Using data mining to improve convective initiation forecasts

Numerical weather prediction models are essential tools for forecasting hazardous weather, such as thunderstorms. However, their initialization, spin-up, and integration time limit their ability to provide the best synthesis of available information in the nowcasting timeframe (0-2 hours). For these short lead-times, statistical methods may provide enhanced forecast accuracy by fusing up-to-the-minute weather observations with the available NWP model forecasts. This is particularly true for the problem of predicting convective initiation -- the growth of new thunderstorms.

This paper describes how empirical data mining may be used to analyze a dataset, determine important predictors, and create a convective initiation forecast. The random forest technique is used to map operational model fields, radar and satellite data, and derived features (including MIT Lincoln Laboratory fields and SATCAST data) to the likelihood of convective initiation. The resulting forecast system is evaluated both statistically and for a number of case studies.