RSI-benchmark-hydro

Forcing and calibration data for a rainfall–runoff benchmark: ten US catchments, hourly radar precipitation in, hourly streamflow out, with the validation storm held out.

The task is to build a hydrologic model from the calibration window and predict discharge through a flood event the model has never seen. The reference bar is EF5/CREST with kinematic-wave routing — the distributed model behind the US operational FLASH flash-flood system — calibrated per catchment with DREAM on exactly this calibration data.

Validation observations are not in this dataset. They are withheld so the benchmark can be scored. Everything needed to build and calibrate a model is here.

Contents

data.tar.gz unpacks to:

data/
  gauges.json           index: coordinates, area, region, window boundaries, hour counts
  pet/                  130 daily global PET grids, shared by every catchment
    etYYYYMMDD.tif        1 degree global, float32, mm/100d, nodata -9999
  static/<gauge>/       0.01 degree (~1 km), EPSG:4269, float32, nodata -9999
    dem_usa.tif           elevation, m
    fdir_usa.tif          D8 flow direction, ESRI codes (1 E, 2 SE, 4 S, 8 SW,
                          16 W, 32 NW, 64 N, 128 NE)
    facc_usa.tif          flow accumulation, upstream cell count
    mask.tif              uint8, 1 = drains to the outlet
    basin.json            outlet row/col, cell count, cell area, delineated area
    wm_usa.tif  b_usa.tif  im_usa.tif  ksat_usa.tif           a priori CREST parameters
    alpha_usa.tif  beta_usa.tif  alpha0_usa.tif  leaki_usa.tif  a priori kinematic-wave
  cal/<gauge>/
    precip/GaugeCorr_QPE_01H_00.00_YYYYMMDD-HH0000.grib2.tif   960 hourly grids, mm/h
    discharge.csv       datetime,discharge_m3s -- 960 hourly rows
    meta.json           coordinates, area, exact window boundaries
  val/<gauge>/
    precip/...          480 hourly grids, mm/h
    meta.json           same, plus the validation period to predict

gauges.json is also uploaded unpacked, so the catchment list is browsable without downloading the archive.

Catchments

gauge	name	area km²	region	window begins
03298000	Floyds Fork at Fisherville, KY	358	NE / Ohio Valley	2018-09-13
02492360	West Hobolochitto Creek near McNeill, MS	452	Southeast	2018-10-18
02435020	Town Creek at Tupelo, MS	560	Southeast	2018-10-18
01473000	Perkiomen Creek at Graterford, PA	721	NE / Mid-Atlantic	2018-09-13
08069000	Cypress Creek near Westfield, TX	726	Central	2018-11-02
07029270	Hatchie River near Walnut, MS	732	South	2018-11-22
02303330	Hillsborough River at Morris Bridge, FL	985	Southeast	2018-10-18
03179000	Bluestone River near Pipestem, WV	1024	NE / Ohio Valley	2018-09-13
02479300	Red Creek at Vestry, MS	1144	Southeast	2018-10-18
07031650	Wolf River at Germantown, TN	1805	South	2018-11-22

Split

Each catchment has its own 60-day window, cut the same way:

day 0 ─────── day 5 ──────────────── day 40 ──────────── day 60
│   warm-up   │   calibration        │   VALIDATION (held out)│
│ forcing only│ forcing + discharge  │ forcing only           │

• Warm-up, 120 h — forcing and discharge given, never scored. Spin up model states.
• Calibration, 840 scored h — forcing and observed discharge.
• Validation, 480 h — forcing only; observations withheld.

Every window contains at least one flood event on each side of the split, so the validation period tests generalisation to a new storm rather than continuation of the calibrated hydrograph.

Conventions

• Timestamps are UTC, tz-naive, on the exact hour. The file-name stamp is the valid time of the accumulation ending at that hour.
• Precipitation grids are clipped per catchment and share the static grids' geotransform exactly, so cell [i, j] is the same place in every raster for a given catchment. The clip is a bounding box; mask.tif selects the catchment.
• PET is mm/100 d — divide by 100 for mm/day. It is a coarse global 1° field.
• Discharge is m³/s at the outlet.
• The two South catchments (07029270, 07031650) have no precipitation grids for 2019-01-20; those 24 hours do not exist upstream. All 480 validation hours are still scored, so the last day runs on recession alone.

Reference performance

EF5/CREST-KW, calibrated per catchment with EF5's native DREAM sampler (20,020 model evaluations × 3 restarts, objective NSCE, calibration window only):

	mean	median
Calibration NSE	0.889	0.938
Validation NSE	0.799	0.829
Validation KGE	0.776	0.815

Uncalibrated EF5 on the a priori grids scores mean validation KGE −1.09. The gap between calibration and validation skill is what the benchmark measures.

Provenance

Derived from chrimerss/hybrid-cali (CC-BY-4.0), with daily PET grids from chrimerss/hydro_cali_agent_example. Precipitation is MRMS GaugeCorr QPE 01H from the Iowa State mtarchive; discharge is USGS NWIS instantaneous values resampled to the hourly grid. Basin masks were delineated by a D8 upstream walk from the outlet snapped on flow accumulation; delineated areas agree with USGS drainage areas to within 1 % for seven catchments and ~13 % for 03298000 and 02435020, matching the source dataset's own delineation.

The catchment pool lies east of −100 °W in two Köppen zones. This benchmark covers event-scale rainfall–runoff modelling in humid and sub-humid basins; it says nothing about arid or snowmelt-driven hydrology.