Fig 1.
Flowchart for developing and implementing the image recognition model.
Methodology behind developing a protocol that defines new-build gentrification traits (Part A) and applying the protocol on Street View Imagery (SVI) pairwise imagery of residential frontages (Part B). Photos were taken by the research team and are for illustrative purposes.
Table 1.
List of built indicators that are connected to new-build gentrification.
Fig 2.
Training data for image recognition model.
Examples of pairwise imagery in our new-build gentrification audit and associated labels of “No Change” versus “Change” which indicates new-build gentrification is apparent. In order to comply with CC-BY copyright, the photos were taken by the research team for illustrative purposes.
Fig 3.
ResNet-50 Model Architecture Flowchart.
ResNet-50 architecture for new-build gentrification prediction using pairwise Street View Imagery (SVI) as data inputs. Building photos were taken by the research team and are for illustrative purposes.
Fig 4.
Comparative new-build development heatmaps.
Comparison between Kernel Density Estimate (KDE) maps of OpenDataPhilly Licenses and Inspections (L&I) Building and Zoning Permits data (a) and audited data points (b). Census tract boundaries and Licenses and Inspections (L&I) Building and Zoning Permits generated using data from Tim Wisniewski (2016), licensed under the MIT License. © Tim Wisniewski. Retrieved from: https://opendataphilly.org/datasets/census-tracts/ and https://opendataphilly.org/datasets/licenses-and-inspections-building-and-zoning-permits/.
Table 2.
Summary of ResNet-50 Fine-tuning results.
Fig 5.
Classification results on gentrified and non-gentrified paired images.
Comparison of the normalized matrix displaying classification probabilities (left) and the actual confusion matrix with counts of correctly and incorrectly classified cases (right).