Convolutional neural networks for predicting the perceived density of large urban fabrics

Abstract

Urban density, along with the associated urban morphology and topology, significantly influences human perception, emotions, and behavior, ultimately affecting our overall well-being. Over the past decades, experts have developed spatial analysis models and tools which evaluate how planning and design impact urban residents and the functionality of cities. One such spatial analyses model is the Urban Spatial Openness Index (USOI) which utilizes ray-casting to conduct 3D visibility analysis predicting the perceived density of entire cities on a macro-scale, represented as 2D heatmaps. In the urban scale, ray-casting analysis is computationally intense and requires significant resources, which hinders its effective application. In this paper, we use a Convolutional Neural Network (CNN) to train a model to predict perceived density in urban fabrics based on 2D heatmap images. The processes described in this paper include creating a dataset of corresponding USOI images and height images from several different cities, training a CNN model, and evaluating the model's performance. The model predicts USOI with a mean absolute error of 1.92 %, which is considered highly accurate for visual perception on the urban scale. This study showcases the capability of CNN models to predict perceived density as measured by the USOI. The use of a predictive model can significantly reduce the processing time of 3D visibility analysis on the urban scale.