Abstract: Addressing the issues of incomplete boundary segmentation and frequent omission of small-scale buildings in remote sensing image building extraction tasks, this paper proposes an enhanced U-Net-based building extraction network named MU-UNet. Building upon the classical U-Net architecture, the proposed method incorporates a Hiera encoder and a densely connected multi-scale output structure to strengthen cross-hierarchical contextual modeling capabilities. It further integrates Squeeze-and-Excitation (SE) modules to enhance adaptive feature representation across channels. Additionally, a Convolutional Edge-Focal Module (CEFM) module is designed to reinforce fine-grained boundary modeling and cross-scale feature fusion. The experimental results show that the method proposed in this paper outperforms the mainstream models in terms of both IoU and F1 metrics on the WHU aerial image dataset, the WHU satellite dataset, and the Massachusetts dataset. Specifically, on the WHU aerial image dataset, MU-UNet improves IoU(Intersection over Union) and F1 by 3.65% and 2.04% respectively compared to the baseline U-Net; on the Massachusetts dataset, the improvements are 4.99% and 3.47% respectively.