prediction2/jpdh-dataset

JPDH dataset

Daily forcing, static descriptors and river discharge observations on a river network covering Japan, prepared for learning hydrological simulations at national scale, together with the simulations of the model described in

Leveraging Japan's streamflow records for end-to-end learning of hydrological simulations at national scale.

Code, trained weights and instructions: https://github.com/TristHas/jpdh.

This is a pre-release snapshot. A versioned archive with a DOI will accompany the publication.

At a glance

River network	8,893 catchments (mean area 36.5 km²) linked by 8,183 downstream edges into 710 river basins
Period	daily, 2008-01-01 to 2020-12-30 (12Z days, see Time convention)
Discharge	daily observed discharge at 1,189 gauges (Ministry of Land, Infrastructure, Transport and Tourism, MLIT)
Reservoirs	daily inflow, outflow and storage of 116 monitored dams (MLIT)
Forcing	catchment means of precipitation (5 products) and MSM / ERA5-Land meteorology, in daily, 12-, 6-, 4-, 2- and 1-hour bins
Static descriptors	382 runoff descriptors (HydroATLAS attributes, JAXA land use, J-FlwDir topography, climatologies) and 17 reach descriptors
Evaluation design	10 folds of whole river basins (PUR), 979 evaluation gauges, 332 training-pool gauges, three 4-year temporal blocks
Simulations	national simulation of every catchment, 2009–2020; out-of-region simulations at the 979 evaluation gauges (10 seeds)

Files

The dataset is distributed as three tar archives; MANIFEST.json lists their sizes and SHA-256 checksums. Each archive extracts into the same dataset root and adds checksums/<archive>.json, the SHA-256 of every file it contains.

Archive	Contents
jpdh_core.tar	everything needed to train and run the model in its headline configuration: network/, statics/, daily/forcing.zarr, daily/discharge.zarr, daily/dams.zarr, splits/, qc/
jpdh_panels.tar	daily/forcing_panels.zarr: the inputs of the sensitivity experiments
jpdh_simulation.tar	simulation/: the model's simulations

With the code repository:

python scripts/download_data.py              # core
python scripts/download_data.py --all        # core + panels + simulation
python scripts/verify_data.py

or with the Hugging Face CLI: hf download <this repository> --repo-type dataset --local-dir jpdh_archives, then tar -xf each archive into one directory.

Zarr stores are in Zarr format 3 (read them with xarray.open_zarr); tables are (Geo)Parquet. Node ids (node_id, spatial) are 64-bit integers shared by every file.

Time convention

Daily values are 12Z days: the label YYYY-MM-DD 12:00 (UTC) covers 12:00 UTC of that day to 12:00 UTC of the next (21:00 to 21:00 Japan Standard Time). An N-hour bin i of a day covers hours [N·i, N·(i+1)) after 12:00 UTC. Accumulated quantities (precipitation) are totals over their window; state quantities (temperature, wind, pressure, cloud cover) are window means.

network/

File	Content
nodes.parquet	one row per catchment, index node_id: lat, lon (outlet, WGS84), catchment_area and drainage_area (= upa, km²), channel_length_km (length of the catchment's reach), is_root (basin outlet), is_leaf (headwater), is_gauge, is_dam, station_id, dam_id, mlit_dam_id
edges.parquet	source → target: each catchment and the one it drains into
catchments.parquet	catchment polygons (GeoParquet, EPSG:4326)
channels.parquet	reach geometries (LineString; Point for one-cell reaches) and river_channel, the raster cells of the reach
stations.parquet	the 1,923 gauging stations and 852 dams on the network: location, names (Japanese), river system, snapping distance, upstream area; for dams, purpose, type, height, capacity and construction dates

The network was derived from the J-FlwDir flow-direction raster with a target catchment area of about 40 km², placing a catchment outlet at every gauging station and dam.

statics/

File	Content
runoff_statics.parquet	382 catchment descriptors, raw physical values (the code standardises them): HydroATLAS attributes (Linke et al., 2019; *_s* for the catchment, *_u* for its upstream area), JAXA land-use fractions (Water bodies, Built-up, Paddy field, Cropland, Grassland, DBF, DNF, EBF, ENF, Bare), J-FlwDir elevation (elv_mean, elv_std), catchment area, ERA5-Land monthly climatologies (era_*_<month>), monthly seasonality indices (season_msm_*, season_oki_*) and mean snow indicators
routing_statics.parquet	17 reach descriptors: elevation, slope and upstream-area statistics, sinuosity, log_upa, log_slope, catch_area, … (NaN for headwater catchments, which have no reach)

daily/

All stores share time (4,748 days, 2008-01-01 to 2020-12-30) and use spatial = node_id.

forcing.zarr (spatial = all 8,893 catchments) — the headline inputs:

Variables	Content	Units
gauge_prcp_4h_bin_{0..5}	in-situ gauge-interpolated precipitation (Hatono et al., 2022), six 4-hour bins	mm per bin
msm_a_temp_4h_bin_{0..5}	JMA MSM analysis 2 m air temperature, six 4-hour bins	°C
gauge_prcp_flag, msm_a_temp_flag	daily QC bitmask: 1 gap-filled hours, 2 masked source defects, 4 partial hours, 8 accumulation rescaled from fewer than 24 observed hours, 16 accumulation summed from the gap-filled series	–

forcing_panels.zarr (optional archive) — 149 variables of the sensitivity experiments:

