Nwp Observational Data Assimilation

[Evo]

I'd best start this off with a disclaimer - I'm not an expert in any of the following so some, or all, of it may be pish!

I've been pondering the issue of model data assimilation since looking at the Kalnay book I mentioned in John Holmes' pinned GFS data thread and reading J07s interesting contribution in that thread.

In the book there are a couple of maps showing the actual observations incorporated into a particular GFS run. It's only recently dawned on me that the observations that are assimilated into the analysis forecast (which forms the basis of the model run) vary by quantity and location from run to run. There is a much higher density of observations near human habitation - which seems intuitive. The main area that attracted my attention, though, was Greenland. There was a scant amount of data input from this region on that particular run.

The potential sources that data can be assimilated from can be summarised as:

i) Land (surface) based observations - both manual and automatic

ii) Sea (surface and sub-surface) observations - such as buoy data

iii) Ascents - observers on land or at sea release 1-2 balloons a day per location

iv) Aircraft - automated and manual observations

v) Satellite - remote sensing of wind and temperature, radiance and vapour

While many stations incorporate automated measuring and reporting equipment, some information is derived from manual observation such as ascents. These require human input obviously and this human input means that occasionally there will be a problem with getting this done in a timely and accurate fashion! Some data is assimilated from aircraft, but on average aircraft spend a large proportion of their flights at cruise level so the information will typically be from 30,000 to 40,000 feet and (this is an educated supposition) be composed of temperature (adjusted to factor in the effect of speed), wind speed/direction and static pressure.

There are some weaknesses with the aircraft observations - aircraft will generally follow set routes called airways. Data will be concentrated along these airways, whilst outside of the airway system there will be little or no data. If you look at the ACARS (aircraft) data on Kalnay's map, you'll notice that the observations over the Atlantic line up vertically. This is because the aircraft crossing the Atlantic follow a movable set of airways called the North Atlantic Track system (or NAT tracks). The NAT tracks are fixed each day and will compose 5 westbound tracks and 7 eastbound tracks. The actual tracks will take into account factors such as the Jetstream. In general the westbound tracks will be more to the North to avoid the Jetstream and in general the eastbound tracks will be more to the South to enlist the help of the Jetstream. This will obviously vary by season and prevailing synoptics. Westbound flights tend to happen in the first half of the day GMT, while eastbound flights tend to happen overnight. Each track is made of up waypoints which are latitude and longitude coordinates. ACARS reports are made at these waypoints, which will usually be at 50W 40W 30W and 20W - you can see the effect of this in the vertical lines in Kalnay's ACARS map over the Atlantic. So, the location, time and quantity of ACARS data will vary each and every day.

Other data are derived from satellite soundings as J07 has illustrated but as far as I (as an interested amateur) can establish the information from these satellites is rather generic and interpolated. The satellite derived data will not be anywhere near as specific and accurate as an ascent.

Finally there are many other issues with the data assimilation process, as J07 has highlighted - getting the data measured and incorporated in a timely fashion, blending different (and potentially conflicting!) observations and the issue of extrapolation into areas where there is a low amount and quality of data available.

Where I'm leading to is this question - Although technology and resources will no doubt have improved since Kalnay's 1997 maps, if there is an inadequate amount and quality of observations in the Greenland region, how can a model hope to resolve the energy balances in the so-called sub-polar field correctly? I find it hard to believe that such an obvious hole would be left, or that this would not have been considered by those more learned that I but it does pose an interesting what-if question, especially given the importance [to our weather] of the energy flows in the GIN corridor!

Sources:

http://mason.gmu.edu/~phouser/houser_files...NAL_22Jan08.pdf

http://www.goes-r.gov/downloads/GOES%2520U...conf5_moore.ppt

http://www.emc.ncep.noaa.gov/gmb/ens/THORP...gfs_thorpex.pdf

http://www.emc.ncep.noaa.gov/mmb/data_proc...cumentation.htm

While not directly relevant, interesting as an under-the-hood look at the GFS model while the Indian met agency implemented the GFS model:

http://www.ncmrwf.gov.in/ncmrwf/gfs_report_final.pdf

[Paul]

You can see details on the data the GFS ingests here:

http://www.netweather.tv/forum/index.php?showtopic=35985

[johnholmes]

this link will give you the surface data for Greenland, some are auto some are manual

http://www.ogimet.com/display_synopsc.php?...8&send=send

Off hand I would imagine that the radio sonde network is about 1 in 10 of the surface reports but I'm not sure about that

and here is the upper air link for Greenalnd, normally they will do a full sonde at 00 and 12z possibly 06 and 18z depending on operational air force requirements I would imagine

http://www.ogimet.com/display_sondc.php?la...8&send=send

Edited December 21, 2008 by johnholmes

[Evo]

Thanks for the replies chaps.

I've attached the image I was referring to so you can see where I was coming from.

