README.md

---
license: cc-by-4.0
pretty_name: A Global Multi-Decadal Convection Tracking Database from ISCCP-H (1983-2017)
language:
- en
tags:
- climate
- atmospheric-science
- earth-science
- satellite
- convection
- cloud-tracking
- mesoscale-convective-systems
- brightness-temperature
- tobac
- ISCCP
size_categories:
- 10M<n<100M
configs:
- config_name: default
  data_files:
  - split: train
    path: data/*.parquet
---

# A Global Multi-Decadal Convection Tracking Database from ISCCP-H (1983-2017)

ISCCP-H CT: a global, 34-year (July 1983 - June 2017) convection-tracking database derived from International Satellite Cloud Climatology Project H-series (ISCCP-H) infrared observations, produced with the Tracking and Object-Based Analysis of Clouds (tobac) framework.

This dataset accompanies the manuscript:

> Luo, Z. J., Wang, L.-P., Selevich, Y., Takahashi, H., Wu, C.-L., Jhang, H., Lin, S.-C., van den Heever, S. C., Machado, L. A., Rossow, W., and Freeman, S. (2026). A Global Multi-Decadal Convection Tracking Database from ISCCP-H: Dataset Description and Convective Lifecycle Analysis. *Earth and Space Science* (in review).

Archival copy with DOI: https://doi.org/10.5281/zenodo.21505419 (activates on Zenodo publication). Please cite the Zenodo DOI, not this repository.

## Description

Convective cloud systems were detected and tracked in ISCCP-H infrared brightness temperature (TB) fields (3-hourly temporal resolution, 10-km spatial resolution) using the tobac framework. Detection used TB minima with dual thresholds of 245 K and 220 K, segmentation at 245 K, and predictive linking with a maximum velocity of 30 m/s.

Each record is one observation of one convective system (feature) at one 3-hourly time step, linked into cell families by the tracker so that the full Lagrangian lifecycle of each system can be reconstructed. For every tracked system the dataset provides position and elliptical geometry, brightness temperature statistics (minimum, mean, maximum, standard deviation, percentiles), size measures (pixel counts, equivalent radius, deep-convective pixel fractions), lifecycle timing, tracking-quality diagnostics, and environmental context (surface type, wind, cloud optical thickness).

## Coverage and Volume

| Property | Value |
|---|---|
| Temporal coverage | July 1983 - June 2017 (34-year record; 1983 is Jul-Dec only, 2017 is Jan-Jun only) |
| Temporal resolution | 3-hourly |
| Spatial coverage | Global, 60S - 60N, 0 - 360E |
| Source resolution | 10 km (ISCCP-H infrared) |
| Records | 67,257,982 system-level observations |
| Columns | 57 documented variables |
| Format | Apache Parquet (Snappy), one file per calendar year |
| Total size | approximately 11.8 GB (35 files) |
| Metadata | CF-1.8 attributes embedded in each Parquet file |

## Usage

Query directly from the Hub without downloading everything.

DuckDB (SQL over all years):

```python
import duckdb

duckdb.sql("""
    SELECT year, COUNT(*) AS n_obs, ROUND(AVG(minTB_feature), 1) AS mean_min_tb_K
    FROM 'hf://datasets/NTU-CompHydroMet-Lab/ISCCP-H-CT/data/*.parquet'
    GROUP BY year ORDER BY year
""").show()
```

Polars (lazy, with predicate pushdown). Note: Polars' native `hf://` scanner currently fails on namespaced repos ("url-encoded slash" error, upstream Polars bug), so route through `HfFileSystem` and PyArrow:

```python
import polars as pl
import pyarrow.dataset as ds
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
files = fs.glob("datasets/NTU-CompHydroMet-Lab/ISCCP-H-CT/data/*.parquet")
dataset = ds.dataset(files, filesystem=fs, format="parquet")

# scan/filter/group_by only build a lazy query plan; nothing is read yet.
# .collect() triggers execution. When scanning remotely, collect small
# results (aggregations, subsets); for bulk row-level extraction, download
# the files first and read locally (see below).
summary = (
    pl.scan_pyarrow_dataset(dataset)
    .filter((pl.col("year") == 2016) & (pl.col("minTB_feature") < 220))
    .group_by("land_water_mask")
    .agg(
        pl.len().alias("n_obs"),
        pl.col("radius").mean().alias("mean_radius_km"),
    )
    .collect()
)
```

