Building Extraction from RGB Satellite Images using Deep Learning: A U-Net Approach

Automatic building extraction from satellite RGB images, is a low-cost alternative to perform important urban planning tasks. Yet, it is a challenging one, especially when natural and non-city block objects interfere in the semantic segmentation of algorithms that extract their key features. In this work we approach the automatic building extraction using a Convolution Neural Network based on the U-Net architecture. In contrast to existing approaches, it successfully encodes important features and decodes the buildings' localization by requiring both reduced computational time and dataset size. We evaluate the U-Net's performance using RGB images selected from the SpaceNet 1 dataset and the experimental results show an accuracy in building localization of 92.3%. Finally, favorable comparison with existing CNN approaches to hyper-spectral images targeting the SpaceNet 1 dataset, demonstrated its effectiveness.