Datasets and Model State for Article: Probabilistic Forecasting of Localized Wildfire Spread Based on Conditional Flow Matching

Datasets and model state used to obtain results presented in the article Probabilistic Forecasting of Localized Wildfire Spread Based on Conditional Flow Matching (https://doi.org/10.48550/arXiv.2603.26975). Model state used to generate results in article is saved in file checkpoint_45000.pth. Training and testing data are constructed from WRF-SFIRE wildfire spread solutions, along with corresponding weather data from NAM, as described in the associated article. File test_cases_for_sensitivity_analysis_uniform_timber_litter_with_heavy_dead_fuel_flat_terrain_80F_circular_start_area_no_wind_mean_SD_scaling.tar.gz contains samples used for performing sensitivity analysis of model from article. recursive_prediction_test_samples_from_ignition.tar.gz contains data used to evaluate recursive application of model for 12 wildfires from ignition to 24h (8 recursive steps) and recursive_prediction_test_samples_fire_extent_subsection.tar.gz contains data used to evaluate recursive application of model for 12 wildfires from 12h to 36h (8 recursive steps) for wildfire subsection, as presented in the manuscript. All datasets are normalized so values are in [-1,1]. To convert data back to physical values use the following normlization: Channel 1, output fire arrival times: Multiply by 3/2 and add 3/2 to get arrival times in [0h, 3h] Channel 2, input fire area: Binary values, no normalization Channel 3, u-wind: Multiply by 2.4372 to get u-wind values in units of m/s Channel 4, v-wind: Multiply by 2.3389 to get v-wind values in units of m/s Channel 5, relative humidity: Multiply by 23.0319 and add 46.1206 to get relative humidity as % Channel 6, temperature: Multiply by 9.2502 and add 292.2049 to get temperature in K Channel 7, terrain height: Multiply by 176.7865 and add 184.6440 to get terrain height in m Channels 8-21, binary fuel category masks: Binary masks, no normalization