AGs-Unet: Building Extraction Model for High Resolution Remote Sensing Images Based on Attention Gates U Network

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Building contour extraction is crucial for urban planning and regional development.
• U-Net networks with skip connections improve segmentation accuracy by integrating multi-level features.
• Attention mechanisms enhance local feature expression in U-Net for better building extraction.

Purpose of the Study:

• To explore the effectiveness of attention gate modules in building extraction.
• To propose a novel Attention Gate Module (AG) and an Attention Gates U network (AGs-Unet).
• To enhance the automatic learning of diverse building structures and efficient contour extraction from high-resolution remote sensing images.

Main Methods:

• Proposed a novel Attention Gate Module (AG) by adjusting the 'Resampler' position within the attention gate.
• Developed the Attention Gates U network (AGs-Unet) by integrating AG modules into the skip connections of a U-Net architecture.
• Integrated AGs-Unet with U-Net to suppress irrelevant features and highlight dominant building features, improving attention map feature selection.

Main Results:

• AGs-Unet demonstrated improved feature learning and attention to small-scale buildings.
• Experiments on WHU and INRIA datasets showed AGs-Unet outperformed classic and state-of-the-art models (FCN8s, SegNet, U-Net, DANet, PISANet, ARC-Net).
• The proposed model effectively improved building extraction quality, prediction performance, and accuracy based on overall accuracy, precision, and intersection over union metrics.

Conclusions:

• The novel AGs-Unet model significantly enhances building extraction from high-resolution remote sensing imagery.
• The integration of attention gate modules improves feature selection and learning capabilities.
• AGs-Unet offers a more effective solution for accurate building contour extraction in urban planning and remote sensing applications.