NWP data assimilation

by Mary Forsythe — last modified Jun 12, 2014 06:47 AM

This page is still under development.  Still need to add work by Kris Bedka, Peter Lean, Renato Galante Negri ++

There is a need for NWP to be clearer on their requirements (spatial, temporal resolution of wind products) and data density. 

There are potentially 2 areas of interest

1. High resolution limited area models - high spatial and temporal wind information within model domain aiming to improve local forecasts and predication of high impact weather events
2. Global models - provision of more detailed wind information in key areas - one area that has been investigated is their potential benefit for improving tropical cyclone (TC) tracks, but we may benefit from improved resolution more widely than this (satellite-derived winds are likely to be most beneficial in the tropics and over the oceans).

Effective NWP resolution

We need to consider both the global and high resolution limited area models.

Skamarock (2006) identifies an important difference between NWP model resolution and grid sampling size; he suggests the effective NWP model resolution to be typically 5-10 times larger than the grid size.  One may expect a similar effect in the vertical dimension due to dynamical closure.  Global NWP models lack convective-scale processes and for example do not show changing winds over the 50 min Metop-A and B time differences, whereas observations show substantial changes near convection over large areas (100 km and more). Limitations in resolution provide a lower limit for the useful scale of flow we could attempt to derive for the AMV and scatterometer winds in global NWP.  Some regional high-resolution NWP models (km scale grid) contain wind variability on the 10 km scale in the free atmosphere i.e. an order of magnitude smaller than global NWP models generally do. 

Deterministic scales

Although the km scale regional models can resolve smaller scales, these tend to change fast, and represent only modest energy conversion, an extension on the observing system would be needed to initialize and evolve these scales in NWP models in a deterministic way.  In fact, while some observation systems exist locally (ocean surface, near airports), the required quantity and coverage of observations remains a daunting challenge.  In the absence of a resolving wind observing system to initialize these turbulent scales in regional NWP models, high resolution NWP will face increased variances, faster changes and smaller scales in background error covariances.  These changes potentially compromise the analysis of the larger scales.  A trade-off between forecast horizon and short-term quality may arise.

Way forward

A good understanding and analysis of effective model resolution and determined scales is needed to proceed successfully.  Can we say anything more on our best guess of preferred resolution and density for AMV and scatterometer obs for assimilation in global and high resolution models from the work carried out so far?  

One aspect relevant to this is how to handle the spatially and temporally correlated error.  This is likely to be particularly important for the AMV data, where studies of normal resolution AMVs have shown significant spatially correlated errors out to 800 km (Bormann et al., 2002).  However, it is also important for scatterometer, radiosonde and aircraft winds for a different reason.  The global NWP effective resolution is currently insufficient to resolve 25 km scale scatterometer winds, high resolution radiosonde data (in the vertical) or aircraft winds.  Therefore a spatial representativeness error is associated with these observations, which is spatially correlated due to atmospheric dynamics (turbulence and convection).

Currently in global NWP both the "measurement" and representativeness observation error correlations are handled by thinning the data (often up to ~ 200 km) and/or inflating the observation errors as the data assimilation itself assumes uncorrelated errors.  It is possible to allow for correlated error in the data assimilation schemes (e.g. EC ref), but this assumes a simplified correlation falling off with distance, which is unlikely to be the case in reality for both AMV measurement and for spatial representativeness errors and it is not clear whether overall it will yield a positive impact (check with Niels, VHAMP study).

Work is needed on:

(1) the characterisation of temporally and spatially correlated errors, and
(2) their reduction (either through post-processing at NWP centres or improvements by producers).

If we can't make progress on (2) then:
(3) either we should not put too much effort into generating them at high space/time resolution,
(4) or we need to learn how to use them better, e.g. observations with correlated errors should contain BETTER information on gradients.

It is very important to analyse the spatial representation of observations and NWP model in order to understand both background and observation error covariances (Vogelzang et al., 2011) before data assimilation.  We can distinguish 3 cases:

1. Details that do not fit with a smooth dynamical closure of a global NWP model cannot be effectively assimilated.  In such cases observations may be integrated to larger scales in order to represent those scales that are effectively assimilated and remove both observation measurement and representativeness error variance that the global NWP model cannot handle and thus would enhance O-B statistics.
2. If observation and model scales match well, the representativeness error may be negligible and assimilation may be straightforward (cf successful OSCAT scatterometer assimilation with effective 100 km resolution at 50 km grid).
3. NWP models may detail features that are not observed, but rather integrated in a measurement.  It is in principle straightforward to implement an integration effect in the observation operator.
   

