README.md

---
license: mit
task_categories:
  - time-series-forecasting
tags:
  - air-quality
  - atmospheric-science
  - pm25
  - pm10
  - east-asia
  - cmaq
  - weather-forecasting
  - world-model
pretty_name: PLUME - East Asia Air Quality (0.25 deg, hourly, 2016-2024)
size_categories:
  - 100B<n<1T
---

# PLUME — East Asia air-quality dataset

**0.25° · hourly · 2016–2024 · 117 × 222 grid**

The preprocessed archive behind **PLUME** — *Physically conditioned correction and ordered uncertainty for multiday particulate forecasting over East Asia*.

KAIST · National Institute of Environmental Research (NIER) · Ajou University

📦 **Code:** [github.com/kaist-cvml/PLUME](https://github.com/kaist-cvml/PLUME)
🔁 **Predecessor:** [`2na-97/FAKER-Air`](https://huggingface.co/datasets/2na-97/FAKER-Air) — the same region at 27 km

---

## What is in here

Two independent trees on one shared grid, plus the trained model weights.

| Path | Contents | Download | On disk |
|---|---|---|---|
| `cmaq_asia_1h_0p25/YYYY.tar.zst` | CMAQ reanalysis — 12-species speciated aerosol and gas fields, surface proxy stack | 9.3 GB/yr | 17 GB/yr |
| `obs_asia_1h_0p25/YYYY.tar.zst` | Ground-station observations gridded to the same cells, with validity and region masks | 0.4 GB/yr | 5 GB/yr |
| `checkpoints/` | Trained PLUME weights, deployed configuration at the four seeds the paper reports | 158 MB | — |

Years available: **2016 – 2024** — 88 GB to download, 196 GB once extracted. One shard per (tree, year), so you can fetch only what you need.

---

## Quick start

```bash
pip install huggingface_hub
git clone https://github.com/kaist-cvml/PLUME.git && cd PLUME

# evaluation years only -- enough to reproduce every reported number (19 GB download, 44 GB on disk)
python scripts/download_data.py --years 2023 2024

# the full training record
python scripts/download_data.py --all

# one tree, one year
python scripts/download_data.py --years 2023 --tree obs

# the trained weights
python scripts/download_data.py --checkpoints
```

The helper downloads, extracts and cleans up the shard. Re-running is safe — a year already extracted is skipped and the Hub download resumes.

<details>
<summary><b>Without the helper script</b></summary>

```python
from huggingface_hub import hf_hub_download

shard = hf_hub_download(
    repo_id="2na-97/PLUME",
    repo_type="dataset",
    filename="obs_asia_1h_0p25/2023.tar.zst",
)
```

then

```bash
mkdir -p data/obs_asia_1h_0p25
tar --use-compress-program=unzstd -xf "$shard" -C data/obs_asia_1h_0p25
```

Several shards at once:

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="2na-97/PLUME", repo_type="dataset",
    allow_patterns=["*/2023.tar.zst", "*/2024.tar.zst"],
    local_dir="downloads",
)
```
</details>

---

## Layout after extraction

```
data/
├── cmaq_asia_1h_0p25/YYYY/MM/DD/NIER_27_01/
│   ├── YYYYMMDD_x_aurora_surf.npy      float32 [25, 13, 117, 222]  surface stack (memory-mapped)
│   ├── YYYYMMDD_x_cmaq_cams_sp12.npy   float16 [25, 12, 117, 222]  speciated aerosol + gas
│   ├── YYYYMMDD_meta.json
│   └── cams_lat.npy, cams_lon.npy, cams_grid_meta.json
└── obs_asia_1h_0p25/YYYY/MM/DD/NIER_27_01/
    ├── YYYYMMDD_y_obs_cams.npz         keys: obs_values, obs_mask, obs_count,
    │                                         region_id, source_id, station_count, lat, lon
    ├── YYYYMMDD_y_obs_cams_pmvals.npy  float32 [25, 2, 117, 222]  PM2.5, PM10 in ug/m3
    ├── YYYYMMDD_y_obs_cams_pmmask.npy  uint8   [25, 2, 117, 222]  1 where a station reported
    ├── YYYYMMDD_y_obs_cams_region.npy  int16   [25, 117, 222]     region ID per cell
    ├── YYYYMMDD_obs_meta.json
    ├── YYYYMMDD_station_map.csv        station-to-cell assignment for that day
    └── cams_lat.npy, cams_lon.npy, cams_grid_meta.json
```

The leading `T = 25` axis is a day's 00–23 hourly slots plus one wrap slot. `*_x_cmaq_cams_sp12.npy` is the precomputed 12-channel selection the model reads; the full 24-channel concentration stack (`*_x_cmaq_cams.npz`) is retained for one sample day per year only, and `data/dataset.py` falls back to it when the `sp12` file is absent.

### Grid

| | |
|---|---|
| Resolution | 0.25° |
| Shape | 117 (lat) × 222 (lon) |
| Domain | 19.6 °N – 48.0 °N, 104.8 °E – 159.6 °E |
| Grid name | `NIER_27_01` |
| Cadence | hourly |

Coordinates ship alongside every day as `cams_lat.npy` / `cams_lon.npy`.

### Channel orders

**Observations** (`OBS_SHORT_NAMES` in `plume/gridio.py`):

