Classifying building energy efficiency from street view and aerial images in Denmark
ABSTRACT
Assessing a building’s energy efficiency is a slow and costly process, as current methods generally require on-site visits of qualified experts. To prioritize buildings in terms of retrofit potential at scale, we propose to classify the energy efficiency of existing buildings from widely available street view and aerial imagery. Based on images for more than 21,000 buildings in Copenhagen’s metropolitan area, this case study evaluates the predictive performance of multiple end-to-end machine learning models to classify buildings as energy efficient (EU rating A-D) or inefficient (EU rating E-G). In addition, we conduct an ablation study to examine the predictive signal of each data source individually. We find that the best end-to-end machine learning model trained on street view and aerial imagery achieves a macro-averaged F1-score of 58.79% and performs on a similar level as models from a case study in the UK. This suggests that the signal in street view and aerial imagery is consistent across geographies and can therefore support large-scale analyses in terms of building energy efficiency.