Implementing a deep-learning model using Google street view to combine social and physical indicators of gentrification.

While physical changes have been empirically recognised as a fundamental component of neighbourhood change, data and modelling constraints have limited the quantification of these indicators. Recently however, the proliferation of Big Data and advancements in deep learning (DL) techniques have enabled mass image processing. It is in this context that we build a Siamese Convolutional Neural Network using Google Street View (GSV) images to detect upgrades to properties as evidence of gentrification. The model achieves 84.8% test accuracy and 74.6% AUC. Building upgrades detected using the model are mapped using Kernel Density Estimation (KDE) and validated against Development Approvals. The DL GSV model is subsequently combined with the socioeconomic-based predictions of seven gentrifying suburbs from a prior Sydney-based gentrification study. The maps of predicted social change are validated against the spatial patterns of building upgrades detected by the DL GSV model. Of the five suburbs with sufficient data, the socioeconomic trends were affirmed with physical indicators of gentrification in four and questioned in one. The paper provides the first machine learning approach to combine social and physical indicators of gentrification. This automated, multi-dimensional approach enables the distinction of gentrification from other forms of neighbourhood change, helping to develop a more comprehensive understanding of gentrification occurring within a city. • Deep learning (DL) model created to detect upgrades to properties using Google Street View (GSV) data. • DL GSV model used to validate socioeconomic trends of previous Sydney-based gentrification study. • Semantic segmentation algorithm implemented to automate data cleaning process. • Study provides first machine learning approach combining socioeconomic and physical indicators of gentrification. [ABSTRACT FROM AUTHOR]