Detection of blueberry based on hyperspectral imaging and deep learning

Area of Science:

• Agricultural Science
• Spectroscopy
• Machine Learning

Background:

• Blueberry sugar content significantly impacts flavor and market value.
• Accurate, rapid sugar detection is crucial for blueberry production and quality control.
• Hyperspectral imaging offers a non-destructive method for analyzing fruit quality.

Purpose of the Study:

• To explore hyperspectral imaging combined with deep learning for detecting blueberry sugar content.
• To evaluate the effectiveness of fractional-order derivative (FOD) and improved Laplacian eigenmap (ILE) for spectral data preprocessing and band selection.
• To develop an accurate and efficient model for predicting sugar levels in blueberries of different cultivars and maturity stages.

Main Methods:

• Hyperspectral data from three blueberry cultivars (F6, L11, L25) were collected.
• Data preprocessing involved Multi-Scale Combination (MSC) and fractional-order derivative (FOD).
• Characteristic bands were selected using the improved Laplacian eigenmap (ILE), followed by shallow convolutional neural network (CNN) model training.

Main Results:

• The optimized CNN model, utilizing MSC + FOD preprocessing (1.10 order) and ILE band selection, achieved high prediction accuracy (Rp² = 0.8597).
• The method demonstrated robustness and accuracy in complex scenarios, outperforming traditional approaches.
• The combination of FOD, ILE, and a lightweight CNN model significantly improved sugar content prediction performance.

Conclusions:

• Fractional-order derivative (FOD) combined with improved Laplacian eigenmap (ILE) is effective for processing blueberry hyperspectral data.
• The developed shallow CNN model offers a robust and accurate solution for detecting blueberry sugar levels in diverse conditions.
• This approach holds significant potential for the rapid, non-destructive quality assessment of blueberries in commercial settings.