False alarm area reduction at DSS event with LightningCast v2

 A simulated DSS event was placed in Garden City, KS, ahead of expected convection forced via cyclogenesis in eastern Colorado and a strong mid-level jet streak.

As severe storms began to approach the vicinity, LightningCast v2 (with Ref. -10C predictor) held probabilities in the 20-30% range, whereas LightningCast v1 (with ABI-only predictors) had probability of lightning in the 50-60% range for an hour, topping out at 75%. Forecasters could see the differences on the DSS lightning dashboard (Figure 1). No flashes were observed within 10 miles of the DSS location.

Figure 1: Lightning dashboard showing LCv1 (red) and LCv2 (green) time series of next-hour probabilities of lightning. No GLM flashes were observed during this time.

When we spatially compare LCv1 and LCv2 probabilities, we see that v1 contours extend much further northeast into the large anvil cloud towards Garden City, KS, whereas v2 probability contours are correctly more conservative.

This example demonstrates the benefit of data fusion: The ABI channels cannot “see” under that very thick ice, but the radar was showing no new convective development. Therefore, the LightningCast v2 model could reduce false alarm area for this DSS event, only needing to factor in anvil lightning potential and (to some extent) the motion of the storms.

Figure 2: Toggle of LCv1 and LCv2 probabilities. LCv1 probabilities extend further northeast towards Garden City, KS, creating more false alarm areas. Note well that these contours are NOT parallax corrected, but the probabilities in Figure 1 are. GOES-19 C02 reflectance and C13 brightness temperatures are plotted in the background, whereas GOES-19 GLM flash-extent density are the blue-to-yellow foreground pixels.

- Hail yeah

LightningCast v1 vs v2 in far southwest Texas

 Today forecasters had an opportunity to look at LightningCast v1 (ABI only) and v2 (ABI + Ref10 inputs) in far southwest Texas, where radar coverage is quite poor.

LightningCast v2 for convection in southwest Texas. Background is GOES-19 C13, foreground blue pixels is GLM flash-extent density. Contour legend: green = 10%, yellow = 30%, orange = 50%.

LightningCast contours v1 for convection in southwest Texas. Background is GOES-19 C13, foreground blue pixels is GLM flash-extent density. Contour legend: green = 10%, yellow = 30%, orange = 50%, red = 70%.

For this area of developing convection, v1 was more bullish on two areas of convection that indeed became thunderstorms.

The signals in the day-cloud-phase-distinction RGB, which are represented as inputs in both versions of LightningCast, were indeed indicative of glaciated convection and high lightning potential.

GOES-19 day-cloud-phase-distinction RGB for southwest Texas (courtesy of College of Dupage NEXLAB).

This region was on the edge of the valid MRMS domain for the Reflectivity -10C predictor (see below).

Reflectivity at -10C. Gray is invalid regions, while the rest of the domain is considered “valid” to the LightningCast v2 model.

It’s possible that LightningCast v2 was expecting good MRMS Reflectivity -10C in that region, but that the co-evolution of Ref -10C with the ABI image predictors was slower than the expected co-evolution based on countless training examples in areas of better radar coverage.

The developers of LightningCast theorize that this could be ameliorated by using the Radar Quality Index (RQI) from MRMS, which quantifies the data quality. Supplying this as a predictor to LightningCast could help in regions with moderate to very poor radar coverage, and provide better uniform guidance throughout the CONUS.

MRMS Radar Quality Index in southwest Texas.

- Hail yeah

Anomalous Cloud-Top Divergence in Absence of Visible Convection

 The top panels show IR and VIS imagery, while the bottom panels display Octane CTD, and GREMLIN reflectivities. GREMLIN agrees well with the other fields/products, though I did not verify it against real radar data (as today is a no-radar day).

One feature that stood out was the Octane CTD product. I marked a blue rectangle highlighting high CTD values (between 2 and 4) in an area where no convection appears in the satellite imagery. I do not know if this is caused by the cloud motion around this area, but it appears to be an artifact caused by the algorithm.

- Iceman

Wednesday Observations 6/4: The Day with No Radar

 Today in the testbed all three forecaster groups were told to act as if they did not have any radar at their disposal, with offices in Albuquerque, Midland, and St. Louis.

Starting the day in Albuquerque CWA...The forecasters were quick to use OCTANE and GREMLIN to identify the strongest storms, and they picked out a storm just west of the city of Albuquerque. Increasing reflectivity values from GREMLIN, along with divergence signals from OCTANE Speed and the Cloud-Top Divergence products. Animation of GREMLIN and OCTANE (with radar included) are included below. The forecasters also leveraged the 12Z sounding from ABQ with its freezing levels, mid level lapse rates, and the wind profile.

