Self-Training Software Aims to Improve Doppler Radar’s Tornado Vision

A type of computer program that learns from its own mistakes could be the next big thing in tornado detection using the National Weather Service’s national network of Doppler radars (NEXRAD). The software is now being tested and could be deployed using existing NEXRAD equipment. If it passes muster, the new software could substantially improve the ability of NEXRAD to detect tornadoes, paving the way for potential warning improvements in the 2020s.

The software, which calls on a technique dubbed machine learning, was discussed in late October at the AMS Conference on Severe Local Storms in Stowe, Vermont. Sponsored every two years by the American Meteorological Society, the conference draws several hundred severe-weather experts from around the world.

“Based on feedback from NWS forecasters, there is much room for improvement in tornado detection algorithms,” says research associate Matt Mahalik (University of Oklahoma/CIMMS and NOAA National Severe Storms Laboratory). “We think that our technique has the potential to make significant strides.” Mahalik presented a conference poster on the work, which is being supported by the NWS Radar Operations Center.

Since their deployment nationwide in the 1990s, the NEXRAD radars have been equipped with a Tornado Detection Algorithm (TDA) that spots potential tornadoes within severe thunderstorms and alerts forecasters. The TDA looks for such red flags as very strong gate-to-gate wind shear, or the contrast in winds between adjacent data points. Strong gate-to-gate shear is often (but not always) an indicator of imminent tornado development.

The TDA is among a set of innovations that led to dramatic improvements in tornado-warning lead times just after NEXRAD was installed. The TDA is not perfect, though. An early study published in 1998 showed that the TDA missed about 57% of tornadoes that occured within radar range. Moreover, roughly half of the TDA alerts for potential tornadoes were false alarms. The new software is designed to improve those statistics.

Figure 1. This early example of the prototype TDA now being developed shows results for a thunderstorm located on the Nebraska-Wyoming border on June 12, 2017. Clockwise, from top left, are 0.5-degree reflectivity (the intensity of rain and hail), velocity (wind speed toward or away from the radar), DivShear (a proxy for divergence), and AzShear (a proxy for rotation). The yellow marker labeled "1" indicates a tornado has been detected. At bottom, there is an output table with attributes of all detections in the entire scan (the one in the image is the second row in the column; detection "0" is not shown). Image credit: Courtesy Matt Mahalik.

Figure 2. This tornado, near Carpenter in far southeast Wyoming, was among three EF2 tornadoes spawned in eastern Wyoming and western Nebraska on June 12, 2017. The tornado developed from the thunderstorm analyzed in Figure 1. Image credit: Max Olson, via NWS/Cheyenne.

Nurturing trees in a random forest

Machine learning is well suited for clarifying processes that involve a number of simultaneous variables, where it’s unclear exactly which ones are most important at particular points in time. Even an expert’s brain can be stymied trying to wrap itself around all the interactions. With machine learning, an algorithm can analyze all of the potential interactions, sniff out the ones that are most important through trial and error, and gradually improve its skill. The myriad permutations of variables are represented in a collection of decision trees known as a random forest—in this case, a forest with 500 trees.

Large amounts of solid data are crucial for an algorithm to “teach itself” through machine learning. “How well it works depends entirely on the quantity, quality, and characteristics of the data used to develop it,” said Kim Elmore, another CIMMS researcher involved with the project.

The new TDA is being developed with more than 17,000 severe weather events from 2014 and 2015 catalogued by the NOAA/NWS Storm Prediction Center. These events are being compared with about 7000 NEXRAD radar scans.

Figure 3. Locations of the severe storm reports from 2014 and 2015 used for initial testing and training of the random forest driving the new tornado detection algorithm. Image credit: Courtesy Matt Mahalik.

Each radar scan of a particular event can be used to either train the algorithm or to test its newfound skill. Ideally, the training and test events are totally independent. That’s not always been possible up to now, but a fresh dataset from 2016 will allow for exhaustive testing with independent data. “When we test the algorithm on 2016 data, we will have a true idea of how well it will work in practice,” said Elmore.

The researchers are optimistic that the new TDA will perform well compared to its predecessor. “If it doesn't [do well], then there is no reason to implement it and replace the current TDA,” Elmore said. “But, if it does as we expect, it will be a marked improvement over the current TDA. What we aim for is something to ease the workload of the warning coordination meteorologist and help them identify the truly threatening characteristics from those that aren't so threatening.”

Another big step will be putting the new algorithm through its paces in 2019 at NOAA’s Hazardous Weather Testbed. Each year, a major spring experiment brings dozens of researchers and forecasters to the testbed, based at the National Weather Center in Norman. The experiment, timed for severe weather season, gives forecasters a chance to issue non-public warnings in real time based on experimental products. In turn, this helps gauge how useful the products would be to NWS meteorologists who have to issue actual public warnings in a real-world severe weather outbreak.

“We don't expect the algorithm to supplant the meteorologist issuing warnings, but we demand that it help improve their confidence in the decisions they make to issue or not issue a warning,” said Elmore. Mahalik added: “A skillful, automated tornado detection algorithm can give forecasters greater confidence when making a warning decision, without having to do much manual analysis. During tornado warning operations, every second counts.”

If next year’s testing and subsequent steps all go smoothly, the new algorithm may become operational by 2021 at the earliest, Mahalik said.

Figure 4. An experimental technique allowed forecasters to warn residents more than an hour ahead of time that a tornado warning might be issued. Image credit: Iowa State University/Iowa Environmental Mesonet, via weather.com.

A brave new world of tornado guidance

While the new TDA is expected to provide more solid evidence of an incipient tornado, there are parallel NOAA efforts to enhance severe weather warnings in other ways. One of these, called Warn-on Forecast, is designed to provide more in the way of advance notice for severe weather when signs of trouble are especially strong. In some cases, tornado warnings may eventually be issued an hour or more before a tornado actually forms.

The prototype Warn-on Forecast system had its first big success on May 16, 2017, as reported by weather.com. That evening, a special weather statement alerted residents of three Oklahoma counties that approaching thunderstorms were likely to result in tornado warnings. The first of several twisters developed roughly 90 minutes later in the alerted region.

As part of Warn-on Forecast, probabilistic values will be assigned to various points in the expected path, showing which areas are most at risk. Much work remains to be done in this long-term project, called Forecasting a Continuum of Environmental Threats (FACETs). Social scientists will be an integral part of FACETs, evaluating how a variety of users process and act on probabilistic guidance, which is challenging to convey in a succinct, timely fashion.

Like the current TDA, the new TDA is designed to serve mainly as a yes/no flag for the presence of a tornado. However, Mahalik and colleagues are planning to translate its output into probabilistic terms, which will help it feed into FACETs and the Warn-on Forecast paradigm. Together, all of these efforts stand to provide better tornado warnings on time scales ranging from a few minutes to an hour or more.