Forecast Soundings: A Look to the Future

Posted by pmarshwx on 26 Jul 2011 | 9 comments

Those who are regular readers of this blog know by now that my research interests are centered on operational meteorology, data visualization, and data mining. For the 2011 Hazardous Weather Testbest (HWT) Experimental Forecast Program (EFP), I was able to combine my interests by creating a means of viewing model forecast soundings from various convection-allowing models, for the same location, simultaneously. Now, I didn’t create the viewing tool, but created a way to properly format millions of lines of data to utilize the BUFKIT sounding program. An example of one of these ensemble soundings is shown below.

As you can see, the ensemble sounding illustrates several different forecast scenarios that are not apparent when looking at a single sounding. In severe weather forecasting, these sometimes subtle differences in an atmospheric profile can lead to vastly different results. By looking at an ensemble of solutions, forecasters can begin to gauge the variability of the numerical guidance, as well as the predictability. Researchers also benefit. In the HWT this year, we learned about how different model physics impact the atmospheric profile, which, in turn, might introduce systematic biases into the forecast. Knowing this allows researchers to develop better ensembles and better models.

Creating this data set was not an easy process. I had to read in text-files containing the sounding information (example located at the end of the post) for each forecast sounding points — 1146 per model — in each of the 18 numerical models. At that point I needed to be able to compute various thermodynamic parameters from the soundings, so I rewrote most of the thermodynamic and kinematic routines in the Storm Prediction Center’s (SPC) sounding viewer, NSHARP (National Skew-T and Hodograph Analysis and Research Program), in Python. This allowed me to compute a thermodynamic and kinematic parameters in a wide variety of ways for the HWT-EFP, all consistent with what Storm Prediction Center forecasters are used to using in operations.

After computing the necessary thermodynamic fields, I then created sounding files in BUFKIT format for each of the 1146 sounding sites for each of the 18 models. Lastly, I combined the 18 model forecasts for each of the 1146 sites into a single file and output 1146 more text files containing the ensemble sounding files which could then be read by E-BUFKIT. A sample image is above. All-in-all this took about 1 hour of computation time on 18 separate computers. That’s a lot of data to crunch through!

After doing this for the HWT-EFP, I thought to myself, “Why stop there?”. Since the end of the EFP, and with the help of John Hart of the Storm Prediction Center, I have begun work on creating an ensemble sounding viewer — written entirely in Python — based on the SPC’s sounding program (NSHARP). Why Python? Because Python allows the program to be cross-platform, meaning anyone who installs Python on their computer can use the program.

The ultimate goal is to create this program and then release it to the atmospheric community as an open source project. This would allow researchers, forecasters, and hobbyists to be able to use the program to view model soundings in both a deterministic and ensemble framework. Additionally, by having a community supported sounding tool, it is my hope that researchers in the climate community will also use this program for their research. If the meteorology community adopts a single sounding tool, scientists, forecasters, and hobbyists will be able to quickly compare parameters from climate models, weather models, and observational data, knowing that the values were computed in the same manner. This would be a huge boost to comparing historical, current, and future events. Not to mention allow for consistency between datasets.

I’m still in the early stages of development. My hope is to be able to present an alpha version of this program at the American Meteorological Society’s Annual Meeting in New Orleans next January. I’ll have my work cut-out for me…

In the mean time, a screen shot from this new sounding tool is shown below. This prototype supports active read-out, with on-the-fly interpolation as one moves the mouse over the sounding. The read-out is listed at the bottom of the sounding.

I’m interested in knowing what people think, so feel free to leave me feedback!

Sample Sounding Text File

Visual Comparison: 3-4 April 1974 and 27-28 April 2011

Posted by pmarshwx on 5 Jul 2011 | 3 comments

By the afternoon of 28 April 2011 it was fully apparent that the unthinkable had happened. In an era of unprecedented communication abilities, a single tornado outbreak took the lives of more people than all the tornadoes over the past several years combined – in broad daylight no less. In the days the followed, many tried to place this event into historical context. Nearly every one defaulted to the 3-4 April 1974 “Super Outbreak”.

The Super Outbreak was nothing short of impressive from a meteorological point of view. 148 tornadoes, 319 fatalities, over 13-states, in 24-hours. Never before, and not until this April, had anything even close to the scale of this tornado outbreak had ever been recorded. By comparison, the tornado outbreak of 27-28 April 2011 has an unofficial count (undertaken by several of us at the Storm Prediction Center) of over 174 tornadoes (done via Public Information Statements) and 259 fatalities attributed to these tornadoes. (Unfortunately, the death toll is considerably higher, I simply have been unable to place all the fatalities to the corresponding tornado at this time.)