Variables	Product
garadar_prcp_4h_bin_*	JMA gauge-adjusted radar precipitation (Radar-AMeDAS analysis)
radar_prcp_4h_bin_*	JMA C-band radar composite precipitation
msm_a_prcp_4h_bin_*	JMA MSM analysis precipitation
era_prcp_4h_bin_*, era_temp_4h_bin_*	ERA5-Land precipitation and 2 m temperature
gauge_prcp, msm_a_temp	daily total / daily mean
gauge_prcp_{1,2,6,12}h_bin_*, msm_a_temp_{1,2,6,12}h_bin_*	1-, 2-, 6- and 12-hour bins
msm_a_u_4h_bin_*, msm_a_v_4h_bin_*, msm_a_sp_4h_bin_*, msm_a_ncld_4h_bin_*	MSM 10 m wind (m s⁻¹), surface pressure (Pa), cloud cover (%)
<product>_flag	daily QC bitmask of each product (as above)

Every forcing variable carries units, long_name, source and the rule used for days with missing hours (completeness) in its attributes.

discharge.zarr (spatial = the 1,189 gauges' catchments): discharge — daily mean discharge (m³ s⁻¹), valid when at least 20 of the 24 hours were observed; discharge_qc — daily screening bitmask (bit k−1 set when an hour was removed by rule k: 1 sentinel value, 2 placeholder value, 3 negative value, 4 outside the physical range, 5 outside the station's envelope, 6 isolated excursion, 7 manual exclusion).

dams.zarr (spatial = the 116 dams' catchments): inflow, outflow (daily means, m³ s⁻¹), storage (absolute volume, 10³ m³, first valid value of the day), storage_rate (%), prcp (mm per day), each with a *_qc bitmask.

splits/

pur_folds_10.json — the evaluation design:

• evaluation_set (979 gauges): every quality-controlled gauge with at least min_valid_days (364) valid days in the window 2009-01-01 to 2020-12-30, dam-affected gauges included;
• training_pool (332 gauges): the evaluation gauges with no dam upstream that are not dam outlets and are not restricted_range;
• folds (10): each holds whole river basins (basin_folds maps a basin's outlet node to its fold); train = training-pool gauges of the other nine folds, val = training-pool gauges of the fold (monitoring only), test = every evaluation gauge of the fold, so each evaluation gauge is tested once;
• gauges_with_discharge: the gauges over which the discharge scale is computed; restricted_range, qc_excluded: the screened gauges; population: the counts at each step.

time_blocks.json — the three temporal blocks (B1 2009–2012, B2 2013–2016, B3 2017–2020) of the PUR-T setting, in which one block is also withheld from training.

qc/

File	Content
station_decisions.parquet	outcome of the station review: decision (exclude: out of the evaluation set; keep_no_train: evaluated, not trained on; keep) and its category
discharge_exclusions.parquet	periods removed from the discharge records before aggregation
forcing_masks.parquet	source-product time steps masked as defective (constant fill fields)

simulation/ (optional archive)

Path	Content
national_simulation.zarr	discharge(time, spatial) (m³ s⁻¹, routed discharge at every catchment outlet) and runoff(time, spatial) (mm day⁻¹, each catchment's own runoff before routing), 2009-01-01 to 2020-12-30, from the headline model trained on all 332 training-pool gauges (seed 0)
pur_evaluation/	series.zarr: sim(seed, spatial, time) and obs(spatial, time) (m³ s⁻¹) at the 979 evaluation gauges, each simulated by the model of the fold that held out its basin, 10 seeds; fold(spatial); scores.csv: per-gauge scores (NSE, KGE' and its components, nRMSD, flow-duration-curve signatures, lag, monthly NSE) per period (window, B1–B3), averaged over seeds; scores_per_seed.csv
purt_evaluation/	the same for the PUR-T setting: every day is simulated by a model that saw neither the gauge's basin nor the day's 4-year block

In the PUR setting the headline model reaches a median NSE of 0.819 (± 0.003 over seeds) and a median KGE' of 0.728 over the 979 evaluation gauges (2009–2020).

Sources

Data	Provider / reference
River discharge, dam records, station metadata	MLIT Water Information System (http://www1.river.go.jp)
In-situ gauge-interpolated precipitation	Hatono, M., et al. (2022), Harvard Dataverse, doi:10.7910/DVN/J215UY
Radar and gauge-adjusted radar precipitation, MSM analyses	Japan Meteorological Agency
ERA5-Land	Muñoz-Sabater, J., et al. (2021), Copernicus Climate Change Service
Flow directions and elevation (network, topography)	J-FlwDir, Yamazaki, D., et al. (2020)
Land use and land cover	JAXA EORC High-Resolution Land Use and Land Cover Map of Japan, Hirayama, S., et al. (2022)
Catchment attributes	HydroATLAS, Linke, S., et al. (2019)

The values distributed here are catchment means and daily aggregates derived from these products; the terms of use of each provider apply to them, and the original providers should be cited. See LICENSE.md.

Citation

Please cite the paper above (in review) and the original data providers.