Unsupervised Classification of Polarimetric SAR Image Based on Geodesic Distance and Non-Gaussian Distribution...

Area of Science:

• Remote Sensing
• Geophysics
• Computer Vision

Background:

• Polarimetric Synthetic Aperture Radar (PolSAR) image classification is crucial for interpreting complex terrain.
• Existing unsupervised methods often struggle with nuanced scattering variations.
• Accurate classification aids in environmental monitoring and resource management.

Discussion:

• A novel unsupervised classification method for PolSAR images is proposed, leveraging geodesic distance and K-Wishart distribution.
• The method defines scattering similarity using geodesic distance between Kennaugh matrices for initial segmentation.
• Further sub-classification within primary scattering categories (surface, double-bounce, volume) is achieved using K-distribution's shape parameter (α) to capture heterogeneity.

Key Insights:

• The proposed method achieves effective initial segmentation into surface, double-bounce, and random volume scattering categories.
• Incorporation of K-distribution's shape parameter allows for detailed sub-classification based on scattering heterogeneity.
• Iterative application of the K-Wishart maximum likelihood classifier refines results and enhances classification accuracy.

Outlook:

• The method demonstrates superior performance across diverse PolSAR datasets (AIRSAR, ESAR, GaoFen-3) with varying resolutions and terrains.
• This unsupervised approach offers a robust alternative for PolSAR image interpretation, particularly where labeled data is scarce.
• Future work could explore adaptive parameter selection and integration with deep learning architectures for further advancements.