@article{Suppa2025, 
author = {Anthony Robert Suppa and Alessandro Aliberti and Marta Carla Bottero and Vincenzo Corrado},
title = {Detecting window-to-wall ratio for urban-scale building simulations using deep learning with street view imagery and an automatic classification algorithm},
year = {2025},
journal = {Building Simulation},
volume = {18},
number = {8},
pages = {2175-2199},
keywords = {building-and district-scale WWR, machine learning, street view imagery, automatic image extraction and stitching, building façade classification, urban building energy modeling (UBEM)},
url = {https://www.sciopen.com/article/10.1007/s12273-025-1301-3},
doi = {10.1007/s12273-025-1301-3},
abstract = {Machine learning techniques can fill data gaps for urban-scale building simulations, particularly gaps around window-to-wall ratio (WWR). This study presents a comprehensive workflow to (1) automatically extract and stitch images from Google Street View (GSV); (2) label images with a custom Rhino-based tool to aid annotation of occluded glazing; (3) detect wall, garage, and glazing objects by training and validating a YOLOv9 deep learning model with three added post-scripts; (4) calculate WWR at façade, building, and district scales; and (5) simulate district energy consumption in an urban building energy model (UBEM). Results include a 96% image-capture rate from GSV, indicating a robust extraction and stitching algorithm. Converting model detections into WWR, 94% and 100% of façades have detected WWRs within ±5% and ±10% of ground truth WWRs, respectively. A novel automatic algorithm upscales façade detection to estimate WWR at non-street-facing sides and rears, resulting in distinct WWRs for each face of each building. For a case study in Turin, Italy, WWR detections are +5.2% and +6.9% greater when upscaling based on OpenStreetMap and municipal GIS data, respectively, compared to TABULA, leading to 1.5% and 35.5% increases in heating and cooling energy need in the UBEM. The workflow is made openly available to support future research in other contexts.}
}