Rethinking how we conceptualize tornado days

This is the first of a multi-part series that introduces a technique developed by the author to classify tornado events and tornado days. We will follow up with additional posts that demonstrate its uses.

Let me introduce myself — I’m a storm chaser. I live in… California.

Chasing tornadoes while living half a country away from Tornado Alley is a challenge, one that many people deal with each year. Fly out for a chase setup? Should I take the risk of buying an expensive plane ticket given the setup might bust? Chasecation in May? What if the tornadoes don’t show up?

As an out-of-region chaser, improving my chances of catching tornadoes is a key motivation. A natural starting point for such hopes is attempting to find the days where tornadoes are most likely in any given year.

But what days should be considered “tornado days?” All such days are not created equal.

Introduction

Much of the current literature focuses on tornado counts or tornado intensities. I think there are a few flaws with that kind of analysis. Below, in the introduction to this method, I’ll give a few examples. For instance, let’s say I mark any day with at least 10 tornadoes as a tornado day. Here are some potential problem cases:

• 10 one-minute tornado touchdowns get treated the same as 10 one-hour long-track tornadoes.
• 10 tornadoes spread out between 10 different states get treated the same as 10 tornadoes spawned from a single cyclic supercell.
• 10 tornadoes spread out over 10 hours get treated the same as 10 tornadoes spawned within 30 minutes.

These kinds of cases are especially important to consider if I’m trying to classify tornado days as a storm chaser. As chasers, we tend to be less concerned about tornado strength or counts, and more about how many minutes of tornado action we can get within a given radius from a certain location. That idea is going to play a central role in this method.

The method

Let’s start here with the idea of tornado-minutes. A naive approach might be to simply count total tornado minutes, but that doesn’t solve the case of tornado minutes being spread out spatially vs. focused in one area. The distribution and density of tornado minutes matters.

My method starts by taking an input of tornado segments and separating each segment into discrete points, with each point representing a position where a tornado has existed for one minute.

Example: an EF5 wedge tornado travels along Interstate 44 from Oklahoma City to Tulsa between 5 and 6 p.m. This setup would separate that tornado into 60 points spaced equidistant between Oklahoma City and Tulsa. Each point represents the interpolated position of the tornado at 5:01 p.m., 5:02 p.m., and up to 5:59 p.m., before the tornado lifts at 6 p.m. In the rest of this text, I’m going to refer to these points as “tornado-minute points”. (Note that in this method, “brief touchdowns” are counted as one minute.)

The next step involves clustering the tornado-minute points using the popular data-mining algorithm known as DBSCAN. DBSCAN finds clusters of densely-packed points from a larger dataset of points. Those points that do not fall within a cluster are labeled as outliers or noise. The parameters that go into DBSCAN are:

• eps: the maximum distance between points for the algorithm to consider to be densely-packed.
• min_samples: the minimum number of densely-packed points for the algorithm to label a cluster.

DBSCAN is primarily designed for spatial data, but it can be tweaked to incorporate geospatial-temporal data. In this variant eps is separated into distance (eps_km) and temporal (eps_min) components, and two points are densely-packed only if they are within both distance and time away from each other. Here are some examples of this in action in the context of tornado-minute points. Let’s say min_samples is 10, eps_km is 10, and eps_min is 10.

• A 5-minute tornado with no other tornadoes on the day is considered noise.
• A 5-minute tornado in Texas and a 5-minute tornado in Nebraska would both be considered as noise.
• A 10-minute tornado with no other tornadoes on the day is considered a “cluster” (because there would be 10 tornado-minute points from that tornado alone and that is sufficient for min_samples).
• A 10-minute tornado in Texas and a 10-minute tornado in Nebraska would be considered two separate clusters.
• Say a cyclic supercell spawns a 5-minute tornado, then another 5-minute tornado six kilometers to the northeast and two minutes after the end of the first tornado. Those tornadoes would be considered one cluster.
• In the previous example, if the last tornado only lasted one minute, it would not be considered a cluster. If the last tornado occurred 12 kilometers to the northeast of the end of the first tornado, all tornadoes in the series would belong to noise. Alternatively if the last tornado happened 15 minutes after the end of the first tornado, all points would also belong to noise.

