Modeling population distribution: A visual and quantitative analysis of gradient boosting and deep learning models for multi-output spatial disaggregation

Spatially aggregated data on socio-demographic groups often fail to capture the population's spatial heterogeneity in cities. This poses challenges for urban planning, particularly when addressing the needs of groups such as migrants or families with children. Moreover, the commonly provided aggregated units, such as census tracts, vary in size and across data sources. Existing literature on disaggregation typically handles individual subgroups separately, ignoring their interrelations in the downscaling process. This article explores the potentials of multi-output regression models for simultaneous spatial downscaling of multiple groups and conducts a detailed spatial error analysis using individualized neighborhoods. We experiment with self-training gradient-boosting trees and fully convolutional neural networks, assessing the quality of results against ground truth data at the target resolution. We show that the evaluation of the disaggregated results at this detailed resolution requires unconventional methods. The methodology proves convenient and achieves high-accuracy results using input datasets of building features.