WSDS-CAN v2: Wildfire Spread Prediction Dataset for Canadian Boreal Forests

WSDS-CAN v2 is a machine-learning-ready dataset of 19,468 Canadian wildfire events of 1 ha or larger, covering the period from 2001 to 2025 and all 14 fire-management agencies. Built using the National Burned Area Composite, the dataset represents each event as a 128×128 pixel tile with a resolution adapted to fire size (median 31 m). Each tile contains the burned-area mask, an ignition layer, terrain, IGBP land cover, FBP fuel type, MODIS leaf area index, and ERA5-Land noon weather data, together with the six codes of the Canadian Fire Weather Index System from 1 April to 31 October. The dataset supports two prediction tasks: final-perimeter prediction for 18,992 events and day-by-day fire-spread prediction for 7,806 events. Of the latter, 7,800 have a day-of-burning raster derived by interpolating MODIS and VIIRS active-fire detections. A total of 15,026 events are flagged as boreal fires. The release includes fixed random, spatial, temporal, and spatio-temporal data splits, quality control reports for every file, and the complete processing code. Start with README.md for an overview and consult DATA_DICTIONARY.md for descriptions of every column and variable. Version 2 is a complete rebuild of version 1: https://doi.org/10.5281/zenodo.20264141.