A vegetation classification method based on improved dual-way branch feature fusion U-net

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Traditional U-Net models struggle with complex parameters and feature extraction for vegetation classification.
• Hyperspectral remote sensing images contain rich spectral information but require efficient processing.

Purpose of the Study:

• To propose an improved U-Net deep network with a dual-way branch input for enhanced vegetation classification.
• To address the limitations of complex structures and low feature extraction capabilities in existing models.

Main Methods:

• Utilized Principal Component Analysis (PCA) for hyperspectral image dimension reduction and effective band selection.
• Integrated depthwise separable convolution and residual connections into the U-Net architecture to reduce parameters and improve feature extraction.
• Developed a dual-way input strategy combining PCA features with fused artificial features (NDVI, GLCM, edge features).

Main Results:

• The improved U-Net with residual structure and depthwise separable convolution achieved 97.13% precision and 92.36% recall.
• The dual-way branch network, incorporating both PCA and artificial features, reached an accuracy of 98.67%.
• Artificial features negatively impacted accuracy in a single-channel network but proved effective complements in the dual-way design.

Conclusions:

• The proposed improved U-Net with depthwise separable convolution and residual connections enhances vegetation classification performance.
• The dual-way branch input strategy effectively integrates diverse features, significantly boosting classification accuracy.
• Artificial features serve as valuable complements to network-extracted features in hyperspectral vegetation classification.