Evaluating spatial statistical and machine learning models in urban dynamic population mapping

Abstract

Understanding population dynamics at fine spatiotemporal granularities are valuable to human-centered studies. With the increasing availability of high-frequency human digital footprint data, the past decades have witnessed numerous efforts in mapping populations at fine spatiotemporal scales. However, such research still lacks a unified standard in modeling strategy and auxiliary data selection, especially a systematic comparison between newly developed machine learning techniques and traditional spatial statistical methods under different covariates provisions. Here, we compared two spatial statistical models, the Bayesian space-time model and geographically and temporally weighted regression, with two machine learning techniques, random forest and eXtreme gradient boosting, in a case study of hourly population mapping at 100 m resolution in Beijing. We evaluated the model performance with varied covariates combinations and found that the Bayesian space-time model achieved the best in conjunction with urban function data. Leveraging the optimal model constructed, we mapped dynamic population distribution and concluded human activity patterns on diverse city amenities. This paper emphasizes the importance of spatiotemporal dependency information in fine temporal scale population mapping and the urban function covariates in urban population mapping.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 5 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.