mvelar/lightning-nowcasting

GOES-16 lightning nowcasting patches over South America

Dataset summary

This dataset contains sequences of co-located GOES-16 lightning and infrared observations over tropical and subtropical South America. It was built to train and evaluate machine-learning models that forecast lightning 10 to 120 minutes ahead.

Each sequence covers 128 × 128 grid cells (about 256 km on a side) and 4 hours at 10-minute steps: 12 input frames followed by 12 target frames. Each frame has eight channels: the flash-extent density (FED) from the Geostationary Lightning Mapper (GLM) and seven infrared bands of the Advanced Baseline Imager (ABI).

There are 77,976 sequences in total, of which 77,778 are used in the accompanying paper: the 198 sequences of the embargo day (2020-02-05) are released but not used. The development set has 71,010 sequences from 2019-04-03 to 2020-03-31, released as twelve monthly files and used for training and validation. The test set has 6,966 sequences from 36 days between September 2023 and February 2024, following the days of the Latin America lightning nowcasting benchmark (Cintineo, 2024).

The dataset accompanies the paper A Satellite Lightning Nowcasting Benchmark for South America: Temporal Cross-Attention in ConvLSTMs (Vela et al., WAIMLAp 2026).

• Curated by: Marco Vela
• License: CC-BY-4.0
• Code: doi:10.5281/zenodo.22979408 (Zenodo)
• Source data: GOES-16 GLM L2 LCFA and ABI L1b radiances, NOAA

Files

The data are hosted in this repository, one folder per month of the development set and one for the test set. Each folder holds three files:

Folder	Files
2019-04/ to 2020-03/	patches_YYYY-MM.nc, fed_true_position_YYYY-MM.nc, provenance_YYYY-MM.csv
test/	patches_test.nc, fed_true_position_test.nc, provenance_test.csv

The files hold about 296 GB in total. The code that builds them is in a separate Zenodo record (doi:10.5281/zenodo.22979408), which contains the scripts, the configuration and this card.

Dataset structure

Files

Each file is NetCDF-4 (HDF5) with three variables:

Variable	Type	Dimensions	Description
sequences	int16	(sample, time, channel, y, x) = (n, 24, 8, 128, 128)	Packed values; scale_factor = 1/32767 gives values in [0, 1]
year	int16	(sample,)	Year of the sequence (UTC)
doy	int16	(sample,)	Day of year of the sequence (UTC)

Sequences are in chronological order. Each one is stored as its own compressed chunk (zlib level 4 with byte shuffle), so reading a single sequence does not decompress its neighbors. Packing to int16 adds an error of at most 1.5 × 10⁻⁵ to the scaled value.

The global attributes of each monthly file give the month, the number of sequences, the splits it contains, the split rule, and the index range of the month in the original single file.

Provenance tables

Each data file comes with a CSV that has one row per sequence, in the same order as the file:

Column	Meaning
index	Position of the sequence in the file
source_index	Position in the single file used for training (test: same as index)
split	train, embargo, validation, or test
year, doy	Date (UTC)
window, start_utc	Window of the day (0 to 10) and the time of its first frame
kind	active (activity-weighted draw) or background (uniform draw)
y0, x0	Grid row and column of the patch's first cell; row 0 is the southern edge
lat, lon	Latitude and longitude of the patch center
pos_frac	Share of the patch's cells with FED > 0 in some frame of the window
activity	Total scaled FED of the patch over the window (the sampling weight is its square root)

The tables were produced by replaying the sampler on the gridded data. The replay reproduced the number of sequences of every day, and the FED channel of every validation, embargo and test sequence, and of the first sequence of every training day, byte for byte.

