Large language model-driven knowledge graph reasoning for enhanced semantic segmentation.

Area of Science:

• Geospatial Artificial Intelligence
• Remote Sensing
• Computer Vision

Background:

• Urban scene segmentation is vital for 3D city modeling and remote sensing applications like urban planning.
• Existing methods struggle with generalizability due to dataset-specific knowledge graphs.

Purpose of the Study:

• To develop a novel framework for remote sensing semantic segmentation using a universal knowledge graph.
• To enhance the adaptability and robustness of urban scene understanding across diverse data sources.

Main Methods:

• Leveraging large language models (LLMs) to construct a universal knowledge graph from multi-source geospatial data.
• Implementing a Graph Construction module for extracting cross-domain semantic relationships.
• Utilizing a Knowledge Graph Fusion module (KGFusion) to integrate the graph into semantic segmentation networks.

Main Results:

• Achieved 70.94% mIoU on the UAVid dataset and 63.23% on a mixed dataset.
• Outperformed baseline methods by 0.43% and 1.04% respectively.
• Demonstrated efficiency and adaptability in cross-domain scenarios.

Conclusions:

• The proposed framework effectively enhances remote sensing semantic segmentation using a universal knowledge graph.
• The method shows significant robustness and potential for broader applications in complex urban environments.