anonymous-dianchi-2026/dianchi-water

---
license: cc-by-4.0
task_categories:
- time-series-forecasting
- tabular-regression
language:
- en
tags:
- water-quality
- imputation
- time-series
- environmental-monitoring
- benchmark
size_categories:
- 100K<n<1M
---
# 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

```python
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:

$$w_{ij} = \max\!\bigl(0,\; 1 - d_{ij} / \tau\bigr)$$

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

```bash
# 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)](https://creativecommons.org/licenses/by/4.0/)
license.