Image-based prediction of residential building attributes with deep learning

Abstract

This study estimates building attributes—floor area and age—using image-based machine learning. Building age and floor area are key inputs to the studies of urban metabolism, material stocks and flows, and embodied greenhouse gases (GHGs) in the built environment. However, these data are challenging to generate and maintain using traditional survey methods, their availability is uneven and often, even when available, very uncertain. Improving our understanding and future management of built environment resource flows and associated environmental impacts requires more complete access to building age and floor area data. The study formulates area prediction as a regression problem and age prediction as a classification problem over six historical periods, achieving a mean absolute percentage error of 19.42% for area prediction and an accuracy of 70.27% for age prediction in Toronto. These results are obtained using an EfficientNetV2 module for feature extraction from Google Street View images, followed by fully connected layers for estimating the two building attributes. The performance of the Toronto-trained model in five other Canadian cities is also reported, highlighting the model's varying effectiveness in different urban contexts and the benefit of local training. Our findings demonstrate the feasibility of using machine learning for building attribute estimation from street-view images, offering a basis for future automated large-scale material flow and stock analysis.