SPDE-Based Spatio-Temporal Models for Satellite-Derived Chlorophyll-a Interpolation in the Red Sea
This dissertation develops and evaluates probabilistic spatio-temporal methods for reconstructing incomplete satellite chlorophyll-a fields while quantifying conditional uncertainty.
Overview
Satellite-derived chlorophyll-a is an important indicator of near-surface phytoplankton dynamics, but cloud cover, atmospheric conditions, and retrieval limitations create substantial spatially structured data gaps. These gaps are particularly important in the Red Sea, where chlorophyll-a varies strongly with season, latitude, and coastal setting.
The analysis uses MODIS-Aqua chlorophyll-a observations at monthly and daily resolutions. Monthly reconstruction for 2003--2022 is evaluated using tensor-product Gaussian and latent normal-inverse-Gaussian SPDE models, non-separable extensions, and DINEOF as a deterministic benchmark. Grouped validation assesses prediction under structured missing-data settings. A seasonally centered anomaly-log response provides a stable scale, while latent-NIG and non-separable extensions do not provide a consistent predictive advantage over the Gaussian tensor-product formulation. The latter is used to construct spatially complete monthly Red Sea fields with posterior summaries of latent-field uncertainty.
The daily analysis models shorter-timescale departures from the reconstructed monthly baseline. Candidate models include a separable tensor-product formulation, diffusion-based non-separable dynamics, and formulations informed by temporally averaged or time-varying geostrophic currents. Depth and distance to coast provide modest predictive improvement. Local-gap and regional space--time block validation show broadly similar performance among the leading formulations, while current-informed dynamics do not provide a consistent predictive advantage. The selected formulation is used to construct spatially complete daily chlorophyll-a fields for 2019--2021.
The reconstructed daily field is used to examine chlorophyll dynamics during 2020. Four regions with contrasting annual chlorophyll cycles are identified, and posterior realizations quantify uncertainty in regional membership and bloom characteristics. The analysis resolves regional bloom onset, peak timing, termination, and duration, tracks the basin-wide extent of bloom conditions, and examines correspondence between the chlorophyll-defined regions and selected physical characteristics.
Across the monthly and daily analyses, increased distributional or physical complexity does not consistently improve predictive reconstruction. The dissertation demonstrates the value of relatively simple, validated probabilistic reconstruction models and of accounting for reconstruction uncertainty in ecological analyses based on spatially and temporally complete chlorophyll-a fields.
Presenters
Brief Biography
Abeer Alzahrani is a Ph.D. student in Statistics at King Abdullah University of Science and Technology (KAUST), under the supervision of Professor David Bolin, and a lecturer at Imam Abdulrahman Bin Faisal University.