A Model-Independent Backtracking Particle Filter Method in the PCRaster Python Framework

Abstract

Abstract Particle filters are an effective way to tackle data assimilation problems in the field of geosciences, especially when the model is either highly nonlinear or the measurement error can not be expressed as a Gaussian curve. The standard particle filter has been implemented into the PCRaster Python Framework which enables researchers to easily employ the method to data assimilation problems. However, the current particle filter method has no way to mitigate filter degeneracy. Spiller et al. (Physica D 237 (2008), 1498-1506) devised a way to repair a degenerated particle filter by reverting the filter to the last time it worked correctly and recalculate the posterior distribution with an increased sample size. In this thesis, a backtracking particle filter method has been constructed within the PCRaster Python Framework. The backtracking particle filter has been designed to be able to function with minimal intervention on the side of the model. The new method has been tested using two nonlinear stochastic models, showing an increase in efficiency. The method has some limitations when the stochastic forcing in the model is too high, but overall backtracking increases filter results. The backtracking algorithm is especially helpful when one of the update steps in the filter has unusually low observational error.