Abstract Author: Huug van den Dool
Abstract Title: The Multi-Model Approach for Soil Moisture Analyses in the Absence of Verification Data
Abstract: Increasingly we are generating, or are being offered, more and more soil moisture analyses. As an operational climate prediction&monitoring center we are obviously interested in this. We now have an ensemble of soil moisture products (real time plus historical data sets) by such models as VIC, Noah, Mosaic etc, as well as the older Leaky Bucket model and the off-line global and regional Re-analyses (R1, R2, NARR, ERA40 and soon CFSRR). But how to weigh the fidelity of one these analyses relative to the other??? Which one is best? And how does one make the best consolidation?
In Multi-Model Forecasts (or Analysis rather in this case) one generally solves a linear algebra problem (Aw=b) in which the weights (vector w) assigned to each model depend on a) the covariance of model generated data among the models (collected in matrix A) and b) the covariance of model data with observations (or verification data) in vector b. For soil moisture item b) is a problem because we have no verification data (unless one restricts the analysis to Illinois only). So how to solve this problem in the absence of verification data?
The central and very important point is to wonder what we really use soil moisture analyses for. In terms of help for the CPC forecasts, we believe that we can be certain only about the local effects of soil moisture on temperature, i.e. that dry (wet) soil leads to higher (lower) temperature. The soil moisture data set that provides the best prediction of temperature (as per local regression) is thus the best.
The above can be generalized to all analyses combined. It is very possible to set up the above consolidation linear algebra problem, where the weights assigned to the various analyses depend on the ability to forecast co-located temperature. The co-linearity among models is in fact unchanged relative to the situation where verification data is available. It is only the rhs (the vector b) that has changed, and as a result the solution, the vector w will change. The approach has elements of ‘hidden modelling’.
We will show how this works over the US for 1979-2007 for about 6 fairly credible and well-known soil moisture data sets.