ProbSevere products over the Southern Plains

May 3rd, 2021 |

The NOAA/CIMSS ProbSevere portfolio contains AI models for nowcasting convective weather. I’ll use Monday’s severe weather over the Southern Plains to highlight several of them.

A strong cold front spawned numerous severe-hail, wind, and tornado producing storms over Texas and Oklahoma, aided by very large values of convective available potential energy (CAPE; > 4000 J/kg).  You can see numerous storm reports in Figure 1.

210503_rpts Reports Graphic

Storm Prediction Center’s preliminary severe storm reports for May 3rd, 2021.

Probsevere version 2 (PSv2) is an operational set of models at NOAA, which predict the probability of severe hail, severe wind, and tornadoes, in the next 60 minutes. The models are storm-centric, and the models’ domain is the entire contiguous United States (CONUS).  These models use MRMS (radar), GOES (satellite), short-term NWP, and terrestrial-based lightning observations to generate probabilistic guidance of severe hazards. Figure 2 shows output from an experimental version (PSv3), which includes additional MRMS, GOES, and NWP fields as predictors in a machine learning model.

Figure 2: ProbSevere v3 contours (colored, around storms), MRMS MergedReflectivity, and NWS severe weather warnings (yellow and red boxes) for storms over the Southern Plains. The second outer contour around some storms is colored by the probability of tornado.


Another ProbSevere product is a convolutional neural network that uses GOES-R ABI and GLM images to detect regions of intense convection, and is often correlated with strong overshooting tops, “bubbly-like” texture in visible imagery, strong lightning cores, and the cold-U/above-anvil cirrus plume signature. The intense convection probability (ICP) can be run on the 1-minute mesoscale scans as well as 5-minute CONUS sector scans aboard the GOES satellites. The ICP does not require radar data, and may also be able to operate on data from satellites with similar intruments (e.g., Meteosat Third Generation). ICP output is being used as a predictor in the experimental ProbSevere v3.


Predicting when and where lightning will occur is also important for many users, such as mariners, aviators, and outdoor event managers. The probability of lightning model (PLTG) is also a convolutional neural network, using images of visible, near-infrared, and longwave-infrared channels to nowcast lightning occurrence in the next 60 minutes. The purple-to-orange shaded regions in the video below show GLM flash-extent density (i.e., flashes passing through a location).

Supercells in the Southeast

May 6th, 2020 |

A cold front with ample moisture and instability ahead of it spawned numerous strong storms in the Southeast U.S. yesterday; particularly one long-lived supercell in South Carolina. A convolutional neural network model (CNN) was deployed in realtime on the 1-min GOES-16 mesoscale sector imagery. The model produces an “Intense Convection Probability” (ICP). The inputs for the model are the GOES-16 ABI 0.64 µm reflectance, 10.3 µm brightness temperature, and GLM flash extent density. It was trained to identify “intense” convection as humans do, associating features with intense convection such as strong overshooting tops, thermal couplets (“cold-U/V”), above anvil cirrus plumes (AACP), and strong cores of total lightning.

The animation below shows the ICP contours overlaid ABI 0.64 µm + 10.3 µm sandwich imagery, annotated with preliminary severe storm reports.

The long-lived supercell in South Carolina exhibited AACP and cold-U features, and produced numerous severe wind and hail reports (up to the size of tennis balls). While the NOAA/CIMSS ProbSevere models handled this storm well, the ICP ramped up on a couple of severe storms in northern Georgia before ProbSevere did. ICP for these cells exceeded 90% 15-18 min before ProbWind reached 50%. The ICP may be able to provide additional lead time and confidence to ProbSevere guidance for certain storms, utilizing spectral and electrical information from geostationary satellites. Incorporating ICP into ProbSevere is an active area of current research.

ProbSevere storm contours and MRMS MergedReflectivity for storms in GA and SC. The main or “inner” ProbSevere contour is shaded by the probability of any severe weather, while the outer contour is shaded by the probability of tornado, which appeared when that value was at least 3%, in this example.

