Probabilistic Approach Useful for Evaluating Cloud System Models
Jakob, C., Monash University
General Circulation and Single Column Models/Parameterizations
Jakob, C., R. Pincus, C. Hannay, and K.M. Xu (2004). Use of cloud radar observations for model evaluation: A probabilistic approach, J. Geophys. Res., 109, D03202, doi:10.1029/2003JD003473.
In evaluating climate models, time and space represent key challenges when extrapolating observations into simulations. Researchers supported by DOE's Atmospheric Radiation Measurement (ARM) Program have explored an alternative method based on "point series data" to arrive at model cloud predictions. Point series data are obtained over time through measurements of cloud properties by various instruments at a fixed location. As described in the Journal of Geophysical Research in February 2004, the authors looked at the relationship between single-point observations and model-averaged output in a fundamentally new way. By treating the observations as a statistical sample, their probabilistic evaluation technique groups observations by classes in the forecast probability, as opposed to averaging the point observations over time.
Using an "idealized" scenario based on large-scale forcing data obtained at ARM's Southern Great Plains site in 1997, the authors compared the performance of time-averaging and probabilistic evaluation methods. They then used probabilistic techniques to test forecasts by various classes of models against a set of remote sensing observations. In their idealized example, time averaging tended to obscure forecast skill, regardless of the size of the model domain. Conversely, probabilistic forecasting using point series data required no assumptions about statistical placement.
Because point observations can be compared to the distribution of forecast quantities within the entire domain, any variable may be evaluated. The authors concluded that probabilistic evaluation may prove a valuable tool for testing model outputs for fields that are not uniform, such as cloud systems.