Multi-scale Meets Active Learning: A Deep Graph Fusion Paradigm for Hyperspectral Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Remote Sensing

Background:

• Deep learning (DL) methods for hyperspectral image classification (HSIC) often struggle with overfitting and oversmoothing, especially with limited labeled data.
• Existing methods fail to fully leverage unlabeled samples and multi-scale structural information, leading to information loss.
• Graph structures in DL models are typically fixed and do not incorporate prior information like labels to adaptively learn node relationships.

Purpose of the Study:

• To propose a novel deep fusion paradigm for multi-scale superpixel graphs (DFSG) to enhance HSIC performance.
• To address the challenges of overfitting, oversmoothing, and information loss in few-sample HSIC.
• To integrate multi-scale graph information and active learning (AL) within a unified framework.

Main Methods:

• Developed a Deep Fusion paradigm for Multi-scale Superpixel Graphs (DFSG) integrating graph-level and feature-level multi-scale information.
• Introduced a re-segmentation based graph correction module within an active learning (AL) process to adaptively learn graph structures.
• Implemented an iterative updating mechanism forming a symbiotic DFSG-AL framework.

Main Results:

• The proposed DFSG-AL framework demonstrated remarkable performance on five real hyperspectral image (HSI) datasets.
• Achieved significant improvements in few-sample hyperspectral image classification tasks.
• The integrated multi-scale approach and adaptive graph learning effectively reduced information loss and overfitting.

Conclusions:

• The DFSG-AL framework offers a powerful solution for few-sample hyperspectral image classification by effectively utilizing multi-scale information and adaptive graph structures.
• The symbiotic integration of AL and multi-scale graph methods enhances model robustness and classification accuracy.
• This approach provides a promising direction for future research in deep learning for HSIC.