Sophisticated Vegetation Classification Based on Feature Band Set Using Hyperspectral Image.

Area of Science:

• Remote Sensing
• Computer Vision
• Ecology

Background:

• Sophisticated vegetation classification (SVC) using hyperspectral imagery faces challenges with spectral information limitations and noise in classification results.
• Existing methods struggle to meet application requirements for detailed vegetation types and are affected by salt-and-pepper noise.

Purpose of the Study:

• To propose a novel SVC strategy by constructing and optimizing a vegetation feature band set (FBS).
• To enhance the separability of different vegetation types and preserve detailed spatial information during image processing.

Main Methods:

• Incorporated 30 spectral indices sensitive to biological parameters into the FBS, alongside original spectral and texture features.
• Developed a spectral-dimension optimization algorithm based on class-pair separability (CPS).
• Implemented a spatial-dimension optimization algorithm using neighborhood pixels' spectral angle distance (NPSAD) for noise reduction while retaining details.

Main Results:

• Experimental results using airborne hyperspectral imagery demonstrated significant improvements in SVC accuracy.
• The proposed method effectively addressed limitations of using spectral information alone and reduced noise artifacts.

Conclusions:

• The developed SVC strategy offers a robust approach for accurate vegetation classification.
• Potential applications include crop species identification, invasive species monitoring, and precision agriculture, benefiting from improved classification accuracy and detail preservation.