Using the Kalman filter for parameter estimation in biogeochemical models
Article first published online: 14 FEB 2008
Copyright © 2008 John Wiley & Sons, Ltd.
Volume 19, Issue 8, pages 849–870, December 2008
How to Cite
Trudinger, C. M., Raupach, M. R., Rayner, P. J. and Enting, I. G. (2008), Using the Kalman filter for parameter estimation in biogeochemical models. Environmetrics, 19: 849–870. doi: 10.1002/env.910
- Issue published online: 20 NOV 2008
- Article first published online: 14 FEB 2008
- Manuscript Accepted: 31 DEC 2007
- Manuscript Received: 30 APR 2007
- Australian Research Council
- Ian Enting's fellowship at MASCOS
- parameter estimation;
- Kalman filter;
- biogeochemical models
We investigate application of nonlinear variants of the Kalman filter (KF) to sequential parameter estimation in biogeochemical models, with particular focus on two components of the statistical model: Q, the covariance of the stochastic forcing which we use to represent model error, and R, the observation error covariance matrix. We explored sensitivity of parameter estimates from the extended and ensemble Kalman filters (EKF and EnKF) to the choice of Q, R, initial parameters and ensemble size using pseudo-data from a simple yet highly nonlinear test model with many characteristics similar to real terrestrial biogeochemistry models. We found for our application that the use of inflated observation uncertainties led to the best and most stable parameter estimates. Although this reduced the rate of convergence to a solution, it also reduced the sensitivity of the solution to model error or ensemble size in the EnKF. Neither the use of model error for the parameters nor inflation of the state error covariance was particularly successful. Copyright © 2008 John Wiley & Sons, Ltd.