From these examples, we can conclude that, given we tune the eps and min_samples parameters properly:

• Clusters in different areas of the country are tallied separately.
• Clusters in similar areas but occur hours apart from each other are also tallied separately.
• Long-track tornadoes can be their only “cluster,” if they are long-lived enough to surpass the min_samples threshold.
• Brief touchdowns tend to be labeled as noise, which is nice because brief, isolated touchdowns tend to be outliers from an atmospheric standpoint as well. But if a bunch of brief touchdowns occur together, say from a single cyclic tornado supercell, the entire group may be significant enough even if any one touchdown is otherwise not noteworthy. And the DBSCAN method will capture that as a cluster.

From there, it is straightforward to assert that a tornado day is one where there is at least one cluster found from all the points obtained from tornadoes on that day.

Finally, the size of the cluster is simply the number of tornado-minute points, which equals the sum of each tornado’s longevity for all the tornadoes in the cluster. So three 10-minute tornadoes in a cluster would yield a cluster size of 30 minutes.

Drawbacks

The clustering method requires complete spatial and temporal data (start/end * coordinates/datetime) to perform the breakdown of tornado segments into tornado-minute points. As of this writing, the Stom Prediction Center (SPC) database does not contain tornado end times, and some are missing end coordinates as well. The Storm Events database does contain the relevant information, but only back to 1996. Thus, the clusterable dataset of tornadoes only goes back 22 years. Nevertheless, this sample size is sufficient for purposes here.

Future iterations of this technique might involve using algorithms to back-fill temporal data before 1996 to expand the sample size.

Finally, the algorithm isn’t perfect and will produce less-than-ideal clusters when a supercell goes long periods without producing a tornado. An example of this will be shown in the next section.

Selected results

Enough talk, what does this clustering look like with real tornado outbreaks? The maps below are generated with the following parameters: eps_km=80, eps_min=60, min_samples=15.

May 3, 1999

(The green cluster isn’t shown here because it’s outside the boundaries of this map.) Note here the Oklahoma City area tornadoes, the Wichita-Haysville area tornadoes, and the Tulsa-Sapulpa area tornadoes are split accordingly.

May 23, 2008

April 27, 2011

The clustering algorithm separates the three stages of the outbreak in Alabama very well: the morning MCS stage is in purple, the midday supercells near Huntsville are in dark blue, and the afternoon violent supercells are in light blue.

May 24, 2011

Feb 5, 2008

This is the example of the “less-than-ideal” case mentioned in the drawbacks section. The Jackson, AL EF4 tornado is awkwardly left out of the yellow cluster since there aren’t many tornado-minute points nearby. That particular supercell went through long phases of not spawning a tornado, which is reflected in the sparse tornado tracks nearby. Additionally, tornadoes associated with the MCS in KY are accidentally added into the magenta cluster, since they are so close in space and time to the supercell tornado northeast of Nashville.

May 4, 2003

The Kansas City area tornadoes, the devastating long-track tornadoes that affected southeast KS and southwest MO, and the tornadoes that included the Jackson, TN EF4 are highlighted nicely here.

April 14, 2012

Clustering can untangle the progression of an outbreak where multiple rounds of tornadoes track over similar areas, and the April 2012 outbreak is a prime example of this.

May 4-5, 2007

This example illustrates clustering on multi-day events. Note the overlapping tracks spanning two days in western Kansas.

Conclusion

This clustering technique offers some positives compared to the typical tornado day classifications. There are also some areas which might be better worked out in the future. In additional examinations of this technique, I will look at how it can be applied to chasing more specifically, and then later we’ll examine an outbreak in depth.