ProbSevere on some early storms near ARX

We noticed a relatively high ProbSeverev3 (53%) on a rather innocuous looking storm (MESH around 0.5”) around 2030z. This was higher than the v2 value of 36%. The individual probs were relatively evenly weighted at lower values near 30%.

(clockwise from top left) MRMS 18dbz echo top, MESH, reflectivity and ProbSevere (storm in center), and low-level MRMS azimuthal shear.
(clockwise from top left) GLM FED, GLM MFA, reflectivity and ProbSevere (storm in center-right), and GLM TOE.

GLM FED was unimpressive, though it’s unclear how much of this is related to lower detection efficiencies in this area. ENI total lightning was halfway decent. High DCAPE values and other environmental parameters may have been sending the ProbSevere v3 higher.

Timeseries for the storm of interest.

Another storm further to the west over SE MN had slightly lower MESH (.39”) but in this case PSv2 was higher at 48% vs Psv3 at 23%.

– Barry Allen

Tags: None

ProbSevere Time Series

I found the ProbSevere time series helpful today as we “triaged” storms and tried to identify storms that may become severe. While the capping inversion stayed strong and therefore prevented storms from becoming severe, it was great to see storms follow a similar intensification process identified by the Prob time series time graph. Most storms intensified in a similar fashion but capped out when ProbSevere reached ~40%. After 40% storms would remained steady state and then gradually weaken. Noticing these trends and seeing them plotted visually helped us pick up on the trends. Any storm that deviated from this and grew upscale faster would be easy to identify on the ProbSevere time series graphs. We knew what the “norm” was for storms in this capped environment because of the time series graphs. We surmised that once the cap broke (which would be after the experiment ended), we could quickly ascertain when storm would finally be able to grow upscale by looking at their respective time series.

– Fear the Shear

Tags: None

ProbSevere and GLM with supercell moving offshore the South Carolina coast

A thunderstorm located east of Charleston appeared to have some supercell characteristics as it moved south-southeastward towards the coast, with a kidney bean shape in reflectivity and a weak mid-level mesocyclone, as well as some deviant motion from the northwesterly flow. As it was over land it appeared to be strong but sub-severe, and maintained generally consistent 20 to 35 percent ProbSevere in v3. ProbSevere v3 seemed more consistent overall, with v2 jumping up and down more often, dropping down into the single digits at times. ProbSevere v3 did jump down below 20 percent briefly when GLM FED really dropped down. But the consistent lower-end probabilities at least indicated that this was a storm to be watched relative to the lower v2, and this may have at least allowed lead time on a low-end special marine warning before it moved offshore and strengthened.

The timeseries is somewhat useful if you just have one storm to look at, but with multiple storms I would probably just look at the loop in the ProbSevere plan view instead.

After it moved offshore, GLM FED increased, slightly in advance of a jump in MESH and associated jump in ProbSevere v3. ProbSevere v2 jumped ahead of v3 in probabilities as often occurs, though at that time MESH around 0.9 inches may have warranted the more conservative ~50-60% v3 approach. Later MESH jumped up to around 1.3 inches, and ProbSevere v3 jumped above 70 percent at this time as well.

Clockwise from top left: MRMS 18dbz echo top, MESH, reflectivity/ProbSevere/low-level azimuthal shear at 2130z.
Clockwise from top left: GLM FED, minimum flash area, reflectivity/ProbSevere, total optical energy at 2130z.

– Barry Allen

Tags: None

ProbSevere v3 for Low End Storms

In the training we reviewed for ProbSevere v3, it was frequently mentioned that overall the values we would see in v3 would tend to be a bit lower than in v2, which is what you’d expect in a better calibrated model. (Given what we know about storms and their potential severity, v2 often seemed a little too sure of itself, right?)

However, in looking at the relatively puny convection over Charleston’s CWA today, I’m seeing a lot of the opposite. That is, instances where v3 is noticeably higher than v2. As I think about it, I believe we’re seeing more evidence here of a better calibrated model, as it’s ultimately showing less certainty than v2 about storms not being severe which seems to make sense intuitively. We’ve all had our share of days where a storm one thinks is relatively benign ends up overperforming and though I don’t have any actual evidence to prove this, it appears that ProbSevere is, in its way, taking that into account. Ultimately, this is giving me a bit more confidence in its output.

