Semantic segmentation of buildings using optical satellite images and deep learning techniques is essential for urban planning and monitoring, especially in suburban areas. In this study, we focused on evaluating the performance of six deep learning models: DeepLabV3 MobileNetV3, DeepLabV3_ResNet50, FCN_ResNet50, EfficientNet-B0, ResNet-101, and UNET. The dataset was collected from the province of Mariscal Cáceres, specifically in the district of Juanjuí, located in the department of San Martín, situated in the northeast of Peru. Our analysis revealed varying levels of precision for each model: DeepLabV3 MobileNetV3 achieved 74.14%, DeepLabV3_ResNet50 reached 83.35%, FCN_ResNet50 attained 83.56%, EfficientNet-B0 yielded 61.37%, ResNet 101 obtained 63.60%, and UNET demonstrated 74.54%. These results provide insights into the effectiveness of different deep learning architectures for semantic segmentation tasks in suburban environments.