Automatic identification of building structure types using unmanned aerial vehicle oblique images and deep learning considering facade prior knowledge

Abstract

ABSTRACT Building structural type information is the foundation for seismic risk assessment and management since it reflects the behavior of buildings under seismic load. However, in earthquake-prone regions, most of this information is out-of-date or nonexistent. This paper proposes a deep learning-based method for automatically identifying building structural types from unmanned aerial vehicle (UAV) oblique images. The method consists of four steps: (1) collect facades of buildings with different structural types by web crawler technology as a sample set; (2) construct a convolutional neural network with a facade prior knowledge attention branch and train the model using the sample set; (3) extract building facades from UAV oblique images based on the georeferencing results of feature points as the test set; (4) identify building structural types by inputting the test set into the trained model. Three cases have been selected to verify the feasibility and applicability of the method. The average recall rate of 85% and the average F1 score of 83% have been achieved in areas with regular building distribution. This method integrates multidisciplinary knowledge to provide a solution for rapid collection of building vulnerability information, and expands the role of oblique photography data in urban management and disaster prevention planning.