HR: 11:20h
AN: B22A-04 [Abstracts]
TI: Comparing Estimators of Microbiological Attributes by Random Subsamples
AU: Li, M
EM: mengyl@umich.edu
AF: Environmental and Water Resources Engineering
Department of Civil and Environmental Engineering
University of Michigan, University of Michigan, 180 EWRE, Ann Arbor, MI 48105-2125
United States
AU: * Adriaens, P
EM: adriaens@umich.edu
AF: Environmental and Water Resources Engineering
Department of Civil and Environmental Engineering
University of Michigan, University of Michigan, 180 EWRE, Ann Arbor, MI 48105-2125
United States
AB:
Reliable characterization of the spatial distribution of sediment site attributes, such as contaminant concentrations or
microbial activity depends on how well sampled values represent all values throughout the entire study site. Whereas
geostatistical tools have been developed to interpolate the attribute values in space and claimed to incorporate multi-scale
estimation, these do not explicitly take into account the data associated with the various spatial scales. The M-scale model
that we developed showed visually the capability of incorporating multi-scale data derived from point measurements, with
cross-validation results quantitatively revealing a less over-smoothing estimate than the ordinary kriging. However, further
validation is needed for the practical ends of the model to real problems, among which the multi-point estimation based on
the data collected.
The M-scale model starts with evaluating relations of mean values over different scales by their covariances, subsequently
make further use of these covariances as basis for a precision-optimized estimator. Unlike conventional geostatistic tools
that are based on the point-to-point spatial structures, the M-scale model introduced a new framework for spatial analysis in
which regional values at different scales are anchored by the correlations of each other that forms an information stack at
each location. The model is developed using least-squares optimization for estimations of different scales, by which the
estimation variance can be evaluated for the estimation of all scales combined. Preliminary results from a comparison to
ordinary kriging of a spatial dioxin dataset from the Passaic River indicate that both estimation models visually agree in
regions with modest values, while M-scale model further preserves the features in the regions where higher and lower
measurements exists. On the basis of the same Passaic River dataset used for validating the M-scale model development, we
performed another validation approach called bootstrapping, which randomly take a designed number of data points out of the
data set, and examine the reproduction of their estimate by the rest of the data. The repeated random selection will be used
to compare the M-scale model to the ordinary kriging, visualized quantitatively by quantile-quantile (Q-Q) plots and scatter
plots.
Estimates and uncertainties evaluated by the M-scale model will be compared using the random subsets, to examine the
unbiasedness of the estimate as well as the appropriateness of the uncertainty evaluated. An additional comparison, using
the dataset collected in Anacostia River, Washington D.C., can further be used to inform further applicability under sparsely
sampled site. For a conclusive test, artificial datasets based on different scenario will then be generated, in order to
examine the general performance and restriction of the models under different data distribution and spatial structures.
DE: 1894 Instruments and techniques: modeling
SC: Biogeosciences [B]
MN: Fall Meeting 2005