An accumulation of ProbSevere storm centroids (white to pink squares, 50% --> 100%), NWS severe weather warnings, and SPC severe local storm reports from 12Z on May 5th to 12Z on May 6th [click to enlarge]

An accumulation of ProbSevere storm centroids (white to pink squares, 50% –> 100%), NWS severe weather warnings, and SPC severe local storm reports from 12Z on May 5th to 12Z on May 6th [click to enlarge]

The probability of “intense convection” using geostationary satellite data

September 27th, 2019 |

Researchers from NOAA and UW-CIMSS have developed an experimental model that predicts the “probability of intense convection” inferred from GOES ABI and GLM fields. The model is a convolutional neural network, which carries the assumption that the inputs are images and have spatial context. It is a great tool for image classification.

GOES-16 ABI CH02 reflectance (a visible channel), ABI CH13 brightness temperature (an infrared window channel [IR]), and GLM flash extent density (FED; generated using glmtools),  were used as inputs to the model. The model learned important features that have been traditionally difficult or expensive to code into an algorithm, such as pronounced overshooting tops (OTs), enhanced-V features, thermal couplets, above-anvil cirrus plumes (AACPs), strong brightness temperature gradients, cloud-top divergence, and texture from visible reflectance.

It is hoped that such a model may be able to one day:

  • provide earlier notice of developing or decaying intense convection
  • provide guidance in regions with no weather radars
  • provide a quantitative way to leverage 1-min mesoscale scans
  • ultimately improve the accuracy and lead time of severe weather warnings

The model is very experimental and is not yet running in real-time. The remainder of this post catalogues some examples of the deployed model on select scenes.

The movies below use as a background the CH02-CH13 “sandwich” product, whereby cloud-top 11-µm brightness temperature and 0.64-µm reflectance can be seen in tandem. This generally helps observers see how changes in storm-top structure correlate with changes in 11-µm brightness temperature. A grid of “probability of intense convection” was generated for each scene with a moving 32×32 pixel window (each pixel = ~2 km), with the model generating one probability for each window. These probabilities were then contoured with the 25%, 50%, and 90% contours as blue, cyan, and magenta. Preliminary severe local storm reports from the SPC rough log are also plotted as circles.

The example below shows that the model handled two separate severe wind threats in Missouri, identifying cold cloud top regions in the IR that also looked “bubbly” from the visible channel. As the sun was setting, a cold front lit up with very intense convection from Oklahoma through Missouri. Again, the model did a decent job highlighting the strongest areas of convection which correlated well with severe local storm reports. It should also be noted that the model does not seem to have significantly degraded output when the visible channel is missing (after sunset).


The next example is at a higher satellite viewing angle in western Nebraska, western South Dakota, and eastern Wyoming. The model again does a good job highlighting the strongest areas of storms. It should be noted that not every identified region has severe reports and not every severe report has a probability of intense convection ? 25%, but that there is generally good correspondence between reports and the model probabilities nonetheless.


This example is from the Southeast U.S. in more of a low-shear “microburst” environment instead of a high-shear “supercell” environment. You can see that instead of predicting high probabilities for all of the convective storms, the model exhibits the highest probabilities for the storm clusters that at least subjectively look the strongest.


This next example from the Central Plains demonstrates the ability to discern decaying convection, as the first storm moves into Missouri and then quickly diminishes in appearance and in probability. It also demonstrates the model’s ability to pick out multiple threat areas within a large cloud mass at night.


This is an example using mesoscale scans. Despite not being trained with 1-min data, the model predictions still look very fluid and reasonable. This could be an excellent way for scientists and forecasters to leverage 1-min observations in a quantitative manner.


Another example using GOES-East 1-min mesoscale scans. The model generally picks out the strongest portions of a MCS in Illinois and Indiana.


At a very high viewing angle, the model predicts probabilities of ?90% for a storm in Arizona. The storm did not generate severe reports, but was warned on by the NWS multiple times.


The model is deployed during an early autumn severe weather outbreak.

This example shows the model deployed on 1-min mesoscale scans over very intense thunderstorms in Argentina. It demonstrates that the model is generally applicable to anywhere in the world where advanced imager and GLM-like observations are present.


Here, a model was trained without GLM data and deployed for an example in the Alaska Panhandle, where GLM data is not available. This storm prompted a severe thunderstorm warning from the Juneau, AK NWS office. Note the change in the values of the probability contours. The maximum probability for this storm was 36% at 02:43 UTC. In a relative sense, the intense convection probability product could still be useful in unconventional regions, such as Alaska.