Bayes in the Environment
Organised by the Environmental Statistics Section of the Royal Statistical Society
Wednesday 20 November 2013
Royal Statistical Society, Errol Street, London
|
| Bayes in the EnvironmentOrganised by the Environmental Statistics Section of the Royal Statistical SocietyWednesday 20 November 2013Royal Statistical Society, Errol Street, London |
|
|---|
Please contact Adam Butler or Richard Wilkinson for further details.
There is a growing field of study which aims to quantify and synthesise all of the uncertainties involved in relating models to physical systems, within the framework of Bayesian statistics, and to use the resultant uncertainty specification to address problems of forecasting and decision making based on the application of these methods. Characteristic features of this methodology are Bayesian emulation for simulator outputs and careful structural discrepancy modelling for the difference between the simulator and the physical system. This talk will describe aspects of this methodology, largely from the Bayes linear viewpoint, in the context of environmental problems. Examples will be chosen from those arising within the RACER consortium, a NERC funded project within the Probability, Uncertainty and Risk in the Environment programme.
Remote sensing is the most cost-effective means of monitoring change on the Earth's surface and in its oceans, but rarely are such observations absolutely precise. Particularly difficult environments may have reflectance spectra - i.e. reflectance values as functions of wavelength - that are highly complex, variable, or hard to distinguish. However, the inference and predictions from remote sensing can be improved by incorporating additional information about the prior state of the environment being studied and the dynamics of its ecosystems. In this talk I will show that integration of disparate data sources through a Bayesian statistical framework can lead to substantially improved monitoring of the Earth. The approach is illustrated using the remote sensing of coral reefs, which are among the most challenging environments to assess spectrally. Specifically, we consider the assessment of hurricane impacts upon coral reefs in both a marine reserve and an area that is open to fishing. The coral cover at the study site was known at a time two years before the imaging. The knowledge of the dynamics of the reef ecosystem takes the form of a stochastic simulation model, developed by coral reef specialists and calibrated against independent field work. I will discuss (a) the results of a standard inversion of the remote sensing data, using an uninformative prior, (b) the prior distribution obtained by running the stochastic simulation and allowing for both model and sampling uncertainty, and (c) the results of a full analysis combining the prior from (b) with the remote sensing data. This is joint work with Dr John Hedley, Environmental Computer Science Ltd, and Prof Peter Mumby, Australian Research Council Laureate Fellow and Head of the Marine Spatial Ecology Lab, University of Queensland, and President of the Australian Coral Reef Society.
Source apportionment is a method in hydrology for attempting to define the runoff sources in river catchments. It involves estimation of the relative proportions of water from different sources, and is often recorded as a time series. Given regular measurements of a chemical tracer on the target water body and, in addition, corresponding measurements for samples of known sources, it is possible to perform source apportionment using compositional analysis taking a Bayesian random effects approach in a hierarchical framework, including covariates if appropriate. This talk considers the case where there are no separate data available for the source components, and develops a model for source distributions via nonlinear regression on the tracer/flow relationship and nonparametric density estimation. We allow these source component distributions to vary from year to year and apply the model to a data set from two streams in central Scotland, comprised of weekly or fortnightly readings of alkalinity over seventeen years. We show that a Bayesian approach enables us to model both the compositional analysis and the nonlinear regression simultaneously, in order to properly account for uncertainty. We conclude there is evidence of a change in source distribution over time; that corresponding to low flow conditions exhibits a gradual increase in alkalinity for both of two streams studied, whereas for high flow conditions alkalinity appeared to be rising for only one stream.
Data from multiple surveys can provide information on common parameters of interest, therefore joint models may yield improved estimates. However, simply fitting a single model to the combined data from multiple surveys is inadvisable without testing the implicit assumption that both the underlying process and its realization are common to all of the surveys. In this talk, I will: . propose a multivariate generalized linear geostatistical model that accommodates two sources of heterogeneity across surveys, variation in quality resulting in biased estimates of prevalence and variation in prevalence when surveys are conducted at different times; . describe Monte Carlo methods for parameter estimaiton and prediction; . apply the methodology to multiple surveys of malaria prevalence conducted in Chikhwawa District, Southern Malawi; . offer an opinion on where Bayesian methods have most to contribute to problems of this kind. Joint work with Emanuele Giorgi, Sanie Sesay and Anja Terlouw
Bayesian theory of model calibration provides a coherent framework for distinguishing and encoding multiple sources of uncertainty in probabilistic predictions of flooding. In this talk we demonstrate the use of a Bayesian approach to computer model calibration, where the calibration data are in the form of spatial observations of flood extent. The Bayesian procedure involves generating posterior distributions of the flood model calibration parameters and observation error, as well as a Gaussian model inadequacy function, which represents the discrepancy between Use of uninitialized value in concatenation (.) or string at E:\listplex\SYSTEM\SCRIPTS\filearea.cgi line 455,line 335. the best model predictions and reality. The approach is first illustrated with a simple didactic example and is then applied to a flood model of a reach of the river Thames in the UK. A predictive spatial distribution of flooding is generated for a flood of given severity.
You need not be a member of the Royal Statistical Society to attend, but registration will be required.
Registration fees, which cover lunch and refreshments, are as follows:
The registration form is available here.
The meeting will take place at the Royal Statistical Society HQ, 12 Errol Street, London.
Directions can be found here.