Aerial wildfire suppression requires not only predicting fire spread, but also designing effective intervention strategies under operational and environmental uncertainty. We present a modeling and optimization framework for aerial wildfire suppression that combines a hybrid neural-cellular automaton wildfire model with gradient-based design of targeted aerial drops. The wildfire model predicts spatially varying spread behavior from terrain, fuel, and wind data, while the intervention module determines binary drop actions with continuous-valued location and orientation parameters mapped to the simulation grid. Water and retardant are represented with distinct suppression effects, corresponding to immediate reduction of active burning and persistent reduction of future spread. To evaluate the robustness of the resulting suppression plans, we quantify both aleatoric uncertainty through Monte Carlo sampling of daily fire-state realizations and epistemic uncertainty through spatially correlated prediction-error perturbations. A case study based on the 2020 Bear Fire shows that the framework can generate coherent aerial suppression schedules for reducing total fire-affected area and can support uncertainty-aware analysis of wildfire intervention strategies.

翻译：空中野火抑制不仅需要预测火势蔓延，还需在运行和环境的双重不确定性下设计有效的干预策略。我们提出了一种结合混合神经-元胞自动机野火模型与基于梯度的定点空中投放设计的建模与优化框架。该野火模型根据地貌、可燃物和风场数据预测空间异质性蔓延行为，干预模块则通过映射至仿真网格的连续空间位置和取向参数确定二元投放决策。水和阻燃剂以不同的抑制效应建模，分别对应活跃燃烧的即时抑制与未来蔓延的持续抑制。为评估所得抑制方案的鲁棒性，我们通过蒙特卡洛采样每日火势状态实现量化偶然不确定性，并通过空间相关的预测误差扰动量化认知不确定性。基于2020年熊火（Bear Fire）的案例研究表明，该框架可生成减少总过火面积的协调式空中抑制调度方案，并为野火干预策略的不确定性分析提供支撑。