Answers are grounded in this paper and cite their sources.
High-resolution UAV imagery lets municipalities map building footprints where no cadastral record exists, and Nepal, among the fastest urbanizing countries, has little such data. Public benchmarks cover Europe, North America, and East Asia, and UAV datasets are mostly private or limited to Chinese cities, so South Asian and Himalayan settlements are uncovered. No study has compared convolutional, transformer, and hybrid architectures across a six-fold ground sampling distance gap, or tested transfer between two settlements of one valley. We annotated two new UAV datasets over Khokana (0.15 m/px) and Lamatar (0.026 m/px), peri-urban settlements in Lalitpur, Nepal, and resampled Lamatar to a common 0.15 m/px grid. We trained a U-Net with ResNet encoders, SegFormer-B2, and TransResUNet from scratch, with full fine-tuning, and with a frozen encoder, then transferred the best model per site in both directions. SegFormer gave the highest within-site IoU (85.9% Lamatar, 80.7% Khokana) and transferred into Lamatar at 82.5% with no adaptation. Transfer into Khokana began at 65.7% and needed full fine-tuning, which lifted all three architectures past their Khokana ceilings. Agencies without cadastral data can annotate one area in full and adapt to a second from a small sample, and the transfer direction determines the strategy.
No takes yet. Share an insight, caveat, or question.
Thapa et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: