Accurate Classification of Multi-Cultivar Watermelons via GAF-Enhanced Feature Fusion Convolutional Neural Networks.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Accurate classification of multi-cultivar watermelons is crucial for industry quality control.
• One-dimensional spectral interference hinders high-accuracy classification of visually similar watermelons.
• VIS-NIR transmittance spectroscopy offers potential for non-destructive quality assessment.

Purpose of the Study:

• To develop an innovative method for rapid online classification of multi-cultivar watermelons.
• To improve classification accuracy by addressing spectral interference issues.
• To integrate Gramian Angular Field (GAF) techniques with deep learning for enhanced feature extraction.

Main Methods:

• Converted one-dimensional VIS-NIR spectra into two-dimensional Gramian Angular Summation Field (GASF) and Gramian Angular Difference Field (GADF) images.
• Designed a dual-input convolutional neural network (CNN) architecture for feature fusion from GASF and GADF images.
• Applied wavelength optimization using competitive adaptive reweighted sampling (CARS) to enhance efficiency and accuracy.

Main Results:

• The GAF-enhanced CNN model achieved 95.1% classification accuracy on the prediction set, outperforming traditional one-dimensional spectral models.
• Wavelength optimization reduced GAF image generation time by 55.19% and improved accuracy to 96.3%.
• The model demonstrated good generalization with 91.9% accuracy on watermelons from different origins.

Conclusions:

• The proposed GAF-enhanced feature fusion CNN significantly improves multi-cultivar watermelon classification accuracy.
• This method offers an innovative approach for fruit quality assessment using VIS-NIR transmittance spectroscopy.
• The integration of GAF and CNNs provides a robust solution for challenges in spectral data analysis.