GREMLIN 4 panel with the MESO scene (upper left), CONUS (lower right), MRMS composite reflectivity (upper right), and ABI MESO Clean-IR (lower left).

OCTANE with Speed (upper left), and the cloud top cooling and divergence flavors.

12Z Sounding from ABQ.

As the forecasters debated on issuing a warning, the 'real' NWS Albuquerque office issued a severe thunderstorm warning for 60 mph winds and 1 inch hail.

After some additional discussion, the HWT 'fake' Albuquerque office issued their own severe thunderstorm warning, also for 60 mph winds and 1 inch hail. They used LightningCast probabilities and ENTLN/NLDN flash locations to identify the thunderstorm 'core'.

Shifting to the Midland CWA...The forecasters in this office were watching for initiating convection just across the US-Mexico border in the higher elevations west of the Big-Bend region. There was discussion of the environment and how supportive it was for deep convection, especially related to the moisture/instability return from the Gulf during the forecast period.

OCTANE speed, cloud top divergence, and cloud top cooling products all showed the first robust updraft tapping into this instability with increasing speeds and strong cloud top cooling signatures followed by cloud top divergence.

The GREMLIN product from the MESO scene also reflected this cooling signature and strong gradients with increasing simulated reflectivity values. These values continued to increase until the HWT MAF office issued a severe thunderstorm warning. With 5-10 minutes of the warning, GREMLIN exceeded 60 dBZ. Another storm to the southeast also initiated and quickly intensified, with GREMLIN values also exceeding 60 dBZ.

We talked about the skill of GREMLIN to identify discrete convection and how realistic the values compared to our expectations from this machine-learning, satellite-based perspective. The takeaway from the forecasters was that GREMLIN has an easier time showing greater reflectivity values in discrete convective cores than multi-cell convection. 

I also looked at the ALPW product, and noted that the surface to 850mb layer showed the returning moisture, though I wonder if forecasters would like to see more frequent updates than hourly. I didn't have the chance to ask about this due to the active discussions during operations, but this seems like it would be the 'ideal' case since there was little cloud cover so sampling the lowest layers would be easier for the polar-orbiting sounders. How much this can help NWS forecasters searching for observations in remote areas is to be determined, but the product seemed to fit the conceptual model for the case.

-Dr. Thunder

Assessing GREMLIN skills

 Around 2000Z, a squall line was observed over Oklahoma and Kansas. GREMLIN successfully reproduces the overall reflectivity distribution seen by the KVNX radar, though it slightly underestimates the reflectivity of the convective cores (Fig. 1).

Fig 1: GREMLIN synthetic radar reflectivity (left) and KVNX NEXRAD actual radar reflectivity (right)

By 2016Z, convection intensifies, with convective cores covering larger areas. GREMLIN reproduces most of these features (Fig. 2), though some are heavily underestimated (black marks). The isolated convection—likely poorly resolved by satellite observations—appears to be particularly challenging for GREMLIN to accurately simulate.

Fig 2: GREMLIN synthetic radar reflectivity (left) and KVNX NEXRAD actual radar reflectivity (right). Black marks represent areas where GREMLIN did not reproduce correctly the intensity of the radar reflectivity field.

- Iceman

Better heads up at DSS event

Forecasters were tasked with alerting event managers in Memphis during a fireworks show of any lightning within 10 miles, at least 30 minutes ahead of time.

Mature and severe storms approached Memphis from the west. The ProbSevere LightningCast model predicts the probability of 1 or more next-hour lightning strikes at any given location. It not only needs to anticipate convective development, but also the short-term movement of thunderstorms.

Figure 1: Animation of LightningCast v2 (contours), GOES-19 ABI day-cloud-convection RGB (background), and GLM flash-extent density (foreground blue-to-orange pixels). The yellow “Home” point and range ring denote the simulated DSS event and 10-mile ring around the event.

In the HWT this year, forecasters are asked to compare LightningCast v1 (ABI-only predictors) and v2 (ABI + MRMS predictors). They also are able to use “lightning dashboards” for on-demand DSS events. The image below demonstrates the time series of LCv1 (red) and LCv2 (green) probabilities at the fireworks show location, along with GLM flash centroids within approximately 10 miles (small blue circles) and 5 miles (large blue circles) of the location. Note how LCv2 probabilities began to  jump up about 20 minutes prior to LCv1. Several nearby flashes were observed between 20:30 and 21:00 UTC. From the 75% threshold, LCv2 provided nearly 40 minutes of lead time to the first flash within 5 miles.

