CMG Collaborative Research: Particle Filters and Ecological Models (PFEM): Application of chainless Monte-Carlo methods to mapping the ecology of the North Pacific Ocean

点击翻译按钮获取中文摘要

Intellectual Merit. Dynamic response of the world ocean ecosystem to interannual and decadal climate variability has been characterized by changes either in rates (e.g. rate of primary and export production) or in community structure. These changes in community structure cannot be easily modeled without introducing additional complexity in the ecosystem model. Similar limitations exist when ecosystem models have to reflect changes in the community from one region of the world ocean to another. The working hypothesis of this study is that advances in data assimilation techniques, coupled physical-biological modeling and high performance computing will help construct maps of parameter estimates for ecological models that will reflect these changes in species and community structure.Parameter estimation using remotely sensed ocean color and in situ observations is especially challenging since the state and/or measurement functions are highly non-linear, and the posterior distribution of the process states is not Gaussian. Several data assimilation techniques have been utilized in the last decade or so for coupled models of ocean physics and biogeochemistry but most of them rely on weakly non-linear and Gaussian systems (e.g. variational methods, Ensemble Kalman Filters) or are prohibitively expensive (e.g. particle filters). In particle methods one estimates the probability density of an ensemble of particles (instances of the model modified by observations). This is usually done by a Bayesian construction, with the resampling needed for accuracy done by Markov chain Monte Carlo. For large-scale systems this procedure can be prohibitively expensive, because the number of particles may need to be substantial and because the MCMC may have an excessive fraction of rejections. This study will employ newly-developed methods, where the probability density is sampled directly, without a Bayesian step, and where the resampling is done without recourse to Markov chains. A chainless Monte-Carlo particle filter is designed to overcome the shortcomings of variational methods and Markov chain Monte-Carlo particle methods for highly nonlinear, non-Gaussian systems with very high state dimensions. These chainless methods have performed well in smaller systems. The aim is to produce spatially and temporally varying parameter estimates for an ecological model of medium complexity coupled to a mixed layer model, based on assimilation of remotely-sensed surface chlorophyll-a observations in the north Pacific basin. Parameter values will be interpreted in terms of species distribution and nutrient cycling. This project will evaluate the ability of this coupled model to simulate observed seasonal to interannual variability, evaluate its sensitivity to forcing, and relate the results to considerations of ocean ecology.Broader Impact. The new data assimilation techniques will be broadly applicable to the highly nonlinear high dimensional problems that appear in interdisciplinary oceanography and applied mathematics. Examination of error estimates will lead to quantitative assessment of the information content of remotely sensed observations, and facilitate planning of future observing programs. The results of the coupled model simulations should provide useful information for assessment of the impact of climate change. The particle methods will be a significant advance in the state of the nonlinear filtering art, and be widely applicable in science and engineering. On the educational side, the project will provide training for future scientists in the field of interdisciplinary modeling and data assimilation.