Development of a simulation method for the integrated prediction of change in both streetscape image and impression evaluation value

Abstract

The impression of the streetscape that constitutes residential areas is crucial to forming a good living environment. However, it is not easy to concretely and quantitatively examine how people's impression of an existing street varies due to landscape change or what landscape changes effectively improve the impression of a street. This paper proposes a simulation method for integrated prediction of changes both in street scene images and impression ratings. The method consists of two deep learning-based models; (1) a ranking learning model to evaluate street impressions based on a large-scale subject questionnaire [Kizawa 2021], and (2) a prediction model of streetscape changes based on CycleGAN [Zhu 2017] and VAEGAN [Larsen 2015]. In addition, using images of streetscapes in residential areas extracted from Google Street View as an example, we demonstrate the usefulness of the proposed method as a policy-making and consensus-building tool in urban planning.