Reproducibility package for multi-year wildfire occurrence susceptibility modelling in Alberta, Canada

This reproducibility package supports the study “Multi-Year Spatiotemporal Generalization of Wildfire Occurrence Susceptibility Models in Alberta, Canada: The Roles of Spatial Context and Control Sampling Uncertainty”. The study evaluates Random Forest, support vector machine, XGBoost, and convolutional neural network models for wildfire occurrence susceptibility in Alberta, Canada, during 2001–2021 at 500 m spatial resolution. All four models were evaluated using matched 5 × 5 predictor patches, spatial block cross-validation, independent temporal testing, and non-fire control-sampling sensitivity analyses. The archive contains processed sample metadata, independent temporal-test predictions, validation-selected thresholds, performance summaries, paired bootstrap statistical comparisons, supplementary Tables S1–S3, final figures, and analysis notebooks used for model fitting, sensitivity analysis, SHAP interpretation, and figure generation. Complete third-party source rasters and large model-input arrays are not redistributed. Original predictor layers can be obtained from the public sources described in the manuscript Methods.