docs/DATA_DESCRIPTION.md

ISCCP_HXG Convective Systems Dataset Description

Formal description of the ISCCP_HXG convection tracking dataset: source data, tracking configuration, data concepts, and statistical characteristics.

This document describes the scientific content of the dataset. For the distributed Parquet format, schema tables, and usage examples, see ../data/README.md. For the complete CF-1.8 metadata specification, see PARQUET_METADATA.md.

---

1. Overview

The ISCCP (International Satellite Cloud Climatology Project) HXG dataset is built from cloud observations by multiple geostationary and polar-orbiting satellites. Deep convective systems were identified, segmented, and tracked in the infrared brightness temperature (TB) fields using the tobac (Tracking and Object-Based Analysis of Clouds) Python package.

• Temporal coverage: July 1983 - June 2017 (34-year record; 1983 contains Jul-Dec only, 2017 contains Jan-Jun only)
• Temporal resolution: 3-hourly
• Spatial coverage: latitude -60 to +60 degrees, longitude 0 to 360 degrees east
• Convection detection threshold: TB < 245 K
• Distributed volume: 67,257,982 records, 11.8 GB (35 year-partitioned Parquet files)

The source data were produced as annual CSV files (54 columns, approximately 1.2 GB per year). The distributed Parquet dataset preserves all original columns and adds three year-prefixed global identifier columns (57 documented columns in total).

2. Core Concepts

Feature (convective system)

A convective system identified at a single time step; the basic unit of the dataset. Each feature represents one convective event at a specific time and location.

Cell (convective family)

A set of convective systems linked in time by the tracker. A single cell ID represents the temporal evolution trajectory of one convective system. Based on the 2017 January-June partition:

• Mean lifetime: 12.4 hours
• Maximum lifetime: 114 hours (4.75 days)
• Mean number of systems per family: 3.7

Frame (time step)

One 3-hourly time slice, used for computing the elliptical geometry parameters.

3. Tracking Configuration (tobac)

Parameter values used in the original tracking runs. Confirm against the accompanying manuscript Methods section before citing.

Feature detection

target = 'minimum'                    # detect TB minima (coldest cloud tops)
threshold = [245, 220]                # dual detection thresholds (K)
n_min_threshold = 2                   # minimum pixel count
position_threshold = 'weighted_diff'  # centroid positioning method
sigma_threshold = 1.5                 # smoothing parameter
n_erosion_threshold = 2               # erosion iterations

Segmentation

target = 'minimum'
threshold = 245                       # segmentation threshold (K)

Linking

method_linking = 'predict'            # predictive linking
v_max = 30                            # maximum velocity (m/s)
adaptive_stop = 2
adaptive_step = 0.95
stubs = 2                             # minimum number of linked time steps
subnetwork_size = 20
time_cell_min = 300                   # minimum duration (seconds)

4. Column Groups

The 57 documented columns fall into the following groups (full schema in ../data/README.md):

Group	Columns	Content
Identifiers	6	frame, feature, cell and year-prefixed global variants
Temporal	9	timestamp components, elapsed time, lifetime
Spatial	9	centroid, bounding box, ellipse center, surface type
Brightness temperature	15	min/mean/max, percentiles, standard deviation, gradient
Convective intensity	6	pixel counts, deep-convective fractions, optical thickness
Wind	3	speed, direction (degrees and 16-point compass)
Geometry	5	ellipse axes, inclination, eccentricity, spatial correlation
Tracking quality	4	overlap percentages, family system count

5. Statistical Characteristics (2017 January-June partition)

The following statistics were computed from the 2017 partition (1,042,051 records, 283,434 cell families) and are representative of the dataset.