From the standpoint of the number of tornadoes recorded and the number of fatalities, these two tornado outbreaks are in a class by themselves (in the “modern” tornado database starting in 1950). In the days that followed, I created a set of figures for internal NWS/SPC/NSSL use to compare the two tornado outbreaks. The images show all reported tornado tracks, color coded based on intensity and the counties are color-filled based on the number of fatalities that occurred within that county’s boundaries. A simple, quick look through the two events shows that the 3-4 April 1974 event covered a much larger area than the 27-28 April 2011 event, although there is considerable overlap between the two events. Several counties experienced fatalities in both events; in fact, Marion County, Alabama was unfortunate to have had a F/EF-5 tornado, and large loss of life, in both of outbreaks (1974: Guin, AL; 2011: Hackleburg, AL). Lastly, each figure has a table of the number of tornadoes and corresponding fatalities, broken down by EF-Scale (the 2011 event is still “preliminary” and subject to change). (Note, higher resolution images, for “zooming” are available by request.)

Meteorologists (and others) can, and will, debate for years as to which event was “more impressive”. I know what my thoughts are, but I’ll spare you those. However, please feel free to leave your thoughts in the comments.

The two images above are on the same background. This means if you download both of them and flip back and forth between the two, the only things that should change are the county colors and tornado tracks. Below is a zoomed in version of the 27-28 April 2011 event, complete with NWS County Warning Areas and County Names denoted.

SPC Day 3 Moderates In Context

Posted by pmarshwx on 8 Apr 2011 | 2 comments

This morning, weather enthusiasts woke to a Storm Prediction Center (SPC) Day 3 Moderate Risk outlook for severe thunderstorms. Most weather enthusiasts already know a Day 3 Moderate Risk outlook is a pretty rare occurrence, but just how rare is it? The simple and misleading answer is that since 2000, and including today’s, only 10 Day 3 Moderate Risk outlooks have been issued. The more precise answer is a bit more complex.

The SPC is continually refining their products based on the state of the science and user feedback. As such the criteria for a Day 3 Moderate Risk outlook has changed over time. Currently, it takes an “Any Severe” probability of 45% and a “Significant Severe” probability of 10% to reach Day 3 Moderate Risk category. Previously it was possible to ascertain the Day 3 Moderate Risk with as little as an “Any Severe” probability as low as 30%, which has happened several more times since the change in probability criteria. Therefore, it is difficult to compare old Day 3 Moderate Risks to current Day 3 Moderate Risks. However, I’ve attempted to break them down below.

Below is a table of the date of issuance for Day 3 Moderate Risk outlooks:

Broken down by year:

For additional information regarding SPC Outlooks, this time focusing on Day 1, please check out the following posts from earlier this year:

• Average Location and Date of First SPC Day 1 Moderate and High Risks
• SPC Day 1 Moderate and High Risk Climatologies
• SPC Day 1 Slight Risks Climatology

Updated Tornado Information Coming Soon!

Posted by pmarshwx on 24 Mar 2011 | 0 comments

Greg Carbin, Warning Coordination Meteorologist at the Storm Prediction Center (SPC), informed me today that he has updated the SPC tornado database up through 2010. Thus, in the coming days, I’ll updated the graphics to include the last 2 years worth of tornadoes. I look forward to playing with the updated data!

SPC Slight Risk Climatology

Posted by pmarshwx on 4 Feb 2011 | 6 comments

Next up in the list of Storm Prediction Center (SPC) outlook climatologies is the yearly climatology of slight risks. Between 1990 and 2008 there were 24,455 slight risk polygons issued. This does not mean there were 24,455 slight risk outlooks issued as it is possible to have more than one slight risk polygon drawn on a given outlook. Furthermore, note that each day consists of multiple outlooks and so it is possible for a grid point to receive multiple “hits” for being located in a slight risk outlook on the same day. In other words, if a location was located within 2 slight risk outlooks, this does not guarantee that two days of slight risks. The location might simply be contained within a slight risk from two separate outlooks issued for the same day.

Below is the slight risk climatology that employs the Kernel Density Estimation technique. Notice how it is smoother along the edges than the raw outlook which is located at the end of this post. The black solid lines are contour intervals of outlooks per year in steps of 25. The color fill contours are in increments of 1 outlook per year. As you can see, the central United States is once again the leader in the number of slight risks to expect per year. This has to do with the repeatability and predictability of the ingredients necessary for severe thunderstorm development. In other words, ingredients come together more frequently and more reliability across “Tornado Alley” than anywhere else in the world. This doesn’t mean that other areas aren’t prone to severe thunderstorms. Some of the largest tornado outbreaks in history have actually occurred well east of traditional Tornado Alley (e.g., 21 January 1999 and 3-4 April 1974).

Below is the raw graphic, with now Kernel Density Estimation applied. Because the edges are not smooth line, rather they are intersections from thousands of polygons, I did not plot any black line contours.

