SpaSSA: Superpixelwise Adaptive SSA for Unsupervised Spatial-Spectral Feature Extraction in Hyperspectral Image

Area of Science:

• Remote Sensing
• Image Processing
• Data Analysis

Background:

• Singular Spectral Analysis (SSA) is used for feature extraction in hyperspectral imaging (HSI).
• Existing SSA methods (1-D and 2-D) have limitations including sensitivity to window size, high computational cost, and inability to capture joint spectral-spatial features.

Purpose of the Study:

• To propose a novel Superpixelwise Adaptive SSA (SpaSSA) method for HSI feature extraction.
• To address the limitations of conventional SSA by exploiting local spatial information.
• To improve classification accuracy and computational efficiency in HSI analysis.

Main Methods:

• SpaSSA combines 1-D SSA and 2-D SSA, applying them adaptively to superpixels segmented from HSI.
• The choice between 1-D SSA and 2-D SSA, and the embedding window size in 2D-SSA, are adapted based on superpixel size.
• SpaSSA is combined with Principal Component Analysis (SpaSSA-PCA) for further performance enhancement.

Main Results:

• SpaSSA significantly outperforms standard SSA and 2D-SSA in classification accuracy across three datasets.
• SpaSSA demonstrates improved computational complexity compared to existing methods.
• SpaSSA-PCA further boosts land-cover analysis accuracy, surpassing state-of-the-art approaches.

Conclusions:

• SpaSSA offers a more effective approach to HSI feature extraction by leveraging local spatial characteristics adaptively.
• The adaptive nature of SpaSSA overcomes the drawbacks of fixed window sizes and improves computational efficiency.
• SpaSSA represents a significant advancement for HSI classification and land-cover analysis.