Figure 2: Lightning dashboard for the Memphis fireworks show.

When we toggle between LCv1 and LCv2 at 19:51 UTC, we clearly see moderate-to-high probabilities (50-75%+) from LCv2 extending further eastward towards the event (“Home” point), compared to LCv1. This is a subtle but impactful difference that can help forecasters alert event managers in a more timely manner.

Figure 3: Toggle between LightningCast v1 and v2 (parallax-corrected contours) outputs, with GOES-19 ABI day-cloud-convection RGB (background) and GLM flash-extent density (foreground).

- Hail yeah

Explosive growth in Oklahoma

Convection is growing rapidly in southern Oklahoma, in the midst of a very unstable environment with ≥4000 J/kg of MLCAPE.

LightningCast version 2, with MRMS Reflectivity -10C (left, below) appears to have a 1-2 minute jump on version 1 (right) with the developing convection. Importantly, LightningCast v2 also diminished some false alarm probabilities to the southeast about 5 minutes prior to v1.

We have already received reports of confirmed tornadoes and significant hail (≥ 2” diameters).

LightningCast v2 contours (left) and v1 contours (right), with GOES-East ABI day-cloud-phase-distinction RGB (background) and GLM flash-extent density (blue, foreground).

- Hail yeah

Large GREMLIN Differences in Western NC

GOES-E Mesoscale and GOES-E CONUS GREMLIN output displayed large differences for a storm in western NC.

GOES-East Meso (left) and GOES-East CONUS (right) GREMLIN output at 1951 UTC 08 May 2025.

A GREMLIN developer attributed this difference to most likely be related to the integration of lightning (1-minute for the mesosector, 5-minute for CONUS).

Severe Tstorm in MRX - Radar Denial

 Though the radar-denial assignment was in GSP for the forecasters, I decided to take a look over at MRX while still avoiding radar. When looking at the OCTANE, there was a large (~50 kt) gradient in the OCTANE Speed Sandwich product, as well as a modest cooling signature and more significant divergence values up to 5*10^-3/s at 2024 UTC 08 May 2025. I’m aware that the smoothing technique can reduce the max divergence values observed, and believe the ‘high’ amount of smoothing may reduce the maximum value to the point that the forecaster will think the storm is not as intense.

OCTANE 4-panel at 2024 UTC 08 May 2025, with the Speed Sandwich (top left) and Cloud-Top Cooling and Divergence with varying smoothing techniques (none - top right, medium - bottom left, and high - bottom right). Please note while sampling the divergence value, it drops from ~5*10^-3/s for no smoothing to ~3*10^-3/s for high smoothing.

Based on OCTANE, I was leaning towards the storm being severe, so I also took a look at the GLM Flash Extent Density (FED) and GREMLIN products. There was a large lightning jump in the FED between 2006 and 2021 UTC 08 May 2025. In addition, the GOES-E Mesosector GREMLIN output showed intensifying reflectivities starting at 2007 UTC 08 May 2025.

LightningCast (V1-left, V2-right) overlaid on GLM FED and GOES-East Day Cloud Phase Distinction RGB imagery at 2006 UTC 08 May 2025.

LightningCast (V1-left, V2-right) overlaid on GLM FED and GOES-East Day Cloud Phase Distinction RGB imagery at 2021 UTC 08 May 2025.

GOES-East Mesosector GREMLIN output at 2007 UTC 08 May 2025 (left) and CONUS GREMLIN output at 2006 UTC 08 May 2025 (right).

I would go ahead and issue a Severe Tstorm Warning. Well enough, at 2024 UTC 08 May 2025, MRX issued the warning for “Ping pong ball size hail and 60 mph wind gusts”.

Mesoanywhere Tracking An Outflow Boundary in JAX

 Mesoanywhere seemed to give a couple of minutes of lead time in detecting the speed and direction of an outflow boundary (possibly aided by sea breeze) pushing inland in JAX.

Mesoanywhere Day Cloud Phase Distinction RGB imagery (left) versus CONUS imagery on the right at 2033 UTC 07 May 2025 (one minute before CONUS-sector update). The cursor on the left is in the ‘middle’ (longitudinally) of the boundary, where on the right, you can see it’s on the leading edge.

One minute later, you can see the CONUS sector update (right) indeed shows the westward progression of the boundary as predicted by Mesoanywhere.

LightningCast v2 under thick ice

 Forecasters were working in Jacksonville and Tulsa this afternoon, but one visiting scientist noticed convection firing in southern Alabama, under thick anvil ice from the massive MCS in the morning and early afternoon along the Gulf Coast.

