Joanne Waller, University of Reading
Diagnosing observation error statistics for AMV observations
Coauthors
M. Cordoba, G. Kelly, S.L. Dance, N.K. NicholsAbstract:
Atmospheric Motion Vectors (AMVs) are wind observations derived by tracking cloud or water vapour features in consecutive satellite images. AMVs are incorporated into Numerical Weather Prediction through data assimilation; for this process, an accurate description of the observation error statistics is required. This work provides novel information by estimating the AMV error variance and horizontal error correlation for observations assimilated in to the Met Office high resolution (1.5km) model. Currently operationally assimilated AMVs are derived by the NWC-SAF HRW algorithm from the 4 SEVIRI channels, IR108, WV062, WV073 and HRVIS. Prior to assimilation the observations are thinned to 20km in attempt to satisfy the assumption of uncorrelated observation errors. In this work we use a diagnostic that makes use of statistical averages of background and analysis residuals to calculate observation error statistics at high (100hPa – 400hPa), medium (400hPa – 700hPa) and low (700hPa – 1000hPa) pressure levels for AMVs from the four different channels.
The results show the observation error variances vary with height ranging from 5 to 14 (m/s)^2. The largest error variances occur at the mid and high levels where the influence of wind speed and shear is largest. The estimated variances are larger than those the derived in previous studies, but lower that the variance used in the operational assimilation system (15 to 65(m/s)^2). The horizontal error correlation length scale for AMVs from IR108, WV062 and WV073 at all pressure levels range between 150km and 210 km. These length scales are similar to results found in previous studies and are likely to be a result of the propagation of height assignment and tracking errors. For AMVs from the HRVis channel the correlations are significantly longer with length scales of approximately 360km. This result suggests that at least one source of error has larger length scales during the daytime compared to the night.