Datasets:
D3P-JP: Dataset for Distributed Discharge Prediction over Japan
Version 1.0.
This repository holds the most recent version of D3P-JP and is the fastest way to download it. The snapshot taken at the time of publication of the paper is archived permanently on Zenodo: https://doi.org/10.5281/zenodo.24212554.
Daily meteorological forcing, static descriptors, quality-controlled river discharge observations and dam records on a river network covering Japan, prepared for learning spatially distributed hydrological simulations at national scale. The dataset is described in
Hascoet, T., Pellet, V., Oishi, S., & Miyoshi, T. Leveraging Japan's streamflow records for end-to-end learning of hydrological simulations at national scale.
The simulation obtained from this dataset is distributed separately as D3P-JP-Q (simulated daily discharge at the outlet of every catchment of the network; https://doi.org/10.5281/zenodo.24212556 and https://huggingface.co/datasets/tristhas/D3P-JP-Q). Code to reproduce the experiments of the paper: https://github.com/TristHas/jpdh.
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 979 quality-controlled gauges (Ministry of Land, Infrastructure, Transport and Tourism, MLIT) |
| Dams | 852 dams with their characteristics (Japan Dam Foundation); daily inflow, outflow and storage of the 116 dams monitored by the 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 catchment descriptors (HydroATLAS attributes, JAXA land use, J-FlwDir topography, climatologies) and 17 channel descriptors |
| Benchmark | 10 folds of whole river basins (PUB), 979 evaluation gauges of which 332 are training gauges, three 4-year temporal blocks (PUB-T) |
Files
Two tar archives; MANIFEST.json lists their sizes and SHA-256 checksums.
Both extract into the same directory and add checksums/<archive>.json, the
SHA-256 of every file they contain.
| Archive | Size | Contents |
|---|---|---|
d3p-jp_core.tar |
1.2 GB | everything needed to train and run the model in its default configuration: network/, statics/, daily/forcing.zarr, daily/discharge.zarr, daily/dams.zarr, splits/, qc/ |
d3p-jp_panels.tar |
13.3 GB | daily/forcing_panels.zarr: the inputs of the sensitivity experiments |
The core archive is enough for the default configuration of the model; the panels archive is needed only for the sensitivity experiments. Download the core archive alone, or both, and extract them into the same directory:
wget https://huggingface.co/datasets/tristhas/D3P-JP/resolve/main/d3p-jp_core.tar
wget https://huggingface.co/datasets/tristhas/D3P-JP/resolve/main/d3p-jp_panels.tar # optional, 13.3 GB
mkdir D3P-JP
tar -xf d3p-jp_core.tar -C D3P-JP
tar -xf d3p-jp_panels.tar -C D3P-JP # only if downloaded
The code repository (https://github.com/TristHas/jpdh) does the same and
verifies the files: python scripts/download_data.py fetches the core archive,
python scripts/download_data.py --panels both.
import pandas as pd, xarray as xr
nodes = pd.read_parquet("D3P-JP/network/nodes.parquet") # one row per catchment
q = xr.open_zarr("D3P-JP/daily/discharge.zarr")["discharge"] # (time, spatial)
Time series are Zarr stores (format 3) of dimensions (time, spatial) in
single precision, missing values being NaN; tables are (Geo)Parquet. Every
catchment is identified by a 64-bit integer, node_id (spatial in the
stores), shared by all files.
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), x, y (outlet pixel in the J-FlwDir grid), catchment_area and drainage_area (= upa, km²), channel_length_km, out_node (downstream catchment, −1 at a basin outlet), is_root (basin outlet), is_leaf (headwater), is_gauge (a gauging station or a monitored dam is at the outlet), is_dam, station_id, dam_id, mlit_dam_id (monitored dams) |
edges.parquet |
source → target: each catchment and the one it drains into |
catchments.parquet |
catchment polygons (GeoParquet, EPSG:4326) |
channels.parquet |
channel geometries (LineString; Point for one-cell channels) and river_channel, the raster cells of the channel |
stations.parquet |
the 1,807 gauging stations and 852 dams of the network, index node_id: location, names (Japanese), river system, snapping distance, upstream area; discharge_series (the series of the station is in daily/discharge.zarr), discharge_status, discharge_status_detail and valid_days_2009_2020 (why it is or is not, see Quality control of the discharge records); experiment_set, the use of the station in the experiments of the paper (train: one of the 332 training gauges, which are also evaluated; eval: evaluated only; dropped: no series distributed; empty for dams); for dams, their characteristics (dam_* columns, see Dams below) |
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.
