Research on Intelligent Wood Species Identification Method Based on Multimodal Texture-Dominated Features and Deep...

Area of Science:

• Wood science
• Computer vision
• Machine learning

Background:

• Traditional wood identification relies on manual expertise, leading to slow speeds and low accuracy.
• Developing automated, robust wood species identification is crucial for forestry and trade.

Purpose of the Study:

• To address limitations in traditional wood identification methods.
• To propose an intelligent wood species identification method using fused spectral and texture features from hyperspectral images.
• To enhance accuracy and robustness in wood classification.

Main Methods:

• Collected hyperspectral images of 10 wood species from Puer, Yunnan, China.
• Extracted spectral features using similarity matrices and representative band selection.
• Extracted texture features from PCA components and fused them with spectral features.
• Developed a deep learning ST-former model for complementary collaborative learning and feature fusion.

Main Results:

• Constructed a hyperspectral wood cross-section database.
• Achieved an overall classification accuracy of 90.27% with the proposed joint model.
• Demonstrated an average improvement of approximately 8% compared to single-modal models.

Conclusions:

• The proposed multimodal deep learning approach effectively fuses spectral and texture features for accurate wood species identification.
• This method offers a robust and efficient alternative to traditional manual identification techniques.
• The study highlights the potential of hyperspectral imaging and deep learning in wood science applications.