Channels

Index	Name	Quantity	Conversion from the scaled value v
0	FED	GLM flash-extent density, flashes per cell per 10 min	FED = 50 v
1	C08	ABI brightness temperature, 6.2 µm	BT = 190 + 90 v (K)
2	C09	ABI brightness temperature, 6.9 µm	BT = 190 + 90 v (K)
3	C10	ABI brightness temperature, 7.3 µm	BT = 180 + 110 v (K)
4	C13	ABI brightness temperature, 10.3 µm	BT = 180 + 140 v (K)
5	C14	ABI brightness temperature, 11.2 µm	BT = 180 + 140 v (K)
6	C15	ABI brightness temperature, 12.3 µm	BT = 180 + 140 v (K)
7	C16	ABI brightness temperature, 13.3 µm	BT = 180 + 110 v (K)

FED is clipped at 50, and brightness temperatures are clipped to the ranges above. On 2019-04-03, fewer than 0.011 % of the cells in any ABI channel hit a limit.

The accompanying paper labels a target cell positive when FED ≥ 1, which is v ≥ 0.02.

Example

import xarray as xr

ds = xr.open_dataset("patches_2019-07.nc")   # applies scale_factor on read
x = ds["sequences"][0].values                # (24, 8, 128, 128), values in [0, 1]
fed = 50.0 * x[:, 0]                         # flashes per cell per 10 min
bt_c13 = 180.0 + 140.0 * x[:, 4]             # K
inputs, targets = x[:12], x[12:, 0] >= 0.02  # 12 input frames; binary targets

Splits

Split	Period	Days	Sequences	Target cells	Positive target cells	Size
Train	2019-04-03 to 2020-02-04	308	60,084	11,812,995,072	812,191,769 (6.88 %)	226 GB
Embargo (unused)	2020-02-05	1	198	n/a	n/a	n/a
Validation	2020-02-06 to 2020-03-31	55	10,728	2,109,210,624	117,672,207 (5.58 %)	41 GB
Test	days 4 to 9 of each month, Sep 2023 to Feb 2024	36	6,966	1,369,571,328	95,578,805 (6.98 %)	26 GB

Target cells are the cells of the 12 target frames, 196,608 per sequence. A target cell is positive when FED ≥ 1. Sizes are those of the single training, validation, and test files used in the paper, in decimal gigabytes; the monthly files hold the same chunks.

Validation is the last 15 % of the development set, rounded to whole days. The day before validation starts is left out so that a single storm is unlikely to appear in both splits. Every monthly file stores this rule in its split_rule attribute.

To rebuild the training and validation files used in the paper, join the monthly files with scripts/join_monthly.py from the code record (doi:10.5281/zenodo.22979408), which copies their compressed chunks into one file in date order, then run scripts/split_dataset.py with VAL_START_YEAR=2020 VAL_START_DOY=37 EMBARGO_DAYS=1. Selecting sequences by year and doy with the dates above gives the same result.

The benchmark (Cintineo, 2024) scores days 5 to 9 of each test month. The test set also includes day 4 of each month.

Contents of each file

A day yields at most 198 sequences (11 windows × 18 patches). Every calendar day of the development period is present. The first file starts on 2019-04-03 because 2019-04-01 and 2019-04-02 fall in the switch to ABI scan mode 6.

Every day with fewer than 198 sequences lost whole windows to gaps in the ABI data; no window was lost for lack of lightning. The sampler skips a window if any of its frames is missing a cell in any ABI channel and the gap cannot be filled. Gaps of up to 3 frames (30 minutes) are filled by interpolation, but only between two complete frames of the same day, so a single incomplete frame at 00:00 or 23:50 UTC is enough to skip a window. There were two kinds of gap, and both come from the NOAA archive, not from our processing:

• Missing data: for the listed times the archive has no file, or a file without data, for at least one of the seven channels.
• Incomplete frames: the files exist, but a band of rows at the edge of our domain has no data (the share of the domain missing is given in parentheses).

Times are UTC. The number of windows skipped is given after each gap.

