DBFA-TSNet: A Three-Stage Building Extraction Network Based on Dual-Branch Fusion and Adaptive Enhancement

With the rapid development of ultrahigh resolution remote sensing technology, using remote sensing image for building extraction is particularly crucial for promoting refined urban planning and accelerating the urbanization process. However, buildings in remote sensing images often face challenges such as large-scale differences, uneven distribution, and unclear contours, which greatly affect the accuracy of building extraction and limit its application. To address these challenges, this article proposes a three-stage network DBFA-TSNet based on dual-branch fusion and adaptive enhancement for building extraction. This network addresses specific challenges in remote sensing images building extraction at each stage, improving the performance of building extraction. Notably, we design an interactive dual-branch feature fusion (IDFF) module that integrates features of different spatial resolutions and semantic levels, and captures comprehensive information ranging from fine-grained textures to large-scale semantics. Additionally, we introduce a dynamic adaptive feature enhancement (DAFE) module, which first enhances high-frequency details in the frequency domain and then captures subtle texture edges and large-scale building contours in the spatial domain. By combining operations in both frequency and spatial domains, we enhance the perception ability of building boundary information and provide comprehensive edge information. Extensive experiments on the WHU, Potsdam, and Massachusetts buildings datasets demonstrate that our method achieves superior performance in building extraction compared to state-of-the-art methods.