Modelling and optimization of unmanned aerial vehicles trajectories using machine learning for collection of data from ground sensors for forest fire scenarios

Abstract

Forest fires are rapidly spreading fires that occur in natural areas, causing significant damage to ecosystems, property, and human lives. Wildfire detection is essential for mitigating the severe consequences of fire outbreaks. This thesis presents a framework integrating unmanned aerial vehicles (UAVs) and ground-based sensor nodes (SNs) to achieve efficient wildfire detection while minimizing resource utilization. The model simultaneously simulates wildfire propagation and UAV movement focusing on refining UAV trajectories, for fire-specific information collecting from SNs while minimizing uplink data transmission energy consumption of SNs. Deep reinforcement learning (DRL) approach is selected to address the wildfire detection problem due to its capability to adapt dynamic environments and optimize decision making in real-time. Problem is structured as Markov decision process (MDP) to effectively model the interactions between the UAVs and the environment.

The primary contribution of this research is the introduction of a boundary curve approach, which evaluates the compromise between transmission energy consumption and the number fire detected SNs that have not transmitted data to UAV, formulated as regret. The model is evaluated in both single-UAV and multi-UAV scenarios. For multi-UAV setups, both centralized multi-agent reinforcement learning (CMARL) and decentralized multi-agent reinforcement learning (DMARL) frameworks are utilized. Simulation results demonstrate the proposed hybrid model incorporating two-UAVs with decentralized action space and centralized state space shows better performance in reducing unserved fire detected SNs and improving energy efficiency. Specifically, at an average per-UAV energy consumption level of 0.2 J, the centralized two-UAV approach achieves a 73% reduction in cumulative regret compared to the single-UAV setup, while the hybrid method further reduces cumulative regret by 90% at the same energy level. These results underscore the adaptability and efficiency of the proposed framework in addressing dynamic wildfire detection scenarios, making it a robust solution for real-world applications.