assimilation and forecast impacts using the expected

ASSIMILATION AND FORECAST IMPACTS USING THE EXPECTED
ERROR IN THE QUALITY CONTROL OF MODIS POLAR WINDS
David Santek, Brett Hoover, James Jung
Cooperative Institute for Meteorological Satellite Studies, University of Wisconsin, Madison, WI USA
Abstract
Atmospheric Motion Vectors (AMV) are routinely generated from geostationary and polar orbiting
satellites and they are incorporated into most global numerical weather prediction models throughout
the world. However, the quality control and thinning strategies applied to the AMVs vary in the
modelling community and are usually based on quality flags since, until recently, no error estimate
was provided with each observation.
The Expected Error (EE) is a new quantity that is computed by applying coefficients from a linear
regression of QI values, AMV parameters, and shear against co-located rawinsondes. Unlike the
normalized QI values, the EE is in units of speed.
We expect that the use of the EE will provide a more quantitative screening of the satellite-derived
winds product, resulting in better assimilation statistics and improved global forecasts. Initial results of
several experiments using polar AMVs in the NCEP Global Data Assimilation System/Global Forecast
System (GDAS/GFS) are presented.
EXPECTED ERROR
Satellite-derived AMVs are assigned individual quality flags. Most of these quality indicators are
normalized scores (e.g., the Quality Indicator, QI), which are not in the units of the wind vector. The
Expected Error (EE), developed at the Bureau of Meteorology in Australia, extends the usability of the
QI (Holmlund 1998) by regressing the QI and other AMV parameters against co-located rawinsondes
(Le Marshall and Rea 2004). This results in coefficients that are applied to individual AMVs to compute
an Expected Error in units of speed, which are more amenable to assimilation systems.
The Expected Error (EE) coefficients result from a least squares regression of the five Quality
Indicator (QI) tests along with other vector and model information against actual AMV and rawinsonde
vector differences (LeMarshall and Rea 2004). The nine predictors are shown in Table 1 for the Terra
MODIS cloud drift winds. The first five predictors are the normalized 0 to 1 scores. The sixth and
seventh predictors, the wind speed and pressure level, are derived from the wind vector itself, while
the two gradient predictors are generated from the model guess. These predictors are generated
using several months of data regressed against the AMV-rawinsonde vector difference in order to
create regression coefficients. The coefficients can then be used to estimate the error for subsequent
vectors.
Component
QI Speed
QI Direction
QI Vector Difference
QI Local Consistency
QI Forecast
Wind Speed
Pressure Level
Wind Shear
Temperature Shear
Constant
Coefficient
-0.4
-0.1
-0.6
-0.6
-2.8
+0.1
-0.003
+0.03
-0.01
+8.4
Table 1: The Expected Error coefficients for the Terra MODIS cloud drift winds.
AMV FILTER CRITERIA
The goal of the project is to use more quantitative-based criteria for the quality control of the winds.
The criteria currently used by NCEP for filtering the MODIS polar winds are empirically based. A wind
is discarded if:
1. It is within 50 hPa of the tropopause
2. It is within 200 hPa of the surface (if the surface is land or ice)
-1
3. (O-B)U > qcU OR (O-B)V > qcV, where qcU = qcV = 7 ms ; O-B is the Observation minus
Background.
The following experiments detail using EE thresholds to quality control the MODIS winds. In all cases,
a 2010 version of the GDAS/GFS is used on the NCEP development system.
GFS EXPERIMENT #1
The first experiment uses a criteria suggested by LeMarshall and Rea (2004): Discard the wind vector
-1
if the EE > 5 ms AND the EE > 0.1 * Observation Speed. The second part of the decision was
included to retain higher speed winds, since fast winds generally have a higher EE value.
This experiment ran for the month of September 2010 and the results of the southern hemisphere 500
hPa Anomaly Correlation Coefficient (ACC) for the day 5 forecast are shown in Fig. 1. For the most
part, the impact is neutral when using the EE to quality control the winds vs. the current NCEP
method. Although on 26 and 27 September 2010, there is noticeable positive impact using the EE.
Figure 1: Day 5 500 hPa geopotential height ACC 08 September to 01 October 2010. Southern Hemisphere: control
(black); experiment (red).
GFS EXPERIMENT #2
nd
Based on the results of the first experiment, the 2 was designed to retain higher winds by using
-1
these criteria: Discard the wind vector if the EE > 9 ms AND ObsSpd < EE.
The results of this experiment, from 01 September to 03 October 2010, are shown in Figure 2. The
solid curves show a neutral impact for this time period. However, the dashed lines show an
improvement using the EE. In this case, the dashed ACC curves are computed using only cases
where the control performed at a standard deviation lower than the mean or less (i.e., the worst
cases). The curves are similar for the northern hemisphere (not shown), but tend to be more neutral.
Figure 2: The southern hemisphere 500 hPa ACC die-off curve for 01 September to 03 October 2010. Solids lines are all
cases; dashed lines are worst cases (more than one standard deviation below the mean). Control in red; experiment in
blue.
GFS EXPERIMENT #3
rd
Since the first two experiments showed essentially a neutral impact, the 3 experiment is a much
different approach to determine how sensitive the assimilation is to EE. In this case, all winds are
-1
retained, but the observation error is assigned the value of the EE, which can range from 0 to 10 m s .
-1
-1
However, if the EE is less than 3 m s , the observation error is set to 3 m s . In the control, the
-1
observation error is essentially a fixed value of 7 m s .
The experiment is run over the same time period, 01 – 27 September 2010. The overall impact is
neutral in northern hemisphere (Fig. 3) and slightly negative in Southern Hemisphere (not shown).
Again, when only examining the worst cases there is an improvement in worst cases in northern
hemisphere (dashed lines in Fig. 3), while neutral in southern hemisphere (not shown).
Figure 3: The northern hemisphere 500 hPa ACC die-off curve for 01-27 September 2010. Solids lines are all cases;
dashed lines are worst cases (more than one standard deviation below the mean). Control in red; experiment in blue.
SUMMARY
The overall impact of using the EE over the current NCEP winds quality control from the three different
experiments is neutral. This is an encouraging result as a more quantitative approach is being used to
filter the polar winds. In addition, there is evidence that the worst cases are improved in the day 6 and
7 forecasts.
These results are considered preliminary, as they are for only the single month of September 2010.
Additional experiments are being run over a longer time period and in an additional season (winter).
REFERENCES
Holmlund, K., 1998. The utilization of statistical properties of satellite−derived atmospheric motion
vectors to derive quality indicators. Weather and Forecasting, 12, pp. 1093−1103.
Le Marshall, J, and A. Rea 2004. Error characterization of atmospheric motion vectors. Aust Met. Mag.
53, pp. 123-131.
ACKNOWLEDEMENTS
This work is supported under NOAA Grant NA10NES4400011.