Structural-Spectral Graph Convolution With Evidential Edge Learning for Hyperspectral Image Clustering.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Hyperspectral image (HSI) clustering is crucial for analyzing unlabeled data.
• Existing graph neural network (GNN) methods struggle with spectral information and graph accuracy.
• Large-scale HSIs pose significant challenges for traditional clustering techniques.

Purpose of the Study:

• To develop an advanced clustering method for hyperspectral images that overcomes limitations of current GNN-based approaches.
• To improve the representation quality of superpixels by co-extracting spatial and spectral features.
• To enhance the accuracy and robustness of HSI clustering on large-scale datasets.

Main Methods:

• Proposed a structural-spectral graph convolutional operator (SSGCO) for improved superpixel representation.
• Introduced an evidence-guided adaptive edge learning (EGAEL) module to refine graph edge weights.
• Integrated SSGCO and EGAEL into a contrastive learning framework for simultaneous representation learning and clustering.

Main Results:

• The proposed method significantly improved clustering accuracy across four HSI datasets.
• Demonstrated superior performance compared to existing state-of-the-art methods, with accuracy gains up to 6.06%.
• Effectively addressed challenges related to spectral information exploitation and topological graph accuracy.

Conclusions:

• The developed SSGCO-EGAEL method offers a robust and accurate solution for hyperspectral image clustering.
• The approach enhances the understanding of complex HSI data by better utilizing spectral and spatial information.
• This work provides a valuable contribution to the field of remote sensing and machine learning for HSI analysis.