Multi-Scale Frequency-Spatial Domain Attention Fusion Network for Building Extraction in Remote Sensing Images

Building extraction from remote sensing images holds significant importance in the fields of land resource management, urban planning, and disaster assessment. Encoder-decoder deep learning models are increasingly favored due to their advanced feature representation capabilities in image analysis. However, because of the diversity of architectural styles and issues such as tree occlusion, traditional methods often result in building omissions and blurred boundaries when extracting building footprints. Given these limitations, this paper proposes a cutting-edge Multi-Scale Frequency-Spatial Domain Attention Fusion Network (MFSANet), which consists of two principal modules, named Frequency-Spatial Domain Attention Fusion Module (FSAFM) and Attention-Guided Multi-scale Fusion Upsampling Module (AGMUM). FSAFM introduces frequency domain attention and spatial attention separately to enhance the feature maps, thereby strengthening the model's boundary-detection capabilities and ultimately improving the accuracy of building extraction. AGMUM first resizes and concatenates attention enhancement maps to enhance contextual understanding and applies attention guidance to further improve prediction accuracy. Our model demonstrates superior performance compared to existing semantic segmentation methods on both the WHU building data set and the Inria aerial image data set.