A geospatial modelling framework to simulate future scenarios of urban form : a case study in Phnom Penh, Cambodia

Abstract

Urbanisation is proliferating globally, with the majority of the world’s population now living in urban areas. This rapid urban growth places significant pressure on cities to expand, often leading to physical expansion rates that outpace population growth. This growth is particularly pronounced in developing regions, where it poses numerous challenges. Urban growth modelling offers decision-makers a valuable tool to under stand the complexities of urban dynamics, anticipate future transformations, and design sustainable and liveable cities. However, existing modelling approaches have largely focussed on large-scale, rasterised, and gridded projections of urban expansion that lack relevance and specificity to the regions being modelled. Furthermore, the models lack a systematic assessment procedure and there has been insufficient attention to generating detailed urban form layouts. To address these gaps, it is essential to use vector data at finer spatial scales to enhance contextual relevance and apply robust assessment methodologies that improve model credibility. Additionally, generating fu ture urban form layouts in developing regions is critical for informed urban planning. Therefore, this PhD thesis aims to develop a geospatial modelling framework to simu late, explore and characterise future urban form scenarios. Phnom Penh, Cambodia, was selected as a case study due to its rapid urban expan sion, data-scarce environment, and the relatively small proportion of built-up area. The framework employs two state-of-the-art techniques: agent-based modelling (ABM) and generative adversarial networks (GAN). The ABM replicates the urban development process basedonacademic literature and governmentreports, representingkeyactors, such as urban planners, property developers, and landowners, involved in this process as agents to simulate future urban growth. This ABM undergoes a thorough and sys tematic evaluation process, including stabilisation analysis, calibration and validation, and sensitivity analysis, to ensure its robustness. Dasymetric mapping is performed to convert the ABM’s development probability outputs into a raster of future develop ment demand density. The GAN then uses this future development density to perform image-to-image translation, converting a density raster into future road network layouts and subsequent building footprints based on the previously generated road network. The generated urban form outputs are then assessed against ground truth urban form using a range of pixel and vector-based accuracy metrics to identify the best-performing GAN experiment for future urban form generation. iii The ABM simulated five urban growth scenarios in Phnom Penh by 2040 which range from low to high growth under regulated and unregulated conditions. These scenarios are based on likely socioeconomic forecasts from grey literature and are variations of the baseline simulation. The evaluation protocol’s stabilisation analysis revealed that the ABM’s stochastic behaviour was effectively captured after 240 model runs. The ABM achieved good performance in calibration, achieving area-weighted fuzzy kappa simulation values of 0.241 in calibration and 0.509 in validation showing sufficient per formance for future simulation. The sensitivity analysis explored 8 model variables and determined their importance, interactions, and polarity with the fuzzy kappa simulation output, ensuring model credibility. The best performing GAN for road network gener ation in the study area achieved a mean absolute error of 40.96%. Whilst the best performing GANs for building footprint generation in the study area achieved mean absolute errors of 54.29% and 29.48%. The baseline or business as usual future sim ulation from the ABM projects 18.04 km2 of new urban area by 2040 in the study area. For urban form, this results in 227.68 km of new roads and 26,521 new building foot prints, this represents a 71.52% increase in road networks and 106.63% in building footprints from the 2023 urban form. These projections indicate large increases in new urban area and urban form over the following decades. The ABM scenarios achieve varied and distinct patterns of growth, with differing loca tions and probabilities of growth, demonstrating the ABM is responsive to variations in the input variables. The evaluation protocol highlighted issues with separating the cal ibration and validation period in equal temporal periods and weaknesses of the fuzzy kappa simulation as a result of this. The protocol could be developed further by employ ing explicit spatiotemporal sensitivity analyses to enhance model understanding in the study area. GANsaredemonstrated to be afeasible and practical technique in creating realistic and convincing future urban form that integrate seamlessly with existing lay outs. However, GAN training performance was sensitive to input image resolution and content and they do generally underestimate urban form features. Overall, this PhD thesis advances urban growth modelling approaches by presenting a novel geospatial modelling framework as a new standard in the field, which can be easily re-applied in other data-scarce areas experiencing rapid growth. It has strong implications for aca demic researchers, policy makers, and practitioners. It offers an innovative approach to generating future urban form with multidisciplinary applications in natural disaster preparedness, urban planning, and transportation modelling