PaperlayerView exact revision on arXiv

Source-linked AI summary

A comprehensive survey of research towards AI-enabled unmanned aerial systems in pre-, active-, and post-wildfire management

Sayed Pedram Haeri Boroujeni, Abolfazl Razi, Sahand Khoshdel, Fatemeh Afghah, Janice L. Coen, Leo ONeill, Peter Z. Fule, Adam Watts, Nick-Marios T. Kokolakis, Kyriakos G. Vamvoudakis

arXiv:2401.02456v1cs.LGcs.AI

TL;DR

Wildfire management lacks a comprehensive review centered on AI-enabled UAV systems across all management stages. This survey synthesizes UAV technologies, AI and computer-vision methods, reinforcement learning, and wildfire modeling from pre-fire through post-fire management, concluding that their integration provides broader management insights and predictive capabilities while remaining constrained by UAV and observation limitations.

  • Problem

    Existing reviews do not comprehensively cover AI-enabled UAV systems and their applications across pre-fire, active-fire, and post-fire wildfire management.

  • Method

    The survey analyzes more than seven hundred studies covering UAV technologies, remote sensing, AI methods, wildfire modeling, and management strategies across the three wildfire phases.

  • Results

    The review identifies AI-enabled UAV applications for fuel monitoring, wildfire detection, classification, segmentation, monitoring, control, recovery planning, damage assessment, and evacuation-related operations.

  • Takeaways & Limitations

    Integrating AI techniques with UAV-based data offers novel insights and enhanced predictive capabilities for understanding dynamic wildfire behavior and improving wildfire-management efforts.

  • Takeaways & Limitations

    Small low-flying single UAVs are not well suited to collect full fire-perimeter observations beyond small fires and face challenges mosaicing coordinated-drone imagery.

Abstract

from arXiv · show

Wildfires have emerged as one of the most destructive natural disasters worldwide, causing catastrophic losses in both human lives and forest wildlife. Recently, the use of Artificial Intelligence (AI) in wildfires, propelled by the integration of Unmanned Aerial Vehicles (UAVs) and deep learning models, has created an unprecedented momentum to implement and develop more effective wildfire management. Although some of the existing survey papers have explored various learning-based approaches, a comprehensive review emphasizing the application of AI-enabled UAV systems and their subsequent impact on multi-stage wildfire management is notably lacking. This survey aims to bridge these gaps by offering a systematic review of the recent state-of-the-art technologies, highlighting the advancements of UAV systems and AI models from pre-fire, through the active-fire stage, to post-fire management. To this aim, we provide an extensive analysis of the existing remote sensing systems with a particular focus on the UAV advancements, device specifications, and sensor technologies relevant to wildfire management. We also examine the pre-fire and post-fire management approaches, including fuel monitoring, prevention strategies, as well as evacuation planning, damage assessment, and operation strategies. Additionally, we review and summarize a wide range of computer vision techniques in active-fire management, with an emphasis on Machine Learning (ML), Reinforcement Learning (RL), and Deep Learning (DL) algorithms for wildfire classification, segmentation, detection, and monitoring tasks. Ultimately, we underscore the substantial advancement in wildfire modeling through the integration of cutting-edge AI techniques and UAV-based data, providing novel insights and enhanced predictive capabilities to understand dynamic wildfire behavior.

1. Introduction

Wildfires impose severe ecological, economic, and human costs, motivating comprehensive AI-enabled UAV strategies across pre-fire, active-fire, and post-fire management. This survey reviews UAV technologies, AI methods, wildfire modeling, and open challenges across these stages.