Pandas (single year; pandas >= 2.0 recommended for Arrow-backed dtypes):

```python
import pandas as pd

df = pd.read_parquet(
    "hf://datasets/NTU-CompHydroMet-Lab/ISCCP-H-CT/data/ISCCP-H_CT_2016.parquet",
    dtype_backend="pyarrow",  # lower memory, correct nullable types
)
```

For repeated analysis, download once and read locally (fastest):

```bash
hf download NTU-CompHydroMet-Lab/ISCCP-H-CT --repo-type dataset --include "data/*.parquet" --local-dir isccp_h_ct
```

```python
import polars as pl

df = pl.scan_parquet("isccp_h_ct/data/*.parquet")  # all 35 years as one table
```

### Requirements

- `pyarrow >= 15` (Parquet engine)
- `polars >= 1.17` or `pandas >= 2.0` (pandas 1.x can read Parquet but lacks Arrow-backed dtypes and robust `hf://` support)
- `duckdb >= 1.0` for `hf://` SQL access
- `huggingface_hub` for `HfFileSystem` and the `hf` CLI

## Data Schema

Key columns (full schema in `docs/DATA_DESCRIPTION.md` and `docs/PARQUET_METADATA.md`):

| Column | Description | Units |
|---|---|---|
| global_cell_id | Cell family ID, format YYYY_cellid (unique within year) | - |
| datetime | UTC timestamp | - |
| latitude / longitude | System centroid | degrees |
| minTB_feature | Minimum brightness temperature (deep convection < 220 K) | K |
| radius | Equivalent circular radius | km |
| lifetime_hours | Total cell lifetime | hours |
| eccentricity | Ellipse eccentricity (0 = circular) | - |
| land_water_mask | Surface type: "land" or "water" | flag |
| percent_overlap | Tracking quality (higher = better) | percent |
| pixel_count | System size in pixels | - |

Missing values are IEEE 754 NaN; no imputation is performed.

## Recommended Quality Control

- `percent_overlap > 50` (reliable tracking)
- `pixel_count > 10` (exclude marginal detections)
- `minTB_feature` between 180 and 300 K (physically valid range)

## Known Limitations

1. `frame`, `feature`, and `cell` IDs reset every year; use `global_*` variants for multi-year work. Cell families are not linked across the year boundary.
2. 1983 covers July - December only; 2017 covers January - June only.
3. Each file contains a legacy pandas index column `__index_level_0__` with no physical meaning; ignore it.
4. No coverage poleward of 60 degrees latitude.
5. The embedded flag_values attribute for `wind_dir_letter` lists 8 directions, but the data contain all 16 compass points.

## Provenance

- Source data: ISCCP H-series Climate Data Record (NOAA NCEI), https://doi.org/10.7289/V5QZ281S ; Young et al. (2018), https://doi.org/10.5194/essd-10-583-2018
- Tracking software: tobac, Heikenfeld et al. (2019), https://doi.org/10.5194/gmd-12-4159-2019

## Citation

Please cite both the dataset and the accompanying manuscript:

```bibtex
@dataset{isccp_h_ct_2026,
  title  = {A Global Multi-Decadal Convection Tracking Database from ISCCP-H (1983-2017)},
  author = {Luo, Zhengzhao J. and Wang, Li-Pen and Selevich, Yuliya and Takahashi, Hanii and Wu, Chun-Liang and Jhang, Heng and Lin, Sung-Che and van den Heever, Susan C. and Machado, Luiz A. and Rossow, William and Freeman, Sean},
  year   = {2026},
  publisher = {Zenodo},
  doi    = {10.5281/zenodo.21505419}
}
```

## Contact

Corresponding author: Zhengzhao Johnny Luo, The City University of New York (zluo@ccny.cuny.edu)

Maintained by the NTU CompHydroMet Lab.