Development set:

File	Sequences	Split	Notes
patches_2019-04.nc	5,382	train	2019-04-03 to 2019-04-30; 2019-04-01 and 2019-04-02 are not included (switch to scan mode 6). Missing data on 2019-04-15, 12:20 to 15:20 (3 windows); 2019-04-23, 12:10 to 13:30 (2); 2019-04-24, 12:20 to 16:20 (4).
patches_2019-05.nc	5,994	train	Missing data on 2019-05-04, 12:10 to 13:00 (2 windows), and 2019-05-30, 12:20 to 18:40 (5). Incomplete frame on 2019-05-26 at 23:50 (under 0.1 %), the last of the day (1).
patches_2019-06.nc	5,796	train	Missing data on 2019-06-27, 13:50 to 17:50 (4 windows), and 2019-06-28, 14:50 to 19:20 (4).
patches_2019-07.nc	6,102	train	Missing data on 2019-07-17, 17:20 to 18:00 (2 windows).
patches_2019-08.nc	6,048	train	Incomplete frames on 2019-08-08, 16:20 to 18:10 (up to 0.1 %; 3 windows). Missing data on 2019-08-20, 15:10 to 15:50 (2).
patches_2019-09.nc	5,922	train	Incomplete frame on 2019-09-13 at 00:00 (under 0.1 %), the first of the day (1 window).
patches_2019-10.nc	6,066	train	Incomplete frames on 2019-10-11, 04:30 to 05:20 (up to 1.8 %; 2 windows), and 2019-10-12, 04:30 to 05:10 (up to 1.6 %; 2).
patches_2019-11.nc	5,868	train	Missing data on 2019-11-06, 01:20 to 03:40 (2 windows), and 2019-11-12, 20:10 to 21:50 (2).
patches_2019-12.nc	6,066	train	Missing data on 2019-12-04, 17:30 to 20:10 except the 18:10 frame (4 windows).
patches_2020-01.nc	6,066	train	Missing data on 2020-01-06, 21:40 to midnight (2 windows), and 2020-01-07, 00:00 to 00:20 (1). Incomplete frames on 2020-01-28 at 23:40 and 23:50 (under 0.1 %), the last of the day (1).
patches_2020-02.nc	5,724	train, embargo, validation	Train on 2020-02-01 to 2020-02-04, embargo on 2020-02-05 (not used), validation on 2020-02-06 to 2020-02-29. Incomplete frame on 2020-02-04 at 23:50 (3.1 %), the last of the day (1 window).
patches_2020-03.nc	5,976	validation	Incomplete frames on 2020-03-01, 05:00 to 05:40 (up to 2.1 %; 2 windows), and 2020-03-02, 05:00 to 05:40 (up to 2.3 %; 2). Missing data on 2020-03-18, 14:50 to 19:30 and 20:30 to 21:20 (5).
Total	71,010		364 days, 2019-04-03 to 2020-03-31.

Test set (patches_test.nc, one file):

Month	Sequences	Split	Notes
Sep 2023	1,188	test	Days 4 to 9, no gaps.
Oct 2023	1,188	test	Days 4 to 9, no gaps.
Nov 2023	1,188	test	Days 4 to 9, no gaps.
Dec 2023	1,026	test	Days 4 to 9. Missing data on 2023-12-04, 17:30 to 19:00 (3 windows), and 2023-12-07, 15:40 to 16:20 (3). Missing data and incomplete frames on 2023-12-08, 09:40 to 10:20 (3).
Jan 2024	1,188	test	Days 4 to 9, no gaps.
Feb 2024	1,188	test	Days 4 to 9, no gaps.
Total	6,966		36 days.

Dataset creation

Curation rationale

Lightning covers a small part of the domain at any given time, and patches placed uniformly would mostly be empty. Patch positions were therefore drawn in proportion to lightning activity, with a smaller share placed uniformly at random. The dataset is meant for training and comparing models. It is not suitable for estimating lightning climatology.

