Image-Based Prediction of Building Attributes with Deep Learning

Abstract

Construction material use causes about 11% of global GHG emissions and is an accelerating driver of global warming. In this research, we use image-based machine learning to predict the floor area and age of buildings which are strongly correlated with embodied GHG emissions. The ability to automatically estimate building attributes from street view images can enable large-scale analysis of the built environment and provide better differentiability compared to patch-wise or pixel-wiseestimation from satellite images. A ResNet-18 model is used for feature extraction, and area and age predictions are formulated as a regression problem and a classification problem, respectively. On area prediction, our model achieves a Mean Absolute Percentage Error of 22.32%. On age prediction, our model achieves a Balanced Accuracy (BA) of 78.05% and Accuracy of 79.05% when there are 3 age classes, but the BA and Accuracy drop to 61.94% and 63.53%, respectively when there are 6classes.