Estimation Method of Soluble Solid Content in Peach Based on Deep Features of Hyperspectral Imagery

Area of Science:

• Agricultural Science
• Remote Sensing
• Computer Vision

Background:

• Soluble solids content (SSC) is a key indicator of fruit quality.
• Hyperspectral imagery offers a non-destructive method for SSC detection.
• Previous methods using spectral or image features alone have limitations in accuracy.

Purpose of the Study:

• To develop a deep learning model for accurate SSC estimation in fresh peaches.
• To leverage hyperspectral image fusion for comprehensive feature extraction.
• To compare the performance of different neural network architectures.

Main Methods:

• Application of deep learning theory, specifically stack autoencoder-random forest (SAE-RF).
• Fusion of hyperspectral image information to create deep features.
• Design and evaluation of various SAE-RF network structures.

Main Results:

• The model utilizing deep features from hyperspectral image fusion outperformed models using spectral or image features alone.
• The SAE-RF model with a 1237-650-310-130 network structure achieved the best prediction (R² = 0.9184, RMSE = 0.6693).
• Demonstrated significant improvement in estimating soluble solids content.

Conclusions:

• The proposed hyperspectral image fusion method enhances the accuracy of non-destructive SSC estimation in peaches.
• This approach provides a theoretical foundation for detecting other fresh peach components non-destructively.
• Deep learning applied to fused hyperspectral data is effective for fruit quality assessment.