Datasets:
File size: 2,852 Bytes
9bd3a33 1a316b8 6960cc9 9bd3a33 1a316b8 973e8a6 1a316b8 4aa38ba 1a316b8 6960cc9 1a316b8 4aa38ba 1a316b8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | ---
license: cc-by-4.0
task_categories:
- image-segmentation
tags:
- wildfire
- remote-sensing
- viirs
- earth-observation
size_categories:
- 100K<n<1M
configs:
- config_name: default
data_files:
- split: train
path: next_day_v3/splits/train.jsonl
- split: validation
path: next_day_v3/splits/val.jsonl
- split: test
path: next_day_v3/splits/test.jsonl
---
# FireComp: Next-Day Fire Spread
Global benchmark for next-day wildfire spread prediction from 375 m VIIRS
active-fire detections, with ERA5 weather, GFS forecasts and Alpha Earth
terrain embeddings. 256×256 patches, 9 regions, 2017–2025.
Code, documentation, loaders and paper: **https://github.com/justuskarlsson/FireComp**
## Paper
This dataset accompanies the FireComp paper, **accepted and presented at GCPR 2026**
(German Conference on Pattern Recognition). It is released so the benchmark results
in that paper can be reproduced and built upon. See the GitHub repository for the
citation entry.
## Contents
| Path | Description |
|---|---|
| `next_day_v3/` | Main dataset: 8 HDF5 shards (`dataset_*.h5`, zstd) + per-shard metadata (`dataset_*.json`), `samples.json`, `stats.json` (normalization) |
| `next_day_v3/splits/{train,val,test}.jsonl` | Flat sample index per split (stratified temporal 60/15/25 within each region) — what the dataset viewer shows |
| `next_day_v3_case_study/` | Small case-study subset used for paper figures |
| `regions/` | Region definitions (`wildfire_regions.json` / `.tif`) |
| `fire_areas.npz` | Per-fire spatial area lookup (fire-size stratification) |
| `vnp14_fires/vnp14_fires_2012.h5` | VIIRS fire-event index, 2012–2025 (~55 GB) — see below |
### `vnp14_fires_2012.h5`
Individual VNP14IMG active-fire detections clustered into **fire events** by a
spatio-temporal BFS, which is what defines the `fire_id` used everywhere else in
the benchmark. Not needed to train on the dataset above, but required for most of
the analysis and visualization: per-fire statistics, size/region breakdowns,
wild-vs-tame classification, fire-shape (solidity) analysis and the paper figures.
| Group | Contents |
|---|---|
| `stats/`, `stats_{10,100,1000,10000}/` | Per-fire records (`id`, `start_date`, `end_date`, bbox, `country`, `avg_xy_neighbors`, `ignition_ratio`, …), the suffix being a minimum-detection-count filter |
| `pixels/` | Every detection: `fire_id`, `date`, position, `cls`, `ignition`, `image_id` |
| `projection_by_fire/`, `projection_by_t/` | Detections rasterized per fire / per timestep (`projection_by_fire` covers fires with ≥100 detections) |
| `images/`, `meta/` | Source granule index and archive date range |
## Usage
Clone/download into `data/` of the GitHub repo and follow its README.
Reading the shards requires `h5py` and `hdf5plugin` (zstd filter).
|