All SPC Moderate and High Risk Climatologies

Posted by pmarshwx on 3 Feb 2011 | 0 comments

After posting the climatology of where the first moderate and high risks occur, I’ve received a couple of requests for additional graphics. One that was extremely easy to produce, and also one most frequently requested, is a climatology of moderate and high risks. Using the same Kernel Density Estimation technique described in the original post, I’ve calculated the number of moderate and high risk outlooks an area might expect during a given year, based on data from 1990 through 2008.

Please note that each day consists of multiple outlooks and so it is possible for a grid point to receive multiple “hits” for being located in a Moderate or High risk outlook on the same day. In other words, if your location is located within 2 high risk outlooks, this does not guarantee that you will have two days of high risks. The location might simply be contained within a high risk from two separate outlooks issued for the same day.

For moderate risks, central Oklahoma appears to be the clear winner with nearly 30 moderate risks averaged per year. Increase probabilities extend both north and east from here.

As expected, the average number of high risks per years is considerably less than the average number of moderate risks. (In fact, I had to change the color scale!) Northeast Kansas, northern Missouri, and west-central Illinois are the most likely areas to experience a high risk in a given year with slightly more than 2 expected. A minor axis of increased probability extends southward from the eastern edges of this highest probability band, reaching portions of eastern Arkansas and far northern Mississippi.

Notice how in both of these climatologies, the maximum probabilities are centered in the central United States — east of the Rocky Mountains and west of the Appalachian Mountains. We’ll leave discussion as to why this is for another blog post.

AOTW: Storm Prediction Center Moderate & High Risks

Posted by pmarshwx on 3 Feb 2011 | 11 comments

Although most of the country is still digging out and cleaning up from a historic winter storm, commonly referred to as the “Groundhog’s Day Storm”, I decided to go ahead and answer the Question of the Week regarding Storm Prediction Center (SPC) Moderate and High Risks.

Before I answer the questions, first let me explain how I arrived at the answers…

I took the polygon outlining the first Moderate [High] risk during a given year, regardless of time of issuance. This means I treated a 12 UTC issuance the same as a 1630 UTC issuance and 2000 UTC issuance. I took the risk polygon, placed it on a 4km grid (specifically grid number 240), and activate all grid points that fell inside the risk polygon. This left me with a grid of 1′s (inside outlook) and 0′s (outside outlook). I created a grid for each year and then summed all the grids together. This gave me a grid containing the number of times each grid point was within the first Moderate [High] risk of the year. I then divided each grid point by the number of years I was examining. This left me with the probability that a grid point would be contained in the first Moderate [High] risk of the year.

The resulting plots were extremely blocky because of the straight edges of the outlook polygons and the relatively small size of the resulting sample size (19 years for Moderate risks and 18 years* for High risks). In an attempt to better retrieve the underlying probability distribution, I employed a technique called Kernel Density Estimation. I used a gaussian kernel with a bandwidth of 120 kilometers (per Brooks et al. 1998), and extended the resulting distributions out to 5 standard deviations.

To determine when the first Moderate [High] risk is typically issued, I converted the day and month of the first Moderate [High] risk into the corresponding day of year (e.g., 1 July is the 182 day of the year). I then took the mean day of the year and converted it back to day and month. This means that the average first issuance of a Moderate [High] risk may not fall on a day that actually had a Moderate [High] risk issued.

For this analysis I only evaluated the years between 1990 and 2008. This is because in the early years of the National Severe Storms Forecast Center (NSSFC) / Storm Prediction Center categorical risks had a different meaning than they do now. Thus, I tried to limit the analysis to relatively current definitions.

So, here are this week’s answers:

In the graphics below, the colored contour intervals are every 1% and the thin black contours are every 5%.

• Climatologically, where is the most likely location to experience the first SPC Moderate Risk of a given year?
There are actually two areas when using the raw probabilities: southwest Arkansas and southern Alabama. However, when applying the kernel density estimator, the area in southern Alabama has a slightly higher probability. As for when this first Moderate risk is issued, the average first issuance is 28 January.
• What about High Risk?
There is actually more variability (both spatially and temporally) in the issuance of the first High risk than the first Moderate risk. This is most likely the result of the greater variability in time of year, and thus greater variability of the large scale pattern. High risks have been issued as early as 21 January (1999) and as late as 1 July (1997). [Technically it has been issued as late as never, because 2000 did not have a single High risk issued.] For early High risks, the favored location is in the southeast United States where the ingredients that favor severe thunderstorms is more often confined to the Gulf Coast areas. Later in the year, more northern locations are favored as the cyclone track, and ingredients for severe thunderstorms, lifts northward. With this said, the first High Risk is typically issued across portions of southeast Arkansas and northern Mississippi between 27-28 March.