Segmentation and classification of hyperspectral images using minimum spanning forest grown from automatically...

Area of Science:

• Remote Sensing
• Computer Vision
• Image Processing

Background:

• Hyperspectral image analysis requires accurate segmentation and classification.
• Existing methods often struggle with spectral-spatial information integration.

Purpose of the Study:

• To propose a novel spectral-spatial method for hyperspectral image segmentation and classification.
• To leverage minimum spanning forest (MSF) and automatically derived region markers.

Main Methods:

• Constructing a minimum spanning forest (MSF) from automatically identified region markers.
• Utilizing pixelwise classification to define reliable markers with class labels.
• Growing MSF trees to form regions and assigning class labels for spectral-spatial classification.
• Refining the classification map through pixelwise classification and majority voting.

Main Results:

• The proposed method achieves improved classification accuracies on hyperspectral airborne images.
• Accurate segmentation and classification maps are generated.
• Investigation into various dissimilarity measures for MSF construction demonstrated effectiveness.

Conclusions:

• The developed spectral-spatial classification scheme offers superior performance over prior techniques.
• The MSF-based approach effectively integrates spectral and spatial information for enhanced analysis.
• This method provides a robust solution for hyperspectral image segmentation and classification.