IntenseStormNet in Kansas and Oklahoma

ProbSevere IntenseStormNet is not being evaluated at HWT this week, but it is a unique application of satellite data for severe weather. On the eve of the 2nd week of the 2024 HWT, severe storms were ravaging the Plains.

IntenseStormNet is an AI model that uses patterns in ABI and GLM image data to predict a probability of "intense" convection. The strongest probabilities are often correlated with strong overshooting tops, bubbly texture in the visible band, storm-top divergence, and lightning cores.

In the movie above, you can see several MCSs traversing the state, producing severe hail, wind, and tornado reports. One the convection becomes cold-pool driven, the probabilities often diminish, sometimes significantly (see eastern Kansas at the end of the period).

Meanwhile in Oklahoma, a monster supercell progressed steadily through the western half of the state, dropping tornadoes, hail, and producing severe wind gusts. One interesting aspect about the product was the drop in probabilities from about 01:40 - 02:10 UTC. This drop corresponded to a short gap in severe weather reports produced by the supercell. Visually, the main overshooting top appeared to diminish.

IntenseStormNet is used in ProbSevere v3, feeding in satellite information at the mature stage to the severe-weather models.

Cold front on the Plains

A potent cold front swept through the Great Plains in late April, spawning severe storms, including supercells, producing large hail, severe wind gusts, and tornadoes. 

Figure 1: Storm Prediction Center local storm reports from 12Z on 30 April to 12Z on 1 May 2024.

ProbSevere LightningCast was able to pick out developing convection along and ahead of the cold front, shown by regions of enhanced probabilities---first in eastern Nebraska, and then later in Kansas and Oklahoma. Toggle back and forth in time in the animation below to investigate model lead time to the first GLM flashes in different cells.

LightningCast is being evaluated at the HWT this spring, with some new features including:

• better probability calibration (as a result of more training data)
• static and on-demand lightning dashboards
• GOES-West version trained on only GOES-18 data
• an additional output product, the probability of ≥ 10 flashes in the next 60 minutes

The lightning dashboards are available from the LightningCast webpage for static locations such as airports and football stadiums, as well as dynamically changing locations like active wild-land fires.

The panel above shows some time series of LightningCast probabilities for the 5-min CONUS sector (red) and 1-min mesoscale sector (yellow) for Wichita Mid-Continent Airport, as storms developed and traveled over the airport. Blue dots are GLM centroid observations within 5 miles (large dots) and 10 miles (small dots) of the airport. 

A new feature being evaluated this year at HWT is the on-demand dashboard capability, whereby a NOAA user can submit location and timing information for an event that they want a dashboard for. They simply fill out this form and receive an automated email with the valid link: https://go.wisc.edu/x16m56. In this way, forecasters can get custom guidance for locations they are serving with decision support. This capability has already been used for events such as NFL games and state fairs.

Below is output from ProbSevere IntenseStormNet, which is a deep-learning model like LightningCast, but uses images of GOES-R ABI and GLM data to predict the probability of "intense" convection, from a satellite perspective. It is useful as storms are maturing. 

From the animation above, one can see how strong overshooting tops, bubbly-like texture in the cloud tops, and above-anvil cirrus plumes correspond well with stronger probabilities, which correlate well with local storm reports. IntenseStormNet works well for deep convection, day and night. Output from this model is used in ProbSevere v3, but can also be useful stand-alone severe-weather guidance in regions without radar coverage. 

The value of data fusion

We had an interesting storm develop in a radar gap in far eastern New Mexico yesterday. This is a great case study to demonstrate the value of data fusion in ProbSevere.

Figure 1 shows where the storm developed (the red circle), which was in a region of very poor "radar quality", as the eastern New Mexico KFDR radar was down. Thus, the closest radar was KAMA in Amarillo, TX.

Figure 1: Radar Quality Index for eastern New Mexico yesterday. The red circle is the approximate location of where the storm first developed. 

