Spectral-Spatial Weighted Kernel Manifold Embedded Distribution Alignment for Remote Sensing Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Transfer learning in remote sensing image classification often suffers from feature distortions.
• Existing methods frequently overlook the crucial spatial information present in remote sensing images, focusing primarily on spectral data.

Purpose of the Study:

• To propose a novel method, spectral-spatial weighted kernel manifold embedded distribution alignment (SSWK-MEDA), for enhanced remote sensing image classification.
• To address feature distortions and effectively integrate both spectral and spatial information within a transfer learning framework.

Main Methods:

• A novel spatial information filter is employed to leverage similarities between adjacent pixels while mitigating the impact of non-sample pixels.
• A composite kernel is developed, combining weighted spatial and spectral kernels to adaptively balance their respective importance.
• Manifold space geometric structures are utilized to resolve feature distortions inherent in transfer learning for remote sensing data.

Main Results:

• The proposed SSWK-MEDA method demonstrates superior performance compared to several state-of-the-art techniques.
• The integration of spectral and spatial information through adaptive weighting significantly enhances classification accuracy.
• The spatial information filter effectively reduces noise and improves feature representation.

Conclusions:

• SSWK-MEDA offers a robust and effective approach for remote sensing image classification, particularly within transfer learning contexts.
• The method successfully combines the strengths of spectral and spatial feature analysis, overcoming limitations of existing approaches.
• The adaptive weighting mechanism and manifold embedding contribute to improved robustness against feature distortions.