IASUNet: building extraction based on impoved attention Swin-UperNet

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Computer Vision

Background:

• Accurate spatial understanding of structures is crucial for environmental, urban safety, resource management, and planning.
• Existing models struggle with global context in complex urban remote sensing, leading to multi-scale challenges due to varying building proportions.

Purpose of the Study:

• To introduce an enhanced Attention Swin-UperNet (IASUNet) model for precise building classification in remote sensing imagery.
• To improve the extraction of global contextual information and address multi-scale challenges inherent in urban landscapes.

Main Methods:

• Integration of a Convolutional Block Attention Module (CBAM) between the Swin Transformer encoder and UperNet decoder.
• Implementation of Focal Cross Entropy (FCE) loss to manage class imbalance and multi-scale issues.

Main Results:

• The IASUNet model demonstrated superior performance compared to the original model.
• Significant improvements were observed in average accuracy (mAcc) and mean Intersection over Union (mIoU) for building identification.

Conclusions:

• The enhanced Attention Swin-UperNet (IASUNet) model effectively improves precise building classification accuracy.
• This advancement supports environmental protection, urban safety, resource optimization, and urban planning through better remote sensing data analysis.