Geographical information about buildings and other spatial objects is crucial for urban management and development, especially in the context of rapid urbanization. In recent years, the use of very high-resolution remote sensing for building segmentation has attracted considerable attention, driven by advancements in deep learning techniques. However, achieving the required accuracy in segmentation tasks using deep learning requires a large, manually labeled dataset, which can vary in characteristics across different areas and regions. To address this challenge, we applied a transfer learning approach for segmentation from satellite images with a small training dataset. In this study, we examined the effectiveness of a pretrained ResUnet architecture, which integrates U-Net and ResNet models, for building segmentation using a limited number of training samples. The experimental results demonstrated that transfer learning consistently outperforms training from scratch in both accuracy and computational efficiency. Specifically, the pretrained ResNet-101 backbone led to an improvement of approximately 4.3% in Intersection over Union (IoU) and reduced execution time by half. With the ResNet-18 backbone, the model achieved a 3.3% increase in precision and a fivefold improvement in processing speed. These findings confirm that acceptable accuracy in segmenting urban spatial features, such as buildings, can be achieved using transfer learning models pretrained on general-purpose datasets like ImageNet, even with a small set of training samples of very high-resolution remote sensing.