Anonymous4AISI/CanFireCast_Dataset

CanFireCast Dataset (Paper Reproduction Sample)

This repository hosts a sampled release used to reproduce the experiments in CanFireCast for Canada Nationwide Wildfire Activity Forecasting with Multisource Spatiotemporal Modeling.

Important: This is not the full nationwide corpus.
The complete dataset (raw products, intermediate layers, annual HDF5 archives, and derived caches) exceeds 10 TB and cannot be redistributed through online platforms.
Access to the full dataset may be requested from the authors after the anonymity period ends (see Full Dataset Access).

Companion code: CanFireCast

---

What This Release Contains

This Hugging Face package provides the paper-reproduction sample / training cache used with --json F (cache_F) in the CanFireCast codebase. Approximate contents (~65 GB):

File	Description
windows_*.h5	Sampled spatiotemporal training windows for train / validation / test
samples_variant_F.json	Sampling metadata for variant F
norm_stats.npz	Train-only normalization statistics (reuse for validation and test)

These files are sufficient to fully reproduce the paper training and evaluation pipeline. They are a carefully constructed subsample of the nationwide data cube, not a dump of every grid cell and day.

---

Task Definition

Next-day wildfire activity forecasting on a 5 km grid over the Canadian land area.

For each target cell, the model receives 10 days of environmental history in a local 13 × 13 patch ending on the issue date, and predicts the probability of a high-confidence active-fire observation at the center cell on the following day.

Each sample is:

[ \mathbf{X}_i \in \mathbb{R}^{54 \times 10 \times 13 \times 13}, \quad y_i \in {0,1} ]

Temporal splits (strict year-based, no leakage across years):

Split	Years
Train	2000–2019
Validation	2020–2022
Test	2023–2025

The test set uses a positive-to-negative ratio of 1:2 to avoid extreme class imbalance.

---

Multisource Inputs

Inputs combine weather and land state, fire-danger indices, vegetation and fuel proxies, terrain, land cover, and human-activity context. Products are spatially interpolated and temporally aligned onto a daily 5 km grid. Cloud-affected, low-quality, or missing remote-sensing observations are filled with the latest valid observation before the target date.

Variable groups

Group	Contents	Main sources
W	Weather, land meteorology, CFFDRS indices, VPD	ERA5-Land; derived FWI System indices
F	Soil moisture, vegetation, satellite land-surface variables, land cover	ERA5-Land; MODIS Terra/Aqua (e.g., MOD/MYD09GA, MCD15A3H, MOD/MYD11A1, MCD12Q1)
B	Water, terrain, and terrain-derived variables	ASTER DEM; OpenStreetMap-derived water density
I	Human-activity variables	WorldPop; OpenStreetMap-derived road / powerline / building density
Label	High-confidence active-fire detection at the center cell	MOD/MYD14A1

Variable summary

Click to expand the variable list

W — weather / meteorology / fire danger

• Temperature 2m (mean, max), Dewpoint Temperature 2m, Skin Temperature Max
• Wind 10m (U, V), Precipitation, Snow Cover, Surface Pressure
• Surface Latent Heat Flux, Net Solar Radiation
• Total / Potential Evaporation, Skin Reservoir Content
• VPD (derived)
• FFMC, DMC, DC, ISI, BUI, FWI (Canadian FWI System, derived)

F — fuel / vegetation / satellite land surface

• Soil Water Layers 1–4 (ERA5-Land)
• Reflectance bands 1, 2, 3, 7 (MOD/MYD09GA)
• NDVI, EVI; LAI, FPAR (MCD15A3H)
• LST Day / Night; Emis 31 / 32 (MOD/MYD11A1)
• Brightness temperature bands 20, 21, 31, 32 (MOD/MYD09CMG)
• Land Cover Types (MCD12Q1)

B — terrain / water

• DEM; Slope; Aspect (sin, cos)
• Hillshade, TPI, TWI
• Water Density (OSM-derived)

I — human activity

• Population (WorldPop)
• Road, Powerline, Building Density (OSM-derived)

Label

• Fire Detection (MOD/MYD14A1 high-confidence active fire)

---

Sampling Notice

Please read this carefully before citing or redistributing this repository:

1. This release is a sampled subset designed for paper reproduction (training windows + metadata + normalization stats).
2. It does not include the full daily nationwide 5 km data cube, all intermediate geospatial products, or every candidate location outside the paper sampling protocol.
3. Results reported in the paper were obtained with this sampling protocol and the associated cache_F configuration.
4. If you need the complete >10 TB corpus for other research uses, request it from the authors after anonymity ends (below).

---

Quick Start

Download

pip install -U "huggingface_hub[cli]"
huggingface-cli download Anonymous4AISI/CanFireCast_Dataset \
  --repo-type dataset \
  --local-dir ./CanFireCast_Dataset

Use with the CanFireCast code

1. Clone the code: https://anonymous.4open.science/r/CanFireCast4AAAI-AISI
2. Set h5_dir in prepare_data_loaders() in train_all_h5.py to the extracted dataset directory.
3. Train / evaluate with --json F so the loader uses this sample cache.
4. Keep norm_stats.npz unchanged for validation and test (train-only statistics).

Example:

export MODEL_SAVE_DIR=./checkpoints/canfirecast
python train_single_model_h5.py \
  --model CanFireCast \
  --json F \
  --gpu 0 \
  --log-dir ./outputs/canfirecast

---

Full Dataset Access

The full nationwide corpus exceeds 10 TB and is not hosted here.

After the anonymity / review period ends, researchers may contact the corresponding author to request access to the complete dataset. Please include:

• name and institution
• intended research use
• which components are needed (raw products / annual HDF5 / caches / etc.)

Access and transfer arrangements may depend on storage capacity and the licenses of the original source products (ERA5-Land, MODIS, ASTER DEM, WorldPop, OpenStreetMap, and derived CFFDRS variables).

Author contact details will be added here after the anonymity period.

---

License and Source Acknowledgments

This release redistributes derived research samples assembled from third-party environmental products. Users remain responsible for complying with the licenses and citation requirements of the original sources, including but not limited to:

• ERA5-Land (ECMWF / Copernicus)
• MODIS Terra and Aqua products (NASA)
• ASTER GDEM
• WorldPop
• OpenStreetMap contributors
• Canadian Forest Fire Danger Rating System (CFFDRS) / FWI System indices derived from weather inputs

The dataset license field is set to other because redistribution terms of the derived sample follow both this research release and the upstream product policies.