LightningCast v2 contours (bottom left), LightningCast v1 contours (bottom right), and MRMS Reflectivity -10C (top). The background on the bottom panels is the GOES-19 ABI day-cloud-phase-distinction RGB, and the blue foreground pixels are GLM flash-extent density observations.

LightningCast v1 uses four GOES-R ABI predictors. While there was a large expanse of cold clouds in Alabama, the low texture and obscuration of any low-level features made it difficult for the model to detect the developing convection (see bottom right panel of the animation above).

LightningCast v2, with MRMS Reflectivity -10C as a predictor, ably picked up on the convection (bottom left), providing 20 minutes of lead time to the first flash in the initial cell (from the 25% probability threshold), and 10-15 minutes of lead time to the southern cells (from the 10% probability threshold). Version 1 of LightningCast provided essentially no lead time from the 10% threshold, and probabilities remained low even while the lightning was ongoing.

I believe this is emblematic of the type of situation where the radar predictor is complementing the satellite predictors in LightningCast v2.

- Hail yeah

LightningCast in morning low-topped convection

 Low-topped convection is evident in southern Kansas under a closed upper-level low.

GOES-19 day-cloud-phase-distinction RGB, courtesy of College of Dupage NEXLAB.

The convection has been ongoing for about 2 hours this morning, with no observed GLM flashes. LightningCast v2, which includes MRMS Reflectivity -10C, has had more moderate probabilities of lightning, topping out at about 50%, whereas LightningCast v1 (GOES-only), has had a much wider range of ≥ 50% probabilities, with maxima > 75%.

Figure 2: LightningCast probability of lightning contours (left: version 2; right: version 1), in southern Kansas, along with the GOES-19 CH13 brightness temperature and C02 reflectance. The blue represents 10%, cyan 25%, green 50%, and magenta 75% probability of lightning in the next 60 minutes.

The MRMS viewer shows a few embedded areas of Ref -10C at 30-35 dBZ, but otherwise all convective cores are < 30 dBZ.

Figure 3: MRMS Reflectivity -10C for southern Kansas.

In this case, the radar predictor is by-and-large correctly moderating and reducing the probabilities of lightning in the region. We’ll see how the low-topped convection develops later in the day.

-Hail yeah

GREMLIN Detecting Convective Initiation in LCH WFO

GOES-East Mesoscale and CONUS GREMLIN output accurately picked up on reflectivities with developing storms where there were radar beam blockages just to the northeast of Houston at 1844 UTC 06 May 2025. By 1902 UTC, the ground-based radar also picked up on the reflectivities as the convection pushed farther to the northeast. The convection continued to develop further, eventually impacting our IDSS location in Beaumont with lightning about an hour later (~2000 UTC 06 May 2025).

Four-panels of GOES-East Mesoscale GREMLIN output, MRMS, GOES-East Mesoscale Channel 07 Band imagery, and GOES-East CONUS GREMLIN output overlaid with GLM Flash Extent Density (FED) at 1844 and 1902 UTC 06 May 2025.

- May052025

LightningCast V1 and V2 in Anvil Cirrus

 Forecasters were down on the Texas and Louisiana Gulf Coast today monitoring the severe weather. While there were ongoing storms, differences were noted between LightningCast V1 (ABI only) and LightningCast V2 (ABI+MRMS). Starting at around 1901 UTC, V1 had much broader probabilities over the thick anvil shield when compared to V2. V2, with the additional input of reflectivity at -10C, was able to provide better predictions over this region, helping forecasters not to erroneously tell DSS partners in the region of impending lightning. Towards the end of the time period, both versions showed higher probabilities as convection started to move into the area.

LightningCast V1 (ABI only) on the left and V2 (ABI+MRMS) on the right from 1901-2011 UTC 6 May 2025 over the Louisiana Gulf Coast.

- AlphaOmega

LightningCast V1 & V2 Differences in ABQ

GOES-East LightningCast Version 1 seemed to pick up on lightning ahead of LightningCast Version 2 (MRMS-trained) for the cell in the southern portion of the ABQ domain (bottom-middle of screen), with higher probabilities (75% + for V1, 50%+ for V2) at 1916 UTC 05 May 2025. However, once the GLM FED detected lightning, LightningCast V2 saw a more expansive 75% contour. A theme I seemed to pick up on was that LxCast V1 would have higher lightning detection probabilities than V2 prior to initiation, but V2 would have a greater footprint of higher probabilities thereafter, more accurately capturing lightning once it was occurring. V1 probabilities seemed to fluctuate more often than V2 as well. 

LightningCast V1 and V2 output in the ABQ WFO area at 1916 UTC and 2006 UTC 05 May 2025.

-May052025