Original title: Label-Efficient Semi-Supervised Building Semantic Segmentation for High-Resolution UAV Imagery
Gahyun Lee, Junseo Baek, Youkyung Han · September 2026
Drones take sharp aerial photos that help map cities and create three-dimensional models. To find buildings in these images, computers usually need people to label every pixel by hand, which takes massive effort. Researchers built a system that learns from a small batch of labeled pictures alongside many unlabeled drone photos from new areas. The system uses a fixed image model to recognize basic shapes and a decoder to outline objects. A student model learns from a teacher model that labels the new images when confident. The team also altered colors and internal features to keep building outlines intact, and checked boundaries only against real labels. It reached an IoU of 0.9197 on Wonju images and 0.8921 on Seoul images.
Imagine you are learning to identify a specific bird species from photos. You have a few clear pictures of that bird from one forest, but you need to recognize it in a different forest where the lighting and background are different. You cannot get new labeled photos from the new forest, but you have many unlabeled photos. Here, the model learns to spot buildings from a few labeled drone images of one city, then uses many unlabeled images from a new city to improve its accuracy, while carefully preserving the building shapes.
Urban planners and disaster response teams need accurate building maps from drone imagery, but manually labeling each pixel is slow and expensive. This work shows that a model trained on just a few labeled images from one city can reliably segment buildings in a completely different city without any new labels. That means faster, cheaper mapping for new survey regions, enabling quicker decisions in urban development and emergency response.
The abstract doesn't say explicitly, but the method relies on a pre-trained DINOv2 encoder, so its performance depends on that model's features. It was only tested on two Korean cities; results may differ for other regions or building styles. Boundary IoU scores are much lower than overall IoU, showing that edge accuracy remains difficult. The framework assumes labeled source and unlabeled target imagery are available; extreme domain shifts might still cause failures.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Label-Efficient Semi-Supervised Building Semantic Segmentation for High-Resolution UAV Imagery — Gahyun Lee, Junseo Baek, Youkyung Han — Remote Sensing — doi:10.3390/rs18183250
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.