nakas/rw2-camels671-modern-corpus

CAMELS-671 Modern Areal Forcing Corpus (1980 → 2026)

Daily, basin-averaged (areal) meteorological forcings and USGS discharge for the 671 CAMELS-US basins, spanning 1980-01-01 … 2026-09-22 on a single, consistent pipeline.

Why this exists

The CAMELS-US forcing archive stops well short of the present: its Maurer forcing ends 2008-12-31 and its Daymet/NLDAS copies end 2014-12-31. Any model trained on CAMELS therefore cannot be run on data from this decade at all. This corpus rebuilds the forcings from products that are still produced, over the full 1980→present span, so that training and inference share one pipeline by construction rather than being bridged by version and convention corrections.

Contents

path	rows/basin	description
gridmet/<gauge_id>.csv.gz	17,067	gridMET (PRISM-anchored), areal
nclimgrid/<gauge_id>.csv.gz	17,067	nClimGrid-Daily (NOAA/NCEI), areal
discharge/<gauge_id>.csv.gz	17,067	USGS daily discharge, one authoritative pull
cohort/m1_cohort.json	—	the 490-basin modern evaluation cohort and its rule
provenance/*.json	—	build sidecars and the forcing-screen results

671 files per directory. <gauge_id> is the 8-digit zero-padded USGS site number.

Columns

Forcing files: date, q_cfs, temperature_2m_mean, temperature_2m_max, temperature_2m_min, precipitation_sum, shortwave_radiation_sum, vapor_pressure Discharge files: date, q_cfs

Units: temperature °C, precipitation mm/day, shortwave radiation MJ m⁻² day⁻¹, vapour pressure Pa, discharge ft³/s. An empty q_cfs means no observation — it is never zero-filled, and USGS -999999 NoData sentinels are masked.

Method

• Areal, not point. Basin means are coverage-fraction-weighted over GAGES-II bas_ref_all polygons (exactextract). Point sampling at the gauge coordinate was measured here to run ~53% dry relative to the catchment mean in mountainous terrain, and the bias is basin-varying, so it cannot be corrected by a constant.
• Registration decided by observed streamflow, not by convention. For each product, peak_k corr(precip[t], relative dQ[t+k]) over k ∈ −2…+2 was computed against USGS discharge, whose dates are unambiguous. Peak-over-lag is registration-invariant, so it measures how much signal a product carries rather than how it is labelled. Both products here were found already on the k=0 convention and are shipped unshifted (shift_days: 0 in each build sidecar).
• Complete daily calendars. Every file is a contiguous daily series over the full span; the assembler refuses to write a corpus with a gap.

Forcing screen (provenance/)

Median peak corr(precip, relative dQ) on 60 cohort basins, 2000-2009, with discharge and the day set held fixed so only the precipitation series varies:

product	median peak r	vs CAMELS-Daymet	modal peak lag
CAMELS Daymet (reference)	0.3554	—	0 (93%)
gridMET	0.3732	+0.0177	0 (67%)
nClimGrid	0.3507	−0.0048	0 (75%)
CAMELS NLDAS	0.3987	+0.0433	0 (92%)
CAMELS Maurer	0.3263	−0.0292	0 (72%)

⚠️ This screen measures single-product same-day precipitation–flow coupling, not ensemble value. Maurer scores −0.0292 yet contributes positively to a multi-forcing ensemble, so the numbers should not be read as a ranking of usefulness.

Known limitations

• nClimGrid ships no radiation or humidity. Its shortwave_radiation_sum and vapor_pressure columns are carried over from gridMET. nClimGrid is therefore independent of gridMET in 4 of its 6 channels — precipitation and temperature, which are the channels that carry the signal — but not in all six. Treat the two as partially correlated.
• Two further products are still building (Daymet V4R1-areal and NLDAS-2-areal) and are deliberately not included rather than published partially. Note that Daymet V4R1 labels precipitation one day earlier than the CAMELS convention; when it is added it will be shipped shifted, because serving it natively hands a model the next day's rain at the step it is scored on.
• The 490-basin cohort is not the 531. Of the 531 Kratzert-subset basins, 24 gauges have no observations at all in 2018-10-01…2025-09-30 and 17 more are below 90% coverage. cohort/m1_cohort.json records the rule, which was fixed before any model score was seen.
• Discharge is a fresh USGS pull, so it reflects USGS revisions and is not bit-identical to CAMELS' archived values (measured over 53,356 overlapping day-pairs: median ratio 1.000000, log-space r 0.998481; the differences are integer rounding in CAMELS plus genuine rating revisions concentrated in a handful of low-flow gauges).

Sources, terms and citations

This corpus is derived from the following openly-available sources. Each retains its own terms; please cite them alongside this dataset.

• gridMET — Abatzoglou, J.T. (2013). Development of gridded surface meteorological data for ecological applications and modelling. International Journal of Climatology 33:121-131.
• nClimGrid-Daily — Durre, I. et al., NOAA National Centers for Environmental Information. U.S. Government work, public domain.
• USGS daily discharge — U.S. Geological Survey National Water Information System. U.S. Government work, public domain.
• GAGES-II basin boundaries — Falcone, J.A. (2011). GAGES-II: Geospatial Attributes of Gages for Evaluating Streamflow. U.S. Geological Survey. Public domain.
• CAMELS-US basin set — Newman, A.J. et al. (2015); Addor, N. et al. (2017).

The derived files in this repository are released under CC-BY-4.0.

Reproducing

Builders and screens live in the scripts_modern/ tree of the RiverWatch2 repository (build_gridmet_areal.py, build_nclimgrid_areal.py, build_q_store.py, assemble_m3.py, screen_products_referee.py). Each build writes a _BUILD_SIDECAR.json recording the product, the applied registration shift and its rationale, the year range and the discharge source; those sidecars are included under provenance/.