Urban-scale estimation of window-to-wall ratio from street view imagery via computer vision for improved building energy modeling

The window-to-wall ratio (WWR) is an important part of building behavior. However, WWR data is not publicly available in most countries. This study proposes a rapid and automated method for estimating the WWR of large-scale buildings using Google Street View (GSV) imagery and computer vision techniques. In contrast to conventional approaches based on field surveys, drawing analysis, or manual modeling, this study offers a scalable and efficient framework that can replace labor-intensive processes. WWR was estimated for 15,740 buildings in a metropolitan district of Seoul, and the entire process was completed in 6 h. The accuracy of WWR estimation was evaluated using manual labeling on 100 building images. As manual calculation of actual WWR is difficult at a large scale, its validity was indirectly assessed based on the change in UBEM accuracy. The estimated WWR showed a median of 17.0%, with significant variation across primary building uses. When all estimated values were used as input, UBEM prediction accuracy improved by 7.42%, increasing to 35.83% for buildings in which the inclusion of estimated WWR improved UBEM accuracy. When envelope area information was used in addition to WWR, prediction accuracy improved by 9.22% for all buildings and by 41.33% for those with improved UBEM accuracy. In the improved buildings, higher WWR led to greater improvements in prediction accuracy. The median improvement was 16.14% for the 30–40% WWR range, 21.49% for 40–50%, and 30.25% for WWR over 50%. Occlusion, glare, low contrast, and image stitching errors were major issues hindering accurate WWR estimation. In addition, the tolerance limits of these four issues were quantified. This study proposes a framework that automatically estimates WWR, incorporates it into UBEM, and indirectly validates the estimates through changes in UBEM accuracy, enhancing the accuracy of urban-scale energy modeling.