Topological Structure and Semantic Information Transfer Network for Cross-Scene Hyperspectral Image Classification.

Area of Science:

• Remote Sensing
• Computer Vision
• Machine Learning

Background:

• Cross-scene hyperspectral image (HSI) classification faces challenges due to domain shifts.
• Existing methods often rely on Convolutional Neural Networks (CNNs) that primarily capture local spatial features, neglecting crucial topological structures.
• This limitation hinders the accurate modeling of underlying data structures and inter-class relationships in HSI data.

Purpose of the Study:

• To develop a novel domain adaptation technique for cross-scene HSI classification.
• To address the limitations of existing CNN-based methods by incorporating topological structure information.
• To improve the robustness and accuracy of HSI classification in varied scenarios.

Main Methods:

• A Topological structure and Semantic information Transfer network (TSTnet) is proposed, integrating graph structures and Graph Convolutional Networks (GCNs).
• Graph Optimal Transmission (GOT) is employed for aligning topological relationships and distribution alignment using Maximum Mean Difference (MMD).
• Dynamic subgraph construction based on CNN features and a consistency constraint between CNN and GCN outputs are utilized.

Main Results:

• The proposed TSTnet demonstrates superior performance compared to state-of-the-art domain-adaptive approaches on three cross-scene HSI datasets.
• Integrating topological information alongside semantic features significantly enhances classification accuracy.
• The method effectively captures both local spatial and nonlocal topological relationships within HSI data.

Conclusions:

• TSTnet offers a robust and effective solution for cross-scene HSI classification by leveraging topological structure and semantic information.
• The integration of GCNs with CNNs provides a powerful framework for addressing domain adaptation challenges in HSI analysis.
• The developed approach advances the field of HSI classification, offering improved performance and broader applicability.