In order to evolve the detail of the local flow field in regional NWP, it is necessary to increase the temporal coverage of observations (Mode-S, OSCAT).  Moreover, thinning distance should be reduced compared to those used for global models. This could be done everywhere or a scheme could be developed to use the data at highest spatial and temporal resolution in regions of interest (e.g. varied or rapidly changing wind fields ++), but we might need to further inflate the observation errors to avoid hitting the assimilation too hard with data which has correlated errors. 

Also discussed under AMV products .  It may be beneficial to make more use of error estimates provided from the derivation when available in the future and reduce reliance on the QI.

To use the data at highest spatial and temporal resolution in regions of interest (e.g. varied or rapidly changing wind fields ++) care needs to be taken of the inherently short error correlation scales in the background.  Currently many observations depicting such elevated atmospheric dynamics are rejected or downweighted by the QC in global NWP.  In regional NWP, we might need to further inflate the observation errors to avoid hitting the assimilation too hard with data with poorly known correlated errors.

Some considerations for QC are likely to be similar for regional and global wind products including spatial blacklisting, threshold checks and modifications to the observation operator (e.g. to treat AMVs as a layer)

Work has been done in the US and Japan on use of rapid scan high resolution AMVs for improving tropical cyclone forecasts. Several papers have shown benefit from assimilating them in global and regional models. Are there plans to produce and disseminate these routinely? Check also if any HFIP results.

References:

Add JMA references

Wu et al., 2014. Influence of assimilating satellite-derived atmospheric motion vector observations on numerical analyses and forecasts of tropical cyclone track and intensity. Monthly Weather Review, 142, 49-71. 

Berger et al., 2011. Impact of Enhanced Satellite-Derived Atmospheric Motion Vector Observations on Numerical Tropical Cyclone Track Forecasts in the Western North Pacific during TPARC/TCS-08. JAMC, 50, 2309-2318.

Langland et al., 2009. Impact of Satellite-Derived Rapid-Scan Wind Observations on Numerical Model Forecasts of Hurricane Katrina. Monthly Weather Review, 137, 1615-1622.

Pu et al., 2008. The Impact of Aircraft Dropsonde and Satellite Wind Data on Numerical Simulations of Two Landfalling Tropical Storms during the Tropical Cloud Systems and Processes Experiment. Weather Forecasting, 23, 62-79.

• CIMSS - 2 studies proposed (awaiting final confirmation of funding)
  

1. rapid-scan AMV derivation using 1 min scans from GOES-14 through GOES-R program.
   

1. application/DA and model forecast impacts on regional hurricane modeling through the Hurricane Forecast Improvement Program (HFIP)
   

• KNMI - spatial and spectral information content; background error structure; Desrozier for scatterometers and Mode-S; Mode-S airplane data; vertical and horizontal sampling for Aeolus data assimilation; representativeness error (correlation) through ESA VHAMP; use EDA in scatterometer analysis and ambiguity removal (weather dependent observation and background error)
  

• Met Office (UK) / Uni of Reading - plan to use the Desroziers approach to look at the AMV correlated errors.  Plan to run further tests with the UKV 1.5 km model to evaluate changes to the derivation and assimilation scheme of UK AMVs produced at the Met Office using the NWC SAF software.

References and Links

• HARMONIE (EUMETSAT fellow)
• EU MyWave
• ESA Aeolus (VHAMP, ..)

The NWC SAF high resolution wind software provides a good option for investigating optimal settings as it has been designed to provide high density winds over relatively large areas and is easily configurable.  The NWC SAF are happy to work with one or more NWP centres to investigate settings e.g.

• The amount of AMV data calculated by HRW can vary up to three/four orders of magnitude depending on the configuration, and the impact of the different densities of AMV data can be evaluated considering the AMV validation or their impact in NWP assimilation studies. A study is proposed in which the separation between tracers and correspondingly the density of AMV data is changed. The NWC SAF would extract the AMVs for the evaluation period using different densities/separations between tracers and evaluate their corresponding characteristics and validation. Some voluntary NWP centre would later assimilate the AMVs in their models to verify the differences in the forecast with the different densities of data.