System size

Pixel count:  median 102, mean 510, maximum 90,731 (strongly right-skewed)
Radius:       median 58 km, mean 90 km, maximum 1,622 km

Convective intensity

minTB_feature: median 218 K, mean 217 K, minimum 160 K (extreme deep convection)
Upper bound:   245 K (detection threshold)

Lifetime

Cell family lifetime: median 9 h, mean 12.4 h, maximum 114 h
Systems per family:   mean 3.7

Environment

• Ocean: 71.5 percent of observations
• Land: 28.5 percent of observations
• Dominant wind directions: W (13.0 percent), WNW (10.1 percent), WSW (10.0 percent)

6. Missing Data

Verified missing-value rates for the 2017 January-June partition:

Column group	Missing rate	Cause
Optical thickness (avg_optical_thickness, max_optical_thickness)	60.2 percent	Satellite retrieval limitations
Wind (wind_speed, wind_dir, wind_dir_letter)	27.2 percent	Reanalysis matching gaps
Overlap (percent_overlap, percent_non_overlap)	36.5 percent	Undefined at the first time step of each track
Geometry parameters	8.8 percent	Require multiple pixels
Basic convective parameters	6.9 percent	Very small system detection limits

Missing values are stored as IEEE 754 NaN; no sentinel values are used and no imputation is performed.

7. Usage Guidance

Quality control recommendations

1. Filter marginal detections: pixel_count < 5 may be unreliable.
2. Screen outliers: inspect minTB_feature < 160 K or radius > 1500 km before use.
3. Handle missing data explicitly: analyze optical thickness on the valid subset; exclude or impute wind fields as appropriate.
4. Tracking continuity: cell tracks can break when systems split or merge.

Suggested feature sets for clustering

Low-missing-rate base set:

features = [
    'pixel_count', 'radius', 'minTB_feature', 'avgTB_feature',
    'std_dev_tb', 'lifetime_hours', 'eccentricity', 'land_water_mask',
]

Extended set (requires missing-value handling):

features = [
    'pixel_count', 'radius', 'minTB_feature',
    'convective_fraction', 'pixels_below_220',
    'lifetime_hours', 'lifetime_num_cs',
    'wind_speed', 'wind_dir',
    'avg_optical_thickness',
    'semi_major', 'semi_minor', 'eccentricity',
]

Example science questions

1. Convective typology: short-lived versus long-lived, weak versus intense, isolated versus organized systems.
2. Environmental contrasts: ocean versus land convection; convective organization under different wind regimes.
3. Spatiotemporal evolution: diurnal cycle, seasonal patterns, regional climate differences.
4. Extremes: long-lived systems (> 24 h), very cold cloud tops (TB < 200 K), large systems (radius > 500 km).

8. Known Limitations

1. 3-hourly sampling cannot resolve rapid convective evolution.
2. Source pixel resolution is approximately 10 km (varying with latitude and satellite).
3. Track continuity breaks when complex systems split or merge.
4. No coverage poleward of 60 degrees latitude.
5. Cell IDs reset annually; systems alive across the year boundary receive new IDs in January (see ../data/README.md).

9. Quick Start

import polars as pl

DATASET = "Dataset/data/ISCCP_HXG_1983_2017_partitioned"

df = (
    pl.scan_parquet(DATASET)
    .filter(
        (pl.col("year") == 2017)
        & (pl.col("minTB_feature") < 220)
    )
    .select(["global_cell_id", "datetime", "latitude", "longitude",
             "minTB_feature", "radius", "lifetime_hours"])
    .collect()
)
print(f"Records: {len(df):,}")

10. References and Contacts

• tobac: https://tobac.readthedocs.io , https://github.com/tobac-project/tobac
• Machado, L.A.T., and W.B. Rossow, 1993: Structural characteristics and radiative properties of tropical cloud clusters. Mon. Wea. Rev., 121, 3234-3260.
• Machado, L.A.T., W.B. Rossow, R.L. Guedes, and A.W. Walker, 1998: Life cycle variations of mesoscale convective systems over the Americas. Mon. Wea. Rev., 126, 1630-1654.
• Schiffer, R.A., and W.B. Rossow, 1983: The International Satellite Cloud Climatology Project (ISCCP). Bull. Amer. Meteor. Soc., 64, 779-784.
• Young, A. H., et al., 2018: The International Satellite Cloud Climatology Project H-Series climate data record product. Earth Syst. Sci. Data, 10, 583-593.

Contacts:

• Yuliya Selevich: yuliya.selevich@gmail.com
• Zhengzhao Johnny Luo: z.johnny.luo@gmail.com
• Hanii Takahashi: hanii.takahashi@jpl.nasa.gov

---

Document version: 2.0 (English translation and revision of ISCCP_HXG_數據說明.md v1.0) Last updated: 2026-07-23 Applies to: ISCCP_HXG 1983-2017 partitioned Parquet dataset