Automatic Building Height Estimation: Machine Learning Models for Urban Image Analysis

Abstract

Artificial intelligence (AI) is delivering major advances in the construction engineering sector in this era of building information modelling, applying data collection techniques based on urban image analysis. In this study, building heights were calculated from street-view imagery based on a semantic segmentation machine learning model. The model has a fully convolutional architecture and is based on the HRNet encoder and ResNexts depth separable convolutions, achieving fast runtime and state-of-the-art results on standard semantic segmentation tasks. Average building heights on a pilot German street were satisfactorily estimated with a maximum error of 3 m. Further research alternatives are discussed, as well as the difficulties of obtaining valuable training data to apply these models in countries with no training datasets and different urban conditions. This line of research contributes to the characterisation of buildings and the estimation of attributes essential for the assessment of seismic risk using automatically processed street-view imagery.

Keywords

• artificial intelligence
• semantic segmentation
• convolutional neural networks
• building height estimation
• seismic exposure
• street view imagery