An adaptive hybrid model for wildfire front forecasting based on cellular automata, multi-agent UAV observations, and binary data assimilation: A case study of the 2021 Dixie Fire

Ramazan Sadvakassov, Kuralay Sadvakassova, Bijeesh Kozhikkodan Veettil , Nurzada Amangeldy, Altynbek Sharipbay, Assel Zhumabayeva, Dzhavdet Suleymanov

This study presents a retrospective case-study evaluation of a hybrid framework for daily wildfire-front forecasting during the 2021 Dixie Fire in California, USA. The framework couples a stochastic cellular automaton (CA) with a multi-agent system (MAS) of simulated UAV observations and a lightweight binary data-assimilation scheme. The model uses topography, vegetation, fuel proxies, and meteorological forcings derived from Copernicus GLO-30/SRTM, ESA WorldCover, MODIS NDVI, MODIS burned-area products, VIIRS active-fire detections, and ERA5-Land data. The CA produces a daily forecast of wildfire evolution, while the MAS directs a configurable set of simulated aerial agents toward spatially distributed high-information regions identified from predicted growth and fire-front structure. Agent observations are assimilated through a binary confirm-deny update that suppresses false ignitions and restores missed burning cells before the next forecast step; the observed footprint can also support local blending of daily weather fields before propagation resumes. In the verified evaluation window from 14 to 18 August 2021, with corrected road and water barrier masks enabled in the CA component, the hybrid CA+MAS framework increased the mean Intersection over Union (IoU) from 0.6664 for the standalone baseline CA to 0.6992, corresponding to an absolute gain of 0.0328, while the mean F1 score increased from 0.7998 to 0.8229. These findings indicate that targeted recurrent aerial observations can improve short-horizon wildfire-front reconstruction while preserving the computational simplicity and interpretability of the underlying cellular automaton.