Study on Quality Detection Methods for Table Grapes Based on Spectral and Imaging Information.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Traditional grape quality assessment is subjective and destructive.
• There is a growing need for rapid, objective, non-destructive quality evaluation methods.
• Hyperspectral imaging (HSI) offers potential for non-destructive analysis.

Purpose of the Study:

• To investigate the feasibility of using HSI combined with machine learning for non-destructive grape quality assessment.
• To develop models for both qualitative (variety classification) and quantitative (Soluble Solids Content - SSC) evaluation.
• To compare the performance of different machine learning algorithms.

Main Methods:

• Acquired hyperspectral data from four table grape varieties.
• Extracted texture features using Gray-Level Co-occurrence Matrix (GLCM) at key wavelengths.
• Built and compared classification and prediction models using Extreme Learning Machine (ELM), Convolutional Neural Network (CNN), and Partial Least Squares (PLS).

Main Results:

• The ELM model using full-range spectra achieved the highest classification accuracy (97.56%) and best SSC prediction (Rp2=0.75, RMSEP=0.81).
• CNN models demonstrated significant robustness in both classification and prediction tasks.
• Texture features combined with spectral data provided valuable information for quality assessment.

Conclusions:

• Hyperspectral imaging combined with machine learning is a viable and effective strategy for non-destructive fresh grape quality assessment.
• The developed models can accurately classify grape varieties and predict SSC.
• This approach overcomes limitations of traditional quality assessment methods.