There is a pressing need for data-driven approaches in urban design. This project is addressing this need by looking at using machine learning to evaluate the visual appeal of streets. The output aims to deepen our understanding of cities and to identify salient visual features of the built environment that inﬂuence an individual’s preferences.
Try out the app
Try out the app and see if your visual preferences for city streets match what the machine learning model thinks.
Having trouble using the app? Try the full web version here.
Explaining the science
The project has created a map that illustrates the latent visual appeal of neighborhoods across Greater London. Using a semi-interpretable model the project researchers are able to extract the marginal eﬀects of the visual appeal of the urban environment on house prices. The visual preference varies from positive (green) through to negative (red).
The project proposes a novel approach to house pricing that leverages visual knowledge of the urban environment to improve predictive power. As the researchers do not have expert annotations of the visual preference of the urban environment, they have learnt feature extractors from Google Street View and satellite imagery by training an end-to-end machine learning model, while controlling for the contribution of the individual housing attributes.
As well as improving the accuracy of standard models, these visual proxies can be interpreted as diﬀerent levels of visual preferences. By comparing the visual proxy extracted from the revealed preference model and the stated preference collected from the survey, it's possible to identify where the algorithm breaks and thus motivating it to improving the model.
The objective of the web-app is to better understand people's individual visual preferences and to engage with the public on the uses of machine learning. The web-app allows people to rate the visual desirability of a random street, followed with a comparison of how the machine learning model rated the streets. The user can then provide feedback as to whether or not the model's ratings make sense.
By comparing the visual proxy extracted from the revealed preference model and its stated preferences, it will be possible to help the public to better understand both the capabilities and limitations of these machine learning models. The result can also help to give more reliable estimation on individual visual preferences of the built environment.
There are three important uses of such models:
The ﬁrst lies in accurately predicting house prices as a guide for realtors and for people looking to put their house on the market; for such individuals, accurate pricing is the most important criteria, where the inclusion of visual preferences can improve accuracy.
The second use of these models lies in econometrics; as well as improving the accuracy of standard models, these visual proxies will be of interest to economists on estimating the willingness to pay for diﬀerent levels of visual preference.
The third use of these models lies in urban planning. The implication is that these models can be used to improve the visual quality of streets and neighborhoods through the implementation of local housing policy. For example, this tool could be developed to identify potential hot spots for regeneration, for example there may be one ‘red’ street (less appealing) in amidst a 'green' (more appealing) area. Why is that? What are the problems there?