  • Motivation: Wildfires cause substantial economic losses, ecological damage, and threats to communities, while a small fraction of fires accounts for most burned area.Only 3–5% of wildfires exceed 100 hectares, yet the largest 1% are responsible for 80–96% of total area burned.
  • Synthesis and outlook: The paper synthesizes research across disciplines to identify monitoring needs, technological capabilities, modeling gaps, and future directions for AI-enabled UAV wildfire management.The authors highlight the integration of observations, UAV hardware and field use, flight optimization, AI, and wildfire modeling, while noting post-fire urgency from secondary disasters.
  • Survey scope: The survey examines AI-enabled UAV systems from pre-fire planning through active-fire detection and control to post-fire recovery and damage management.Its scope includes prevention, early warning, monitoring, firefighting, recovery planning, evacuation, damage assessment, and operational strategies.
  • UAV technologies: The review analyzes UAV-based visual remote sensing systems, including their device types, strengths, weaknesses, and wildfire-management applications.It also considers how UAV technologies and wildfire modeling can support more efficient firefighting efforts.
  • Active-fire AI: The survey evaluates machine-learning and deep-learning computer-vision methods for wildfire detection, classification, and segmentation in active-fire management.It also investigates reinforcement learning for wildfire monitoring and describes this as a comprehensive assessment of RL-based UAVs in wildfire management.

2. Background and Related Literature

Wildfire behavior is shaped by ignition, fuel, moisture, weather, and other factors that can operate below the resolution of satellites and mesoscale networks. The survey reviews prior wildfire-management research to position AI-enabled UAV systems within these observational and technical gaps.

  • Wildfire stages and influential factors: Lightning-ignited fires can smolder undetected for days, especially when cloud cover obscures detection and later drying enables spread.
  • Wildfire stages and influential factors: Whether an ignition becomes self-sustaining depends on heat release overcoming fuel moisture and limitations in fuel amount or continuity.
  • Wildfire stages and influential factors: Many fire-controlling factors and thresholds occur beneath the spatial scales detected by satellites or mesoscale meteorological networks, or are obscured by canopies.
  • Review of existing surveys: The review systematically searched academic databases for wildfire-management surveys published from 2015 to 2023 using terms spanning UAVs, remote sensing, computer vision, detection, and monitoring.
  • Review of existing surveys: The authors critically evaluated prior surveys and summarized their advantages, limitations, and technical gaps in a comparative table.
  • Review of existing surveys: Earlier surveys covered UAV platforms, sensors, remote sensing, and fire-detection methods, but differed in their treatment of machine learning, UAV technologies, datasets, and technical challenges.

3. UAV Technology and Device Specifications

UAV wildfire systems combine platform selection with specialized sensors to collect aerial, high-resolution, and real-time information for monitoring, detection, and response. The survey compares UAV configurations, sensor categories, capabilities, and operational constraints.

  • UAVs provide aerial perspectives, broad-area coverage, and operation in challenging environments for wildfire monitoring and detection.
  • UAV Types: Fixed-wing UAVs offer long endurance and large payloads, rotary-wing UAVs provide low-altitude maneuverability, and hybrid systems combine horizontal and vertical takeoff capabilities.
  • UAV Types: UAVs are also classified as bicopters, tricopters, quadcopters, hexacopters, or octocopters according to rotor and propeller counts.
  • Sensor Types: Optical, thermal, gas, acoustic, meteorological, navigation, and chemical sensors support complementary wildfire observations, including visual signs, hotspots, temperatures, and combustion gases.
  • Sensor Types: UV cameras can detect flame-related ultraviolet wavelengths near 200 nm and hidden fire risks, but sunlight interference limits daytime discrimination.
  • Challenges and Future Directions: UAV deployment remains constrained by limited endurance and range, payload capacity, large multispectral datasets, adverse weather, and differing regulations.

4. Pre-Fire Management

Pre-fire management uses AI and UAVs for fuel monitoring, fire-risk analysis, and detection, while addressing limitations in data coverage, model validation, and UAV operations. Existing work shows accurate fuel estimation but still largely relies on post-flight processing.

  • Pre-fire management applies AI-enabled UAVs to fuel monitoring, fire hazard modeling, and wildfire detection.
  • Fuels Monitoring and Management: UAV imagery and machine learning support fuel-condition monitoring, but current imagery is generally processed after flight rather than during mission planning or decisions.
  • Fuels Monitoring and Management: R2 = 0.87 and RMSE = 11.3% were reported for deep-learning estimates of biomass from UAV-derived imagery across five fuel types.
  • Fire Hazard and Risk: Fire-risk modeling remains limited by scarce comprehensive datasets, changing hazard conditions, and the extensive observations required for predictive-model validation.
  • Fire Hazard and Risk: AI-enabled UAVs can collect fine-scale weather, fuel, ignition, and infrastructure data needed for wildfire behavior and risk modeling.
  • Challenges and Future Directions: U.S. UAV operations are constrained by visual-line-of-sight rules, a 122-meter altitude limit, and continuous operation by one pilot, while broader deployment requires further safety and practicality testing.

5. Active-Fire Management

Active-fire management organizes AI methods around detection, monitoring, and control. The survey distinguishes supervised, unsupervised, and agent-based learning and reviews ML, DL, and RL applications supported by computer vision.

  • Active-fire management uses UAV-equipped AI and computer vision to support wildfire detection, monitoring, and control.
  • Algorithmic Framework: Active-fire algorithms are broadly categorized as supervised, unsupervised, or agent-based learning.
  • Wildfire Detection: Detection studies include ML and DL approaches, with DL tasks covering wildfire classification, segmentation, and object detection.
  • Monitoring and Control: RL-based and agent-based learning methods are reviewed for wildfire monitoring, while control methods address effective wildfire management.

5.1. Wildfire Detection

Wildfire detection research applies ML and DL to image and video data, with approaches spanning classification, segmentation, and object detection. The reviewed methods emphasize timely detection, challenging small-fire recognition, and deployment considerations for UAV and resource-constrained systems.

  • Wildfire Detection: Early wildfire detection uses ML and DL algorithms to analyze image and video data for smoke, flames, and vegetation changes.These models can be deployed on AI-enabled UAV systems for automated data processing and monitoring.
  • Machine Learning Approaches: Supervised learning supports wildfire classification and regression, while unsupervised learning uses clustering and dimension reduction for wildfire data analysis.The reviewed tables summarize these algorithms for detection, prediction, mapping, and classification tasks, alongside their benefits and limitations.
  • Deep Learning Approaches: Deep learning methods are reviewed for wildfire classification, segmentation, and object detection, providing timely information for active-fire detection, monitoring, and post-fire analysis.The survey emphasizes these tasks as central applications of DL-based wildfire computer vision.
  • Wildfire Segmentation Approaches: Deep-RegSeg achieves an F1-score of 94.46% for wildfire segmentation under smoke, changing luminosity, and diverse weather and brightness conditions.Its RegNet backbones and alternative loss functions provide flexibility for segmenting fire pixels and identifying small fire areas.
  • Wildfire Segmentation Approaches: U-Net with a ResNet50 backbone provides the highest segmentation accuracy among FCN, U-Net, PSPNet, and DeepLabV3+ models evaluated for forest-fire imagery.DeepLabV3+ with ResNet50 also demonstrates satisfactory segmentation performance with faster running speed.
  • Wildfire Object Detection Approaches: Small-fire detection remains challenging for YOLO-based systems, motivating added small-target layers, attention modules, and feature-extraction changes.Fire-Yolo reports real-time detection at an average of 0.04 s per frame at 416 resolution and outperforms Faster R-CNN and unimproved YOLOv3 for very small targets.

5.2. Wildfire Monitoring

Wildfire monitoring is framed as active exploration supported by onboard detection, path planning, coverage, coordination, and learning methods. The review emphasizes that monitoring objectives must balance coverage, computation, communication, path length, and vehicle constraints.

  • Monitoring is defined as active exploration to locate ignited areas using data processed and annotated by an onboard detection module.
  • Wildfire monitoring aggregates objectives including sufficient coverage, low computation and communication overhead, and short paths between local destinations.
  • Monitoring tasks: Single-UAV monitoring combines constrained path planning with subsequent coverage maximization, while multi-UAV systems divide missions into navigation, coverage, communication, learning, recharging, and alarming tasks.
  • Path planning: Trajectory optimization determines efficient continuous or discrete motion and higher-order dynamics, whereas path planning selects optimal waypoints.
  • Control-based optimization: One reviewed system uses a tower-based thermal alarm to generate safe take-off, orbit, and return waypoints guiding a UAV toward detected fires.
  • Reinforcement learning: Reinforcement learning is presented as suitable for multi-constraint monitoring, but many algorithms remain untested for wildfire path planning, identifying an algorithmic research gap.

