NOAA and CIMSS are developing a product that uses a deep-learning model to recognize complex patterns in weather satellite imagery to predict the probability of lightning in the short term. Deep learning is a branch of machine learning based on artificial neural networks, which have the ability to automatically learn targeted features in the data by approximating how humans learn.

A convolutional neural network (CNN) was trained on over 23,000 images of GOES-16 Advanced Baseline Imager (ABI) data to predict the probability of lightning, within any given ABI pixel, in the next 60 minutes. The CNN was trained using 118 days of data collected between May and August of 2018. Images of GOES Geostationary Lightning Mapper (GLM) flash-extent density (created with glmtools) were used as the source of lightning observations. Note that GLM is an optical sensor that observes both in-cloud and cloud-to-ground lightning.

The CNN currently uses four ABI channels: band 2 (0.64-µm), band 5 (1.6-µm), band 13 (10.3-µm), and band 15 (12.3-µm). Bands 2 and 5 are only utilized under sunlit conditions. Utilization of additional channels and time sequences of images is under investigation. The model uses a semantic image segmentation architecture to assign the probability of lighting in the next 60 minutes to each pixel in the image. The model is very computationally efficient, only needing 30 seconds to process the ABI CONUS domain and 3 seconds to process an ABI mesoscale domain using multithreading on a 40-CPU linux server.

Currently, the model only utilizes satellite radiances. Thus, it can be applied to nearly any spatial domain covered by the ABI or an ABI-like sensor (e.g. AHI). Based on near-real-time testing, the model routinely nowcasts lightning initiation with 10-30 minutes of lead-time. We expect the skill and lead-time will increase as new predictors (e.g. more ABI fields, NWP, radar where available) are added to the model.

Below are a sampling of recent examples. The base images are 0.64-µm reflectance, with GLM-derived flash-extent density overlaid as filled semi-transparent polygons. The flash-extent density is the accumulated number of flashes within the previous 5 minutes. The CNN-derived probabilities are displayed as contours at select levels (near-real-time output is available through RealEarth).

The overall objective is to improve lightning nowcasts in support of aviation, mariners, and outdoor events/activities. Beyond improving the CNN, our work will focus on packaging the output into actionable information for forecasters and other decision makers.

A cold front in Iowa

 

Thunderstorm development on sea-breeze boundaries in Florida and the Bahamas

 

Diurnally and orographically forced storms in the Southwest U.S. and Rocky Mountains

 

Storms in central Oklahoma, on the edge of Hurricane Laura’s cloud shield

 

A couple of examples over the Northeast U.S.

 

A boundary of convection on the southern bank of Lake Ontario

 

Storms in a warm sector in IL/IN/OH, perhaps along an outflow boundary

 

Southeast U.S. offshore region

The background image in this example transitions from 10.3-µm brightness temperature to 0.64-µm reflectance, while the flash-extent density enhancement also changes, in an attempt to enhance contrast.

View only this post Read Less

1-minute Mesoscale Domain Sector GOES-16 (GOES-East) “Red” Visible (0.64 µm) and “Clean” Infrared Window (10.35 µm) images — with and without an overlay of GLM Flash Extent Density (above) showed the rapid development of thunderstorms along a cold front across eastern South Dakota on 30 August 2020. One particularly well-defined and long-lived Enhanced-V signature with an Above-Anvil Cirrus Plume (reference | VISIT training) was seen in the Visible and Infrared imagery, which extended northeastward from the core of a thunderstorm where large hail and tornadoes were occurring.

1-minute GOES-16 Visible images with time-matched plots of SPC Storm Reports are displayed above, with the corresponding Infrared images below.

There were some overshooting tops which exhibited infrared brightness temperatures as cold as -72ºC — a plot of 00 UTC rawinsonde data from Aberdeen, South Dakota (below) indicated that this was about 7ºC colder than the tropopause temperature of -66.1ºC.

View only this post Read Less

1-minute Mesoscale Domain Sector GOES-16 (GOES-East) “Red” Visible (0.64 µm) and “Clean” Infrared Window (10.35 µm) images — with and without an overlay of GLM Flash Extent Density (above) showed Category 4 Hurricane Laura as it made landfall near Cameron, Louisiana around 0600 UTC on 27 August 2020. The GLM data showed intermittent lightning activity along the inner eyewall region of the hurricane.

Strong outer convective bands ahead of Laura’s landfall produced isolated tornadoes as it moved onshore (SPC Storm Reports). Peak wind gusts included 116 knots or 133 mph at Lake Charles at 0642 UTC (in addition, Lake Charles reported another peak wind gust of 113 knots or 130 mph at 0703 UTC). Strong winds associated with the northern portion of the eyewall destroyed the Lake Charles radar (YouTube video) — the final reflectivity and velocity images at 0553 UTC (12:53 am CDT) are shown here (the 0.5-degree inbound and outbound radial velocity values were as high as 160-162 mph).

#HurricaneLaura has made landfall near Cameron, LA with max winds of 150 mph – the strongest #hurricane to make landfall in Louisiana since the Last Island Hurricane of 1856 (also 150 mph max winds). pic.twitter.com/QeiHwigs74

— Philip Klotzbach (@philklotzbach) August 27, 2020

A toggle between Suomi NPP VIIRS Day/Night Band (0.7 µm) and Infrared Window (11.45 µm) images at 0751 UTC (above) revealed the nighttime glow of lights from Lake Charles (since that city was near the inside edge of the eye of Hurricane Laura at that time) — in other locations across Louisiana and far eastern Texas, the signature of city lights was muted to varying degrees by the storm’s dense cloud cover and precipitation.

The corresponding Suomi NPP ATMS Microwave (88.2 GHz) and MiRS Rainfall Rate images at 0751 UTC (below) depicted the pattern of precipitation that was spreading inland.

DMSP-17 and GMI Microwave (85 GHz) images from the CIMSS Tropical Cyclones site (below) showed the structure of Laura several hours before landfall.

An animation of the MIMIC-TC product during the 26-27 August period (below) showed the deterioration of the eyewall structure after landfall.

Prior to making landfall, Laura had been moving across the warm waters of the Gulf of Mexico — however, it began to encounter an environment characterized by increasingly unfavorable deep-layer wind shear as it approached the Gulf Coast (below) which likely prevented further intensification.

View only this post Read Less

1-minute Mesoscale Domain Sector GOES-16 (GOES-East) “Red” Visible (0.64 µm) and “Clean” Infrared Window (10.35 µm) images (above) showed Laura during the the 12-hour period after it intensified from a Tropical Storm to a Hurricane in the southern Gulf of Mexico at 1215 UTC on 25 August 2020. Numerous convective overshooting tops were observed, some exhibiting cloud-top infrared brightness temperatures as cold as -87ºC.

A comparison of NOAA-20 MiRS Microwave (88 GHz), GOES-16 “Red” Visible (0.64 µm) and GOES-16 “Clean” Infrared Window (10.35 µm) images at 1851 UTC (below) revealed a curved convective band wrapping around the eye of Laura.

In a toggle between Infrared Window images from Suomi NPP (11.45 µm) and GOES-16 (10.35 µm) at 1943 UTC (below), the coldest cloud-top infrared brightness temperature was -91.9ºC in the Suomi NPP image (compared to -87.0ºC on the GOES-16 image). The northwestward parallax displacement associated with GOES-16 imagery over the southern Gulf of Mexico was also apparent.

View only this post Read Less