Assimilation of High-Latitude Atmospheric Motion Vector Winds in GEOS-5

A series of experiments has been conducted to test the impact of assimilating high-latitude (polar) wind observations (atmospheric motion vectors) based on imagery from the Advanced Very High Resolution Radiometer (AVHRR) on the quality of analyses and forecasts produced by the GEOS-5 atmospheric data assimilation system. The AVHRR winds were provided by the Space Science and Engineering Center (SSEC) at the University of Wisconsin. Among other things, the study compares the impact of the AVHRR observations with the impact of polar winds based on the Moderate Resolution Imaging Spectroradiometer (MODIS), which are assimilated routinely in GEOS-5 and in most operational forecast systems.

The experiments were conducted for the period June-July 2009 using version 5.6.1p4 of the GEOS-5 atmospheric data assimilation system at 1/2-degree horizontal resolution with 72 vertical levels. A control experiment (including MODIS winds but no AVHRR winds) was run, as well as experiments with both MODIS and AVHRR winds, and no polar winds, as summarized in the table below. Additional experiments with winds derived from the Multi-angle Imaging SpectroRadiometer (MISR) are also planned, but require modifications to GEOS-5 to accommodate the height-based vertical coordinate of these data.

Experiment MODIS AVHRR MISR Polar Winds Name Obs System Summary
0 CTL x MODIS only std1_d72
1 CTL + AVHRR x x MODIS+AVHRR amw1_d72 June 2009 July 2009
2 CTL + AVHRR - MODIS x AVHRR only apw1_d72
3 CTL + MISR x x MODIS+MISR mmw1_d72
4 CTL - MODIS no polar winds npw1_d72
5 CTL + MISR - MODIS x MISR only mpw1_d72

Observation Increments

The distributions of observation-minus-background forecast (OMF) and observation-minus-analysis (OMA) departures (often referred to as observation increments, see Daley 1991) provide an indication of the relative agreement between the observations and model background state before and after the assimilation procedure. Ideally, the background state should provide a reasonably accurate estimate of the analyzed state, and the observation and analysis increments should provide only a minor adjustment to the background state. The OMF and OMA values should be unbiased and normally distributed, with the OMA having smaller standard deviation than the OMF reflecting the improved "fit" of the observations to the analyzed state.

The figures below show results for the experiment amw1_d72, in which both AVHRR and MODIS winds are assimilated. Overall, the OMF and OMA values for both data types have reasonable distributions, although the distributions for the AVHRR winds have longer tails and larger standard deviations than those for the MODIS winds, indicating a slightly worse fit between the AVHRR winds and other sources of information in the analysis compared with the MODIS winds.

MonthQCAVHRRMODIS
U-windV-windU-windV-wind
JuneAll obs
Passed QC
JulyAll obs
Passed QC

Forecast Skill Scores

The figures below show forecast skill scores in terms of the anomaly correlation coefficient (ACC) for 500 hPa height. This is a standard metric for measuring the skill of numerical weather forecasts, where ACC=1 denotes a perfect forecast relative to the GEOS-5 analysis valid at the same time as the forecast. The "die-off" curves on the left show the average ACC values of the different experiments as a function of forecast lead time for the Northern Hemisphere (NH), Southern Hemisphere (SH), Europe (EU) and North America (NA). The lower portion of these figures shows the statistical significance of the experimental results in terms of their difference from the control run. The significance test is computed using paired difference testing as described here. The figures on the right show time series of ACC for the individual forecasts during the test period at various forecast lead times.

Overall, the forecast skill scores show small but insignificant differences between the experiments. In most cases, the experiment without polar winds exhibits the largest ACC skill score at a forecast lead time of five days although there are some cases where using the AVHRR winds, or using both sets of polar winds, perform better than the 'no polar' or control experiments.

Month Die-Off Curves Time Series
NHSHEUNA NHSHEUNA
June
July
June-July

Adjoint-based Observation Impacts

The adjoint of a data assimilation system provides an accurate and efficient means of estimating the impacts of any or all assimilated observations on measures of short-range forecast skill. The impacts are computed for all observations simultaneously, without having to add or remove selected subsets of observations from the data assimilation system. A description of the adjoint technique, and how it compares with traditional data-denial experiments, is given by Gelaro and Zhu (2009).

The adjoint of the GEOS-5 data assimilation system was used to compute the impact of observations on 24-h forecasts during the study period in terms of the reduction of a global (energy-based) measure of error combining temperature, wind and surface pressure from the surface to 150 hPa. The figures below show, for the various data types assimilated in GEOS-5, their total impact, impact per-observation and fraction of observations that improve (versus degrade) the 24-h forecasts during the study period. Note that for the total impact and impact per-observation, negative values indicate a reduction in the forecast error measure and thus an improvement in the forecast due to assimilation of the observations. Also shown for reference are the observation counts for each data type, which also helps explain, for example, the difference between the total impact and impact per-observation of a given data type.

In June, the AVHRR winds have near zero impact globally, although the majority of the observations have a beneficial impact on (reduce the error of) the 24-h forecast. In should be kept in mind that, being restricted to the polar region, the AVHRR (and MODIS) winds might have a larger impact in terms of regional (high-latitude) forecast measures as opposed to the global measure used here.

In July, the AVHRR winds do not perform as well as in June and the statistics show that, overall, they increase the 24-h global forecast error measure. The negative impact of AVHRR during July is especially clear in terms of the impact per-observation. Note also that the majority of the AVHRR observations during July degrade the forecast, indicating that the overall negative impact during this month is not likely attributable to just a few "outlier" observations with unusually large negative impacts.

MonthImpactImpact per Ob% Beneficial ObsObs count
June
July

References

Daley, R., 1991: Atmospheric Data Analysis. Cambridge University Press. 457 pp.

Gelaro, R. and Y. Zhu, 2009: Examination of observation impacts derived from observing system experiments (OSEs) and adjoint models. Tellus, 61A, 179-193.