Forecasting Wildfire Ignitions: A Two-Step Machine Learning and Expert-Based Fire Hazard Model

Short-term fire hazard forecasting is a critical component of wildfire preparedness, yet widely used operational indices such as the Fire Weather Index (FWI) primarily represent meteorological fire danger and do not explicitly model ignition likelihood. We present a two-step, data-driven fire hazard modelling approach that combines machine learning with expert-based refinement. In the first step, a machine learning model learns the relationship between environmental fire drivers and observed wildfire ignitions to generate probabilistic fire hazard maps at a coarse spatial scale. In the second step, these base-level hazard maps are upsampled to 1 km resolution using an expert system that incorporates high-resolution susceptibility information, enabling operationally relevant fire hazard forecasts.The machine learning component is trained on OroraTech’s proprietary six-year global active wildfire dataset, which provides a best-in-class trade-off between spatial resolution and revisit frequency. This dataset enables robust learning of ignition-relevant patterns across diverse fire regimes. Input features combine environmental variables derived from climate reanalysis, remote sensing products such as digital elevation models, and large-scale spatio-temporal dynamics capturing seasonal and regional fire behaviour. The model integrates spatial and temporal information to produce fire hazard estimates at 0.1° spatial resolution.To support operational use, the hazard estimates are refined to 1 km spatial resolution using an expert system that applies susceptibility masks derived from aggregated vegetation indicators, infrastructure information, and additional static and dynamic constraints. This allows the generation of high-resolution fire hazard maps with lead times of up to one week.Across the study regions, the proposed model correctly predicts up to 30 times more fire ignitions than the Fire Weather Index under comparable conditions. The model is currently being rolled out for selected users within OroraTech’s wildfire solution platform to support short-term preparedness and operational planning.