As climate change increases the magnitude and occurrence of coastal hazards, coastal ocean modelling has become critical to minimising damage to coastal communities. It is therefore important to improve coastal ocean model performance by incorporating external data sources. This work investigates methods for obtaining hydrodynamic information from image sequences and subsequent use for coastal ocean model calibration.



Deep learning methods have outperformed computer vision benchmarks at motion estimation. We investigate the ability of various deep learning methods, trained on synthetic images, at inferring velocity fields from images of real hydrodynamic flows. This is achieved through estimation on hydrodynamics laboratory images, and further on drone imagery representing real-world conditions. The results indicate that the deep learning methods recover complex flow features and generalise to flow types outside of their training datasets. Additionally, we demonstrate improved alignment with classical methods using limited unsupervised fine-tuning.



Given the successful inference of velocity fields from laboratory images, we demonstrate their use for numerical model calibration. This is achieved through the development of shallow water models of flow types from the laboratory experiments, within the adjoint-capable \textit{Thetis} coastal ocean modelling framework. We utilise the adjoint method to recover optimal parameters through dual twin experiments and extend to model calibration using observations obtained from image analysis, demonstrating how both approaches can be combined to improve shallow water models.



Following laboratory-scale experiments, we investigate velocity field estimation from large-scale satellite imagery, presenting an underutilised opportunity. We develop a methodology for identifying and processing satellite image pairs from the PlanetScope archive and apply this to the novel estimation of submesoscale eddies across a range of contexts.



The investigations in this thesis form a framework facilitating identification and analysis of large-scale image pairs to yield velocity fields, which can then be used to improve coastal ocean model performance through calibration.