CanadaWildFireDaily/CanadaWildFireDaily-v1

🔥🔥 CanadaWildfireDaily: A Large-Scale Dataset for Daily Wildfire Spread in Canada 🔥🔥

Folder Structure

This section provides the details needed to understand and use the released dataset. We release both raw data and training/val/test ready samples.

CanadaWildFireDaily Train/Validation/Test Samples (data_samples/)

The final training samples are generated through a two-step process. First, the CSV files and metadata mappers are used to assign fire IDs to the training, validation, and test subsets. Second, individual samples are constructed by iterating through the tile-specific keys in the metadata JSON files. For a given tile and day of burn, the following rules are applied:

• Mask merging: If the metadata indicates that multiple fire events, or parts of large fires, are active within the same tile, the pipeline accesses each relevant .h5 file, extracts the corresponding fire masks, and merges them. This creates a unified ground-truth mask representing the total burning area within the 256 × 256 tile.
• Feature extraction: Since environmental features are consistent across the same geographic tile and day, they are extracted from a single reference .h5 file.

Each split has its own compressed folder of train/val/test, that you can decompress using: tar --use-compress-program="zstd -d" -xf NAME.tar.zst

Raw Data (to generate training/val/test samples)

he raw dataset includes three main components: Fire Growth Points Data (.csv), yearly fires metadata files (.json), and main data arrays (.h5). The following subsections explain what each file contains and how these components work together to build the final samples used for model training. The values in the raw data (the environmental features and the satellite bands) are not normalized or altered in any way.

Fire Growth Points Data (raw_data/fire_growth_points_cfsds)

Fire growth points data is provided as a collection of yearly CSV files. These files are retrieved from the publicly available Canadian Fire Spatial Dataset (CFSDS) and are provided exactly as-is; we did not alter or curate them in any way.

For end users, the main purpose of these files is to support dataset splitting. They contain the fire IDs and geographic coordinates for each fire, and are used to determine which fires belong to the training, validation, and test sets.

Yearly Fires Metadata (raw_data/metadata_fires_per_year)

The metadata consists of JSON files that organize fire activity by year, geographic location, and time. Since the Canadian map is divided into fixed tiles, each identified by a tile ID, the JSON files are structured as dictionaries where each key represents a specific tile, year, and day of burn:

tileID_year_DOB

Each key maps to a list of fire IDs and their corresponding day keys that are active within that tile on that date. This structure serves two main purposes:

• It supports dataset splitting by ensuring that all fire activity within the same tile is assigned to the same split (train, validation, or test), preventing spatial data leakage.
• It guides sample construction by allowing the sample generator to identify and merge masks from multiple overlapping fires within the same tile into a single ground-truth mask.

The list of fire IDs within each split is also provided as a JSON file (raw_data/trainvaltest_splitting_IDs.json).

Fire Data (raw_data/fire_data)

The core dataset is provided as HDF5 (.h5) files, with one file per unique fire event named according to its ID, for example:

fire_2024_188.h5

Each file contains the following levels:

• Global attributes: Metadata stored at the root level, including the coordinate reference system, pixel resolution, fire ID, and fire year.
• Tile groups: Data is organized by the tiles that the fire overlaps.

Each tile group contains:

• static_features: Topographic and SCANFI features that do not change on a daily basis.
• coords: Coordinate grids for the tile, including longitudes, latitudes, eastings, and northings.
• days: Daily folders for each fire day where dynamic data was captured.

Each day folder contains:

• features: Daily weather variables, vegetation indices, and the fire state for that specific day.
• satellite: Raw Sentinel-2 bands and the Scene Classification Layer (SCL). This group also contains JSON-formatted metadata describing the acquisition date and cloud coverage for each group of pixels in the image.