AUTOMATIC BUILDING EXTRACTION FROM UAV-BASED IMAGES AND DSMs USING DEEP LEARNING

Abstract

Accurate and automatic building footprint extraction from single UAV images has become essential in many photogrammetry and remote sensing applications such as 3D building modeling, smart city, monitoring, disaster management, and urban planning. In this paper, the capability of U-Net architecture with ResNet as the backbone of the network is investigated to extract the building footprints from UAV-based orthophotos and normalized Digital Surface Models (nDSMs) considering the complementary nature of RGB and height information. The data has been captured from five non-overlapping rural scenes of Yazd province, Iran. After pre-processing, the training and test datasets are prepared to evaluate the performance of U-Net using different hyperparameters and input channels such as RGB (only orthophotos) and RGBD (orthophotos and nDSMs). The experiments highlight the effectiveness of height information to detect and extract the building footprints with significant improvements in precision from 89% to 97% and in recall from 77% to 91%.