(back)


Initialization Assessment
The goal in this phase of a model diagnosis is to answer the question "How well did this model capture the initial state of the atmosphere ?".  You should assume there is error in the initialization (find it !)".

1. Manual Analysis
Call it old fashioned, but it is excellent practice to perform a manual analysis at the following levels while waiting for model data to become available:

It does not take long to perform a manual analysis and its a critical component of model diagnostics. This will get you familiar with the pattern and intimate with some of the observations going into the model initialization.

2. ADMNFD message
This is a key piece of information for you to use in the initialization assessment - read it !!!  The ADMNFD message from the NCEP NCO SDM describes both data availability and quality control measure taken on the data intended for model initialization.

Note in this message the number of upper air sites whose observations where either missing or needed to be truncated (note that 10144 means due to power problems no upper air balloon was launched).

Note if areas locations where data is missing or deleted corresponds to where significant weather is expected to evolve FROM (rather than develop OVER).  For example - if energy over the Pacific Northwest Coast is expected to spawn a storm east of the Rockies, you should be very concerned about  upper air obs over the Pacific Northwest Coast that are missing.

In general, missing data in regions of sharp contrasts in temperature, humidity and wind should be taken very seriously.

3. Comparison of Model Initialization to Observations
The next step is to compare a model's initialization (fhr 00 output) against observations to discern any glaring discrepancies.

Parameter Assessment and Discrepancy Thresholds
As a general rule a model discrepancy can be considered significant if it varies more than 10% of the observed value for moisture and wind speed.  For example, a model's initialized PW value should not vary by more than .2" where the observed value is 2".  If wind speed is observed to be 70 kts, the model's value should be + or - 7 kts of the observed value.

Temperature, height, and pressure are slightly different.  In general the model initialized value should NOT differ from an observation by more than 2 degrees C for temperature, 2-5 mb for surface pressure, and half a standard contour interval for height.  For example, a model's initialized value for 500 mb height should not differ by more than 3 dm since the standard contour interval at that level is 6 dm.

These are not rigid thresholds.  Some discrepancies noted in regions where significant weather is not expected to either evolve from or over may not be detrimental to the model output.  Pressure initialization can be particularly tricky to assess over complex terrain and near rapidly deepening systems over oceans.  In these cases, observations may not be as dense as required for a proper assessment and a larger threshold may be acceptable. Conversely, more stringent thresholds may apply to different meteorological sensitivities.  During winter it may be prudent to use a 1 degree C thresholds near freezing lines.

Domain of Assessment
Over what domain should an assessment be made ?  If, for example, the pattern is west to east, assessing the model initialization well east of your area of interest is not as critical as assessments made west of your area of interest.

As a general rule in mid latitudes you should look upstream approximately 20 degrees longitude per forecast day to assess initialization.

To put this rule into perspective, it is equivalent to the distance from the western Ohio valley to New England traversed in one day.. or from the pacific northwest coast to the front range of the Rockies traversed in one day.  This assumes a mean upper level flow of 50 kts.  Certainly upper level jets can be up to four times stronger than this, but they usually are not as such over the entire domain.

Large scale systems down stream of your area of interest can also influence local meteorology, so a good initialization assessment will also include features about 20 degrees longitude east of your easternmost area of interest per forecast day.  For example, if you are concerned about a snow storm over the Ohio Valley 48 hours from now, you would likely have to pay close attention to energy initialized just off the West Coast under relatively zonal flow.

Meridional transport in the atmosphere also varies by pattern - for upper level flow if the pattern is zonal and low amplitude, north south transport is on the order of 10 degrees latitude or less per day.  This is may be greater for high amplitude patterns.

Optimally, assessment of initialization intended for the CONUS should cover an area from the International Date Line (for systems impacting the west coast of the CONUS by fhr 72) and east to about 50 W longitude over the Atlantic between the equator and 70N latitude.

This roughly corresponds to the North American scale on AWIPS and NAWIPS displays.

CAUTION - this is a rule of thumb, there will be times  you need to view data (if possible) further east/west and/or north/south during regimes of fast flow or progressive patterns.

