Due to the complex environment surrounding buildings and the varying scales of buildings，smaller buildings are prone to issues such as incomplete or missed segmentation，which leads to a reduction in the accuracy of extraction results. To address these challenges，an improved model based on the Unet++ architecture （REUnet++） was proposed. This model enhanced performance by incorporating a ResNet34 residual network as the encoder structure. Additionally，an attention module，SE，was integrated into the model to strengthen its ability to extract important features from the dataset. Experiments were conducted on the public xBD dataset. The experimental results demonstrate that the REUnet++ model outperforms existing models in both feature extraction and segmentation accuracy in complex scenarios. Compared to the Unet++ model，the F1 score increases by 3.08%，and the intersection over union score improves by 4.68%，with significant improvements observe in other relevant performance metrics. Finally，the model’s generalization ability was further validated using the WHU building dataset.