KokosDev/nwm-hydrology-benchmark

NOAA National Water Model (NWM) — Hydrology ML Benchmark (Zarr)

This benchmark is derived from the NOAA National Water Model (NWM) Retrospective v2.1 Zarr data on AWS Open Data.

• Upstream source (public, no auth):
  • s3://noaa-nwm-retrospective-2-1-zarr-pds/chrtout.zarr (streamflow)
  • s3://noaa-nwm-retrospective-2-1-zarr-pds/precip.zarr (RAINRATE forcing)
• Variables:
  • streamflow (m³/s): CHRTOUT streamflow at basin outlet reach
  • precipitation_rate (mm/hr): RAINRATE sampled at the nearest forcing grid cell to each outlet
• Time: 2018-01-01 → 2019-12-31 (hourly)
• Format: Zarr (consolidated metadata), chunked for weekly windows

Basin selection methodology (10–20 major basins)

We select 15 basin outlets automatically and reproducibly from the NWM routing network:

• Start from NWM reaches that have a non-empty gage_id (USGS gage linked in NWM).
• Filter to a rough CONUS bounding box.
• Rank candidates by NWM stream order (order) and keep only order ≥ 6.
• Greedily choose outlets that are geographically separated (minimum degree separation) to avoid selecting many nearby points.

This yields 10–20 large-river outlet points that are both “major” (high order) and geographically diverse.

Reproduce selection + extraction:

python build_nwm_hydrology_benchmark.py --out nwm_hydrology_benchmark.zarr --n-basins 15

Selected basins (outlets):

basin_feature_id	usgs_gage_id	lat	lon	stream_order	elevation_m
5092616	07022000	37.22028	-89.46666	10	96.00
19088319	07374000	30.46205	-91.19692	10	4.97
7474830	07032000	35.12103	-90.10248	10	57.62
20721986	09404200	35.77279	-113.36486	9	414.73
4391417	06893000	39.12090	-94.57079	9	219.69
17219468	06610000	41.27217	-95.91407	9	293.41
14523623	06342500	46.81313	-100.82358	9	494.94
1170022657	12399500	48.98944	-117.63982	9	400.90
1543497	07249455	35.39183	-94.43327	9	119.57
10010638	09429000	33.73117	-114.50939	9	82.96
6013168	06935450	38.56324	-91.00963	9	145.00
947020558	12472800	46.64091	-119.66222	9	149.17
947070191	14105700	45.60490	-121.16764	9	23.64
13260207	06295000	46.27838	-106.67533	8	765.87
4931276	09315000	38.99018	-110.14662	8	1234.77

Forecasting task definition

We define a multivariate time-series forecasting task at basin outlets.

• Input: past (L) hours of [streamflow, precipitation_rate]
• Target: next (H) hours of streamflow

Recommended defaults:

• (L = 168) (1 week)
• (H = 24) (next day)

Splits should be time-contiguous (no leakage).

Comparison to existing hydrology benchmarks

This benchmark is complementary to common hydrology datasets:

• CAMELS / CAMELS-US: observed streamflow + basin attributes at ~hundreds of basins; typically daily.
• HYSETS: large multi-basin hydrometeorological dataset (observations/reanalysis-based).
• LamaH / Caravan: large-scale hydrology benchmarks with basin delineations and attributes.

How this differs:

• Model-based (NWM retrospective), hourly resolution, CONUS-wide routing network.
• Includes a forcing precipitation rate field (sampled at outlet), suitable as an exogenous driver.
• Outlet selection is fully reproducible from NWM metadata (no external shapefiles required).

Usage

import xarray as xr

ds = xr.open_zarr("nwm_hydrology_benchmark.zarr", consolidated=True)
print(ds)

PyTorch DataLoader

See examples/pytorch_dataloader.py.

Benchmark (local vs streaming-from-source)

Run:

python bench/throughput_benchmark.py --local nwm_hydrology_benchmark.zarr --stream-s3

Example results (this machine):

mode	samples/sec	MB/sec	first_batch_sec
local	22.588	0.465	0.071
streaming_s3	0.102	0.003	10.452