Techniques toward Optimizing a Geostationary Satellite-Based Convective Initiation Nowcasting Algorithm

Forecasting the placement and timing of convective initiation (CI) remains a significant challenge for present day meteorologists despite the significant advancements in numerical weather prediction (NWP) and in-situ and remote sensing measurements. Using geostationary satellite data can provide important growth characteristics about convective clouds, including cloud-top height, cloud-top glaciation, and updraft strength. The development of the SATellite Convection Analysis and Tracking (SATCAST) system by researchers at the University of Alabama in Huntsville (UAHuntsville) has combined these growth characteristics to create 0-2 hour forecasts of convective initiation. Early versions of SATCAST were binary yes/no forecasts based on strict thresholds (Mecikalski and Bedka, 2006; Walker et al. 2012) and provided no information to forecasters on the “probability” of CI. Based on forecaster feedback from the GOES-R Proving Ground at the Storm Prediction Center and multiple National Weather Service Forecast Offices (NWSFOs), updates have been made to SATCAST to provide “Strength of Signal” satellite-only forecasts, along with Random Forest generated “probabilistic” forecasts using both satellite information and NWP.

The researchers at UAHuntsville developed a database of thousands of CI and NON-CI events that consisted of the satellite indicators along with ~20 NWP variables, including CAPE, CIN, and shear. The “Strength of Signal” product uses logistical regression of the satellite indicators to develop a forecast of CI. Evaluation by forecasters at the GOES-R proving ground and an intense evaluation period during a month-long period during the summer by several NWSFOs yielded very positive results with forecasters developing a sense of how to use the product within the convective regime that they were observing. Ultimately, the use of Random Forest includes the convective environment information provided by NWP. The Random Forest technique employed at NCAR uses a series of decision trees to determine what satellite indicators and NWP variables are of utmost importance and creates a probabilistic forecast. Examples of output and results from the Random Forest and “Strength of Signal” approaches will be presented.