Deep Learning for Building Extraction from High-Resolution Remote Sensing Images

The extraction of buildings from satellite images of very high spatial resolution is an important issue for many applications mainly related to the implementation of public policies, urban development, land use planning, and the updating of geographic databases, and during the last two decades, it has been the subject of much research. Several existing classical techniques have been proposed in remote sensing images, but they have several limitations that prevent segmenting buildings with high accuracy. We propose, in this paper, a U-net architecture with a ResNet50 encoder for the extraction of buildings from Massachusetts building datasets, in order to automate the production chain of urban three-dimensional models. The results obtained in this study show a very promising segmentation with huge accuracy, it outperforms many presented models with 82.63% of intersection over union (IoU).