Use this URL to cite or link to this record in EThOS:
Title: Particle filtering for applications in data assimilation
Author: Pons Llopis, Francesc De Borja
ISNI:       0000 0004 9350 8937
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Access from Institution:
In this thesis, several important topics in the area of particle filtering for applications in Data Assimilation are covered. The main focus of my research has been to explore ways to overcome the weight degeneracy problem that Importance Sampling (IS) faces in high dimensional problems or for very informative observations. It is also considered the case where it is necessary to determine what sensor took the observations before formulating the filtering problem. Here are the main findings of the research: • In the filtering problem where we have a continuous time signal following a stochastic partial differential equation (SPDE) and discrete time observations, I have found that the combination of tempering the likelihood to bridge the sequence of posterior distributions, Markov Chain Monte Carlo (MCMC) steps to reintroduce diversity of the particles and the use of IS to guide the particles to regions of interest allows us to use particle filters to tackle harder problems that are intractable otherwise. • When both the signal and the observation are continuous in time, the methods proposed in discrete time cannot be applied straightforwardly. I study an IS method based on modifying the drift of the signal, focus- ing on how to approximate the drift. I have found that this term could be approximated via the solution of a Partial Differential Equation (PDE) or via approximating the Smoothing density and concluded nu- merically that the second option was more promising computationally. I also show how this importance sampling improves the performance of particle filtering when compared to a basic implementation. • Finally, I show that the use of machine learning algorithms has the potential to considerably improve the accuracy of previous algorithms based on a deterministic classification tree for the problem of classifying historical sea temperature and salinity data according to which sensor was used to collect this data.
Supervisor: Kantas, Nikolas ; Crisan, Dan ; Brindley, Helen Sponsor: Engineering and Physical Sciences Research Council
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral