Research Analyzer
← Back ICRA 2026

Wildfire Containment Using Multi-drone Systems: A Predict-Then-Optimize Approach

Aoran Cheng, Shijie Pan, Yiqi Sun, Kai Kang, Cristobal Pais, Yulun Zhou, Zuo-Jun Max Shen

PDF

AI summary

Key figure (auto-extracted from paper)
A predict-then-optimize framework combining deep learning fire prediction with chance-constrained robust optimization significantly outperforms existing methods in both wildfire spread accuracy and large-scale multi-drone task success.
wildfire suppression multi-drone systems predict-then-optimize chance-constrained optimization deep learning prediction drone task allocation

Problem

Multi-drone wildfire firefighting is hindered by complex, dynamic environments and the difficulty of integrating accurate fire spread predictions with efficient, robust task allocation algorithms.

Approach

The method first predicts future fire spread using a deep residual network (Firefighter-DResNet), then optimizes multi-drone task assignments and routing via a Mixed-Integer Programming model enhanced with Dynamic Programming and Chance-constrained Robust Optimization.

Key results

  • Firefighter-DResNet surpasses state-of-the-art models by 78.1% and 86.4% in IoU for 20×20 and 40×40 environments
  • Achieves over 54.7% and 27.3% higher precision than leading baselines
  • Chance-constrained optimizer secures a 30% success rate in large-scale 40×40 scenarios where all baselines fail
  • Seamlessly integrates prediction and optimization to enable robust, real-time multi-drone coordination

Why it matters

Offers emergency responders and disaster management agencies a scalable, data-driven solution for rapidly coordinating autonomous drone swarms in dynamic wildfire emergencies.

Abstract

A multi-drone system coupled with data in- telligence can be the future of wildfire fighting. However, multi-drone firefighting faces considerable challenges, such as complex environmental conditions, rapidly changing wildfire spread, and computational complexity in multi-drone opera- tions. We address these challenges by developing a predict-then- optimize approach to enable multi-drone firefighting. First, we create a wildfire spread prediction model Firefighter-DResNet based on previous wildfire data. Next, we propose a Mixed- Integer Programming (MIP) model coupled with Dynamic Programming (DP) to facilitate efficient multi-drone task plan- ning. We further use Chance-constrained Robust Optimization (CCRO) to ensure robust performance under varying condi- tions. After training with 75 varied wildfire environments, our approach demonstrates exceptional performance in controlling wildfires through multi-drone coordination. Our prediction model Firefighter-DResNet outperforms state-of-the-art models by 78.1% (20×20 environment) and 86.4% (40×40 environment) in terms of Intersection over Union (IoU), and achieves more than 54.7% and 27.3% higher precision. Subsequently, our chance-constrained optimizer succeeds in large-scale scenarios, achieving a 30% success rate in N=16 40×40 scenarios where all other baselines fail. This integrated solution significantly enhances the robustness and efficiency of autonomous multi- drone firefighting, marking a substantial leap towards practical deployment.

Index terms

Path Planning for Multiple Mobile Robots or Agents Planning Scheduling and Coordination Aerial Systems: Applications

Related papers