Study On Adjoint Data Assimilation Applied To A Simple Marine Ecosystem Dynamical Model

In present paper, variational adjoint assimilation technique is applied to a simple 3-dimension nutrient-phytoplankton-zooplankton-detritus (NPZD) marine ecosystem dynamical model for the Bohai Sea and the Yellow Sea. Twin experiments are carried out to study the parameter correlation issue and the influences of synthetic random errors on the recovery of control variables.According to the results of twin experiments, if all the parameters are chosen as control variables, the cost function reduces, but about half of the biological parameters in the model can not be retrieved. By sensitivity analysis, a new approach which can appraise the correlation among the parameters qualitatively by analyzing the linear combination of the Lagrange multiplies in the gradient equation of the cost function with respect to the parameters is brought up. The new method is more convenient than the traditional method, and the result of the two methods are accordant. Based on the result of sensitivity analysis, the parameters are separated into 5 groups. One parameter which has the largest cost function value in each group is selected as the control variable, and the number of control variables is reduced to 5 from 12. The following experiments indicate that the estimated result of the parameter in group 2 (Zooplankton maximum grazing Rate at 10℃, Zooplankton Ivlev constant, Zooplankton temperature coefficient for growth) is not good enough. The reason is that in the process of recovery, the gradient value of cost function with respect to the parameter in group 2 is much less than the others, so the parameter can not gain valid step size to retrieve. The recovery of 4 control variables is success, and the relative error between estimated value and base value is less than 1.0×10-2. Then the random synthetic errors are added in the observation, and the capability of adjoint assimilation to recover the control variables is tested. If the range of synthetic error is from -30% to +30%, the influence of the observation errors on the recovery of control variables is not large. But if the range of synthetic error is from 0 to +30%, the influence is...