Developing a pathway to improve large-scale gridded population modelling based on the World Settlement Footprint suite

The field of large-scale human population modelling has emerged as a response to the increasing demands for actionable, consistent and comparable population data, needed to support a large number of applications related to sustainable development. Nowadays, Earth Observation-derived, top-down large-scale gridded population datasets describing the extent and spatial distribution of the human population as continuous surfaces, are openly accepted by many governments and institutions around the world, who use them as an alternative source of information to complement/supplement conventional census/estimate-based population data. Given the wide range of applications where gridded population products are being employed, research performed to improve the accuracy and spatial resolution of large-scale population models has become of utmost importance. For the last decade, the scientific community has constantly leveraged the increasing availability of Earth Observation data, the improvements made on Remote Sensing techniques, and––more recently––the developments made on Machine Learning algorithms, to produce large-scale population datasets with higher usability and reliability than ever before. Today, some of the most accurate and spatially explicit products available at a global scale are produced, for example, by harnessing the inclusion of remotely-sensed derived proxy layers with improved thematic and spatial resolution, especially those describing the characteristics of the built-up environment such as built-area layers and building footprint datasets, respectively. However, a systematic literature review undertaken within this PhD research has revealed that existing top-down large-scale population models still suffer from a number of limitations that affect the final accuracy and usability of their corresponding derived population datasets. In particular, it has been identified that existing techniques to produce large-scale gridded population maps are still affected by the quality and recency (age of the data) of the underlying census/estimate-based population data on the one hand; and by the still low spatial resolution of the geospatial proxies used for disaggregation, the persistent inaccuracy in identifying populated areas from remotely-sensed data, and the lack of information on the functional use and 3D characteristics of the built-up environment, on the other hand. Overall, it has been concluded that if some of these limitations still exist in the field, is because data or methods that can help overcome these issues at very local–or fine scales are yet not available or transferable at large-scale (e.g. continental or global scales). Therefore, to address these challenges, this PhD thesis explores the capabilities and effectiveness of the new World Settlement Footprint (WSF) suite in the production of large-scale gridded population models with improved accuracy and spatial resolution In detail, it presents a methodological framework that investigates the effectives of each WSF-layer, namely the WSF2015, the WSF2019 and the WSF3D, respectively, in terms of 1) their ability to improve population estimates compared to binary-dasymetric models, 2) their ability to produce consistent and comparable accuracies across large territorial extents, 3) their ability to produce accurate population estimates acting as single proxies for population modelling, 4) their ability to reduce the technical complexities of multi-layer weighed-dasymetric models, 5) their ability to discriminate large industrial areas using a simple and spatially transferable machine learning approach, and finally 6) their ability to improve population estimates through the integration of volume and settlement use information. Within this methodological framework, a comprehensive set of spatial and statistical analyses were designed to evaluate the uncertainties of a large number of population distribution maps at local, national and continental scale. For a robust assessment, population models were produced at varying currencies, qualities and spatial scales of the input population data, with the purpose of analysing how the differences in the level of spatial granularity of the available administrative boundaries and the variability in the morphology of built-up landscapes influence the accuracy of each WSF layer. The main findings of this PhD thesis demonstrate that the WSF-layers are capable of tackling some of the main limitations identified in the field of large-scale population modelling. First, the independent weighting framework provided by the non-binary WSF layers allowed outperforming the mapping accuracies of widely employed binary-dasymetric models on the one hand, and reduce the technical complexities of (multi-layer) weighted dasymetric models on the others. Second, as single proxy layers used for dasymetric disaggregation, each WSF layer was also capable to deliver consistent and systematic accuracies across space; where the robustness of each layer was consistent under varying qualities and spatial resolutions of the input population data. Finally, spatial metrics derived solely from the WSF3D dataset showed to be extremely effective in classifying the built-up environment into industrial and non-industrial uses; which ultimately allowed incorporating, –for the first time ever––, settlement use and settlement volume information into large-scale models of population disaggregation. In view of these promising results, the main contributions of this PhD research can be summarised as follows: 1. Quantitative and qualitative demonstration of how employing the WSF-suite for population modelling overcomes some of the most prominent limitations in the field. 2. First in-depth quality assessment aimed at evaluating the effectiveness and suitability of each WSF-dataset as proxy layer for large-scale population modelling. 3. Design and implementation of the Settlement Size Complexity (SSC) index as a robust metric to evaluate the uncertainty of population models based on built-up area layers. 4. Improving understanding on the “fitness for use” of large-scale gridded population datasets. 5. The development of a highly accurate, semi-automatic and globally transferable method for the identification of industrial and non-industrial areas using only the WSF3D dataset in combination with a Machine Learning approach. 6. First time delivery of large-scale population datasets produced on the basis of the WSF-layers, to serve as actionable data for a large number of ongoing-projects