10.5 Spatial Aligned Mean Ensemble Consensus Method Applied to CAM Precipitation Forecasts in the 2023 FFaIR Experiment

Wednesday, 31 January 2024: 11:45 AM
302/303 (The Baltimore Convention Center)
Chang Jae Lee, CAPS, Norman, OK; and K. A. Brewster and P. Spencer

It is common for an ensemble of high-resolution forecasts of precipitation to have differences in propagation speed and initiation location of storms among the ensemble members, resulting in spatial offsets of features among the member forecasts. The differences can grow through the length of the forecast. The Spatial Aligned Mean (SAM) technique for finding ensemble consensus directly addresses such spatial offsets by moving fields toward a common central location. This method is proposed as an alternative to more commonly used methods for determining ensemble consensus, such as the simple mean, Probability-Matched (PM) mean and Localized PM (LPM) mean. In fact, the LPM rescaling method can be applied to the SAM repositioned members to create a SAM-LPM method, potentially capitalizing on the advantages of both algorithms.

Our previous work demonstrated this advantage using the High Resolution Ensemble Forecast (HREF) from the summer of 2022. In this experiment, SAM and SAM-LPM are applied to an ensemble designed by Center for Analysis and Prediction of Storms (CAPS), using FV3-Limited Area Model (FV3-LAM) for the 2023 NOAA/WPC Hydrometeorology Testbed (HMT) Flash Flood and Intense Rainfall Experiment (FFaIR). SAM and SAM-LPM ensemble consensus for 6-h precipitation fields were produced in real-time and are evaluated over the contiguous United States (CONUS) using Stage IV precipitation data as verification. The point-wise verification metrics (Frequency Bias, POD, FAR, and ETS) and spatial feature verification are calculated using several thresholds in the Meteorology Evaluation Tools (MET) and MET-MODE programs, respectively.

The overall verification results and flood cases during the FFaIR period, such as flash flood cases in Kentucky and Vermont during the summer of 2023, will be presented and discussed. The figure below is an example from the western Kentucky flash flood at 12 UTC on 19 July 2023. Though the members differed on the location of the high precipitation area, SAM and SAM-LPM were able to bring all the forecasts to a common central position, similar in structure to the observed rainfall. In addition, SAM-LPM preserved the forecasts maxima close to the observed maximum.

From the preliminary verification results, Spatially Aligned Mean (SAM) ensemble consensus technique outperformed the simple ensemble mean. The results show that the spatial alignment technique improves the ensemble consensus in common verification metrics such as ETS. Also, SAM-LPM improves the structure of the mean while preserving the ensemble forecast maxima, thus seems to be the best candidate for calculating an ensemble consensus for these fields.

Within the SAM algorithm offset vectors for all possible pairs of members are calculated, so as the number (N) of members increases, computation time will increase substantially, on the order of NxN calculations. A modified SAM technique, which scales according to N rather than NxN, thus using less resources and time, will be presented and discussed from the perspective of running operationally, with some preliminary evaluations.

Figure Caption
Observed precipitation (Stage IV, upper left panel) and four ensemble consensus 60h forecasts (standard mean, SAM, standard LPM, SAM-LPM) of 6-hour precipitation valid at 12 UTC 19 July 2023

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner