Automating Storm Identification and Classification for Extreme Storm Catalogs to use with Stochastic Storm Transposition

Advances in computational power and increased availability of high-resolution spatiotemporal quantitative precipitation estimates (QPE) have allowed for increased use of Stochastic Storm Transposition (SST), a data-driven method for generating precipitation-frequency relationships. Teams supporting FEMA as part of their Future Flood Risk Data (FFRD) initiative are evaluating SST as the hydrometeorological basis for updates for flood hazard models across the United States. To perform SST, a large catalog of historical storms is needed for transposition from their location of occurrence to a randomly selected location within an area of similar hydroclimatic conditions, called a ‘transposition domain’. In transposition domains that are sufficiently large, where multiple local to mesoscale storm events can occur simultaneously, or storm events have high seasonal variability in precipitation mechanisms, an understanding of storm type is helpful in making informed decisions on initial sampling conditions.

Methodology was developed to automatically identify large storm events by precipitation magnitude and areal extent by shifting a basin outline to every possible combination of locations within a transposition domain and identifying the location shift which maximizes precipitation over the basin. This is repeated for a dataset’s entire period of record to produce an extreme storm catalog that is unique to a given transposition domain and associated basin of interest. This initial storm identification and cataloging process has been completed for the Trinity, TX basin – a high-interest and high-consequence watershed with unique basin characteristics, climatology, and storm types that produce elevated flood risk. The Trinity is a long and narrow, north-south oriented basin which extends from the Dallas-Fort Worth Metro Area to the Gulf Coast and is regularly impacted by mid-latitude and tropical cyclones.

To identify storm types within a basin’s extreme storm catalog, the Four-Kilometer Long-Term Regional Hydroclimate Reanalysis over the Conterminous United States (CONUS404) dataset was leveraged as part of a two-part typing process: first, manually typing a sample dataset of storms using CONUS404 data and second, training a model to automatically classify storms. Meteorological parameters such as accumulated precipitation, surface-based convective available potential energy (CAPE), mean sea-level pressure, storm relative helicity, precipitable water, and 500 mb geopotential height from CONUS404 were visualized in a Jupyter notebook and used to manually type a sample of storms in the catalog. For the Trinity, these manual types were mid-latitude cyclones, tropical cyclones, mesoscale storms, local storms, and hybrid storms - which can be a combination of multiple storm types.

These meteorological parameters were then used along with manual storm types to train and test a multinomial log-linear regression model before applying to the full suite of storms in the extreme storms catalog. The result is an enhanced catalog of extreme storm events to be used for SST in hydrologic modeling, with information regarding individual storm type and seasonality of all storms in the dataset. This presentation will discuss the process of developing storm specific meteorological datasets for manual storm typing and development of regression models using open-source software with gridded QPE and climate reanalysis for the purpose of automatically typing an extreme storm catalog.