Optimising deep learning for building extraction: Dataset efficiency and model backbones under data constraints

Accurate building extraction from very high-resolution (VHR) satellite imagery is critical for urban planning, disaster response, and environmental monitoring. However, the performance of deep learning models remains highly sensitive to training sample size, model complexity, and learning strategy, especially in data-scarce scenarios. Here, we systematically evaluated the DeepLabV3+ architecture with ResNet backbones (i.e., ResNet-18, 50, 101, 152) across varying training sample sizes (i.e., 20–120% of the primary WHU dataset) and three training approaches: random initialisation, ImageNet pre-training, and pre-training with data augmentation. Our results revealed a U-shaped relationship between dataset size and performance, with segmentation accuracy peaking at full dataset usage and declining when additional samples introduce redundancy. Peak performance reached 88.0% IoU and 93.2% F1-score under optimised configuration, while shallower models achieved optimal performance under limited data availability. We found that hybrid learning strategies are essential for mitigating overfitting and achieving high accuracy, with transfer learning improving accuracy by 7.9% IoU (5.8% F1-score), and data augmentation offering an additional 1–3% IoU (0.2–2.3% F1-score) gain in low-data settings (≤40%). Deeper networks (ResNet-101/152) achieved superior performance only when trained with ≥60% of the dataset and appropriate regularisation. The stability of these data-dependent model selection patterns was further confirmed through external validation on the Japan Building Dataset, demonstrating transferability across geographic contexts. Our findings yield practical and generalisable guidelines: (i) avoiding unnecessary dataset expansion, (ii) prioritising transfer learning and augmentation when data is scarce, and (iii) aligning model depth with data availability. By explicitly linking model selection to data availability, this study supports efficient and reliable deployment of deep learning for urban analytics and disaster response under realistic annotation constraints. • Benchmarked DeepLabV3+ with ResNet backbones for building extraction tasks. • Found U-shaped accuracy trend across 20–120% training data regimes. • Pretraining and augmentation boosted IoU by up to 9.1% in limited data settings. • Early stopping reduced overfitting by 10% and improved model efficiency. • Provide practical guidelines for deep learning under data constraints.