Historical city changes are most often observed by assessing the change in
survey and census data over the years. Such an approach neglects the clues that visual aesthetics of a neighbourhood carry with relation to change, such as the role of green space in gentrification. In this research we want to explore the potential for social media (Flickr) images to find such patterns.
Historical city change has often been observed through the lens of surveys and censuses. While these can have reliable year-to- year coverage, using such data sources neglects the role of visual information that accompanies urban change. Urban change studies sometimes try to link change patterns to visual aspects. For instance, large-scale neighbourhood renovations can be linked to gentrification. The replacement of local stores such as corner shops and small grocery stores is also believed to be a sign of ongoing neighbourhood change. Computer vision methods have been used to find patterns in image materials in a similar manner for images of the current urban environment, but not for historical image data. The topic of interest is therefore to be able to use historic image data to be able to quantify changes in the neighbourhood.
In this research the student will explore the use of social media data (i.e. Flickr) to find such historic change patterns. The student will download and organize images to assess which neighbourhoods can be studied using image data. The student will then use deep learning models that have been pre-trained on the Place Pulse dataset to determine how wealthy, active, safe, beautiful, and friendly each neighbourhood is based on the images available. The neighbourhood perception is then evaluated throughout the years in order to determine its trend. Lastly, the student will evaluate their findings against literature on urban change patterns such as gentrification and impoverishment to validate their findings.
- Create scripts that download and filter images of Dutch neighbourhoods from Flickr
- Use deep learning algorithms to determine perceived urban quality in Flickr images using pre-trained models
- Validate your findings with literature on urban change
- Discuss potential issues and benefits of Flickr images
- N. Naik, J. Philipoom, R. Raskar, and C. Hidalgo, “Streetscore – Predicting the Perceived Safety of One Million Streetscapes,” in 2014 IEEE Conference on Computer Vision and Pattern Recognition Workshops, Jun. 2014, pp. 793–799.
- L. Ilic, M. Sawada, and A. Zarzelli, “Deep mapping gentrification in a large Canadian city using deep learning and Google Street View,” en, PLOS ONE, vol. 14, no. 3, 2019.
- C. Hochstenbach and W. P. van Gent, “An anatomy of gentrification processes: Variegating causes of neighbourhood change,” en, Environment and Planning A: Economy and Space, vol. 47, no. 7, pp. 1480–1501, Jul. 2015
- Programming experience, or the strong motivation to learn it
- Experience with machine/deep learning, computer vision, or advanced statistics is highly recommended (e.g. Completion of GRS-34806 Deep learning course or equivalent)
Theme(s): Sensing & measuring; Integrated Land Monitoring