Click for Powerpoint Articulate Training Module
Download Training Slides
Supplemental Table of Contents
The NOAA/CIMSS ProbSevere model is a
Naive Bayesian classifier. The Naive Bayesian classifier predicts, using any number of datasets, the
probability of a 'yes' class event will occur based on the datasets used. In applying the classifier to whether a thunderstorm will first produce severe weather
in the next 60 minutes it is necesary to define the 'yes' class and 'no' class--in this application 'yes' class are storms that produce severe weather and 'no' class
are storms that do not produce severe weather.
The datasets used within the model (RAP composite MUCAPE and EBS (effective bulk shear), satellite observational predictors (normalized vertical cloud growth rate and cloud-top glaciation rate), and radar observational predictor (MRMS MESH) are collected for a set of training data--a population of severe thunderstorms ('yes' class) and a population of non-severe thunderstorms ('no' class). Class-conditional probabilities are computed for the range of values within each dataset for the yes and no classes. It is this training dataset that is used to compute real-time probabilties for all storms.
In real-time, the data values for each predictor for a given storm are input into the ProbSevere model. The differences between class-conditional probabilties for the datasets used by the model are mathematically combined to generate a final probability--the probability viewed in AWIPS-II and on this web site. The two examples below illustrate how the class-conditional probabilities are used in the ProbSevere model to compute the final probability.
The full mathemtical equation for the naive Bayesian model is given by:
Where the final probability of a 'yes' event is given by the ratio of the product of yes-class conditional probabilities and yes-class prior probability to the product of the yes-class conditional probabilities and yes-class prior probability plus the product of the no-class conditional probabilities and no-class prior probability.
The following figures are the RAP composite MUCAPE/EBS yes-class prior probability and the class-conditional probabilitiy distributions from the
training datasets for the three observational parameters (satellite: normalized vertical growth rate and cloud-top glaciation rate) and radar: instantenous MRMS MESH).
Plotted on these figures (black) are example values for a hypothetical severe thunderstorm.
Hypothetical severe thunderstorm with 3,000 J/kg of MUCAPE and 20 m/s EBS from the RAP compsite.
Yes-class prior probability = 0.17. (No-class prior probability = 1 - 0.17 = 0.83)
Hypothetical severe thunderstorm normalized vertical growth rate of 3.0%/min.
Hypothetical severe thunderstorm glaciation rate of 0.06/min.
Hypothetical severe thunderstorm MRMS MESH of 0.75".
Combining all the class-conditional probabilities and prior probabilities yields an end probability of 97%. Or this hypothetical severe thunderstorm has a
97% chance of first producing severe weather in the next 60 minutes. The very high probability is attributed to a favorable environment, strong satellite growth rates,
and a large, yet still sub-severe MRMS MESH value.
ProbSevere currently uses the Rapid Refresh model (RAP) from NCEP. The effective bulk shear (EBS) and most-unstable convective available potential energy (MUCAPE) are used as predictors in the statistical model. New RAP forecast and analysis data are available approximately every hour. Once new RAP data are available, the EBS and MUCAPE fields are computed for every gridpoint for the analysis grid, as well as the 1-, 2-, and 3-hour forecasts.
Next, for each grid point, the maximum EBS (or MUCAPE) is taken over the previous hour analysis (computed previously), the current analysis, and the 1-, 2-, and 3-hour forecasts (just computed). This "off-centered" approach is implemented since RAP data have about a one hour latency. Thus, the 1-hour forecast is approximately valid at the time the data are delivered. After the temporal compositing is complete, a spatial filter is applied, which has a Gaussian kernel. The temporal compositing and spatial smoothing of the RAP data is performed in an effort to mitigate placement and phasing errors inherent in NWP data.
The NOAA/CIMSS ProbSevere model utilizes GOES satellite imagery for both identifying and tracking cloud clusters as well as quantifying growth rates of clouds. Below
is an example of GOES cloud cluster identification and tracking for a developing thunderstorm over the Texas panhandle. The top row of images is IR-window brightness temperatures and the
bottom row are the same images with an illustration of how cloud clusters are identified and tracked by the computer.
The satellite tracking uses infrared data--so the tracking is the same day and night. Within each cloud cluster the computer identifies and tracks, two satellite growth rates are computed: the normalized vertical growth rate and the cloud-top glaciation rate. The normalized vertical growth rate uses a field known as the top of troposphere emissivity (Pavolonis 2010). The vertical growth rate computed using this field is analagous to brightness temperature cooling rates, except the growth rates are normalized for varying tropospheric depth/tropopause height, while raw brightness temperature data are not. The cloud-top glaciation rate uses GOES/GOES-R cloud phase/type algorithm output to characterize how quickly the cloud-tops change from water phase to ice phase. The statistical model details shows how these two growth rates, along with radar and NWP data are used within the NOAA/CIMSS ProbSevere model.
The NOAA/CIMSS ProbSevere model heavily leverages the Multi-Radar Multi-Sensor (MRMS) products developed at NOAA-NSSL and OU-CIMMS. By using multiple radars to sample weather, gaps in radar coverage due to things such as terrain blockage, the "cone of silence", and the radar beam overshooting weather at far ranges may be mitigated. Furthermore, combining multiple estimates of radar moments at any particular point can give a better final estimate. Multiple radar surveillance of weather can also provide more frequent updates. The ProbSevere model updates at the MRMS frequency, which is approximately every 2 minutes.
MRMS merged reflectivity is used to identify and track storms in radar imagery, using the Warning Decision Support System -- Integrated Information (WDSS-II). WDSS-II automatically identifies storms using an enhanced watershed algorithm and tracks storms by using methods to match identified objects in consecutive image pairs.
Figure is adapted from OU-CIMMS images.
Once storm objects are tracked in radar imagery, the Maximum Expected Size of Hail (MESH) is extracted from the spatial bounds of the objects. MESH is empirically derived from the Severe Hail Index (SHI), which is essentially a thermally weighted vertical integration of reflectivity above the melting level. Several recent studies have shown that MESH has some skill for identifying the presence of severe hail in storms. Please see Witt et al. (1998) for a more complete description of MESH and SHI. The instantaneous maximum MESH is used as a predictor in the ProbSevere model.