Source data

All data is from GOES-16 and was downloaded from the NOAA public archive: GLM Level 2 Lightning Cluster Filter Algorithm (LCFA) flashes, and ABI Level 1b full-disk radiances in scan mode 6. The development period starts on 2019-04-03, the first full day in scan mode 6.

Gridding

GLM and ABI share one grid, regular in the GOES-16 fixed-grid scan angle:

Property	Value
Spacing	55.9 µrad (2.0 km at the sub-satellite point, up to 2.7 km at the edge of the domain)
Projection	+proj=geos +h=35786023 +lon_0=-75.2 +sweep=x +ellps=GRS80
Size	2,017 × 1,849 cells
Extent	about 85° W, 25° S to 45° W, 10° N
Time step	10 minutes, 144 steps per UTC day

FED is accumulated over each 10-minute step by clipping each flash's event polygons onto the grid. A flash therefore adds to every cell its footprint touches, not only to the cell that holds its centroid. ABI radiances are converted to brightness temperature with the Planck coefficients in each file and mapped to the grid by nearest neighbor.

Lightning channel position

An error in our gridding places the FED channel 0.2° of longitude, about 20 km, west of its true position. lib/GlmGrid.jl converts flash positions to scan angles for a sub-satellite longitude of 75.0° W, while the grid is defined for 75.2° W. On our grid the offset is 7.5 to 11.1 cells east-west and under 0.4 cells north-south. The seven ABI channels are placed correctly. The patches_*.nc files keep FED where the paper's models saw it during training.

Each month folder therefore includes a companion file, fed_true_position_<month>.nc (fed_true_position_test.nc for the test set), with the FED channel moved to its true position:

• It holds only FED (fed, (sample, time, y, x)), with the same sequences in the same order, the same times and patch positions, and the same int16 packing as the companion patches_*.nc file. year and doy are repeated.
• Each cell is moved east by its own offset, computed with pyproj and rounded to whole cells, so every flash lands within about 1 km of its true position.
• Cells at the western edge of the grid have no source after the move and are set to 0. The attribute sequences_touching_west_edge_fill gives the number of sequences affected in each file.
• Rebuilt without the move, the channel equals the released FED channel byte for byte.

To use lightning and infrared at their true positions, replace channel 0 of sequences with fed from the companion file.

Sequence and patch sampling

Each UTC day is cut into 11 windows of 24 frames. The windows start every 2 hours from 00:00 to 20:00 UTC, so consecutive windows share 12 frames. They never cross midnight, which means no target frame falls between 00:00 and 02:00 UTC.

Each window yields 18 patches of 128 × 128 cells. A position counts as a candidate when at least 2.5 % of its cells have FED > 0 in some frame of the window. Fifteen patches are drawn from the candidates with replacement, with probability proportional to the square root of their total FED over the window. The other three are placed uniformly at random anywhere in the domain. A window without any candidate yields no patches.

ABI gaps of up to 3 frames (30 minutes) are filled by linear interpolation in time, and windows with longer gaps are skipped. The random stream is seeded with (42, year, day of year), so rerunning the sampler gives the same patches.

Reproducibility

The code record contains the whole pipeline:

Step	Code
Gridding	lib/GlmGrid.jl, lib/AbiGrid.jl, scripts/grid_day_*.jl
Patch sampling	scripts/build_patches.py, lib/glm_dataset, scripts/patches.yaml
Train/validation split	scripts/split_dataset.py
Monthly files	scripts/make_monthly_release.py
Provenance tables	scripts/patch_provenance.py, then scripts/make_monthly_release.py --provenance
FED at its true position	scripts/build_corrected_fed.py

We rebuilt two days from the gridded data, 2020-02-29 (validation, 198 sequences) and 2023-09-07 (test, 198 sequences). Both match the released sequences byte for byte in all eight channels.

The monthly files copy the compressed chunks of the original single file without decoding them. Each file was checked against the original chunk by chunk.

Where the code expects the files

The scripts read and write under one data root. It is data/ in the repository, run from the repository root, unless the environment variable LIGHTNING_DATA points elsewhere.

