Day-ahead ensemble probabilistic forecasts of hail produced by (a) a U-net model based upon Burke et al. (2020), (b) the random forest model of L20, and (c) the random forest model of Burke et al. (2020). For comparison, also shown is (d) a practically perfect forecast (an “optimal” forecast calculated from the actual locations of observed hail reports). Contours indicate predicted probability of hail at 5% intervals (5% and 15% contours are highlighted). Observed hail reports are overlaid on each panel, with green circles indicating reports of severe hail [diameter 1–2 in. (25.4–50.8 mm)], and black circles indicating reports of significant severe hail (diameter 2 in. or greater).
Figure 14 from Steinkruger et al. (2020), showing an example of localized tornado warning output from an AI system.
(a)–(f) A select sequence of images depicting the evolution of LightningCast probabilities along a cold front in Iowa, superimposed on 0.64-μm reflectance (grayscale) and GLM flash-extent density from GOES-16. Lead times to the initial GLM flashes for several areas of interest are annotated in (f), showing lead times in minutes from both the 75% and 25% probability thresholds [reproduced, with permission, from Fig. 14 in Cintineo et al. (2022)].
The 24-h maximum HREF ensemble mean (a) most-unstable CAPE × 10 m–500-hPa wind shear, (b) 10-m wind speed, (c) 0–1-km storm relative helicity (SRH), (d) upward vertical velocity, (e) 2–5-km updraft helicity, (f) 1-km reflectivity, (g) supercell composite parameter (SCP), (h) significant tornado parameter (STP), and (i) significant hail parameter from the 0000 UTC 11 May 2022 initialization. These are used (along with other variables, not shown) to produce RF-based (j) tornado, (k) wind, and (l) hail probabilities, valid from 1200 UTC 11 to 1200 UTC 12 May 2022 using the procedure outlined in L20, L22. In (j)–(l), observed SPC tornado, wind, significant wind, hail, and significant hail reports are indicated by red dots, blue dots, black squares, green dots, and black triangles, respectively.
Representative example comparing the WoFS–ML–wind product vs a carefully calibrated probability of exceedance (80-m wind speed > 40 kt) during 2021 HWT SFE. This forecast is valid for 0030–0100 UTC 5 May 2021. Objects shaded in blue are ensemble storm tracks (Flora et al. 2019, F21) where the probability value is the likelihood of a severe wind report occurring within the object. For reference, the WoFS probability-matched mean of composite reflectivity is shown as a shaded underlay and observed severe wind reports (blue dots) are overlaid.
Contoured (a) version 1 and (b) version 2 NCAR NN-based tornado probabilities (S20) tested during the 2022 HWT SFE, valid from 1200 UTC 18 to 1200 UTC 19 May 2022. Severe thunderstorm and tornado warning polygons are overlaid in blue and red, respectively. The figure is adapted from the SFE Viewer website (SPC 2022).