An example of a somewhat low end thunderstorm where v3’s percentage (29%) is noticeably higher than v2’s (2%). V2 seems a bit too certain that this storm is not severe. As forecasters, we know better.

– Earl Grey Tea

Tags: None

ProbSevere in Jacksonville

The 0.5 degree SRM from JAX shows a concentrated area of wind within the cluster of storms in St. John County, FL with radar estimates of the wind near 50kts. Given the lightning jump illustrated by FED values quickly rising to ~170 flashes per 5 min, the storm is intensifying.

However, ProbSevere values are rather low with version 3 showing only 22% and version 2 showing 33%. ProbWind surprisingly was even lower with only 19%. This is a reflection of the lack of base radar data involved in the ProbSevere and ProbWind algorithms. Especially for ProbWind, base radar velocity data needs to be included in ProbWind for this product to be useful in identifying wind producing severe thunderstorms.

It would be useful to integrate base radar data from multiple single radars and combine these values into one algorithm. It may be useful to identify notable/sharp changes within velocity data between pixels which could help in picking out downdrafts.

– Fear the Shear

Tags: None

Storms in northern and central Florida


On the afternoon of June 15, 2 different convective regimes were noted across Florida, with different GLM lightning characteristics. A cold front was sinking southward towards the Florida panhandle, with convection developing along the Gulf sea breeze along the FL panhandle. Convection of more uncertain forcing developed in Central Florida.

Convection along the FL panhandle had higher MLCAPE and DCAPE due to mid-level drier air and steeper lapse rates, with somewhat lower PWATs. SBCAPE in excess of 5000 J/kg and MLCAPE in excess of 3000 J/kg was unusually high for this region. Convection in central FL was in a more tropical air mass, with PWATs at or above 2 inches and more saturated profiles. Convection in the FL panhandle developed in an area with very high microburst composite parameter values, indicating conditions very favorable for microbursts and localized damaging wind gusts.

12z TAE sounding:

15z XMR sounding:


19z DCAPE:

19z PWATs:

Microburst composite parameter 19z:

The FL panhandle convection was more intense on radar and also had higher flash extent densities. It also tended to have lower minimum flash areas, centered on locally strong updrafts. Notable hail cores were observed aloft, and melting of these hailstones caused strong downdrafts and damaging wind reports, and in a few instances quarter size hail made it to the ground, with one instance of golf ball size hail..


The central FL convection was weaker and had also been going on for longer, so there was some convective debris stratiform precipitation with larger minimum flash areas. Flash extent densities were lower than in the FL panhandle. There were still areas of lower minimum flash area centered on the updrafts.

The GLM flash points were very useful and lined up with the NLDN and ENTLN strikes and flashes. The parallax correction was especially useful for DSS purposes as partners often request notification on lightning strikes within a particular radius on the order of 8 to 20 miles, so an accurate location is important. At first glance there were much less flash points but this appeared to be due to the data only being 1 minute data without having the 5 minute accumulation that the NLDN and ENTLN offers. Having this similar 5 minute accumulation would be imperative for using the GLM flash points in operations. The sampled metadata for the flash points appeared less useful operationally. The flash area would be more of interest than the duration, but with a large number of flash points some sort of graphical depiction would be needed, and flash extent density seems to serve this purpose.


One interesting thing that was noted was v3 had much lower ProbHail than v2, while still having decent ProbSevere (mainly wind-driven values). We speculated that this was due to some of the machine learning based on environment and climatology, since severe hail would be less likely this time of the year with higher freezing levels/hot surface temperatures causing melting. However, in this case a golf ball size hail LSR was issued at 19:59z (report time appeared to be incorrect) for 2 ENE Saunders in Bay County, FL. This was comparable to MRMS MESH which maxed out at 1.89”.

On the technical side, I did want to note that typically I have sampling turned off in AWIPS, but then double-click on something that I want to sample. Since the ProbSevere timeseries plugin is also opened by double-clicking on the object, sometimes when I meant to double-click to sample the ProbSevere values I accidentally ended up opening a time series. And then I would double-click outside the ProbSevere area to sample something else or turn off sampling and I would get a black banner. Perhaps the timeseries doubleclick function could be turned on and off by making the ProbSevere product editable or not editable in the legend.


