Improvement method for tea leaf moisture content prediction using VIS-NIR spectrum based on transfer learning.

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Tea plant growth and quality are significantly influenced by moisture content.
• Traditional moisture detection methods are destructive, slow, and labor-intensive.
• Accurate and rapid moisture assessment is crucial for tea quality control.

Purpose of the Study:

• To develop a non-destructive method for detecting tea leaf moisture content using visible-near infrared (VIS-NIR) spectroscopy.
• To investigate and improve the generalization ability of predictive models across different sample batches.
• To explore the effectiveness of transfer learning and ensemble methods for optimizing moisture prediction.

Main Methods:

• Utilized visible-near infrared (VIS-NIR) spectroscopy in the 500-870 nm range for spectral data acquisition.
• Developed a prediction model (FD + SNV + CARS + ε-SVR) for tea leaf moisture content.
• Applied stacking ensemble learning and instance-based transfer learning to enhance model generalization.

Main Results:

• The FD + SNV + CARS + ε-SVR model showed excellent prediction for 2024 data (R_p=0.903) but poor generalization to 2022 data (R_p=0.138).
• Instance-based transfer learning achieved a high prediction accuracy (R_p=0.851) using only 55 transfer samples.
• Stacking ensemble learning achieved R_p=0.85 with 60 samples, demonstrating improved prediction with fewer samples.

Conclusions:

• VIS-NIR spectroscopy is a viable technique for rapid, non-destructive tea leaf moisture assessment.
• Transfer learning and ensemble methods significantly enhance model generalization and reduce the need for extensive sample data.
• Optimized models improve the practical application of spectroscopic methods in tea quality management.