【摘要】 【目的】森林火灾具有强动态性和多要素耦合特征，复杂地形、植被阻力及火势蔓延过程共同影响人员通行效率，使传统基于静态代价栅格的路径规划方法难以同时兼顾安全性与效率。【方法】针对这一问题，本文构建了融合地形坡度、植被密度、消防员速度场和火场蔓延风险的多源环境模型，提出一种面向森林火灾动态场景的速度场驱动安全路径规划方法（HNP-DuelingDQN）。该方法设计了由目标引导、通行速度、安全因子及任务完成项组成的复合奖励机制，并在强化学习框架中引入HAS启发式动作选择与N-PER经验回放结构，以提高策略学习的稳定性与效率。【结果】基于福建省真实地形与近五年内两场典型森林火灾数据开展的仿真实验表明，所提方法在静态与动态火灾实验环境中均表现出稳定优势，相较于HNP-DQN、A*与ACO算法，该方法能够显著提升路径整体安全性能，使火场蔓延风险水平平均降低约17.4%～57.8%，平均通行速度提升约13.4%～28.3%，路径的平均坡度降低约11.0%～29.9%。【结论】实验结果表明，本文所提方法实现了路径安全性与通行效率的协同优化，可为山地森林火灾情景下的应急救援路径决策提供更具实时性、可解释性和可量化的技术支撑。更多还原

【Abstract】 [Objectives] Forest fires are characterized by strong dynamics and multi-factor coupling. Complex terrain, vegetation resistance, and fire spread processes jointly affect personnel mobility, making traditional path planning methods based on static cost grids unable to simultaneously balance safety and efficiency. In mountainous wildfire scenarios, a path that is spatially feasible may still become unsafe when the fire perimeter evolves over time, which highlights the necessity of incorporating both traversability and fire-induced risk into a unified decision framework. Therefore, this study aims to develop a safety-aware and efficiency-oriented path planning method that can explicitly represent the coupled effects of terrain-vegetation constraints on human movement and the dynamic fire spread threat, and provide a practical decision-making basis for emergency rescue operations. [Methods] To address this challenge, this study develops a multi-source environmental modeling framework that integrates terrain slope, vegetation density, firefighter velocity fields, and fire spread risk, and proposes a safetyaware path planning method driven by environmental velocity fields for dynamic forest fire scenarios, termed HNP-DuelingDQN. A composite reward mechanism composed of goal guidance, traversal speed, safety factor, and task completion terms is designed to reduce sparse feedback and guide policy learning toward both efficient movement and risk avoidance. In addition, Heuristic Action Selection(HAS) and an N-PER experience replay structure are incorporated into the reinforcement learning framework to improve learning stability and efficiency. The overall design emphasizes that mobility is governed by the environmental velocity field, while safety is constrained by fire spread risk, enabling the agent to learn a balanced path planning policy under dynamic wildfire conditions. [Results] Simulation experiments based on real terrain data from Fujian Province and two representative forest fire cases within the past five years show that the proposed method exhibits stable advantages in both static and dynamic fire environments. Compared with HNP-DQN, A*, and ACO algorithms, the proposed method significantly improves overall path safety, reducing fire spread risk levels by approximately 17.4%～57.8% on average, increasing average traversal speed by 13.4%～28.3%, and reducing average path slope by 11.0%～29.9%. These results suggest that the learned policy consistently steers paths away from high-risk areas while preserving favorable traversability, and maintains reliable performance under time-varying fire conditions, which is essential for emergency scenarios requiring the simultaneous consideration of route feasibility and safety margins. [Conclusions] The results demonstrate that the proposed method achieves coordinated optimization of path safety and traversal efficiency, providing a more real-time, interpreter and quantitatively supported technical solution for emergency rescue path decision-making in mountainous forest fire scenarios. Overall, the proposed framework offers an operationally meaningful approach to wildfire-related path planning by coupling multi-source environmental traversability and fire-induced risk within a reinforcement learning decision process, and it can serve as a practical reference for safety-aware route selection under dynamic wildfire threats.更多还原