Quantifying uncertainty in precipitation estimates

Background/Question/Methods

Measuring the amount and chemistry of rainfall at a precipitation station is relatively straightforward.  However, estimating the input of rain water and solutes to ecosystems requires interpolation between the precipitation stations.  Various methods of interpolation are used in precipitation and atmospheric deposition studies, but the uncertainty in the interpolation is rarely reported or used in estimating uncertainty in deposition estimates. 

A preliminary analysis of spatial uncertainty in rainfall amounts using data from the Hubbard Brook Experimental Forest in New Hampshire showed model selection uncertainty to be small: interpolation using Thiessen polygons, spline, inverse distance weighting, kriging, and regression modeling differed by < 1% across the methods for annual precipitation volumes.  However, the error within the models (e.g. model or parameter error in the regression) has yet to be estimated.  Accounting for uncertainty in solute deposition is further complicated by the spatial and temporal mismatch between volume and chemistry samples, with fewer samples typically collected for solute chemistry than for rainfall volume. 

This work was conducted by a Synthesis Working Group supported by the LTER Network Office.

Results/Conclusions

We compared model selection error across four LTER sites:  Hubbard Brook, Sevilleta, H.J. Andrews, and Coweeta.  We found greater sensitivity of precipitation volume at sites with greater variation in deposition and at sites with fewer precipitation collectors.

We present a hierarchical regression model that can accommodate the spatial and temporal mismatches in precipitation volume and chemistry observations.  The regression estimates monthly and annual wet deposition of solutes for each site and probability distributions for inference on predictive covariates. The model also partitions uncertainty due to measurement error, missing data, and poor model fit and provides estimates of environmental stochasticity.  Methods for spatial interpolation of regressions can be compared for impacts on annual fluxes, using spatial analyses packages in ArcView, R (e.g, ModelMap package) and Bayesian kriging methods in OpenBUGS (i.e., geoBugs).

The long-term goal of QUEST (Quantifying Uncertainty in Ecosystem Studies) is to raise consciousness about the value of uncertainty analysis, provide guidance to researchers interested in uncertainty analysis, and support developers and users of uncertainty analyses (www.quantifyinguncertainty.org).