The US is the most active nation in the world for severe weather. Each year brings reports of about 1300 tornadoes, 6200 episodes of severe hail, and 13 000 severe wind events.
It’s a meteorologist’s job to integrate observations, model predictions, and his or her own experience to generate forecasts that keep people safe from extreme weather. That’s becoming an increasingly difficult task. Over the past few decades, the evolution of numerical weather prediction (NWP) models has led to a huge increase in the amount of information available to forecasters. Although increased data can provide more insight, meteorologists cannot issue forecasts in a timely manner if they consider every piece of available information.
Researchers at the National Severe Storms Laboratory (NSSL), the University of Oklahoma, and the Storm Prediction Center (SPC) are attempting to control the fire hose of model information by distilling several model-produced variables, particularly the strength of rotating updrafts, into easy-to-interpret severe weather probabilities. Those probabilities can be considered an NWP “first guess” for forecasters to build upon, using their experience and current observations. Perfecting such a data-distilling technique would enable meteorologists to issue rapid, decisive forecasts that protect lives and property.
The forecasting of severe weather—defined as a tornado, hail greater than an inch in diameter, or wind greater than 50 knots—is a complex and evolving field. Forecasters at the SPC create daily probabilistic Convective Outlooks for the contiguous US for each of the next eight days. The outlook for the current day differentiates hazard types, generating individual probabilities for hail, wind, and tornadoes (see figure 1). Often, forecasters will use an ensemble of NWP models, which provide a range of outcomes and illustrate a weather system’s variability, to help generate those forecasts. Ensembles can comprise the same NWP model with different starting conditions, for example, or multiple NWP models with identical starting conditions.
Figure 1. The Day 1 Convective Outlook on 27 April 2011, the day of a historic tornado outbreak over the southeastern US: (a) the overall category of risk, (b) the probability of a tornado, (c) the probability of severe wind, and (d) the probability of severe hail. Pink and purple designate the highest level of risk. Credit: National Oceanic and Atmospheric Administration
Forecasters are increasingly using convection-allowing models, which typically have horizontal grid spacing of 3–4 km, small enough to represent features of organized thunderstorms. Coarser models cannot explicitly simulate thunderstorms, so they rely on relatively simple parameterization schemes that do not resolve storms. The high resolution of convection-allowing models provides significant insight into the timing and storm mode. Knowing whether to expect individual supercell storms (which produce nearly all of the strongest tornadoes and severe hail events in the nation) or large conglomerates of storms, such as squall lines (which commonly produce damaging winds), allows forecasters to anticipate particular storm hazards.
Meteorologists can also use ensembles of convection-allowing NWP models to estimate the uncertainty inherent in a given forecast scenario. If models produce a wide range of solutions, the forecaster knows that there is greater uncertainty than if all of the ensemble members produce similar storm evolution. The information can then be combined with observations of the current atmospheric state and with forecasters’ experience from previous, similar scenarios to produce a well-informed forecast.
My research distills output from the 10-member NSSL–WRF (Weather Research and Forecasting) ensemble into tornado probabilities that meteorologists can use as a first guess for their forecasts (see figure 2).
Figure 2. The probability of a tornado within 25 miles of a point, generated using the NSSL–WRF ensemble for 19 May 2015. Tornado reports are overlaid as inverted triangles.
The probabilities are based on a proxy for severe thunderstorms called updraft helicity, which integrates the product of vorticity and updraft speed over a selected atmospheric layer. The presence of large updraft helicity indicates a rotating updraft, or mesocyclone. Persistent mesocyclones occur in supercells, which produce the most reports of severe hail, wind, and tornadoes of any storm mode.
My colleagues and I took updraft helicities computed from the NSSL–WRF ensemble and used them to generate probabilities. The maps in figure 3, which are similar to SPC Convective Outlooks, display those probabilities using different levels of statistical smoothing. Then we put our maps to the test, hoping to show that they effectively discern tornadic events from nontornadic events.
Figure 3. These maps use the same data as the map in Figure 2 but use different levels of statistical smoothing. The top map has a small sigma of 20 km; the bottom, a larger sigma of 200 km.
Our results revealed that, of the updraft helicity thresholds tested, 75 m2s-2 maximizes the ensemble reliability—it produces the best match between the forecast and observed probability. Furthermore, updraft helicity ≥ 75 m2s-2 yields plots with similar smoothness as Convective Outlooks and generates a high area under the receiver operating characteristic (ROC) curve (figure 4). The ROC curve is constructed by plotting the probability of detection (number of “yes” forecasts of a binary yes/no event that actually occurred divided by the total number of occurrences of the event) versus the probability of false detection (number of “yes” forecasts of a binary yes/no event that did not occur divided by the total number of “yes” forecasts of the event) over a range of probability thresholds. A perfect forecast produces a score of 1, and a random forecast produces a score of 0.5.
