About RUNOFF
What the dataset is, why it exists, and how it is built.
Why RUNOFF
Flash floods develop in small basins within minutes to a few hours of intense rainfall. The community datasets that accelerated rainfall-runoff modeling, such as CAMELS and Caravan, are daily or hourly, basin-mean, and centered on basins larger than the ones that flash flood. No open benchmark pairs sub-hourly radar precipitation with gage discharge at event scale on an operational catchment network. RUNOFF fills that gap. Any flash flood modeling effort in the United States, inside or outside the NextGen framework, can use it for calibration, evaluation, and benchmarking.
How the dataset is built
Flood reports, water years 2021-2025
We start from 19,993 flash flood observations in the NOAA Storm Events Database. The period is chosen deliberately: it overlaps the dual-polarization algorithm era of the MRMS quantitative precipitation estimation product, so every event pairs with radar rainfall of consistent quality.
Report aggregation into events
Reports describing the same storm are merged on two criteria: overlapping spatial buffers built from each report's start and end locations, and the shared Storm Events episode identifier. This aggregation also reveals the regions most exposed to repeated flash flooding.
Pairing with USGS gages at flash flood scale
Events are kept only when an upstream or downstream USGS gage from the GAGES-II inventory, in a basin smaller than 1000 km2, can serve as the reference observation. Gages are snapped to the NextGen HydroFabric flowpath network to make the pairing hydrologically consistent. 7,990 events survive.
Radar quality screening
Radar coverage of each event's upstream catchment is used as a proxy for QPE uncertainty: it measures how well the radar sees precipitation reaching the ground, not whether data exist. Events with at least 80 percent coverage are retained. The final catalog holds 5,097 flash flood events.
Forcing and observations for every event
Each event carries 2-minute MRMS PrecipRate radar precipitation, hourly AORC meteorology, USGS instantaneous discharge in UTC, and reported impacts (fatalities, damage). A one-time crosswalk maps every MRMS grid cell onto the NextGen HydroFabric, covering 98.8 percent of CONUS catchments, so radar forcing drops directly into NextGen simulations.
Event separation
Observed hydrographs are separated into individual flood events with the NOAA-OWP HydroTools decomposition method, using parameters tuned for flashy, small-basin response (one-hour smoothing half-life, two-day window, one-hour minimum event duration). Events are then filtered against the flow duration curve: only events whose peak exceeds a chosen exceedance threshold, Q25 by default, enter the dataset. For every event the pipeline reports peak flow, peak time, runoff volume, duration, time to peak, and the Richards-Baker flashiness index, all timestamped in UTC.
engine/events/. Point it at any HUC8 watershed and it
rebuilds the event record from live USGS data:
python engine/events/extract.py --huc8 03020201 --wy-start 2021 --wy-end 2025
What is in the dataset
| Component | Source | Resolution | Notes |
|---|---|---|---|
| Radar precipitation | MRMS PrecipRate (radar-only) | 2 min, 0.01 deg | near real time, no gauge correction |
| Gauge-corrected precipitation | MRMS MultiSensor Pass 1 / Pass 2 | 1 h, 0.01 deg | gauge-bias corrected, higher latency |
| Meteorology | NOAA AORC v1.1 | 1 h, 0.8 km | temperature, humidity, wind, pressure, radiation |
| Streamflow | USGS NWIS instantaneous values | 15 min | reference observations, UTC |
| Events and impacts | NOAA Storm Events Database | event scale | fatalities, damage, cause |
| Network | NextGen HydroFabric v2.2 | catchment | 831,777 CONUS catchments |
Status
Team and credits
Built by Team "Don't Runoff!" at the Summer Institute 2026: Jessica Keiser, Leo Lonzarich, Azizur Rahman, with theme leaders Mohamed Abdelkader and Humberto Vergara.