447 Hagelslag: Scalable Object-Based Severe Weather Analysis and Forecasting

Monday, 11 January 2016
Hall E ( New Orleans Ernest N. Morial Convention Center)
David John Gagne II, CAPS/Univ. of Oklahoma, Norman, OK; and A. McGovern, N. Snook, R. A. Sobash, J. D. Labriola, J. K. Williams, S. E. Haupt, and M. Xue

Handout (3.3 MB)

Hagelslag is an open-source Python package for generating and evaluating calibrated, object-based forecasts of severe weather from convection allowing model ensembles. The package features modules for identifying and tracking storm objects from gridded model output and observation data, training and running machine learning models to predict the probability of severe weather hazards within the storm objects, and evaluating the machine learning predictions with a set of appropriate statistics. Scientific Python packages and multiprocessing are used extensively to make the data processing and prediction operations fast and scalable. Multiple ensemble members and model runs can be processed simultaneously. Potential threat objects can be identified and tracked through time before being matched with observed tracks. Parallelizable formulations of contingency tables, receiver operating characteristic curve, performance diagram, and reliability diagram have been developed to make statistical evaluation of months of gridded ensemble forecasts faster and less memory-intensive. The package has been able to generate probabilistic hail forecasts in real time from the Center for Analysis and Prediction of Storms Storm-Scale Ensemble Forecast system and the National Center for Atmospheric Research Ensemble. The code has run successfully on laptops, servers, and supercomputers. The hagelslag package will be released at the AMS Annual Meeting on Github and on the Python Packaging Index.

Supplementary URL: https://github.com/djgagne/hagelslag

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner