Operational Dust Prediction - Mineral Dust: A Key Player in the Earth System - page 242

It is important to remember that most of these aerosol analysis systems solve
an initial condition problem: the analysis is used to obtain the initial conditions
in the aerosol fields, so that the subsequent forecast matches the observations.
In some cases, finding the optimal initial conditions for the atmospheric aerosol
concentrations is not sufficient as the actual aerosol amounts may be due to sources
that are not accounted for. Studies which include direct estimation of emissions
have been promising both for dust (Sekiyama et al. 2011 ) and other aerosol types
(Huneeus et al. 2012 ), and it is likely that future aerosol analysis systems will
include emission parameters in their control variables.
The other aspect which is peculiar to aerosol assimilation is that the problem
is severely under-constrained due to the fact that several aerosol species have to
be constrained with a total column-integrated observation for radiometric mea-
surements or a profile of backscattering for lidar measurements. This implies that
there is no one-to-one correspondence between the observations and the control
variable. There are various approaches to get around this problem using sensible
assumptions. For example, ECMWF formulates the control variable in terms of a
total aerosol mixing ratio and distributes the increments from this variable into the
single species mixing ratios in order to avoid defining the error statistics for all
species, which would be heavily reliant on the model. Other centres, for example,
MRI/JMA, use a method where the emission intensity is treated as a poorly known
model parameter defined at each model surface grid point and estimated in the
analysis. The control vector then consists of the dust emission parameters and
model variables such as aerosol concentrations and meteorological components. In
the end, it has to be accepted that no matter how complex and sophisticated the
aerosol assimilation system is, a lot of the information comes from the model rather
than the observations . This is a special limitation for dust because retrieval is often
impossible over the bright deserts that are dust sources.
Even with its limitations, dust forecasts from systems with aerosol analysis
have been shown to have reduced bias and improved correlations with respect to
independent observations, when compared to forecasts from the same systems with
no aerosol analysis, in particular for dust events (Benedetti et al 2009 ; Zhang et al.
2008 ). Moreover, aerosol reanalyses, in particular of dust and biomass burning
aerosols, are becoming increasingly valued to assess annual and seasonal anomalies
and to monitor the state of climate (Benedetti et al. 2013 ).

10.5 Evaluation of Atmospheric Dust Prediction Models

10.5.1 General Concepts

An important step in forecasting is the evaluation of the results that have been gen-
erated. This process consists of the comparison of the model results to observations
on different temporal and spatial scales. In this framework, there are three primary
objectives in forecast evaluation: