Multi-Scale Guided Context-Aware Transformer for Remote Sensing Building Extraction.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Building extraction from high-resolution remote sensing imagery is crucial for urban planning and disaster management.
• Existing methods face challenges due to high intra-class variability and multi-scale building distributions.

Purpose of the Study:

• To develop an advanced deep learning model for robust building extraction.
• To address limitations in current building extraction techniques using a novel network architecture.

Main Methods:

• Proposed the Multi-Scale Guided Context-Aware Network (MSGCANet), a Transformer-based framework.
• Integrated a Contextual Exploration Module (CEM) with dilated convolutions for feature enhancement.
• Designed a Window-Guided Multi-Scale Attention Mechanism (WGMSAM) for cross-scale dependency modeling.
• Employed a cross-level Transformer decoder with deformable convolutions for feature alignment.

Main Results:

• MSGCANet achieved high Intersection over Union (IoU) scores: 75.47% (Massachusetts), 91.53% (WHU), and 83.10% (Inria).
• The model obtained excellent F1-scores: 86.03% (Massachusetts), 95.59% (WHU), and 90.78% (Inria).
• Demonstrated robust and accurate building extraction performance across diverse datasets.

Conclusions:

• MSGCANet effectively addresses challenges in building extraction from remote sensing data.
• The proposed network architecture significantly improves the accuracy and robustness of building detection.
• The findings support the application of MSGCANet in urban planning and disaster management scenarios.