Examining geographical generalisation of machine learning models in urban analytics through street frontage classification and house price regression

The use of machine learning models (ML) in spatial statistics and urban analytics is increasing. However, research studying the generalisability of ML models from a geographical perspective had been sparse, specifically on whether a model trained in one context can be used in another. The aim of this research is to explore the extent to which standard models such as convolutional neural networks being applied on urban images can generalise across different geographies, through two tasks. First, on the classification of street frontages and second, on the prediction of real estate values. In particular, we find in both experiments that the models do not generalise well. More interestingly, there are also differences in terms of generalisability within the first case study which needs further exploration. To summarise, our results suggest that in urban analytics there is a need to systematically test out-of-geography results for this type of geographical image-based models