Object-based data mining in forecasting and verification: Adventures in the 2011-2012 Hazardous Weather Testbed (HWT) Experimental Forecast Program (EFP)

Summarizing large storm scale ensemble data sets has mostly been limited to geographic displays focused on individual variables of interest. In an operational setting this approach can be time consuming and results in an equally large number of maps and variables. For the synoptic scale this approach is reasonable as the features of interest scale with the display and are easily recognized. However, in convection allowing models the features of interest can be much smaller than the display scale and cover areas on the synoptic scale. Statistical summaries of large data sets can thus be used to focus on convective storms and severe convective storm proxies.

A real-time, simple, and fast algorithm was developed to identify objects within convection allowing models run during the HWT-EFP. A spread-growth algorithm using a so-called double threshold and double pixel requirement was used. The two threshold values were chosen to capture the minimum value needed for detection and an upper threshold meant to capture features of particular interest. The pixel counts are meant to enhance the robustness of detected features adding value and avoiding noise at the grid scale. The algorithm has been utilized to analyze large storm scale ensembles for: 1. convective storms (using simulated composite reflectivity) including the storm environment and severe attributes, 2. Hourly Maximum Updraft Helicity (UH) tracks (with a slight modification), and 3. transient, potentially severe phenomena such as heat bursts.

The talk focuses on applications of this summary method as developed specifically within three contexts: 1. Convective storm environments and applicability to distinguishing severity, 2. severe weather location, timing, and magnitude via severe proxies such as UH, and 3. verification of individual storms in data assimilation comparisons and forecasting.

Lessons learned and future directions for the analysis and forecast utilization of this approach are discussed. These include evaluating ensembles, model physics, and data assimilation systems in real-time, and utilizing these statistical databases for the purposes of easing the data overload problem. In the latter case, the summary method allows forecasters to look at all the data and then focus in on a few members and variables.