Building segmentation through a gated graph convolutional neural network with deep structured feature embedding.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Automatic building extraction from optical imagery is challenging due to complex shapes.
• Deep convolutional neural networks (DCNNs) excel at pixel-level classification but struggle with precise boundary delineation.
• Progressive down-sampling in deep architectures limits fine-grained segmentation accuracy.

Purpose of the Study:

• To develop a generic framework for overcoming boundary delineation issues in semantic segmentation.
• To enhance the precision of building footprint extraction from optical imagery.
• To propose a novel approach integrating graph convolutional networks (GCNs) and deep structured feature embedding (DSFE).

Main Methods:

• Integration of a gated graph convolutional network (GCN) with deep structured feature embedding (DSFE) into an end-to-end workflow.
• Development of a gated GCN architecture to refine coarse semantic predictions and generate sharp borders.
• Comparative analysis of different feature embedding architectures and GCNs for building footprint segmentation.

Main Results:

• The proposed framework with the gated GCN architecture significantly outperforms state-of-the-art approaches in building footprint extraction.
• The method achieves fine-grained pixel-level classification with accurate boundary delineation.
• Demonstrated effectiveness in refining weak semantic predictions for improved segmentation results.

Conclusions:

• The novel framework effectively addresses the challenge of precise boundary delineation in semantic segmentation tasks.
• The proposed gated GCN architecture offers a significant advancement over classic GCNs for fine-grained image analysis.
• The method is broadly applicable to various binary and multi-label segmentation tasks beyond building footprint extraction.