Levels to Assess
Technically you have a three dimensional grid at fhr 00 over at least the size of North America for the Eta model - and over the entire northern hemisphere for the GFS model.  Further, each model has a gazillion parameters that can potentially be assessed - and at a bazillion vertical levels.  How in God's name are you supposed to perform a complete analysis of initialization error ?

You can't !  You simply can not assess every field at every level over the horizontal domain of the model.  Ideally you would want some kind of automated routine to show you in 3D where the initialization departs from the analysis for temperature, moisture, density, and wind.

But until that becomes available, the next best thing would be to view initialization departures from observations horizontally at specific levels.
That we CAN do.  Further, you can focus your assessment on those parameters and levels either typically used when forecasting or those expected to be critical to the event that is expected (cyclogenesis, convection, tropical development, etc.).

To date one of the most efficient tools to perform an initialization assessment at HPC is AWIPS.  Additionally a web based Model - Observation Comparator (MOC) is available for previous cycles.  Both allows you the ability to plot the most observational data against the Meso and Global model initialization.

AWIPS requires the use of the Volume Browser and the Sampling Feature (at the end of this tutorial is an appendix describing how to obtain data from the volume browser and how to use the sampling feature).  The web based tool will show the observations that were actually ingested into the data analysis systems of both models.  It is found at http://www.hpc.ncep.noaa.gov/mdd/moc

Below are techniques that can be employed to discern initialization errors in moisture, wind, height/pressure, and temperature using AWIPS and the MOC.

Initialization Assessment Techniques - Large and Small Scale Circulations
In AWIPS, on the North American scale, overlay WV imagery with fhr 00 400 mb heights centered + and - 3 hours around fhr 00. "Why WV imagery and why 400 mb heights ?" you ask.  WV imagery is the best channel to view atmospheric circulations - large and small.  Since the WV channel can only sense down to about 600 mb (with its strongest response about the 400 mb level), we view heights at a level consistent with the WV channel.

This loop will give you a quick indication if the large scale has been captured by a model's initialization.  If the model has not captured observed large scale circulations, the model should immediately be considered suspect and even discarded.

The 6 image loop below shows WV imagery in 1 hour increments centered over 12Z.  Note the large circulations over the Pacific.  Embedded in this pattern are also small scale circulations over the central Texas and western New Mexico.

Now is a good time to overlay the 400 mb heights from the GFS and Eta initialization to see if the models captured the large scale pattern (below).

Note both models did indeed capture the large scale at this level.. for the most part.  The short wave over the Pacific seems underdone by both models.  Although the Eta domain on AWIPS gets cut off near the vicinity of that short wave, it is easy to see that it does not exhibit sufficient amplitude as suggested by WV imagery.

There is also a strong anti cyclonic circulation over Louisiana.  Not terribly important perhaps, but it can be used as a gauge in determining which model may have a better handle on the initial analysis.

It is good practice to overlay 400 mb absolute vorticity on the WV imagery to discern if smaller scale circulations were captured by the models.   It is true that some circulations evident on WV may be above or below 400 mb, but 400 mb is a good place to start when comparing to WV imagery.

Loosely speaking the Eta did pick up on the smaller scale circulations in TX/NM and LA, but be careful! The Eta as you see it below is plotted on a 40 km grid.  So although it may not look like has picked up some of the small scale circulations it may actually have done so.  A comparison to the GFS will shed some light on how well the Eta picked up on these features (because the GFS is a lower resolution model than the Eta).

The GFS seemed to better resolve the small scale vorts.  Although there is one vort max depicted in TX/NM, it actually is depicting two maxima - one in NM and one in TX which corresponds closely to what is evident on WV imagery.  The GFS depicts a stronger anticyclonic circulation over LA than the Eta.  Also, the GFS seems to have only caught a fraction of the strength of the Pacific short wave.  Over all the GFS seems a little better initialized than the Eta with the vorticity pattern, but is still underdone with the Pacific short wave.

(cont)