Although the area under the ROC curve was higher for some of the smaller updraft helicity thresholds, the ensemble reliabilities at those lower thresholds indicated vast over-forecasting. In other words, using too low an updraft helicity threshold to produce probabilities tended to create overly high and expansive probabilities across the US. That’s not very helpful to a forecaster trying to determine which regions are more vulnerable to tornadoes (and therefore warrant greater attention) on a given day.
Figure 4. An example receiver operating characteristic (ROC) curve for probabilities aggregated over spring 2014–15 using five different updraft helicity (UH) thresholds (colored curves). The dashed line illustrates a random forecast.
Although updraft helicity is a good proxy for supercells, predicting tornadoes specifically requires more information. Supercells frequently produce all types of severe weather, and distinguishing the likelihood of each hazard on a given day is still a significant forecast challenge. The incorporation of other environmental information beyond updraft helicity can greatly improve probabilities.
My research has tested environmental parameters that help isolate the tornado threat from hail and wind threats. The lifted condensation level (LCL) height, or the height of the cloud base, reflects the amount of moisture over the lowest 1–3 km of the atmosphere. If abundant moisture is present, cloud bases will be low and tornadoes will be more likely. We also looked at the ratio of the surface-based convective available potential energy (CAPE) to the most unstable CAPE. This measure indicates how much of the storm’s buoyancy-generated potential energy is located close to the ground. If a storm is not drawing most of its inflow from near the surface, tornadoes are much less likely.
Our final environmental parameter, the Significant Tornado Parameter (STP), is a composite of five other measures and was developed by SPC forecasters. The STP is a function of the LCL height, the mean parcel CAPE from the lowest 100 mb of atmosphere (another measure of the storm’s potential energy), the convective inhibition (the amount of energy required to overcome negative buoyancy near the ground), the 0–6 km wind shear (the difference in wind speeds between the surface and 6 km above ground), and the 0–3 km storm-relative helicity (the amount of environmental vorticity available to be ingested into the storm). Together, these parameters describe the favorability of a given environment for tornadoes.
My colleagues and I evaluated the success of our probability maps for the months of April, May, and June in 2014 and 2015. In addition, we received feedback from participants in the 2015 Spring Forecasting Experiment (SFE2015) in the Hazardous Weather Testbed, a collaboration between the research and forecasting community to test and produce cutting-edge methods of forecasting severe weather. The experiment lasted five weeks in May and June 2015.
Figure 5. An example receiver operating characteristic (ROC) curve, also derived from 2014–15 data, using an updraft helicity (UH) threshold of 75 m2s-2. The four colored lines represent four different probability sets, which incorporate different environmental information. The dashed line illustrates a random forecast.
Statistically, all sets of probabilities performed well. The addition of the environmental information resulted in increased reliability but a decrease in the area under the ROC curve (Figure 5). Subsequent analysis concluded that the decrease was caused by the more restricted probabilities occasionally missing events, rather than by the large span of area where no event was forecast and no event occurred. However, ROC areas were still quite high, with values of about 0.85–0.90 for the updraft helicity threshold of 75 m2s-2. We are confident that our forecasts can distinguish well between areas of tornado occurrence and nonoccurrence.
Feedback from SFE2015 scientists was extremely valuable. Participants noted that the probabilities were more accurate on days with strong forcing for severe weather than on marginal days, when the large-scale weather patterns were not as conducive to severe weather. Some participants thought that the magnitude of the probabilities was too high compared with Convective Outlooks.
The combination of statistical results and feedback from SFE2015 supports using updraft helicity as a proxy for severe weather, along with incorporating environmental information to improve the tornado forecasts. The feedback is helping my colleagues and I develop new methods of distilling the information given by high-resolution ensembles into rapidly digestible probabilities for forecasters. Through our collaboration with the SPC, we hope to generate products that the forecasting community can use to protect Americans from the dangers of severe weather.
Burkely Gallo is a PhD student at the University of Oklahoma. She uses convection-allowing ensembles to create probabilistic tornado forecasts and evaluates the resulting forecasts with her adviser, Adam Clark.
To learn more about the tornado research discussed here, see B. T. Gallo, A. J. Clark, S. R. Dembek, “Forecasting tornadoes using convection-permitting ensembles,” Weather and Forecasting 31, 273 (2016).
