Nathon-t/dianchi-water

Dianchi Water

A high-frequency (4-hourly) multi-station surface water quality dataset covering 22 monitoring stations, 9 water quality variables, and 3 years (2022–2024) in the Dianchi Lake basin, China.

Natural missing rate: 19.8% (>99% block-structured).

Contents

data/
  dianchi_data_df.parquet            # Main dataset (116,783 records)
  dianchi_station_distance_km.csv    # 22×22 pairwise Haversine distance (km)
scripts/
  build_adjacency.py                 # Generate adjacency matrices + heatmaps

Quick start

import pandas as pd

df = pd.read_parquet("data/dianchi_data_df.parquet")
print(df.shape)                 # (116783, 11)
print(df.columns.tolist())      # ['tm', 'station', 'TEM', 'PH', ...]
print(df["station"].nunique())  # 22

Column descriptions

Column	Type	Description
tm	datetime64[ns]	Timestamp (4-hourly cadence)
station	string	Monitoring station name (English)
TEM	float64	Water temperature (°C)
PH	float64	pH
DO	float64	Dissolved oxygen (mg/L)
CON	float64	Electrical conductivity (μS/cm)
NTU	float64	Turbidity (NTU)
IMN	float64	Permanganate index (mg/L)
NH_N	float64	Ammonia nitrogen (mg/L)
TP	float64	Total phosphorus (mg/L)
TN	float64	Total nitrogen (mg/L)

Dataset scale

• Records: 116,783
• Stations: 22
• Variables: 9 target water quality variables
• Time range: 2022-01-01 to 2024-12-30
• Frequency: 4-hourly (6 observations/day)
• Full 4h grid per station: 6,568 time steps
• Aggregate missing rate: 19.8% (on full 4h grid, 9 variables)

Station observation rates

Observation rates after reindexing to the full 4-hourly grid (6,568 steps per station):

Station	Records	Obs rate
Daguanhe Inlet	5,866	89.3%
Chuanfang Bridge	5,866	89.3%
Duanqiao	5,848	89.0%
Caohai Center	5,844	89.0%
Xinhecun Inlet	5,837	88.9%
Guanyinshan West	5,788	88.1%
Wangda Bridge	5,767	87.8%
Huilong Village	5,762	87.7%
Dianchi South	5,758	87.7%
Luojiaying	5,752	87.6%
Baofengcun Inlet	5,747	87.5%
Haikou West	5,676	86.4%
Jiangwei Lower Sluice	5,627	85.7%
Dayuxiang Tuluocun Inlet	5,512	83.9%
Huiwan Central	5,497	83.7%
Baiyukou	5,446	82.9%
Yanjiancun Bridge	5,429	82.7%
Guanyinshan East	5,354	81.5%
Guanyinshan Central	4,835	73.6%
Dongdahe Dianchi Inlet	4,686	71.3%
Cigang River Inlet	2,876	43.8%
Xiyuan Tunnel	2,010	30.6%

Variable summary statistics

Variable	Unit	Missing%	Mean	Std	Min	P5	Median	P95	Max	Skew
TEM	°C	19.3%	18.73	4.23	0.00	11.64	19.12	24.80	36.10	−0.2
PH	—	20.0%	8.26	0.71	0.00	7.36	8.30	9.13	10.99	−4.7
DO	mg/L	19.3%	7.60	3.16	0.00	2.89	7.43	13.22	29.99	1.1
CON	μS/cm	19.3%	514.59	131.85	0.00	348.10	490.30	753.12	1780.86	1.8
NTU	NTU	19.5%	20.71	44.97	0.00	2.60	14.10	51.99	9918.65	103.0
IMN	mg/L	20.5%	4.64	2.27	0.00	1.43	4.36	8.15	31.51	0.6
NH_N	mg/L	20.2%	0.21	0.49	0.00	0.03	0.04	0.78	16.73	9.6
TP	mg/L	20.1%	0.08	0.07	0.00	0.02	0.07	0.17	3.55	10.2
TN	mg/L	20.1%	3.24	2.24	0.00	0.90	2.41	7.66	45.03	1.3

Adjacency construction

The script scripts/build_adjacency.py reads the distance matrix and constructs adjacency matrices using a linear-decay weight:

$$w_{ij}=\max \text{\hspace{0.17em}⁣}(0,1-d_{ij}\mathrm{/}\tau )w\_\{ij\}\; =\; \backslash max\backslash !\backslash bigl(0,\backslash ;\; 1\; -\; d\_\{ij\}\; /\; \backslash tau\backslash bigr)$$

where $d_{ij}$ is the geodesic distance between stations $i$ and $j$, and $\tau$ is a user-specified threshold in kilometres.

# Default thresholds (10, 15, 20, 25, 30 km)
python scripts/build_adjacency.py

# Single threshold
python scripts/build_adjacency.py --threshold-km 20

# Custom thresholds, custom output directory
python scripts/build_adjacency.py --thresholds-km 5,10,20 --output-dir ./outputs

Dependencies: numpy, pandas, matplotlib

Privacy note

Raw station coordinates are not included. The pairwise distance matrix preserves all information needed for distance-based graph construction without exposing exact locations.

Citation

If you use this data, please cite the accompanying paper:

@article{anonymous2026dianchiwater,
  title   = {A High-Frequency Multi-Station Surface Water Quality
             Dataset and Mask-View Augmentation Benchmark for
             Time-Series Imputation},
  author  = {Anonymous},
  year    = {2026},
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY-4.0) license.