Path under the data root	Holds	Made by
release/	the files of this repository, all in one folder: patches_YYYY-MM.nc, fed_true_position_YYYY-MM.nc, provenance_YYYY-MM.csv, and the same three for test	join_monthly.py --download (patches files only), or copied from this repository
patches-128-2yr/patches.nc	the twelve months joined, 71,010 sequences	join_monthly.py
patches-128-2yr/train.nc, val.nc	training and validation splits	split_dataset.py with PATCHES_DIR=$LIGHTNING_DATA/patches-128-2yr
patches-bench/patches.nc	the test set (a link to release/patches_test.nc)	join_monthly.py --test

Training reads train.nc and val.nc from PATCHES_DIR, so set it to $LIGHTNING_DATA/patches-128-2yr for train.py as well. evaluate.py takes the validation file from --split (default $LIGHTNING_DATA/patches-128-2yr/val.nc). The gridded GLM and ABI data, needed only to rebuild the patches from the raw archives, go in glm-grid-jl, abi-grid-jl, glm-grid-bench and abi-grid-bench under the same root. In the fed_true_position_* files, only the lightning channel is stored; the paper's models were trained on patches_*.nc.

Uses

Direct use

The dataset suits training and evaluating lightning nowcasting models for lead times up to 2 hours, and comparing architectures on a fixed split that separates training and test in time. The seven ABI channels also make it possible to test which bands matter as predictors.

Out-of-scope use

• Lightning climatology or frequency estimates. The sampling favours active regions, so lightning is far more common in these sequences than in the domain as a whole.
• Diurnal cycle studies, since no target frame falls between 00:00 and 02:00 UTC.

Bias, risks, and limitations

• Between 5.6 and 7.0 % of target cells are positive, well above the rate for the domain as a whole: over all cells and all frames of the same days, 1.34 % are positive. Patches drawn with replacement can overlap, and one storm can appear in several sequences.
• GLM detection efficiency decreases beyond about 4,000 km from the sub-satellite point (Virts and Koshak, 2025). Only the southeastern corner of the domain, 1.3 % of its cells, lies beyond that distance, at up to 4,640 km, so labels there may be less complete.
• FED in the patches_*.nc files sits about 20 km west of its true position; the fed_true_position_*.nc files hold it at the true position (see Lightning channel position).
• ABI frames filled by interpolation carry no flag.
• FED above 50 flashes per cell and brightness temperatures outside the channel ranges are clipped.
• The development set covers a single year, and March appears only in validation. The test set covers September to February only.
• All data comes from GOES-16 and from infrared bands only. There are no visible bands, so every input channel is available by day and by night.

Citation

If you use this dataset, please cite:

@inproceedings{vela2026benchmark,
  author    = {Vela, Marco and Villanueva, Edwin and Takahashi, Ken},
  title     = {A Satellite Lightning Nowcasting Benchmark for South America:
               Temporal Cross-Attention in {ConvLSTMs}},
  booktitle = {WAIMLAp 2026},
  address   = {Lima, Peru},
  year      = {2026}
}

and the code record, which holds the pipeline and links to this repository:

@misc{vela2026code,
  author    = {Vela, Marco},
  title     = {Code for: A Satellite Lightning Nowcasting Benchmark for South America:
               Temporal Cross-Attention in {ConvLSTMs}},
  publisher = {Zenodo},
  year      = {2026},
  doi       = {10.5281/zenodo.22979408}
}

References

Cintineo, J. (2024). A benchmark dataset for lightning nowcasting in Latin America. Zenodo. doi:10.5281/zenodo.14261376

Virts, K. S., and Koshak, W. J. (2025). Bayesian analysis of the detection performance of the Geostationary Lightning Mappers. Journal of Atmospheric and Oceanic Technology, 42(8), 935 to 947. doi:10.1175/JTECH-D-24-0130.1

Contact

Marco Vela, mvelar@uni.pe