J10.4
Evaluation of a probabilistic convective nowcast for CoSPA

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Thursday, 21 January 2010: 11:45 AM
B305 (GWCC)
D. Ahijevych, NCAR, Boulder, CO; and J. Williams, S. Dettling, H. Cai, and M. Steiner

Presentation PDF (739.1 kB)

This paper describes the evaluation of a new probabilistic nowcast system developed as part of the FAA's Collaborative Storm Prediction for Aviation (CoSPA) effort to provide improved short-term convective forecasts for NextGen. While numerical weather model ensembles can be a vital tool for assessing uncertainty in longer-term forecasts, their initialization, spin-up, and integration time limit their usefulness in the nowcasting timeframe (0-2 hours), where extrapolation of existing storms (including their growth and decay) often out-performs all models. The random forest technique produces a probabilistic nowcast model for each lead-time that maps a large number of operational model fields, radar and satellite data, and derived features (including some provided by MIT) to the likelihood that a storm intensity threshold will be exceeded. The model accounts for growth and decay, and may predict initiation even in locations where no radar echo is present at the analysis time.

In the summer of 2009, random forest probabilistic nowcasts of VIP level 1 and VIP level 3 exceedance were developed for 1 and 2-hour lead-times and run in real-time. The RF nowcasts were evaluated by setting a fixed probability threshold and scoring the resulting deterministic predictions alongside other forecasts. They were also scored as probabilistic forecasts, enabling reliability and resolution comparisons to the LAMP thunderstorm guidance. These evaluations demonstrate that the RF nowcasts add value to CoSPA by providing an independent analysis of the input data along with valuable quantitative uncertainty estimates that could be essential to users.