A Multi-Scale attention network for building extraction from high-resolution remote sensing images.

Area of Science:

• Computer Science
• Remote Sensing
• Artificial Intelligence

Background:

• Remote sensing images often yield incomplete building contours and struggle with complex urban scenes.
• Existing semantic segmentation networks face limitations in accurately extracting building boundaries.

Purpose of the Study:

• To propose a novel multi-scale network with dual attention mechanisms for precise building boundary extraction.
• To enhance the adaptability and accuracy of building extraction in complex remote sensing scenarios.

Main Methods:

• Implemented a multi-scale network incorporating Squeeze-and-Excitation (SE) and Atrous Spatial Pyramid Pooling (ASPP) modules.
• Integrated channel grouping shuffle and dual attention mechanisms in the decoding phase for feature interrelation analysis.
• Developed a hybrid loss function to mitigate class imbalance and stabilize network training.

Main Results:

• The proposed method achieved superior performance compared to PSPnet, U-net, and DAnet on two high-resolution datasets.
• Achieved a high F1 score of 83.23% and a Mean Intersection over Union (MIoU) of 73.56%.
• Demonstrated marked improvements in accuracy, recall, F1 score, and MIoU.

Conclusions:

• The developed multi-scale attention network effectively extracts clear building boundaries from remote sensing images.
• The method shows significant potential for practical applications in automated building extraction.
• The proposed approach offers enhanced accuracy and robustness for complex building scenes.