Back to search

School of Geography and Environmental Sciences

Using georeferenced building-level administrative data to enhance settlement classification methods for population mapping

Abstract

dc:description.abstract

Accurate and granular population data are fundamental to urban planning, public health, and disaster response, especially in rapidly urbanising low- and middle-income countries (LMICs) where censuses are infrequent. Consequently, these modelled estimates provide a necessary bridge between decennial censuses, and their accuracy relies significantly on the quality of the underlying settlement information used. Yet, despite the proliferation of high-resolution geospatial products, a fundamental gap in characterization persists: most global frameworks continue to treat the urban fabric as a monolithic 'built-up' entity, ignoring the socio-spatial variations critical for precise modelling. This masks the complex functional heterogeneity of megacities and limits the accuracy of small-area demographic models. This doctoral thesis addresses these challenges by establishing a systematic framework to evaluate satellite-derived settlement data, advancing functional classification methods, and testing their integration into population modelling. Using Lagos, Nigeria, as the central case study, the research tests the hypothesis that semantically enriched settlement data distinguishing between residential, non-residential, and mixed-use areas, substantially improves the spatial realism and predictive accuracy of top-down population disaggregation.<br/>Chapter 3 establishes the empirical foundation by systematically assessing the accuracy of leading global building footprint datasets (Microsoft, Google, Ecopia/Maxar, OpenStreetMap) and 3D settlement layers (World Settlement Footprint 3D, Global Human Settlement Layer (GHSL), Google 2.5D). Using over 180,000 geo-referenced buildings from the Lagos Building Sample Points (LBSP) survey received from Lagos land use agency as ground truth. The study benchmarks their spatial completeness, footprint area, and building height reliability of the 3D settlement datasets. Results show that while Microsoft and Google datasets exhibit strong correlations with ground truth, notable intra-urban inconsistencies persist. Also, the vertical accuracy of 3D settlements varies with World Settlement Footprint 3D (WSF3D) showing stronger correlation with ground truth. This chapter highlights both the opportunities and limitations of openly available settlement products in LMIC megacities, underscoring the necessity of validation against local data before use in population models.<br/>Building on this foundation, Chapter 4 addresses the technical gap regarding the lack of functional granularity in current settlement layers. It is critical to distinguish residential, non-residential, and mixed-use areas at the pixel level because binary "built-up" masks fail to capture the socio-economic diversity of urban land use, often leading to significant population misallocation in dense, polycentric hubs. Both Random Forest and Bayesian Hierarchical models were implemented to predict proportional shares of functional uses at 1 km² grid resolution. Results demonstrate strong predictive performance, with Random Forest achieving higher correlations for residential and non-residential classes, while the Bayesian model provided superior uncertainty quantification and better representation of mixed-use areas. The findings reveal the spatial heterogeneity of Lagos, highlighting the prevalence of mixed-use development in dense neighbourhoods that traditional categorical classifications overlook. <br/>Chapter 5 empirically tests the operational value of these classifications by integrating them into a dasymetric disaggregation framework. Four models were compared against a baseline benchmark model (MOD1) that utilised standard WorldPop covariates (night-time lights, road proximity, and points of interest) without functional settlement information. Results demonstrate that incorporating functional classifications consistently improved predictive accuracy relative to this baseline. Specifically, the functional layers ranked as top predictors, markedly reducing common errors of population over-allocation in non-residential city centres and under-estimation in residential districts.<br/>Chapter 5 tests the central proposition of the thesis: whether incorporating functionally classified settlement data improves top-down population models. Using a dasymetric disaggregation framework adapted from WorldPop, four models were compared, integrating standard covariates (night-time lights, proximity to roads and points of interest) with functional settlement classification datasets from Random Forest, Bayesian models, and Global Human Settlement Layer (GHSL) residential/non-residential layers. Results reveal that functional classifications consistently improved predictive accuracy. The settlement classification datasets ranked among the most influential predictors of population density. Errors of over-allocation in non-residential city centres and under-estimation in residential districts were markedly reduced, demonstrating the operational value of semantic enrichment in population redistribution.<br/>Collectively, this thesis advances the field of spatial demography by (i) validating global settlement datasets in an LMIC megacity, (ii) introducing scalable methods for functional settlement classification using geospatial modelling, and (iii) showing the potential to improve population predictions at fine spatial scales. The findings not only strengthen methodological pipelines for population mapping but also provide practical, policy-relevant evidence for enhancing demographic visibility in rapidly urbanising, data-scarce contexts such as Lagos.<br/>

Degree

thesis:*
Name dc:type.qualificationname
Ph.D.
Level dc:type.qualificationlevel
doctoral
Grantor dc:publisher.institution
School of Geography and Environmental Sciences
Year dc:date.issued
2026

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Adewole, Wole Ademola
Advisors dc:contributor.advisor
  • Tatem, Andrew
  • Cockings, Samantha
  • Lloyd, Christopher T
ORCID person:identifier.orcid

Rights

Language dc:language
English, eng

Identifiers

dc:identifier.*

Chain of custody

source
Harvested from
University of Southampton
Base URL
eprints.soton.ac.uk/cgi/oai2
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
related terms
citation

Adewole, Wole Ademola. Using georeferenced building-level administrative data to enhance settlement classification methods for population mapping. doctoral thesis, School of Geography and Environmental Sciences, 2026. https://doi.org/10.5258/SOTON/PG/T197