Machine Learning Estimation of Maximum Vertical Velocity from Radar

TL;DR

The paper tackles the absence of real‑time convective updraft measurements by training a U‑Net with SHASH parametric regression to estimate the distribution of the maximum vertical velocity from 3D radar reflectivity. Trained on synthetic WoFS data and evaluated on transfer domains GridRad and C3LOUD‑Ex, the approach achieves $R^{2}\approx0.65$–$0.75$, $IoU\approx0.45$–$0.51$, and $\text{RMSE}\lesssim4.5\,m\,s^{-1}$ on WoFS test data, while real‑world case comparisons show a consistent underestimation of extreme updrafts (median ~50% of dual‑Doppler values) and smaller IoU for updraft areas (≈0.25 at 5–10 m s⁻¹). The study demonstrates that even with single time steps of 3D reflectivity, the model can rapidly distill radar structure into meaningful updraft proxies, offering a potentially valuable diagnostic for storm severity while highlighting the need for further real‑world vetting and inputs. Overall, this work lays groundwork for near‑real‑time updraft diagnostics that could support hazard assessment, with clear paths for improvement in resolution, additional input channels, and broader validation.

Abstract

The quantification of storm updrafts remains unavailable for operational forecasting despite their inherent importance to convection and its associated severe weather hazards. Updraft proxies, like overshooting top area from satellite images, have been linked to severe weather hazards but only relate to a limited portion of the total storm updraft. This study investigates if a machine learning model, namely U-Nets, can skillfully retrieve maximum vertical velocity and its areal extent from 3-dimensional gridded radar reflectivity alone. The machine learning model is trained using simulated radar reflectivity and vertical velocity from the National Severe Storm Laboratory's convection permitting Warn on Forecast System (WoFS). A parametric regression technique using the sinh-arcsinh-normal distribution is adapted to run with U-Nets, allowing for both deterministic and probabilistic predictions of maximum vertical velocity. The best models after hyperparameter search provided less than 50% root mean squared error, a coefficient of determination greater than 0.65 and an intersection over union (IoU) of more than 0.45 on the independent test set composed of WoFS data. Beyond the WoFS analysis, a case study was conducted using real radar data and corresponding dual-Doppler analyses of vertical velocity within a supercell. The U-Net consistently underestimates the dual-Doppler updraft speed estimates by 50$\%$. Meanwhile, the area of the 5 and 10 m s^-1 updraft cores show an IoU of 0.25. While the above statistics are not exceptional, the machine learning model enables quick distillation of 3D radar data that is related to the maximum vertical velocity which could be useful in assessing a storm's severe potential.

Figures (12)

• Figure 1: (a) Schematic summarizing past work on relating storm updraft proxies and their physical location. The black circle represents the approximate location of midlevel mesocyclones Sessa2020, the red circle illustrates the typical height of Zdr columns Wilson22, the blue shaded region is the approximate region for lightning charging Deierling2008, and the yellow circle shows the location of a overshooting top Marion2019 (b) Contours of measured radar reflectivity volumes and how it could outline the full storm updraft (inspired by chisholm1972, reproduced in trapp2013). The location of a Bounded Weak Echo Region Browning1967 is annotated. The background supercell illustration for both (a) and (b) was provided by the National Severe Storms Laboratory.
• Figure 2: Training data schematic. (a) the 3d slices of reflectivity used for the input to the machine learning middle (center). (b) column maximum vertical velocity. Tensor shapes are found in the titles.
• Figure 3: Warn on Forecast System domain locations. Top shows all domains where WoFS was run in 2019, middle shows 2018 and bottom shows 2020.
• Figure 4: Training distributions schematic. The red line depicts a gridpoint at the beginning of training. The circle marker at 18 $\mathrm{m \ s^{-1}}$ is the assumed true updraft value for that pixel. The blue line depicts the distribution after training. Example parameter values are noted in the boxes.
• Figure 5: One to one comparison of the median machine learning updraft prediction to the WoFS updraft prediction for the test dataset. The statistics for this comparison can be found in Table 1. (a) data for the 2dmax model (b) data for the 2d24f model and (c) data for the 3d model. The colorbar is the log of the counts.