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 comprehensive AI-enabled UAV reviews spanning pre-fire, active-fire, and post-fire stages. This survey synthesizes UAV systems, sensors, and AI methods across these stages, reporting broad progress in wildfire detection, segmentation, monitoring, modeling, and management.

  • Problem

    Wildfire modeling and management still lack detailed initiation data and a comprehensive review of AI-enabled UAV applications across management stages.

    Source: §2.1 · p. 10

  • Method

    The survey systematically reviews UAV systems, sensor technologies, and AI approaches for pre-fire, active-fire, and post-fire wildfire management.

    Source: §8 · pp. 72–73

  • Results

    The survey finds substantial advances in UAV- and AI-based wildfire detection, classification, segmentation, monitoring, recovery planning, and damage assessment.

    Source: §8 · pp. 72–73

  • Takeaways & Limitations

    The synthesis identifies AI-enabled UAV technologies as relevant across wildfire prevention, response, recovery, and future wildfire-management research.

    Source: §8 · pp. 72–73

    Small low-flying single UAVs are not well suited to collecting full-perimeter observations for larger wildfires and face image-mosaicing challenges.

    Source: §7.4 · p. 72

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 cause extensive human, economic, ecological, and infrastructural harm, motivating integrated management across pre-fire, active-fire, and post-fire stages. This survey reviews AI-enabled UAV systems, sensing, modeling, and open challenges across those stages.

1. Introduction

  • Wildfires threaten communities, ecosystems, air, water, and soil, while a small fraction of fires produces most burned area.Only 3–5% of wildfires exceed 100 hectares, yet the largest 1% accounts for 80–96% of total burned area.Source: §1 · pp. 3–4
  • The survey examines AI-enabled UAVs across pre-fire, active-fire, and post-fire management, including sensing, monitoring, modeling, and integration challenges.It emphasizes limitations involving UAV technology, flight time, onboard processing, and closed-loop vision-based control.Source: §1 · pp. 5–6

1.1. Motivation of This Study

  • The review analyzes over seven hundred studies to address gaps in comprehensive AI-enabled UAV research for multi-stage wildfire management.Its interdisciplinary scope connects fire processes, monitoring needs, observations, instrumentation, UAV operations, and modeling from anticipation through impact assessment.Source: §1.1 · p. 6

1.2. Contributions of This Survey

  • The survey covers risk analysis, prevention, prediction, detection, monitoring, control, recovery, evacuation, damage assessment, and AR/VR-supported post-fire operations.It specifically reviews computer vision, deep learning, reinforcement learning, UAV remote sensing systems, and wildfire modeling.Source: §1.2 · p. 7, §1.1 · p. 7
  • It contributes detailed reviews of UAV visual remote sensing, deep-learning wildfire analysis, RL-based UAV monitoring, post-fire strategies, wildfire modeling, and open problems.The authors describe the RL review as, to their knowledge, the first comprehensive exploration and assessment of RL-based UAVs in wildfire management.Source: §1.2 · pp. 7–8

1.3. Organization of This Paper

  • The paper is organized into reviews of prior surveys, UAV technologies, pre-fire, active-fire, post-fire management, AI-based wildfire modeling, and conclusions with future directions.These topics are presented in Sections 2–8, respectively.Source: §1.3 · p. 8

2. Background and Related Literature

Wildfire behavior spans interacting fuel, weather, topographic, and combustion processes that are often unresolved by existing observation systems. This survey therefore reviews wildfire stages and prior surveys to motivate AI-enabled UAV systems for improved, multi-stage understanding and management.

2.1. Fire Stages-Ignition, Spread, and Aftermath

  • UASs can complement existing wildfire-monitoring capabilities through higher-resolution, more controllable data collection for early-growth mapping and process modeling.Lightning-caused ignitions may smolder undetected for days, and geostationary satellites may detect early fires without geographically locating them; rapid repeated mapping remains challenging.Source: §2.1 · p. 10
  • Wildfire ignition, spread, and extinction depend on interacting fuel, moisture, weather, topography, combustion modes, and thresholds across spatial and temporal scales.Fine-scale near-fire conditions remain insufficiently observed, while satellite and mesoscale networks may miss or obscure controlling processes.Source: §2.1 · p. 10

2.2. Review of The Existing Survey Papers

  • The review systematically searched academic databases for wildfire-management surveys published from 2015 to 2023, critically evaluating their relevance and limitations for AI-based UAVs.The selected surveys were assessed by titles, abstracts, findings, contents, and citations, with Table 1 summarizing their strengths, weaknesses, and technical gaps.Source: §2.2 · pp. 10–11, Table 1 · p. 12
  • Existing surveys address selected UAV, remote-sensing, detection, or deep-learning topics but leave gaps in integrated UAV technologies, datasets, sensors, reinforcement learning, and multi-stage wildfire management.The reviewed papers variously omit technical UAV and sensor detail, wildfire datasets, real-time monitoring, classification or segmentation, RL-based methods, challenges, and wildfire-stage coverage.Source: §2.2 · pp. 11, 13

3. UAV Technology and Device Specifications

The section surveys UAV-based remote sensing for wildfire monitoring, covering platform types, sensor categories, operational roles, and deployment challenges. It emphasizes that effective wildfire response depends on matching UAV capabilities and integrated sensing systems to mission requirements.

