Predicting B-Carotene Content in Apricots Using Artificial Intelligence

Strategic Insights -

Research conducted by the University of Ruse in Bulgaria has explored the application of artificial intelligence in predicting the b-carotene content in apricots using hyperspectral images captured in the near-infrared region during the drying process. The study focused on comparing the performance of various machine learning models, including Partial Least Squares Regression (PLSR), Stacked Autoencoders (SAEs) combined with Random Forest (RF), and Convolutional Neural Networks (CNNs). The results indicated that the PLSR model showed excellent results with an R-squared value of 0.97 for both training and testing, while the SAE-RF model performed well with R-squared values of 0.82 and 0.83 for training and testing, respectively. The CNN models displayed varying results, with the 1D-CNN achieving moderate performance and the 2D-CNN and 3D-CNN exhibiting signs of overfitting.

Key Takeaways:

  • The study focused on predicting b-carotene content in apricots using hyperspectral images captured in the near-infrared region during the drying process.
  • Four machine learning models were compared: Partial Least Squares Regression (PLSR), Stacked Autoencoders (SAEs) combined with Random Forest (RF), 1D-CNN, 2D-CNN, and 3D-CNN.
  • The PLSR model showed excellent results with an R-squared value of 0.97 for both training and testing.
  • The SAE-RF model performed well with R-squared values of 0.82 and 0.83 for training and testing, respectively.
  • The CNN models displayed varying results, with the 1D-CNN achieving moderate performance and the 2D-CNN and 3D-CNN exhibiting signs of overfitting.
  • The study suggested that PLSR and SAE-RF models deliver more reliable and robust predictions for b-carotene content in hyperspectral imaging.

Statistics:

  • R-squared value for the PLSR model: 0.97 for both training and testing.
  • R-squared values for the SAE-RF model: 0.82 for training and 0.83 for testing.
  • Mean Absolute Error (MAE) for the PLSR model: not specified.
  • Root Mean Squared Error (RMSE) for the PLSR model: not specified.
  • 1D-CNN, 2D-CNN, and 3D-CNN models achieved moderate to high R-squared values, but exhibited signs of overfitting.

Sources:

  • "Hyperspectral Analysis of Apricot Quality Parameters Using Classical Machine Learning and Deep Neural Networks." Engineering Proceedings, 2025, 104(1):24. doi: 10.3390/engproc2025104024
  • NewsRx. University of Ruse Researcher Discusses Research in Machine Learning (Hyperspectral Analysis of Apricot Quality Parameters Using Classical Machine Learning and Deep Neural Networks). Journal of Engineering. October 13, 2025; p 5436.