango; models were calibrated on one side and evaluated on the opposite side as an independent validation set. Pipelines combined pre-processing, wavelength selection (k = 10), optional data augmentation, PCA compression, and regressor tuning via cross-validated grid search. Across dates, SSC achieved independent validation R² of 0.67–0.71 (RMSE 2.4–2.6 °Brix) and pH achieved independent validation R² of 0.63–0.81 (RMSE 0.36–0.51 pH units). Latent-space analysis indicated a date-related spectral shift consistent with ripening. Future work should extend validation across seasons and instruments to further improve calibration transfer.</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

Principal component analysis (PCA)-integrated machine learning (ML) pipelines for spectral modelling of Harum Manis mangoes

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Anam Irshad,
  • Jelena Muncan,
  • Ahmad Fairuz Omar

摘要

Visible–near-infrared (Vis–NIR) reflectance spectroscopy (500–1050 nm) is a proven non-destructive tool for fruit quality assessment, yet calibration robustness can be limited by spectral collinearity, scatter effects, and date-to-date drift. Here, Harum Manis mangoes were analysed using chemometric pre-processing, principal component analysis (PCA), and modern machine-learning regressors. Four datasets (two acquisition dates × two targets: soluble solids content (SSC) and pH) were acquired from both the front and back sides of each mango; models were calibrated on one side and evaluated on the opposite side as an independent validation set. Pipelines combined pre-processing, wavelength selection (k = 10), optional data augmentation, PCA compression, and regressor tuning via cross-validated grid search. Across dates, SSC achieved independent validation R² of 0.67–0.71 (RMSE 2.4–2.6 °Brix) and pH achieved independent validation R² of 0.63–0.81 (RMSE 0.36–0.51 pH units). Latent-space analysis indicated a date-related spectral shift consistent with ripening. Future work should extend validation across seasons and instruments to further improve calibration transfer.