Façade feature extraction for urban performance assessments: Evaluating algorithm applicability across diverse building morphologies

Abstract

Urban feature extraction methods have presented multiple opportunities in the field of environmental building performance studies, and urban assessments. With reference to façade features, window size and position are key parameters that influence occupant perception and environmental performance of indoor spaces. A myriad of methods have been proposed to automatically detect façade window layouts from street view images yet it is still unclear how to assess the strengths and limitations of these methods or how they can be combined for higher detection performance. This paper aims to add clarity for those aiming to use computer vision techniques for façade layout extraction for urban level sustainability assessments. An automated pipeline to enable the extraction and detection of WWRs (Window to Wall ratios) is introduced that is based on two fundamentally different computational approaches; a grammar-based edge detection framework (Method 1) and a learning-based method (Method 2) that utilizes CNNs (Convolutional Neural Networks). The paper then compares the detection efficacy of both methods, focusing on WWR accuracy, in New York and Lisbon, including their ability to extract more detailed façade properties such as floor-to-floor heights. The study finds that the learning-based method shows lower error scores across the two cities. In Lisbon, 69 % of conditions were detected within the 10 % error range and 91 % were within the 20 % range under Method 1. Under Method 2, 82.5 % of conditions were within the 10 % error range and 95 % were within the 20 % range. In New York, 66 % of conditions were within the 10 % error range and 90 % were within the 20 % range under Method 1 while 77.5 % of conditions were within the 10 % error range and 93 % were within the 20 % range under Method 2. Finally, a hybrid method is proposed to leverage the strengths of the two models, and higher accuracies are obtained in both the New York and Lisbon dataset. In New York, 96.5 % are now detected within the 20 % error range and 81.5 % within the 10 % error range. In Lisbon, 96 % are now detected within the 20 % error range and 83.5 % within the 10 % error range. With reference to the total building height extraction formulated under Method 2, the results show a relative error of 3.5 % in height estimation in the sample set.