Knowledge and Geo-Object Based Graph Convolutional Network for Remote Sensing Semantic Segmentation.

Area of Science:

• Remote Sensing
• Computer Vision
• Geographic Information Systems

Background:

• Pixel-based semantic segmentation models struggle with geographic object representation and topological relationships, leading to salt-and-pepper effects and low accuracy in remote sensing.
• Traditional graph neural networks (GNNs) rely heavily on sample context for information aggregation, which can be distorted and reduce node classification accuracy.

Purpose of the Study:

• To address the limitations of existing semantic segmentation models for remote sensing images.
• To propose a novel object-based model that effectively captures geographic objects and their relationships.
• To improve the accuracy and reduce artifacts in remote sensing image segmentation.

Main Methods:

• A knowledge and geo-object-based graph convolutional network (KGGCN) was developed.
• Superpixel blocks were utilized as nodes in the graph network.
• Information aggregation combined prior knowledge with spatial correlations, extending the receptive field to the entire study area.

Main Results:

• The KGGCN model demonstrated improved performance in semantic segmentation of remote sensing images.
• The model effectively overcame the distortion of sample context by incorporating broader prior knowledge.
• Experimental results showed a 3.7% improvement over Cluster GCN and a 4.1% improvement over U-Net.

Conclusions:

• The proposed KGGCN model offers a more robust solution for semantic segmentation in remote sensing.
• Integrating prior knowledge and geo-object information enhances the representation of geographic objects and their topological relationships.
• The KGGCN model effectively mitigates the salt-and-pepper effect and achieves higher segmentation accuracy.