Urban Population Prediction with Deep Learning for Enhanced 5G Simulations

Sammanfattning

Precise knowledge of population distribution at fine spatial resolutions is essential for numerous socio-economic, environmental, and infrastructure planning tasks, including city planning and telecommunication simulation. However, traditional census data are often outdated, infrequently collected, and aggregated at coarse administrative levels, making them unsuitable for applications requiring high-resolution population estimates. To address this challenge, this thesis - developed in collaboration with Tietoevry - presents the Multimodal Architecture for Population Synthesis (MAPS), a deep learning framework for estimating residential building-level populations. The approach integrates heterogeneous data sources in a completely new dataset - OpenStreetMap (OSM) features, building geometries, building heights, and Google Street View imagery - with supervision from two ground truth sources: building-level population from the Swedish state personal address register (SPAR) and aggregated zone-level population from Statistics Sweden. A multitask learning setup combining a building-level regression loss with a zone-level loss is proposed to bridge the granularity gap between these sources, optimizing accuracy over two different granularities. Additionally, two auxiliary models are developed to fill in missing data: one for building type classification, HouseCat, achieving 95% accuracy, and another for floor count estimation, FloorCast, reaching a mean absolute error of 1.0. The proposed method is the first of its kind predicting population on the building level with easily accesible geospatial data while using a combined zone- and building-level loss training strategy. The MAPS method achieves a 64% reduction in mean squared error on the building-level compared to Tietoevry’s baseline model and attains a mean absolute percentage error of just 21% at the zone level.