A Random Forest Sensitivity Test Using Radar Signatures to Forecast Wet Downbursts at Spaceport Florida

The 45^th Weather Squadron (45WS) U.S. Air Force Unit at Patrick Air Force Base, Florida provides weather support and issues weather warnings to the space program at Spaceport Florida, defined as the Cape Canaveral Air Force Station (CCAFS) and NASA Kennedy Space Center. The warnings for winds are ≥ 35 kt with a desired lead time of 30 minutes and ≥ 50 kt with a desired lead time of 60 minutes. This study focuses on improving the warnings for convective storms that produce ≥ 35 kt winds since they occur much more often than storms that produce ≥ 50 kt. The emphasis is on differentiating between convective storms that do and do not produce winds ≥ 35 kt and increased lead times and lower false alarm ratios on the warnings. In this study, we identified convective events that produced winds ≥ 35 kt (hereafter “wind” events) and winds < 35 kt (hereafter “null” events) using data from 29 instrumented weather towers in the CCAFS area for the 2015 and 2016 warm seasons. Data from a C-band dual-polarization radar were used to identify and track storms related to each of the wind and null cases and to calculate a number of radar signatures for every storm cell. These radar signatures describe physical properties in storms that could predict the occurrence of downburst winds ≥ 35 kt. Initially, 84 wind events and 126 null events were analyzed, and roughly 50 radar signatures were calculated for each event. A correlation analysis was performed to eliminate highly correlated signatures, and a Principal Component Analysis was performed among the remaining signatures to remove parameters that did not offer much variance explanation. The final eight selected predictor radar signatures were: (1) the vertical extent of the 1 dB differential reflectivity (Z_dr) contour in a Z_dr column in the presence of radar reflectivity (Z_h) ≥ 30 dBZ at temperatures colder than 0°C, (2) the vertical extent of co-located values of Z_h ≥ 30 dBZ and Z_dr ~0 dB at temperatures colder than 0°C (hereafter called the “precipitation ice signature”), (3) the peak Z_h at any temperature within a storm (4) the peak Z_h at temperatures colder than 0°C, (5) the height of the peak Z_h in the storm, (6) vertically integrated liquid (VIL), (7) vertically integrated ice (VII), and (8) density of VIL (DVIL). These signatures were integrated into a classification random forest method in which thousands of sensitivity tests were performed in order to obtain the optimal set of random forest settings. Random forest sensitivity tests included: varying the number of signatures considered at each split in a tree, varying the fraction of wind events in the training dataset, considering a single radar volume scan versus a set of radar volume scans prior to the occurrence of the downburst, and considering either the maximum or mean value of a parameter over a given time period. For each random forest run applied to a set of test storms, performance metrics such as the probability of detection (POD), probability of false alarm (POFA), true skill statistic (TSS), etc., were calculated. Using a large ratio of wind to null events (0.4 to 0.5 wind) to train the random forest provided better statistics than using a lower ratio. Varying the number of variables considered at each split in a tree did not greatly affect the skill score results, but, on average, using a smaller number of variables (such as 2) provided slightly better results than using a larger number of variables. Skill scores were higher during the time periods of 15 to 30 minutes and 20 to 30 minutes before the downburst compared to times closer to the downburst. This is likely due to the nature of most of the radar signatures, which develop and reach a peak tens of minutes before a downburst occurs. This is the case for signatures like the Z_dr column, the precipitation ice signature, and the height of the peak Z_h in the storm. The results from the above-mentioned tests show that the method can be used to predict the downburst potential, and thus could in principle increase warning skill and lead-time for convective events over the CCAFS area.