5.3. Wildfire Control

Wildfire control extends detection and monitoring into suppression, using UAVs to deliver retardants or water and coordinate aerial interventions. The reviewed approaches expose trade-offs among delivery altitude, heat exposure, payload, speed, and fire coverage.

  • Effective detection and monitoring require a suppression framework, with control theory, optimization, and reinforcement learning identified as useful control approaches.
  • Aerial suppression: UAVs can deploy retardants or water over fires to help prevent spread toward industrial or residential borders.
  • Learning-based control: Distributed multi-agent reinforcement learning models UAV movement and retardant dumping jointly on a lattice whose trees occupy healthy, fire, or burnt states.
  • Suppression evaluation: Ortho-rectified airborne infrared imagery can measure drop dimensions, perimeter proximity, and effects on fire spread for comparing suppression tactics and delivery systems.
  • Control-based optimization: Retardant delivery must balance dissipation at high altitude against heat exposure at low altitude, while heavier payloads constrain maximum speed.

5.4. Challenge, Discussion, and Future Directions

Active-fire AI and UAV systems support wildfire detection, monitoring, and control, but practical deployment remains constrained by real-time processing and robustness across environments. The review calls for optimized systems, interdisciplinary development, and rigorous validation.

  • Active-fire management integrates ML, DL, and RL with UAVs for wildfire detection, monitoring, classification, segmentation, and control.
  • Challenges: Limited real-time processing and computational constraints hinder rapid decision-making in dynamic wildfire scenarios.
  • Challenges: Algorithm reliability and accuracy must be maintained across diverse weather and terrain conditions.
  • Future directions: Future work should combine optimized algorithms, hardware improvements, and collaboration among computer scientists, wildfire experts, and UAV engineers.
  • Conclusion: The review concludes that addressing current challenges can support more resilient, adaptive, and efficient wildfire management systems.

6. Post-Fire Management

Post-fire management uses UAVs and related immersive technologies for damage assessment, recovery monitoring, evacuation planning, rehabilitation, and workforce training. The survey highlights high-resolution sensing, autonomous planning, and simulated testing as important directions.

  • UAV-assisted post-fire management supports damage assessment, evacuation planning, rehabilitation, and ecosystem recovery after wildfires.
  • Forest recovery monitoring: High-resolution UAV imagery reveals fire damage, vegetation regrowth, and ecosystem dynamics during forest recovery monitoring.
  • Damage assessment: UAV-SfM and multispectral sensing provide methods for measuring forest impacts and assessing fire severity, including comparison with airborne laser scanners.
  • Evacuation planning: UAV imagery and mapping support evacuation planning by supplying information on road conditions, traffic congestion, hazards, and potential routes.
  • AR/VR training: AR and VR support workforce training and safe wildfire operations, including situational-awareness assessment for air attack supervision.
  • Experimental platforms: MAR-CPS enables controlled indoor testing of planning and learning algorithms in dynamic simulated wildfire environments using motion capture and projected displays.
  • Future directions: Future directions include automated recovery assessment, AI-based evacuation planning, autonomous UAV guidance, and AR-based firefighting training.

7. Wildfire Modeling

Wildfire models support understanding observed mechanisms, exploring hypothetical conditions, accessing unobserved processes, and predicting future behavior.

  • Wildfire models are used to test explanations of observed fire mechanisms and explore what-if scenarios under hypothetical conditions.
  • They also provide information about processes or variables that cannot be directly observed and support forecasts of future wildfire behavior.

7.1. Physics-aware Approaches to Fire Behavior and Effects

Physics-aware wildfire modeling has progressed from traditional fire-behavior approaches toward more physically based dynamic models that represent fire responses to environmental conditions and associated effects.

  • Recent physically based dynamic models represent how fire spreads through fuel strata and what phenomena it produces in response to environmental conditions.
  • Modeling has evolved from separate kinematic calculations of fire behavior and effects toward computational fluid dynamics, physics-based, data-driven, and machine-learning approaches.
  • Figure 19 combines CAWFE-modeled heat flux and near-surface winds with VIIRS active-fire detections and Landsat 8 post-fire imagery for the 2020 Calwood Fire.
  • Modeled burn-related variables are presumed to correlate with analogous products, but this relationship has only been loosely examined.

