Cooperative Drone and Water Supply Truck Scheduling for Wildfire Fighting Using Deep Reinforcement Learning

Lin-Yuan Bai, Xin Chen, Hai-Feng Ling, Yu‐Jun ZhengDrones 2025 · DOI

TL;DR

This paper tackles the critical challenge of efficiently fighting wildfires by proposing a novel deep reinforcement learning (DRL) approach for coordinating drones and water supply trucks. The DRL model learns to dynamically schedule these resources to minimize the total time required to extinguish a wildfire, outperforming traditional optimization methods.

What Problem Does This Paper Address?

Wildfires are a growing threat, spreading rapidly and causing immense damage. While drones offer a promising solution for delivering water or fire retardants, their limited payload capacity necessitates frequent reloads. This paper addresses the complex problem of optimizing the coordinated deployment of multiple drones and water supply trucks to overcome this limitation. Specifically, it focuses on scheduling drones to extinguish fires in different subareas and ensuring they can efficiently reload water from mobile trucks or fixed water sources. The core challenge lies in determining optimal drone paths, truck routes, and refueling locations in real-time to minimize the overall firefighting operation duration, a crucial factor in emergency response.

Methods

The authors propose a Deep Reinforcement Learning (DRL) framework to solve this complex scheduling problem. The DRL model consists of two main components:

  • Encoder: This part takes various input features, including roadway and airway networks, environmental conditions (temperature, wind), subarea characteristics (vegetation density, heat), water source availability, and the number of drones and trucks. It uses Graph Neural Networks (GNNs) for network data and Recurrent Neural Networks (RNNs) for sequential and environmental data to generate high-level embeddings of the problem instance.
  • Decoder: This component iteratively constructs the solution. It uses an attention mechanism to predict the next subarea to be targeted for firefighting. At each step, it considers the current state, including updated subarea information and available water resources. The decoder learns to make sequential decisions, assigning drones to subareas and determining their reload strategies (from trucks or water sources) to minimize the completion time.

The training of the DRL model uses a policy gradient method with a rollout baseline, aiming to minimize the expected objective function (total firefighting time).

Datasets

The study utilizes test instances constructed upon three real-world wilderness areas in Zhejiang Province, China: Hangzhou West Mountain Forest Park and two locations within the Tianmu Mountain Nature Reserve. The paper provides a table (Table 1) detailing the number of ignitable subareas, water supply subareas, and the specific test instances created for each area, varying in the number of drones and trucks. These instances are generated by varying ignition points, environmental conditions, and resource availability.

Key Insights and Ideas

  • Drone-Truck Synergy is Crucial: The paper highlights that simply using drones isn't enough; their effectiveness is significantly amplified by the mobile water supply provided by trucks, enabling continuous operations.
  • Real-time Dynamic Scheduling: The problem is inherently dynamic, with fire spread and resource availability changing. The DRL approach excels at adapting to these changes and making real-time decisions, which is critical for emergency response.
  • DRL for Complex Combinatorial Problems: This work demonstrates the power of DRL in tackling complex combinatorial optimization problems like scheduling, where traditional methods struggle with the vast solution space and dynamic nature.
  • End-to-End Solution Generation: Unlike methods that break down the problem into sub-problems, the DRL approach learns an end-to-end mapping from problem instance to a complete solution, potentially leading to more globally optimal outcomes.

Key Results

The proposed DRL method demonstrated significant performance advantages over several heuristic and metaheuristic optimization methods (NEH, Suliman, DE, EBO, WWO, VNS, and Meme).

  • Superior Solution Quality: For most test instances, the DRL method achieved significantly better (shorter) firefighting times compared to all seven comparative algorithms.
  • Fast Solution Generation: Once trained, the DRL model could produce solutions very quickly, typically within 30 seconds, which is considerably faster than the one to three minutes required by the heuristic/metaheuristic algorithms.
  • Scalability: While heuristics struggled with larger instances, the DRL method showed better performance and faster solution times, with its computational complexity being more linearly proportional to the instance size. For example, on the largest instance (C-5), DRL achieved a solution in 424 minutes, while the best comparative algorithms took 520-621 minutes.

Limitations and Future Work

  • Battery Consumption Not Explicitly Modeled: The current formulation does not explicitly consider drone battery consumption. However, the authors note that battery replacement can often be done simultaneously with water reloading, having a trivial effect on the overall schedule.
  • Extension to Other Drone Types: The current study focuses on drones carrying water capsules. Future work will extend the scenario to include drones carrying fire-extinguishing bombs and those with water hoses.
  • Integration with Rescue Operations: The authors plan to integrate firefighting operations with rescuing victims, requiring cooperative scheduling of drones, ground vehicles, and human rescuers.
  • Further Scenario Exploration: Future work will also explore scenarios involving drones acting as mobile base stations for communication enhancement.

Summary updated

Abstract

Wildfires often spread rapidly and cause significant casualties and economic losses. Firefighting drones carrying water capsules provide an efficient way for wildfire extinguishing, but their operational capabilities are limited by their payloads. This weakness can be compensated by using ground vehicles to provide mobile water supply. To this end, this paper presents an optimization problem of scheduling multiple drones and water supply trucks for wildfire fighting, which allocates burning subareas to drones, routes drones to perform fire-extinguishing operations in burning subareas and reload water between every two consecutive operations, and routes trucks to provide timely water supply for drones. To solve the problem within the limited emergency response time, we propose a deep reinforcement learning method, which consists of an encoder for embedding the input instance features and a decoder for generating a solution by iteratively predicting the subarea selection decision through attention. Computational results on test instances constructed upon real-world wilderness areas demonstrate the performance advantages of the proposed method over a collection of heuristic and metaheuristic optimization methods.
BibTeX
@inproceedings{bai2025cooperative,
  author = {Lin-Yuan Bai and Xin Chen and Hai-Feng Ling and Yu‐Jun Zheng},
  title = {{Cooperative Drone and Water Supply Truck Scheduling for Wildfire Fighting Using Deep Reinforcement Learning}},
  booktitle = {Drones},
  year = {2025},
  doi = {10.3390/drones9070464},
}
Figure