Operational Dust Prediction - Mineral Dust: A Key Player in the Earth System - page 241

MODIS data are assimilated in the ECMWF, NRL and NASA forecasting
system, although each centre adopts individual strategies for filtering and bias
correcting the MODIS observations. As mentioned, the standard MODIS product
provides no information on dust over the sources, but it does, for example, over the
Atlantic Ocean where dust outflow from the Sahara is the main contributor to the
aerosol load. In regions that are not observed, therefore, the model plays a large
role in generating that information. Recently, enhanced efforts have been made at
several centres to include other observations, for example, the land AOD product
from the SEVIRI instrument on board of the MSG payload at the UK Met Office,
MODIS Deep Blue at NRL, OMI data at NASA and lidar backscatter at ECMWF,
NRL and JMA. JMA/MRI has been pioneering the possibility of assimilating lidar
data, with proven benefits on the dust prediction with their offline assimilation and
forecasting system (Sekiyama et al. 2010 , 2011 ). In what follows, the main concepts
of data assimilation are briefly discussed with focus on some specific aspects related
to dust/aerosol assimilation. Technical details are provided in Appendix A .

10.4.2 Main Concepts

Data assimilation is the process of finding the most likely estimation of the true sys-
tem state via the combination of observations and any available a priori information.
In the particular case of forecasts, this a priori information corresponds to the output
of model simulations. In other words, data assimilation is an objective way of filling
in information gaps and finding the optimal estimate of the true state by minimising
the variance of the a posteriori probability distribution function describing the state.
This approximation of the true state is the analysis . The method most commonly
used to obtain an analysis is the least-squares method, which is based on minimising
the variance of the a posteriori distribution according to a weighted-mean calcula-
tion. One of the simplest forms of analysis is, for example, the interpolation of the
differences between model and observations from the locations where the observa-
tions have been taken back to the model grid points in order to correct the prior
model state ( first guess ). Usually, the weights assigned are inversely proportional
to the errors in the observations and the errors in the model first guess. No matter
how complicated the schemes and systems involved are, the basic concepts of data
assimilation can always be illustrated by this weighted-mean calculation.
Most of the current dust prediction systems rely on assimilation developments
already in place for the meteorological models: for example, ECMWF uses the
incremental 4D-Var formulation with augmented control vector to include an
aerosol total mixing ratio variable (Benedetti et al. 2009 ). At the Met Office, 4D-
Var assimilation of dust observations follows a similar approach, using total dust
concentration as the analysis variable to be optimised (control variable). NRL and
NASA GMAO use 2D- and 3D-Var approaches. In the case of the regional NMME-
DREAM8 dust model (Pejanovic et al. 2010 ), an assimilation method based on
Newtonian relaxation is applied using background dust concentrations from the
DREAM dust model and the ECMWF dust analysis.