```
pm2p5, pm10, so2, no2, o3, co
```

**CMAQ concentrations** (`CMAQ_CONC_ORDER` in `plume/species.py`):

```
SO4_25, NH4_25, NO3_25, ORG_25, EC_25, MISC_25, PM2P5,
tcso2, tcco, tcno2, O3, NO, NOx,
SO4_10, NH4_10, NO3_10, ORG_10, EC_10, MISC_10, PM10,
ISOPRENE, OLES, AROS, ALKS
```

**Surface proxy stack** (`*_x_aurora_surf.npy`):

```
2t, 10u, 10v, msl, pm1, pm2p5, pm10, tcco, tc_no, tcno2, gtco3, tcso2, source_mask
```

### Regions

`*_y_obs_cams_region.npy` carries a region ID per cell: the 19 Korean provinces (`KR_SEOUL`, `KR_INCHEON`, … `KR_JEJU`) plus Chinese and other regions appended at runtime. The paper reports three strata — the broad East-Asian domain, Korea, and a prespecified Chinese dust-source corridor.

---

## Loading a day

```python
import numpy as np

day = "data/obs_asia_1h_0p25/2023/03/21/NIER_27_01"
obs = np.load(f"{day}/20230321_y_obs_cams.npz")
print(obs.files)

pm = np.load(f"{day}/20230321_y_obs_cams_pmvals.npy")   # [25, 6, 117, 222]
region = np.load(f"{day}/20230321_y_obs_cams_region.npy")
lat = np.load(f"{day}/cams_lat.npy")
lon = np.load(f"{day}/cams_lon.npy")
```

In practice you want `DirectLeadDataset` from the code repository, which assembles issue-time inputs, history windows, validity masks and the motion estimate for you.

---

## Intended use, and one caveat that matters

The task is **exceedance forecasting** from +6 h to +120 h in 6 h steps, at the public-guidance boundaries PM<sub>2.5</sub> > 35 and PM<sub>10</sub> > 80 µg/m³.

> **CMAQ at the target valid time must never be a model input.**
> In PLUME it appears only as a training-time teacher label; everything the model reads is available at issue time. The archive contains the full hourly record, so it is *possible* to build a leaking pipeline from it by accident. The `--causal-student` path in the reference code enforces the split — if you build your own loader, enforce it yourself, or your scores will not mean what you think they mean.

Suggested protocol, matching the paper: train on **2016–2022**, select on **2023**, evaluate once on **2024**.

---

## Citation

```bibtex
@article{plume2025,
  title   = {Physically conditioned correction and ordered uncertainty
             for multiday particulate forecasting over East Asia},
  author  = {Kang, Inha and Ryu, Wonjeong and Hong, Sung-Chul and
             Lee, Jae-Bum and Ban, SooJin and Jeong, Seongeun and
             Kang, Yoon-Hee and Kim, Soontae and Shim, Hyunjung},
  year    = {2025},
  note    = {Code: https://github.com/kaist-cvml/PLUME}
}
```

## Acknowledgements

Observations come from the Korean national air-quality monitoring network and regional partners; the reanalysis is CMAQ. Preprocessing and grid conventions began as a fork of the [Aurora](https://github.com/microsoft/aurora) stack.

## License

MIT for the packaging and derived arrays. Underlying observations and reanalysis remain subject to the terms of their original providers; check those before redistribution.