justuskarlsson/FireComp

1---2license: cc-by-4.03task_categories:4  - image-segmentation5tags:6  - wildfire7  - remote-sensing8  - viirs9  - earth-observation10size_categories:11  - 100K<n<1M12configs:13  - config_name: default14    data_files:15      - split: train16        path: next_day_v3/splits/train.jsonl17      - split: validation18        path: next_day_v3/splits/val.jsonl19      - split: test20        path: next_day_v3/splits/test.jsonl21---22 23# FireComp: Next-Day Fire Spread24 25Global benchmark for next-day wildfire spread prediction from 375 m VIIRS26active-fire detections, with ERA5 weather, GFS forecasts and Alpha Earth27terrain embeddings. 256×256 patches, 9 regions, 2017–2025.28 29Code, documentation, loaders and paper: **https://github.com/justuskarlsson/FireComp**30 31## Paper32 33This dataset accompanies the FireComp paper, **accepted and presented at GCPR 2026**34(German Conference on Pattern Recognition). It is released so the benchmark results35in that paper can be reproduced and built upon. See the GitHub repository for the36citation entry.37 38## Contents39 40| Path | Description |41|---|---|42| `next_day_v3/` | Main dataset: 8 HDF5 shards (`dataset_*.h5`, zstd) + per-shard metadata (`dataset_*.json`), `samples.json`, `stats.json` (normalization) |43| `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 |44| `next_day_v3_case_study/` | Small case-study subset used for paper figures |45| `regions/` | Region definitions (`wildfire_regions.json` / `.tif`) |46| `fire_areas.npz` | Per-fire spatial area lookup (fire-size stratification) |47| `vnp14_fires/vnp14_fires_2012.h5` | VIIRS fire-event index, 2012–2025 (~55 GB) — see below |48 49### `vnp14_fires_2012.h5`50 51Individual VNP14IMG active-fire detections clustered into **fire events** by a52spatio-temporal BFS, which is what defines the `fire_id` used everywhere else in53the benchmark. Not needed to train on the dataset above, but required for most of54the analysis and visualization: per-fire statistics, size/region breakdowns,55wild-vs-tame classification, fire-shape (solidity) analysis and the paper figures.56 57| Group | Contents |58|---|---|59| `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 |60| `pixels/` | Every detection: `fire_id`, `date`, position, `cls`, `ignition`, `image_id` |61| `projection_by_fire/`, `projection_by_t/` | Detections rasterized per fire / per timestep (`projection_by_fire` covers fires with ≥100 detections) |62| `images/`, `meta/` | Source granule index and archive date range |63 64## Usage65 66Clone/download into `data/` of the GitHub repo and follow its README.67Reading the shards requires `h5py` and `hdf5plugin` (zstd filter).68