It is certainly interesting how much preprocessing of this raw data has to happen before an analysis forecast can be produced - upon which the model run is initiated.

[Evo]

After some more digging it turns out my hunch, while not totally correct, had some mileage. It turns out that a research project called "The Greenland Flow Distortion Experiment" has recently completed field observations to investigate this issue (well not specifically but to get a better handle on the mesoscale processes that occur around Greenland).

First, there are two background Met Office research papers which are worth a look. You need to complete a free registration to view them but I’ve summarised what I consider to be the salient points.

Dumelow, R - "The impact of conventional observations on global and regional NWP forecasts" Available at http://www.metoffice.gov.uk/research/nwp/p...reports/516.pdf

This paper looked at the relative importance in terms of forecasting skill of the differing types of observations that are assimilated into a model analysis. The author is coming from the angle of what would happen to the accuracy of the system if resources were concentrated on mainly satellite-derived data. The paper broke the observational categories into these groups:

i) All available observations - COMB

ii) Satellite derived data, GCOS, GUAN and buoys - BASE

iii) BASE + Aircraft data - BPAIR

iv) BASE + non GUAN radiosonde temperature and wind obs. - BPGTW

v) BASE + non GUAN radiosonde wind observations - BPNGW

GCOS = Global Climate Observing System

GUAN = GCOS Upper Air Network

The author points out that other papers have looked at denying entire classes of data or targeted field observations.

An experiment was conducted during two periods - one in winter and one in summer. At T+48 in winter there is a marked improvement in forecasting skill by using COMB compared to BASE in a belt at 60N. This is explained as likely to be due to the density of observations in this region and the lesser difference between BASE and COMB data volume elsewhere. At T144 the areas of error are more widespread with only 30N to 30s showing little difference, however counter-intuitively the influence of the difference in observation use declines with increasing forecast range.

In general using BASE over COMB produces a 5-30% decrease in accuracy depending on the time differential, location and season.

The paper then looks at the effect of BPAIR, BPGTW and BPNGW. The order of effectiveness is:

BPGTW (radiosonde temperature and wind observations)

BPNGW (radiosonde wind observations)

BPAIR (Aircraft data)

So, adding aircraft air data is the least effective way of improving forecast skill over the base observations.

However this has all using the UK Met Global model. When using the more localised UK Met NAE model, during winter no improvement can be seen by adding any of the extra data in! During summer adding the aircraft data is the most effective - giving a result similar to the complete data set. In the timeframe that most "model watchers" are looking at the NAE model would be of little interest even if it were published, though this is worth bearing in mind.

Next,

Bovis K, Dow G, Dumelow R & Keil M - "An assessment of deployment strategies for targeting observations" Available at http://www.metoffice.gov.uk/research/nwp/p...reports/515.pdf

This paper looks at the potential benefit of deploying additional radiosonde observations within a targeted area to improve forecast accuracy for a specified region. In this particular context, they are looking at hurricane track prediction.

There are two potential strategies for identifying areas that would benefit from additional observations - using a human forecaster and an analysis chart or by applying an Ensemble Transform Kalman Filter or ETKF, which in laymen’s terms identifies regions where forecast uncertainty is highest. It turns out that EKTF is far better at picking regions that would benefit from additional targeted observational data.

Not surprisingly, when targeting these regions an improvement in forecast skill is observed. The authors note that that the single most important category in terms of forecasting skill is radiosonde data.

Finally they note that additional work in this area has been undertaken called the Greenland Flow Distortion Experiment (GFDex). Rather than repeat the details of what GFDex is, I'd recommend a read of http://lgmacweb.env.uea.ac.uk/e046/researc...description.htm. The field observations have now completed and investigation of the collected data is underway.

The key theme of each of these links is that improving upstream observations improves downstream forecast accuracy. This is obvious really when you think about it. The good news is that we can look forward to improvements in this area in the next couple of years as well as the continued improvements in model resolution. Whether this means more resources are allocated to creating fixed ascent stations or aircraft are used for dropsondes I’m not sure. In my opinion efforts to obtain more efficient and accurate use of satellite data will be the most likely course of action.

Whether this will enable the GFS to accurately predict whether it will snow at T240 is anyone's guess!

[johnholmes]

tks for that Evo, I'll have a read of the links you have posted.

Like you say, more data available from every area of the world will obviously improve the model accuracy; what we have to bear in mind is, will the cost outweight the benefits that the model centres have to fork out for this?

[BrickFielder]

Satellite monitoring has come a long way in the last few years and if you want to read up on it then the EUMETSAT site might be a good place to start. Once specific area where satellite measurements may be less accurate is in very cloudy conditions. This typically occurs in the north and to the west of us which is why models go through a particular pattern of adjustment as model lead times lessen.

Three days out and models tend to give temperatures too cold at 850hPa for the UK. This adjusts to give temperatures a touch too high at just over two days out and approaches actuals at one day out.

Eumetsat article on satellite monitoring