Dams
852 dams are outlets of catchments of the network. They are the rows of
network/stations.parquet with is_dam true (is_dam also marks their
catchments in nodes.parquet). Their characteristics come from the Dam
Handbook (ダム便覧) of the Japan Dam Foundation and are the dam_* columns of
stations.parquet:
| Column | Content |
|---|---|
dam_id |
identifier of the dam in the Dam Handbook |
mlit_dam_id |
identifier of the dam in the MLIT Water Information System, for the 116 monitored dams (empty for the others) |
dam_name |
name of the dam (romanized) |
dam_purpose |
purposes of the dam, one letter each: F flood control, N river maintenance flow and unspecified uses, A irrigation, W municipal water supply, I industrial water supply, P hydropower, S snow melting, R recreation |
dam_type |
G concrete gravity, R rockfill, A arch, E earthfill, GF combined concrete gravity and fill, HG hollow concrete gravity, GA gravity arch, CSG trapezoidal CSG, FA asphalt-facing fill, B buttress, FC asphalt-core fill, MA multiple arch |
dam_height_m, dam_crest_length_m |
height and crest length of the dam (m) |
dam_volume_1000m3 |
volume of the dam body (10³ m³) |
dam_catchment_area_km2 |
catchment area of the dam (km²) |
dam_water_surface_ha |
water surface of the reservoir (ha) |
dam_reservoir_capacity_1000m3 |
total storage capacity of the reservoir (10³ m³) |
dam_construction_start, dam_construction_end |
years in which the construction started and was completed (empty when not reported) |
The same rows give the location of each dam (lat, lon) and the area
upstream of it in the network (upstream_area_km2). The MLIT publishes records
for 116 of these dams: their daily inflow, outflow and storage
are in daily/dams.zarr (see daily/).
statics/
| File | Content |
|---|---|
runoff_statics.parquet |
382 catchment descriptors, raw physical values: 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 of the MSM analysis and in-situ precipitation (season_msm_*, season_gauge_*) and mean snow indicators (msm_snow_avg, gauge_snow_avg) |
routing_statics.parquet |
17 channel descriptors: elevation, slope and upstream-area statistics, sinuosity, log_upa, log_slope, catch_area, … (NaN for headwater catchments, which have no channel) |
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 inputs of the default configuration:
| 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 (second 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 catchments of the 979 quality-controlled gauges):
discharge — daily mean discharge (m³ s⁻¹), the mean of the hourly values of
the day, given when at least 20 of the 24 hours are valid; a daily value equal
to zero is set to missing; discharge_qc — daily screening bitmask (bit k−1
set when an hour of the day, or the day, was removed by rule k of the table
below). See Quality control of the discharge records.
dams.zarr (spatial = the catchments of the 116 monitored dams): inflow,
outflow (daily means, m³ s⁻¹) and storage (held volume, 10³ m³, first valid
value of the day; negative values are set to missing), each with a *_qc
bitmask.
Quality control of the discharge records
Stations. daily/discharge.zarr holds only the series that pass the
quality control of the paper (Supporting Information, Text S3): the
979 stations with a calibrated rating curve whose record is not
removed by a manual rule, is not excluded by the review of the records, is not
calibrated over a restricted range of flows, and holds at least
364 valid days between 2009-01-01 and 2020-12-30.
