Improving particle filter performance with a generalized random field model of observation errors
share
This article shows that generalized random field (GRF) models of additive observation error can reduce the ensemble size required to avoid collapse in particle filtering of spatially-extended dynamics. This kind of random field model has increasing variance at small scales, and realizations appear `jagged.' Particle weights depend on how well a particular ensemble member agrees with the observations, and collapse occurs when a few ensemble members receive most of the weight. The GRF observation error model reduces the incidence of collapse by de-emphasizing small-scale differences between the ensemble members and the observations. This observation error model smooths the posterior mean, though it does not smooth the individual ensemble members. Two options for implementing the observation error model are described. Taking discretized elliptic differential operators as an observation error covariance matrix provides the desired jagged property of process realizations. This choice also introduces structure exploitable by scalable computation techniques, including multigrid solvers and multiresolution approximations to the corresponding integral operator. Alternatively the observations can be smoothed and then assimilated under the assumption of independent errors, which is equivalent to assuming large errors at small scales. The method is demonstrated on a linear stochastic partial differential equation, where it significantly reduces the occurrence of particle filter collapse while maintaining accuracy. It also improves continuous ranked probability scores by as much as 20%, indicating an improvement in the quality of the probability distribution associated with the particle weights. The method is compatible with other techniques for improving the performance of particle filters.
READ FULL TEXT
