Spectral-Spatial Shared Linear Regression for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Hyperspectral image (HSI) classification is crucial for many applications but faces challenges due to high dimensionality and limited data.
• Subspace learning (SL) methods are commonly used to address the high-dimensional small-sized problem in HSI classification by reducing pixel dimensions while retaining discriminant information.

Purpose of the Study:

• To propose a novel spectral-spatial shared linear regression (SSSLR) method for HSI feature extraction and classification.
• To enhance the discriminant capability of linear projection matrices by incorporating spatial structure and a shared learning model.

Main Methods:

• Developed a spectral-spatial shared linear regression (SSSLR) framework inspired by ridge linear regression (RLR).
• Utilized a convex set to explore spatial structure for computing the linear projection matrix.
• Employed a shared structure learning model with original and hidden feature spaces to learn a more discriminant projection matrix.
• Designed an efficient iterative algorithm for optimizing the proposed SSSLR method.

Main Results:

• Experimental results on the Indian Pines and Salinas HSI datasets demonstrated the effectiveness of the proposed SSSLR method.
• SSSLR significantly outperformed several existing subspace learning methods in HSI classification tasks.
• The method successfully leveraged both spectral and spatial information for improved feature representation.

Conclusions:

• The proposed spectral-spatial shared linear regression (SSSLR) method offers a powerful approach for hyperspectral image classification.
• Incorporating spatial information and shared structure learning enhances the discriminant power of feature extraction.
• SSSLR provides a robust and efficient solution for the high-dimensional small-sized problem in HSI analysis.