The records of the other stations are not distributed. The reason is recorded
for each station in network/stations.parquet, column discharge_status,
assigned in the order of the table below (Table S5 of the paper);
discharge_status_detail gives the sub-category of Table S5 for the stations
removed by a manual rule or by the review, and valid_days_2009_2020 the
number of valid daily values of the station between 2009-01-01 and
2020-12-30. The rules and decisions themselves are in qc/.
discharge_status |
Stations | Meaning |
|---|---|---|
stage_only |
508 | the station reports stage only (no rating curve) |
no_digital_series |
39 | registered stage-discharge station without a digital discharge series |
removed_by_manual_rule |
71 | whole record removed by a manual rule (qc/discharge_exclusions.parquet, action drop_station) |
no_data_2009_2020 |
105 | no valid daily value between 2009-01-01 and 2020-12-30 |
fewer_than_364_valid_days |
21 | fewer than 364 valid daily values in that period |
restricted_range |
26 | rating curve calibrated over a restricted range of flows |
excluded_by_review |
58 | excluded by the review of the records (qc/station_decisions.parquet) |
released |
979 | series in daily/discharge.zarr |
The last seven rows are the 1,299 registered stage-discharge stations of the network.
Values. Before aggregation to days, the hourly records pass the manual
rules of qc/discharge_exclusions.parquet and then an automatic screening,
applied in the order below (rules 1 to 7). Two rules act on the daily values:
a daily discharge equal to zero is set to missing (rule 8), and so is a negative
held volume of a dam (rule 3). A removed value is set to missing, never infilled,
clipped or rescaled. A is the drainage area (km²), the larger of the value
reported by the MLIT and the upstream area of the network; a and b are the
nearest valid values before and after an hour or block.
| Code | Rule | Records | Removes |
|---|---|---|---|
| 1 | Missing-value code | all | values ≤ −10⁶ or ≥ 2·10⁸, and the provider's codes (e.g. −9999.99, 777777.8) |
| 2 | Placeholder | dams | zeros under the provider's per-variable mark of a missing value (-, #, $) |
| 3 | Negative | discharge, dam inflow, outflow, held volume | values below 0 |
| 4 | Range | held volume | volume above 1.2 times the capacity |
| 5 | Flood envelope | discharge, dam inflow, outflow | Q > 100 C(A) A, with C(A) = 1.303 (0.386 A)^(0.936 A^(−0.048) − 1) m³ s⁻¹ km⁻² |
| 6 | Isolated excursion | discharge and dam inflow (A ≥ 20 km²); held volume | one hour at least 10 times above, or 10 times below, both a and b when these agree, by at least 0.1 A m³ s⁻¹ (flows); a block of up to 6 hours away from both a and b by more than the volume carried in one hour by the envelope flow (held volume) |
| 7 | Manual exclusion | discharge | the periods of qc/discharge_exclusions.parquet |
| 8 | Zero | discharge | daily values equal to zero |
The largest genuine hourly peaks of the records reach 58 on the scale of C(A), so that the envelope at 100 leaves a wide margin. C(A) is written as it is computed; it differs from the textbook Creager expression by the constant factor 1/0.386.
splits/
pub_folds_10.json — the benchmark of the paper (Section 4.2):
evaluation_set(979 gauges): the gauges ofdaily/discharge.zarr, dam-affected gauges included (645);training_pool(332 gauges): the evaluation gauges with no dam upstream that are not dam outlets;folds(10): each holds whole river basins (basin_foldsmaps the outlet node of a basin to its fold);train= training gauges of the other nine folds,val= training gauges of the fold (monitoring only),test= every evaluation gauge of the fold, so that each evaluation gauge is tested once;gauges_with_discharge: the gauges over which the scale of the discharge is computed (=evaluation_set);window,min_valid_daysandpopulation: the period and the counts of the benchmark.
The column experiment_set of network/stations.parquet gives the same
assignment station by station: train for the gauges of training_pool,
eval for the other gauges of evaluation_set, dropped for the stations
without a distributed series.
time_blocks.json — the three temporal blocks (B1 2009–2012, B2 2013–2016,
B3 2017–2020) of the PGB-T and PUB-T settings, in which one block is withheld from
training.
qc/
| File | Content |
|---|---|
discharge_exclusions.parquet |
the 97 manual rules applied to the discharge records: 71 remove a station as a whole (drop_station), 26 mask dated segments or implausible values of a record |
station_decisions.parquet |
outcome of the review of the records, 185 stations: decision (exclude: series not distributed; keep_no_train: evaluated, not trained on; keep), its category, and discharge_series (the series is in daily/discharge.zarr) |
Citation
Please cite the paper above and this record.
- Downloads last month
- 35