Estimating parameters for a stochastic dynamic marine ecological system
Estimating parameters for a stochastic dynamic marine ecological system
复制标题
DOI：
10.1002/env.1083
复制
发表时间：
2011-06-01
期刊：
影响因子：
1.7
通讯作者：
Dowd, Michael
中科院分区：
文献类型：
--
作者：
Dowd, Michael
随机动力系统的参数估计是环境生态科学的一个核心问题。本文研究了一个简单的海洋生物地球化学非线性数值模型的参数估计。我们提出了一个基于非线性随机微分方程的模型，用于估计沿海海洋观测站收集的非高斯海洋测量数据的参数。序贯蒙特卡罗过程，或粒子滤波，提供了时间演化状态的估计，也是参数估计的基础。对比了两种估计系统静态参数的方法。第一种方法基于似然计算，第二种方法基于用静态参数增强系统状态。敏感性分析确定了两个生态参数（在微分方程模型中）和一个统计参数（控制动态噪声水平）作为估计的候选参数。由于基于样本的计算，发现计算的似然面是粗糙的；它们也表明了普遍存在的生态参数依赖和可识别性问题。本文采用了一种改进的状态增强过程，其中包含了一个平滑的自举步骤，用于参数估计。该方法提供的参数值的实现允许计算矩和密度估计，这些估计与似然的性质很好地匹配。在这方面也考虑纳入有关参数的先前资料。这种改进的状态增广方法为数值模型的参数估计提供了一条很有前途的途径。版权所有：John Wiley & Sons， Ltd。
Parameter estimation for stochastic dynamic systems is a core problem for the environmental and ecological sciences. This study considers parameter estimation for a simple nonlinear numerical model of marine biogeochemistry. We present a nonlinear stochastic differential equation based model for estimating parameters from non-Gaussian ocean measurements collected at a coastal ocean observatory. A sequential Monte Carlo procedure, or particle filter, provides for estimation of the time evolving state and also the basis for parameter estimation. Two approaches for estimating static parameters of the system are contrasted. The first is based on likelihood calculations, and the second on augmenting the system state with the static parameters. Sensitivity analysis identified two ecological parameters (in the differential equations model) and one statistical parameter (governing the level of dynamical noise) as candidates for estimation. Computed likelihood surfaces were found to be rough due to the sample based calculations; they also indicated the ubiquitous problem of ecological parameter dependence and identifiability. A modified state augmentation procedure, incorporating a smoothed bootstrap step, was used here for parameter estimation. Realizations for the parameter values provided by this method allowed for calculation of moments and density estimates that matched well the properties of the likelihood. Incorporation of prior information on the parameters was also considered within this context. It is concluded that such a modified state augmentation procedures provides a promising avenue in parameter estimation in numerical models. Copyright (C) 2011 John Wiley & Sons, Ltd.