7.2. Data Driven Fire modeling

Data-driven wildfire modeling uses observations and learned representations to infer fire spread, but depends on sufficient, diverse data and must balance spatial and temporal resolution across sensing systems.

  • Data-driven wildfire modeling has expanded with computational resources and deep learning, while data quantity and quality remain essential prerequisites.
  • The modeling task may predict continuous fire intensity or spread rate after classifying fire versus no-fire or segmenting areas by ignition probability.
  • Wildfire spread modeling infers fire state at a location and time, then predicts one or more future time steps using probabilistic or non-probabilistic approaches.
  • The reviewed literature includes cellular automata, machine-learning models using satellite and UAV data, and probabilistic approaches with Bayesian updates.
  • Cellular automata simulate spatio-temporal spread through recursive neighboring-cell rules, but simultaneous cell updates may not match actual fire behavior.
  • Deep learning is suited to extracting abstract spatio-temporal features from complex, high-dimensional, multi-source, and multi-scale Earth-system data.
  • GEO satellites support long-term observation sequences but generally provide lower spatial resolution than lower-altitude LEO satellites.
  • Satellite and UAV spread-modeling studies use deep networks, thermal imagery, environmental inputs, fire-perimeter detection, and similarity metrics such as Jaccard, SDI, and Sørensen indices.

7.3. The role of UAVs in wildfire behavior and effects modeling

UAV observations support fine-scale investigation of fire behavior and effects, while operational landscape-scale modeling combines repeated simulations with updated weather and fire-mapping data.

  • UAVs can capture pre-fire conditions, combustion modes, fuel consumption, remaining fuels, and fine-scale burn-severity variability.
  • Fine-scale atmospheric simulations have limited predictability of roughly 1–2 days, motivating sequential weather-forecast-style landscape simulations.
  • These sequential simulations are initialized and validated with the latest weather and fire-mapping data, including VIIRS, airborne, incident, and polar-orbiting satellite sources.

7.4. Challenges, gaps, and future directions

Wildfire modeling requires diverse, timely information, but existing data and sensing systems have resolution, coverage, and deployment limitations. Future directions include improved high-resolution observations, integrated monitoring and spread modeling, and computationally efficient onboard systems.

  • Data requirements: Fire behavior models require ignition timing or location, fire extent, weather, fuel moisture, terrain, and later observations for validation.Datasets collected for legacy models may be incomplete or mismatched for higher-dimensional models and differing spatial or temporal resolutions.
  • Remote sensing gaps: Small low-flying UAVs cannot generally capture complete fire perimeters except for small fires, while coordinated-drone mosaicing remains challenging.High-flying aircraft and Predator-class UAVs can encompass most fire perimeters in a single image, unlike smaller single UAVs.
  • Modeling trade-offs: Coupled weather-fire models require spin-up and lose skill over time, whereas data-driven approaches are better suited to short-term prediction.Weather-station observations alone are too sparse to represent all conditions driving a fire, including conditions generated within the fire itself.
  • Future directions: Research gaps include non-satellite spread modeling, high-resolution generalizable models, integrated monitoring and tracking, and low-complexity systems deployable onboard aircraft.Large modular systems must meet computational constraints for real-time aerial monitoring.

8. Conclusion

The survey reviews AI-enabled UAV wildfire management across pre-fire, active-fire, and post-fire phases. It synthesizes prevention, computer vision, reinforcement learning, recovery, and damage-assessment research while identifying future directions for wildfire management.

  • Survey scope: The survey organizes UAV and AI wildfire-management research into pre-fire, active-fire, and post-fire phases.The pre-fire review covers preprocessing, prevention strategies, and early-warning systems.
  • Active-fire management: Active-fire coverage evaluates UAV computer-vision studies and deep-learning algorithms for detection, classification, and segmentation.The survey also examines reinforcement-learning algorithms for wildfire monitoring.
  • Post-fire management: Post-fire coverage reviews recovery planning, damage assessment, and strategies for mitigating post-fire impacts.The paper also discusses open problems and future directions for researchers, policymakers, and professionals.
Loading 2401.02456v1…