Predicting intra‐urban well‐being from space with non‐linear machine learning

There is a growing need to analyze welfare at an intra-urban level because cities often evince stark divisions. It is therefore important to identify inequalities within them. However, data are hardly available – or very expensive. The purpose of this article is to test whether non-linear machine learning algorithms provide more accurate predictions of intra-city well-being than the linear models. In addition, we aim to check if freely available and easily accessible data from Open Street Map offer an alternative to high-resolution daytime satellite images from Google Maps in accurately predicting well-being on a local level. Inspired by the Local Human Development Index (UNDP, 2012) we construct a well-being index based on three dimensions: health, education, and welfare. Potential predictors of well-being include indicators related to the urbanization rate, access to natural amenities, the transportation system, and access to public transport. Four non-linear machine learning algorithms (support vector regression with polynomial and radial kernel, random forest and xgboost) are compared with the linear LASSO approach for the 18 districts of Warsaw, Poland. In addition, we apply innovative tools of explainable artificial intelligence (XAI) to identify the most important predictors of well-being (measuring model-agnostic feature importance) and to disclose the shape of relationships between well-being and its most important predictors. We conclude that the application of non-linear machine learning algorithms to modelling well-being not only allows us to reach higher predictive accuracy, but also to better identify and explain the impact of its predictors.