Leveraging the Recent Advances of Virtual Reality and Artificial Intelligence to Analyze Peoples’ Preferences for Green Spaces in an Urban Setting

Abstract

This thesis is motivated by (i) predictions that the expansion of urban land globally will reshape both natural and human-made landscapes, such as cityscapes, and (ii) the potential value of integrating choice models with new tools such as virtual reality (VR) and artificial intelligence (AI), the combination of which could more accurately assess people's preference for urban greening projects. First, an introduction to the thesis is given, discussing the background and context of economic valuation studies and, in particular, choice modelling. This chapter also introduces VR and AI tools before presenting the aims and objectives of this dissertation. Second, the immersive VR experiment is described. In particular, I conduct a splitsample experiment that quantifies the economic value of urban greenery (e.g., trees) with the following three different presentation formats: text only, video and VR. In this way, it can be verified whether a VR-enhanced discrete choice experiment (DCE) improves respondents' evaluation and interpretation of complex information and, therefore, improves their certainty when responding to a DCE survey. Thus, this thesis empirically demonstrates whether a VR-enhanced DCE can be used to address some of the biases that exist in DCE but also to elicit more accurately respondents' preferences and willingness to pay (WTP). To achieve our objectives, this thesis (i) identifies the scale parameters in the treatment groups through a mixed logit model that accounts for both preference heterogeneity and scale heterogeneity, and (ii) uses respondents' self-reported choice certainty expressed after each set of choices for all three subgroups. The results show that stated certainty is significantly higher for participants in the VR experiment compared to the text version. In addition, both forms of multimedia presentation (video and VR) show lower error variance compared to the text group with VR showing the lowest. This dissertation’s results suggest that this could be due to the improved evaluability of the video and VR format leading to reduced respondent uncertainty when making choices. Finally, it is showed that the WTP estimates are significantly influenced by the DCE presentation tool. The third chapter is motivated by the findings from Chapter 2 showing that a VRenhanced DCE can improve respondents' certainty and more accurately elicit their preferences and WTP. Therefore, this chapter analyzes the type of ecosystem services (ES) provided by urban greenery that are most important to the respondents, and derive their attitudes towards the ES provided by urban greenery. Based on the attitude data collected, the ordinal logit model shows that additional urban greenery can improve aesthetics, biodiversity and increase property values. To better understand people's preferences and in order to assess whether their attitudes about ES relate to their choices in a DCE, a controlled (classroom) experiment that illustrates the alternative designs in an immersive VR environment is conducted. The results of the study show that respondents who stated that urban greening could improve biodiversity are more likely to choose a greener street. Finally, respondents report the highest WTP for a symmetrical street design with few high canopy trees and lots of planters. In chapter four, this thesis combines traditional discrete choice models with machine learning (i.e., convolutional neural network for image segmentation) to understand people's preferences for urban planning. A choice model is informed with empirical preference data derived from an online pairwise choice experiment from the MIT Place Pulse platform, where the features of the photos (e.g., trees, lakes, buildings) were extracted using a deep neural network (DNN). Next, through the lens of a conditional logit model, the utility coefficients for different landscape features are estimated. The findings reveal that respondents strongly prefer urban green and blue, have a positive but slight preference for impervious surfaces and a disinclination to grey space (e.g., buildings) and sky. Chapter four concludes by comparing the predictive ability of the DNN-informed choice model with the predictive ability of an approach that is entirely grounded in deep neural networks (standalone DNN). These tests lead to the recognition that the DNN- nformed choice model is useful, as it combines the flexibility of a DNN for attribute selection with the interpretability of a choice model. Synthesizing the information presented above, chapter five of this dissertation presents the overall conclusions and policy recommendations. That chapter also proposes avenues for further research.