README.md

---
license: mit
task_categories:
- image-classification
- time-series-forecasting
language:
- en
tags:
- wildfire
- remote-sensing
- earth-observation
- canada
- benchmark
- hard-negative-mining
size_categories:
- n>1T
---

# FireMPC: A Pan-Canadian Wildfire Forecasting Benchmark

FireMPC is a pan-Canadian wildfire risk benchmark covering approximately one
billion hectares across all fifteen Canadian terrestrial ecozones at 1 km daily
resolution from 2000 to 2025, integrating 55 drivers across fuel, terrain,
anthropogenic, and meteorological families.

This release provides four pre-built sample caches that share the same
underlying data cube but differ in their training and test sample construction
strategies, enabling controlled study of FWI-guided Hard Negative Mining
(FWI-HNM) versus random negative sampling.

## Repository Contents

```
MPCFire/
├── cache_A/             # Train: random sampling  | Test: FWI-HNM
├── cache_Y/             # Train: random sampling  | Test: random sampling
├── cache_G/             # Train: FWI-HNM          | Test: FWI-HNM
├── cache_H/             # Train: FWI-HNM          | Test: random sampling
└── Entire_Canada_Maps/  # Raw driver rasters, 13 modalities x 26 years (yearly tar archives)
```

Each cache directory contains exactly three files:

| File | Size | Description |
| --- | --- | --- |
| `windows_<hash>.h5` | ~14 GB | Pre-extracted 10-day input windows + labels for every sample (positives and negatives). One HDF5 file per cache. |
| `samples_variant_<X>.json` | ~4.4 MB | Sample index: train / val / test split assignments, sample identifiers, and metadata. |
| `norm_stats.npz` | ~2 KB | Channel-wise mean and standard deviation used for input normalisation. Skips the fire-mask channel and the categorical land-cover channel. |

## Variant Design

The four caches form a 2x2 ablation grid that decouples the negative-sampling
strategy used during training from the strategy used during evaluation:

|         | Test = FWI-HNM | Test = Random |
| ------- | -------------- | ------------- |
| **Train = Random** | `cache_A` | `cache_Y` |
| **Train = FWI-HNM** | `cache_G` | `cache_H` |

* **FWI-HNM (FWI-guided Hard Negative Mining)** scores every non-fire candidate
  with a calibrated six-component CFFDRS composite (FFMC, DMC, DC, ISI, BUI,
  FWI), partitions the pool at the median, and combines hard negatives
  (fire-weather-matched non-ignitions) with representative negatives
  (low-danger baseline) in equal proportions.
* **Random sampling** draws negatives uniformly from the non-fire candidate
  pool (firemask = 5).

Comparing rows isolates the effect of the training-pool construction; comparing
columns isolates the effect of the evaluation-pool construction. The diagonal
pair (`cache_A`, `cache_G`) corresponds to the standard production setup; the
off-diagonal pair (`cache_Y`, `cache_H`) is used to verify that any
FWI-HNM advantage reflects genuine boundary hardening rather than train-test
distributional alignment.

## Raw Driver Maps (`Entire_Canada_Maps/`)

`Entire_Canada_Maps/` provides the underlying raster stack used to build the
sample caches above. It covers the full Canadian landmass for 2000-2025 (26
years) and is organised into 13 modality subfolders that correspond to the
drivers listed in Table 1 of the paper. Each subfolder contains one tar
archive per year; each archive holds that year's daily (or annual / static)
GeoTIFFs.

All rasters are stored with **integer scaling (scaledInt)** to reduce volume.

