|
|
|
|
|
|
|
![](images/dian9.gif) |
Time-series analysis of ocean biogeochemical and ecological data
|
|
|
The Adaptive Metropolis algorithm as a tool for model selection given irregular and imperfect time-series data (Invited) Tuesday 8th @ 1440-1515, Conference Room 5 Sherwood. Lan Smith* , Marine Ecosystem Dynamics Research Group,Japan Agency for Marine-Earth Science and Technology (JAMSTEC) Yokohama JAMSTEC, Presenter Email: lanimal@jamstec.go.jp |
Oceanographic and other ecological observations often yield irregular time-series of data, including substantial uncertainties. Models, which express quantitatively our hypotheses about the underlying processes, are useful and in many cases essential tools for interpreting such observations. However, given the uncertainties in both the data and models, rigorous model-data comparison is a daunting task. Here I will present details from a previously published example (Smith et al. J. Plankton Res. 38, doi: 10.1093/plankt/fbv038, 2016) using the Adaptive Metropolis (AM) algorithm (Haario et al. Bernoulli 7, p. 223-242, 2001) to combine oceanic time-series observations with plankton ecosystem modelling. Time-series observations consisted of nutrients, chlorophyll, and primary production, from the K2S1 project (https://ebcrpa.jamstec.go.jp/k2s1/en/). The Bayesian statistical foundation of the AM algorithm allows for systematically combining: 1) prior knowledge of parameter values, 2) irregular time-series data of different types, each with its own uncertainties, and 3) different model formulations. The algorithm generates an ensemble of model simulations tuned to match the range of observations. It thus provides: 1) quantitative metrics for determining which model formulation is best supported by the available data (i.e., model selection), 2) posterior distributions of model parameter values and corresponding model outputs, and 3) uncertainty estimates (from Gibbs sampling) of the model-data mismatch for each observation type, respectively. Such Bayesian approaches provide a systematic means of quantifying uncertainties and evaluating competing hypotheses given irregular time-series data. This is important for interpreting large, modern data sets consisting of multiple data types, each having different units and associated uncertainties. It also works well with complex models having non-linear dynamics. However, even these computationally efficient statistical methods require thousands to many millions of model simulations, depending on data quality and the number of parameters being optimized. Until recently this has limited their use to 0-D models, but fast modern computers have recently made it possible to apply them with 1-D ocean models as well (Chen and Smith. Geosci. Model Dev., doi: 10.5194/gmd-11-467-2018, 2018), expanding their potential for studying the combined effects of physical, ecological, and biogeochemical processes. |
|
|
|
|
|
|
|
|