Radar extrapolation vs. radar–NWP blending: who wins?
Three contrasting U.S. cases, one blending method, and a benchmark package written so that the comparison can be repeated honestly.
// In short
- In the fast-evolving Rolling Fork supercell (2023), STEPS blending beat every baseline at all lead times from 1 to 6 h.
- In a slow stratiform nor'easter (2023), it lost to plain extrapolation at +1 h.
- A storm-structure rule for predicting the better method, derived from those two cases, failed on a third, out-of-sample flash-flood case (2026).
Setup
NEXRAD extrapolation nowcasts were blended with HRRR and ECMWF open-data forecasts using pysteps STEPS, and compared with radar-only, NWP-only and linear blends. CSI differences were checked against multi-seed ensemble noise before being interpreted.
nowcast-bench
The evaluation code became a package: every score is reported with its sample size, undefined cases are flagged instead of silently dropped, and results were cross-checked against pysteps.
Related
Analyses of the Moore (2013) and Rolling Fork (2023) tornadic storms with Py-ART: CAPPIs, VIL, azimuthal shear, dual-polarization debris signatures and interactive 3D renderings. Code
Questions or data requests welcome · Contact