Figure 2: ProbSevere IntenseStormNet contours with GOES-16 ABI vis-IR sandwich product for a rapidly developing storm in eastern New Mexico.

One input into ProbSevere v3 is the probability of "intense" convection, as computed from IntenseStormNet. This is a deep-learning model that uses images of ABI 0.64-µm reflectance, 10.3-µm brightness temperature, and GLM flash-extent density to compute a probability of how "intense" the storm looks from a satellite perspective [paper]. 

The rapidly increasing IntenseStormNet probability, along with a favorable environment, and increasing total lightning flash rates helped jump the probability of severe despite poor radar reflectivity.

As the storm moved south and east into better radar coverage, radar reflectivity increased and the probabilities of severe further increased to above 70%. 

Figure 3: ProbSevere contours, MRMS MergedReflectivity, and NWS severe weather warnings.

Forecasters at the HWT have noted numerous times how ProbSevere v3 has increased before v2, particularly noticeable in the regime we've experienced this week, where the storms have had a dearth of lightning at the developing stages. At the time in Figure 4, this storm had PSv3 of 36% vs PSv2 of 12%. 

Figure 4: ProbSevere and MRMS MergedReflectivity for a developing storm in a radar gap in eastern New Mexico.

Later on, this storm produced numerous large hail, severe wind, and several tornado reports. Interestingly, the ProbTor v3 was much higher than ProbTor v2 prior to the first tornado report. In Figure 5, we can see PTv3 is 47% while PTv2 is only 9%. Looking into this deeper, we found that the environmental information such as the 0-1 km storm-relative helicity (~ 30 m^2/s^2) and the 1-3 km mean wind (~15 kt) were very low. The HRRR values in PTv3 were much better (~100 m^2/s^2 for SRH and 27 kt for the low-level mean wind). I believe this is an indication that PTv2 was too dependent on environmental information, compared to PTv2. This also demonstrates that the HRRR had a better handle on the environment than the RAP. You can see the low 0-1 km storm-relative helicity in the SPC mesoanalysis (Figure 6).

Figure 5: ProbSevere contours (the outer contour is colored by the probability of tornado), MRMS MergedReflectivity, and NWS severe weather warnings.

Figure 6: 0-1 km SRH (contours) and storm motion (vectors) prior to tornadogenesis. The red circle shows where the approximate location of the storm prior to producing tornadoes. 

Figure 7 demonstrates how ProbTor v3 was much higher than ProbTor v2 early on. The vertical black lines in the top-left two panels represent the times of the first and last tornado reports. The interactive version of these time series have been saved off and are available here.  

Figure 7: Time series of ProbSevere probabilities and radar, satellite, lightning, and HRRR attributes for the tornadic storm in Figure 6. 

IntenseStormNet and ProbSevere

The deep-learning model IntenseStormNet uses only image patches of GOES-R ABI and GLM data to detect the most intense parts of storms. Specifically, it uses the 0.64-µm reflectance, 10.3-µm brightness temperatures, and the GLM flash-extent density. In this way, IntenseStormNet tries to holistically use spatial patterns and textures to predict the probability of intense storms from a purely satellite perspective. Figure 1 shows the output of IntenseStormNet for several supercell or supercell-like storms in Texas. Notice how the model predicts high probabilities in areas of cold cloud tops, bubbly cloud-top texture, and strong overshooting tops.

Figure 1: Output from the IntenseStormNet (probability of intense convection contours), GOES-16 sandwich product (0.64-µm reflectance and 10.3-µm brightness temperature), and severe storm reports.

Though the IntenseStormNet model isn't being evaluated explicitly in the HWT this month, it serves as an input into ProbSevere v3 (PSv3), so that the PSv3 models can utilize the rich spatial and spectral information that GOES-R provides.  The IntenseStormNet output is often in the top third of contributing predictors for ProbSevere predictions.

Figure 2: ProbSevere contours, MRMS MergedRef, and NWS severe weather warnings for two supercells in Texas during the HWT. The outer contours are colored by the probability of tornado. The time series window is for the rightmost storm.