Classifying Proximity Soundings with Self-Organizing Maps toward Improving Supercell and Tornado Forecasting

Abstract

The self-organizing map (SOM) statistical technique is applied to vertical profiles of thermodynamic and kinematic parameters from a Rapid Update Cycle-2 (RUC-2) proximity sounding dataset with the goal of better distinguishing and predicting supercell and tornadic environments. An SOM is a topologically ordered mapping of input data onto a two-dimensional array of nodes that can be used to classify large datasets into meaningful clusters. The relative ability of SOMs derived from each parameter to separate soundings in a way that is useful in discriminating between storm type, location, and time of year is discussed. Sensitivity to SOM configuration is also explored. Simple skill scores are computed for each SOM to evaluate the relative potential of each variable for future development as a method of probabilistic forecasting. It is found that variance in SOM nodes is reduced compared to the overall dataset, indicating that this is a viable classification method. SOMs of profiles of wind-derived variables are more effective in discriminating between storm type than thermodynamic variables. The SOM method also identifies meteorological, geographic, and temporal regimes within the dataset. In general, conditional probabilities of storm-type occurrence generated using SOMs have higher skill when wind-derived variables are considered and when forecasting nonsupercell events. Stormrelative wind variables tend to have better skill than ground-relative wind variables when forecasting nonsupercells, whereas ground-relative variables become more important when forecasting tornadoes.

Linking large-scale circulation patterns to low-cloud properties

“…The neighborhood radius, which determines the number of nodes surrounding the winning node that nudge toward the input vector, slowly reduces to one (only the winning node is nudged) through the training period. Overall, our choices are similar to and follow guidelines outlined in prior SOM studies that focus on vertical sounding classification problems (e.g., Jensen et al, 2012;Nowotarski and Jensen, 2013;Stauffer et al, 2017) and synoptic meteorology pattern recognition (e.g., Cassano et al, 2015;Ford et al, 2015;Mechem et al, 2018). Additionally, we find that changing the initial parameters (iterations and neighborhood radius to 25 000 and 2, respectively) has a relatively small impact on the final node topology similar to other studies (e.g., Cassano et al, 2006;Skific et al, 2009).…”

Linking large-scale circulation patterns to low-cloud properties

“…The neighborhood radius, which determines the number of nodes surrounding the winning node that nudge toward the input vector, slowly reduces to one (only the winning node is nudged) through the training period. Overall, our choices are similar to and follow guidelines outlined in prior SOM studies that focus on vertical sounding classification problems (e.g., Jensen et al, 2012;Nowotarski and Jensen, 2013;Stauffer et al, 2017) and synoptic meteorology pattern recognition (e.g., Cassano et al, 2015;Ford et al, 2015;Mechem et al, 2018). Additionally, we find that changing the initial parameters (iterations and neighborhood radius to 25,000 and 2, respectively) has a relatively small impact on the final node topology similar to other studies (e.g., Cassano et al, 2006;Skific et al, 2009).…”

Tropospheric ozonesonde profiles at long‐term U.S. monitoring sites: 1. A climatology based on self‐organizing maps

“…The SOM algorithm is configured via a user‐specified map (network) size and shape that dictates the number and relationship of the nodes that will represent the data. The map can be of any dimension, with 2‐D maps (e.g., a 4 × 4 map of 16 nodes) preferred in recent related meteorological applications [e.g., Jensen et al , ; Nowotarski and Jensen , ]. The initial values of the nodes, analogous to cluster centroids in k ‐means, can be obtained in a number of ways.…”

“…Jensen et al [] performed a cluster analysis on over 900 tropical ozonesonde profiles. They employed self‐organizing maps (SOM) [ Kohonen , ], which have been used as a clustering algorithm across many disciplines, including several recent meteorology and climate studies [ Hewitson and Crane , ; Hong et al , ; Liu et al , ; Nowotarski and Jensen , ]. Jensen et al [] used SOM to describe influences dictating O 3 variability at two SHADOZ stations, Natal, Brazil, and Ascension Island, from 1998 to 2009.…”