Bostick, W. McNair (University of Florida, Dept. of Ag. and Bio. Engineering, 1 Frazier Rogers, Gainesville, FL, 32611-0570; Phone: 352-392-1864 x 292; Fax: 352-392-4092; Email: mcnair2@ufl.edu)

 

Stochastic Simulation and Data Assimilation for Estimation of Soil Carbon Dynamics

 

W. M. Bostick*, J. Koo, J. W. Jones, G. Hoogenboom, V. Bado, W. D. Graham

 

The large land areas needed to sequester tradable amounts of carbon (C) will lead to uncertainty in estimates of soil C dynamics in sequestration projects. The objective of this paper is to present a method that uses the Extended Kalman Filter (ExKF) algorithm for assimilating simulations and measurements to reduce uncertainty in estimates. The method uses a stochastic soil C model to propagate model states, their variance and the covariance between states at different locations. When measurements are made, the ExKF updates model states at locations that are measured and unmeasured. The degree to which simulated states are updated depends the difference between simulations and measurements and on the magnitudes of the measurement variances, the simulated variances and simulated covariances of model states. In general, low measurement variance, relative to the simulation variance, can yield large state updates when measurements are assimilated. On the other hand, if simulation variance is low, relative to the measurement variance, measurements may have little effect on state estimates. Larger covariances between states at measured and unmeasured sites may yield larger updates at the unmeasured site. We demonstrate the use of this method with a long-term rotation experiment from Burkina Faso. Soil C samples from the experiment were used to initialize the initial C states. A semi-variogram model was also developed from these data and used with multi-Gaussian stochastic simulation to initialize the covariance matrix. The ExKF performance was analyzed for different measurement strategies. For all measurement scenarios considered, aggregate C estimates were more accurate than measurement or simulation results alone. In addition, the variance of the aggregate filtered estimates decreased as data were assimilated. In contrast, the variance of simulation estimates alone increased over the period of estimation.