| # TandemFire — Sample Dataset (NeurIPS 2026 Submission) |
|
|
| This is a **40-sample representative subset** of the TandemFire dataset, created for reviewer access during the NeurIPS 2026 Datasets & Benchmarks submission process. |
|
|
| ## Sample composition |
| - 20 samples from the training split |
| - 10 samples from the validation split |
| - 10 samples from the test split |
|
|
| Samples were selected as the first entries from each split, covering diverse geographic regions including Africa (Madagascar, Chad, Mozambique, Guinea, Kenya, South Sudan, Tanzania, DRC, Côte d'Ivoire), Asia (China, Mongolia, India, Kazakhstan), Europe (Ukraine), Russia (Russian Federation), and the Americas (Bolivia, Brazil, Guatemala, Venezuela). |
|
|
| ## Full dataset |
| The complete dataset (1,781 triplets, ~49 GB) is available at: |
| 👉 https://huggingface.co/datasets/ali97/fire-progression-dataset |
|
|
| ## Structure |
| Each sample contains 7 GeoTIFF files: |
| - `pre_s2.tif` — Pre-fire Sentinel-2 imagery (10 bands, 10m) |
| - `t1_s2.tif` — T1 active-fire Sentinel-2 imagery (10 bands, 10m) |
| - `t2_s2.tif` — T2 active-fire Sentinel-2 imagery (10 bands, 10m) |
| - `pre_rf.tif` — Pre-fire RF-derived fire mask (binary) |
| - `t1_rf.tif` — T1 RF-derived fire mask (binary, model input) |
| - `t2_rf.tif` — T2 RF-derived fire mask (binary, prediction target) |
| - `aux.tif` — Auxiliary data (8 bands: elevation, slope, aspect, landcover, temp, wind_u, wind_v, precip) |
|
|
| ## License |
| CC BY 4.0 |