### Subfolder Index

| Folder | Channels | Source | Cadence | Notes |
| --- | --- | --- | --- | --- |
| `DEMs/` | DEM, Slope, Aspect (sin/cos), Hillshade, TPI, TWI (7 channels) | ASTER GDEM | Static | Elevation and derived terrain indices; broadcast across the daily axis |
| `ERA5/` | temperature_2m, u/v_component_of_wind_10m, snow_cover, total_precipitation_sum, surface_latent_heat_flux_sum, dewpoint_temperature_2m, surface_pressure, volumetric_soil_water_layer_1-4, temperature_2m_max, skin_temperature_max, potential_evaporation_sum, total_evaporation_sum, skin_reservoir_content, surface_net_solar_radiation_sum (18 channels) | ERA5-Land Daily Aggregate (`ECMWF/ERA5_LAND/DAILY_AGGR`) | Daily | Atmospheric reanalysis fields |
| `FWI/` | FFMC, DMC, DC, ISI, BUI, FWI (6 channels) | CFFDRS (ERA5-driven) | Daily | Canadian Forest Fire Weather Index components |
| `MCD09CMG/` | Coarse Resolution Brightness Temperature Bands 20 / 21 / 31 / 32 (4 channels) | MOD/MYD09CMG | Daily | Coarse-resolution composite brightness temperatures |
| `MCD09GA/` | Bands 1, 2, 3, 7 (4 channels) | MOD/MYD09GA | Daily | QA-filtered, gap-filled surface reflectance |
| `MCD11A1/` | LST_Day_1km, LST_Night_1km, Emis_31, Emis_32 (4 channels) | MOD/MYD11A1 | Daily | QA-filtered, gap-filled land surface temperature and emissivity |
| `MCD12Q1/` | Land Cover Class | MCD12Q1 | Annual | Land cover / land use class; broadcast across the daily axis |
| `MCD14A1/` | Active Fire (binary) | MOD/MYD14A1 | Daily | Supervision target only; **not** included as a model input channel |
| `MCD15A3H/` | LAI, FPAR (2 channels) | MCD15A3H | Daily (from 4-day composite, interpolated) | Leaf area index / fraction of absorbed PAR |
| `NDVI_EVI/` | NDVI, EVI (2 channels) | MODIS-derived | Daily | Vegetation activity indices |
| `OSMs/` | Road / Powerline / Building / Water Density (4 channels) | OSM-derived | Static | Infrastructure accessibility; broadcast across the daily axis |
| `VPD/` | Vapor Pressure Deficit | ERA5-derived | Daily | Atmospheric moisture demand |
| `Worldpop/` | Population Density | WorldPop | Annual | Human population density; broadcast across the daily axis |

### File Naming

```
Entire_Canada_Maps/<modality>/<modality>__<YYYY>.tar
```

After extraction, individual files are named `YYYY_MM_DD.tif` for daily
modalities or `<modality>_YYYY.tif` for static / annual modalities.

### Download and Extraction

```python
from huggingface_hub import hf_hub_download
import tarfile, pathlib

tar_path = hf_hub_download(
    repo_id="AnonymousData4NeurIPS/MPCFire",
    repo_type="dataset",
    filename="Entire_Canada_Maps/ERA5/ERA5__2020.tar",
)
out = pathlib.Path("./ERA5_2020")
with tarfile.open(tar_path) as tf:
    tf.extractall(out)
```

To pull an entire modality:

```python
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="AnonymousData4NeurIPS/MPCFire",
    repo_type="dataset",
    allow_patterns=["Entire_Canada_Maps/ERA5/*"],
)
```

## Splits

A temporal hold-out is used throughout:

| Split | Years |
| ----- | ----- |
| Train | 2000 - 2019 |
| Validation | 2020 - 2022 |
| Test  | 2023 - 2025 |

The three-year test window deliberately covers the record-setting 2023 fire
season alongside the more typical 2024 and 2025 seasons. All splits maintain a
fixed 1:2 positive-to-negative ratio.

Sample identifiers and split assignments are stored in
`samples_variant_<X>.json` inside each cache.

## File Format

`windows_<hash>.h5` is a single HDF5 file with pre-extracted 10-day driver
windows and binary labels. The hash in the filename identifies the windowing
configuration (10-day backward window, 1 km patches) and is shared across all
four caches because the underlying input cube is identical; only the
positive/negative selection differs per variant.

`norm_stats.npz` provides per-channel mean and standard deviation arrays. The
fire-label channel and the categorical land-cover channel are excluded from
z-score normalisation.

## Reading the Data

Quick example using `huggingface_hub`:

```python
from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="AnonymousData4NeurIPS/MPCFire",
    repo_type="dataset",
    allow_patterns=["cache_G/*"],   # download a single variant
)
```

Then load the HDF5 file and the sample index:

```python
import h5py, json, numpy as np

cache = f"{local_dir}/cache_G"
with open(f"{cache}/samples_variant_G.json") as f:
    samples = json.load(f)
norm = np.load(f"{cache}/norm_stats.npz")
h5 = h5py.File(next(p for p in __import__('os').listdir(cache) if p.endswith('.h5')), "r")
```

## Citation

This dataset accompanies a paper currently under anonymous peer review. A
citation entry will be added on acceptance.

## License

Released under the MIT License. The dataset is built from publicly available
products (MODIS, ERA5-Land, ASTER, WorldPop, OpenStreetMap); please consult the
licenses of the upstream sources for redistribution of derivative products.