Skip to content
Publications

Dissertation · King Abdullah University of Science and Technology Repository · April 1, 2026

SPDE-Based Geostatistical and Point Process Models for Environmental and Urban Network Applications

Abstract

Spatial data analysis increasingly requires methods that operate on complex domains beyond standard Euclidean settings, including networks constrained by road topologies. This thesis develops statistical models based on stochastic partial differential equations (SPDEs) for geostatistical and point process data, with applications to oceanographic monitoring, traffic safety, and urban air quality. We first develop a framework for bivariate spatial models that incorporates correlated measurement errors within a Matérn-SPDE formulation. Applied to global Argo oceanographic data, the methodology explicitly separates correlation in the nugget effect from the spatial dependence of the latent bivariate process, encompassing both Gaussian and non-Gaussian driving noise. A moving-window analysis demonstrates that accounting for this correlated noise structure reduces bias in cross-field dependence estimation and improves uncertainty quantification for joint temperature–salinity prediction. We then introduce log-Gaussian Cox processes on general compact metric graphs, modeling the log-intensity through Whittle–Matérn fields defined as solutions to fractional-order stochastic differential equations on the graph, with computationally efficient likelihood-based inference. The methodology is implemented in the MetricGraph R package and integrated with R-INLA. Applied to traffic accident data on a road network in Al-Ahsa, Saudi Arabia, the approach identifies road segments with unexplained elevated accident risk through excursion set analysis of the latent spatial field. Finally, we apply bivariate Gaussian random fields on metric graphs to urban air quality mapping using a linear model of coregionalization built from independent Whittle–Matérn components, fitted via integrated nested Laplace approximations. Applied to mobile monitoring data from Central London’s Ultra Low Emission Zone, we illustrate hotspot identification through excursion sets for NO2, then jointly model CO2 and NOx under a bivariate specification. To our knowledge, this is the first application of an LMC-based bivariate Whittle–Matérn model on a metric graph to urban air quality mapping. Collectively, these contributions advance spatial modeling through methodological improvements in bivariate covariance structures with correlated measurement errors, new model classes for point processes on non-Euclidean domains, and scalable computational workflows demonstrated on large-scale oceanographic and urban network datasets.

Spatial and Panel Data AnalysisPoint processes and geometric inequalitiesGaussian Processes and Bayesian Inference

More by Damilya Saduakhas