Multiscale Segmentation-Guided Fusion Network for Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) excel at feature extraction in Euclidean spaces for hyperspectral image classification (HSI).
• Graph Convolutional Networks (GCNs) capture spatial-contextual information in non-Euclidean spaces, improving HSI classification (HSIC).
• Current GCN methods for HSIC are limited by single-scale graph structures, hindering multi-range feature extraction.

Purpose of the Study:

• To propose a novel multiscale segmentation-guided fusion network (MS2FN) for enhanced HSIC.
• To overcome the limitations of single-scale graph structures in existing GCN-based HSIC methods.
• To improve feature representation by processing features from different spatial scales distinctly.

Main Methods:

• Constructing pixel-level graph structures using multiscale segmentation data to enable GCNs to extract features across various spatial ranges.
• Implementing distinct processing strategies for different feature types to enhance overall feature representation.
• Developing a multiscale segmentation-guided fusion network (MS2FN) integrating CNNs and GCNs.

Main Results:

• The proposed MS2FN method demonstrates superior performance compared to several state-of-the-art (SOTA) approaches in HSIC accuracy.
• The multiscale graph structure effectively extracts features across different spatial ranges, addressing limitations of prior GCN methods.
• Distinct feature processing strategies contribute to enhanced feature representation and improved classification outcomes.

Conclusions:

• The MS2FN offers a significant advancement in hyperspectral image classification by effectively leveraging multiscale spatial information.
• The method's ability to extract and fuse features from multiple scales leads to improved classification accuracy.
• The proposed approach provides a robust framework for future research in GCN-based HSIC.