Understanding the land use intensity of residential buildings in Brazil: An ensemble machine learning approach

Abstract

The verticalization of cities impacts the quality of urban life. The empirical investigation of the determinants of the floor-area ratio (FAR) of lots using the traditional econometric approaches, however, has little explanatory power, and research about it using machine learning (ML) is almost nonexistent. This study applies two ensemble machine learning strategies, random forest (RF) and extreme gradient boosting (XGBoost), to investigate the determinants of the FAR of all formally registered multifamily residential lots in the city of Recife, Brazil. Taking into account a collection of key determinants influencing the floor area ratio (FAR), which encompass structural, accessibility, environmental, amenity, and policy variables, the findings reveal that the ensemble random forest approach significantly enhances the explanatory ability of these determinants when compared to conventional strategies like ordinary least squares (OLS) or locally weighted regression (LWR). Although generally in line with traditional urban economic arguments, the evidence also reveals important non-linearities in the effects of the variables on the FAR that are useful for urban planning and public housing policy.