This project addresses the task of building segmentation on the pre-disaster subset of the xBD dataset using a U-Net with a ResNet-34 encoder. Throughout a series of experiments, we inves- tigated factors that limited the model’s ability to produce accurate building masks, including data imbalance, boundary precision, and the model’s reliance on contextual cues. We evaluated several modifications—learning-rate tuning, normalization, boundary-aware loss, Grad-CAM analysis, data augmentation, and hard negative mining—but found that these had little effect compared to the choice of sampling strategy. The main performance gains came from rebalancing the training data through building-centered random crops, which exposed the model to more informative examples and improved the building IoU from 0.57 to 0.67. Our results highlight that, for this dataset, data selection plays a more decisive role than architectural or loss-function adjustments, and that controlling what the model sees is essential for achieving reliable segmentation performance.