Automated building extraction from very-high-resolution satellite imagery is a core task for urban planning, land-use monitoring, and disaster management, yet dense construction, occlusion, and spectral confusion with paved and roofed surfaces continue to limit the accuracy of automated pipelines. This study evaluates the accuracy and efficiency of YOLOv11-L, the large variant of Ultralytics YOLO release, for detecting buildings in high-resolution satellite imagery of Karaj, Iran, a rapidly urbanizing city with dense residential and industrial texture. A dataset of 2,517 satellite images was labeled for two land-cover classes, buildings and vegetation, on the Roboflow platform, and expanded through augmentation to 7,552 images split 83/9/8% into training, validation, and test sets. The model was trained on the Ultralytics platform for 150 epochs with a learning rate of 0.001, batch size of 4, input resolution of 1280 pixels, the SGD optimizer, and a patience of 140. The trained model reached mAP@0.5 = 0.946, mAP@0.5:0.95 = 0.598, precision = 0.932, recall = 0.895, and F1-score = 0.913 on the held-out test set, indicating strong and stable performance in localizing and distinguishing buildings from surrounding urban texture. Benchmarked against recent YOLO-based building-extraction studies published in 2024-2025, the reported accuracy is competitive with, and in several respects exceeds, comparable pipelines built on earlier YOLO releases. The paper discusses the sources of residual error - principally shadow, glare, and off-nadir geometry - and outlines directions for future work, including multi-sensor fusion and hybrid transformer-CNN architectures.