Will a perfect global model agree with perfect observations?

Abstract:

Global aerosol models and observations differ strongly in their spatio-temporal sampling. Model results are typical of large gridboxes (200 by 200 km), while observations are made over much smaller areas (e.g. 10 by 10 km for MODIS, even smaller for ground sites). Model results are always available in contrast to observations that are intermittent due to orbital constraints, retrieval limitations and instrument failure/maintenance. These twin issues of temporal sampling and spatial aggregation are relevant for any observation, be it remotely sensed, or in-situ. We ask this question: will a perfect model agree with perfect observations? The short answer is: unlikely.

Using two different modelling frame-works (year-long global model runs collocated with actual observations and month-long high resolution regional models runs) we show that significant errors can be introduced in a model to observation comparison due to different spatio-temporal sampling. These sampling errors are typically larger than observational errors and are of comparable size as true model errors. While the temporal sampling issue can be dealt with by properly resampling model data to observation times, the spatial aggregation issue introduces noise into the comparison. We propose and evaluate several strategies for mitigating this noise.

The most succesfull strategy is further temporal averaging of the data. However, this seems to have a less benefical effect on surface in-situ observations than on remotely sensed column-integrated measurements. For instance, monthly averaged black carbon mass concentrations measured at ground sites still allow significant (~ 30%) noise into the comparison. Furthermore, flight campaign data, by its nature, are not open to long-term (monthly, yearly) averaging and allow sampling errors of 50% or more in black carbon mass concentrations. Other observables (AOT, extinction profiles, number densities, PM2.5, CCN) will also be discussed.