Integrating initial fractal dimension with LightGBM for enhanced wildfire burned-area prediction

Traditional wildfire prediction models rely heavily on environmental variables, such as temperature, humidity, and vegetation indices, yet they often fail to capture the geometric complexity of the initial fire spread. This study introduces the Initial Fractal Dimension (IFD) derived from satellite-detected active fire masks during the first 72 h post-ignition as a novel predictor of burned areas. The IFD was calculated using an alpha-shaped boundary approximation and box-counting methods on MODIS data and integrated into a LightGBM regression framework alongside conventional predictors. Using a dataset of 589 large U.S. wildfires across three regions, the results showed that including IFD consistently improved the model performance and robustness compared with models based solely on environmental variables. Importantly, leakage tests and correlation analyses demonstrated that the IFD contributes unique, independent information, rather than redundantly encoding fire size. These findings establish IFD as an effective descriptor of early fire perimeter complexity, offering substantial gains in wildfire burned-area prediction. This study highlights the value of fractal geometry in remote sensing applications and provides a foundation for more climate-resilient fire management strategies.