Gridded NUCAPS and individual NUCAPS soundings at 1840z showed steeper 700-500mb lapse rates than what was shown on the SPC mesoanalysis and some of the morning soundings, in areas away from convection. It’s hard to say which one was right, but the hail cores observed do seem more consistent with 700-500 mb lapse rates of near 7 C/km or greater. (Note that it would be useful to have contours to go with the images on the gridded NUCAPS plots.)

NOAA-20 sounding availability and example sounding (1823z)

1840z gridded NUCAPS 700-500mb lapse rate:

18z SPC mesoanalysis 700-500mb lapse rate:

NUCAPS also indicated the more saturated profiles/weaker lapse rates in the central FL convective regime.

NUCAPS did indicate some of the higher CAPE values, but with missing data in much of the area of interest as convection had already initiated when the pass occurred.

– Barry Allen

Tags: None

Prob Severe V 3

See a significant improvement in the Prob Severe with V3.  Looking at a particular storm in the panhandle of FL shows a significant difference between the two versions.  The image below is the cell to the right of the screen and the new prob severe v3 time series along with a cross section from GRAnlysist.

What can’t be seen is the sample over the prob severe v3 outline which shows both the v2 and v3 output.  In this case, the v2 showed a 65% chance for severe hail but the v3 only showed 7% for severe hail.  Severe wind was 52% and 58% respectively.  Looking at the cross section and knowing the PW values are in the 1.7-1.8 range shows the main threat would be more of a wind/rain rather than hail.  This is a big improvement.  To be honest, the issue with over forecasting hail on the v2 is a big reason why I usually don’t use prob severe.  Seeing this change with V3, I am much more likely to be looking at it as it seems to be more refined and takes the climate, area and conditions into account before producing significant hail or wind values.  While I don’t think it was looking at the cross section to help make its decision, this cross section is a very good example of a wind risk over hail.


Tags: None

ProbSevere Time Series & Cell Merger

Two separate thunderstorms gradually merged to the WSW of New Bern between 1845-1900 UTC. As ProbSev polygons merged into 1, some of the probabilities took a small dip. Eventually the eastern storm cell intensified slightly, bringing ProbSev values back up.

– Guillermo

Tags: None

ProbSevere Data Readout/Sampling

The ProbSevere readout is useful with all the details shown, but when in 4-panel mode, it takes up a lot of space, sometimes with some data being omitted by the frame. Would it be possible to have the option of a simplified readout displaying only the top 2 line probabilities and then have the option of displaying the additional data? (Perhaps akin to the double left click on a ProbSev polygon for the time series chart)

– Guillermo

Tags: None

Optical Wind, GLM, and ProbSevere Use in Convective Environments in North Carolina and South Carolina

Looking at multiple levels of optical winds can be useful in analyzing the amount of wind shear over an area in near-real time. In this case, the tool shows limited wind shear, so one would expect storms to be a bit more short lived. Would it be possible to add wind shear fields directly into this tool for quicker analysis?

Optical winds for the ILM CWA on 6/15 at 18Z showing little difference in the winds between the 800-600mb and 600-400 mb levels.
4 panel of the GLM data at ILM around 19 UTC illustrates Flash Extent Density (top right), Minimum Flash Area (bottom left) and Total Optical Energy (bottom right). We adjusted the colormap of the minimum flash area so that we could identify the updrafts more easily since the minimum flash areas were under 100km^2 and the default map was set to cover images up to 2000km^2. This allowed us to identify which storms featured the strongest updrafts which when combined with data from the Flash Extent Density, we could watch for storms that were strengthening and thus posed a greater need for a warning.

Three Body Scattered Spike & ProbHail

Three body scattered spike is visible in the storm in the top right panel.
ProbHail shows values of ~65% when the three body scattered spike appears with MESH values over 2” supporting the likelihood of at least severe size hail in the discrete cell.

Watching the meteogram on this storm, we can see the probhail values jumped up to 65% over the last 15 minutes. It’s probably best to have ProbHail values of 60% or more last for a few volume scans because that suggests the residence time in the hail growth zone is long enough for hail to grow and become 1” in diameter or larger.

–Earl Grey Tea and Fear the Shear

Tags: None