3. UAV Technology and Device Specifications

  • UAV remote sensing supports real-time wildfire data collection and monitoring, with platform selection depending on endurance, payload, maneuverability, and task requirements.UAVs can measure fire-relevant variables including temperature, humidity, wind speed, and smoke density while reducing human exposure in dangerous terrain.Source: §3 · p. 14

3.1. UAV Types

  • UAVs are categorized as fixed-wing, rotary-wing, or hybrid-wing systems, trading endurance and payload capacity against maneuverability, speed, complexity, and wind susceptibility.Fixed-wing systems suit long-range surveillance, rotary-wing systems support low-altitude close-range response, and hybrid systems combine horizontal and vertical takeoff capabilities.Source: §3.1 · p. 16
  • UAV configurations ranging from bicopters to octocopters provide differing stability, maneuverability, payload capacity, flight time, and speed for wildfire mapping, detection, tracking, prediction, and monitoring.Equipped cameras and sensors also support fire-behavior assessment, resource allocation, and, for some models, targeted delivery of water or retardants.Source: §3.1 · p. 16, Figure 7 · p. 17

3.2. UAVs’ Sensor Types

  • Wildfire UAVs integrate perception, acoustic, meteorological, navigation, and chemical sensors, combining imaging, environmental, positioning, and gas measurements for comprehensive monitoring.Optical, thermal, and gas sensors respectively support fire and smoke analysis, hotspot and temperature detection, and combustion-gas monitoring.Source: §3.2 · pp. 17, 19
  • RGB, IR, multispectral, hyperspectral, UV, LiDAR, and RADAR sensors extend wildfire observation from visible smoke and flames to thermal radiation, vegetation fuel, terrain, fire behavior, and adverse-weather detection.IR supports night-time and smoky operations; multispectral and hyperspectral sensing supports vegetation-risk and recovery analysis; LiDAR provides 3D structural data; and RADAR operates in adverse weather.Source: §3.2 · pp. 19–20

3.3. Challenge, Discussion, and Future Directions

  • UAV deployment remains constrained by limited endurance and range, payload capacity, sensor-data processing demands, adverse weather, and diverse regulatory frameworks.Proposed responses include improved batteries and lightweight materials, AI-enabled real-time analytics, autonomous systems, sensor fusion, miniaturization, 5G, and edge computing.Source: §3.3 · p. 22

4. Pre-Fire Management

Pre-fire management centers on fuel treatment, fire-hazard and risk modeling, and detection, with AI-enabled UAVs offering detailed data collection, monitoring, and predictive capabilities. The field remains constrained by limited integrated studies, insufficient fine-scale datasets, operational restrictions, and the need for field validation.

4. Pre-Fire Management

  • Fuel management is central to pre-fire decisions because vegetation is the most manageable wildfire-behavior factor, while thinning, prescribed fire, and invasive-species removal reduce hazards and support resilience.Fuel conditions evolve over years to decades, requiring treatment well before fire events.Source: §4 · p. 22

4.2. Fire Hazard and Risk

  • Pre-fire AI-UAV research remains limited: hazard modeling has received less attention than detection, and integrated UAV-AI processing is still evolving despite comparable performance to traditional methods [48].Comprehensive datasets are rare, and predictive models require extensive in situ observation or ex-situ replication for validation.Source: §4 · p. 23, §4.1 · p. 23, §4.2 · p. 24

4.1. Fuels Monitoring and Management

  • UAV-derived biomass estimates reached R2 = 0.87 and RMSE = 11.3% with deep learning, while PointNet-based vegetation classification extracted tree metrics with up to 95% accuracy.UAV imagery and laser-scanning data supported classification of heterogeneous fuel types and individual tree metrics, although imagery is currently processed post-flight.Source: §4.1 · p. 23

4.2. Fire Hazard and Risk

  • AI-enabled UAVs could improve fire-risk modeling by supplying sub-meter, fine-scale weather, fuel, ignition, infrastructure, and thermal data needed for simulation and validation.Existing studies use weather, fuel, and anthropogenic inputs but often label outputs as fire severity, hazard, or ignition indices rather than the broader natural-hazards definition of wildfire risk.Source: §4.2 · p. 24

4.3. Challenge, Discussion, and Future Directions

  • UAV deployment is constrained in the United States to line-of-sight operation, below 122 meters, and continuous piloting, while BVLOS standards, UAV fleets, and AI assistance could expand coverage and first-responder access [FAA's BEYOND program].Field testing is still needed to assess the safety and practicality of these options.Source: §4.3 · pp. 24–25

5. Active-Fire Management

Active-fire management uses UAVs with AI-enabled computer vision to support wildfire detection, monitoring, and control through machine learning, deep learning, and reinforcement learning. The section organizes these approaches into detection techniques, RL-based monitoring, and wildfire control methods.

5. Active-Fire Management

  • Active-fire management applies UAV-enabled AI and computer vision to detect, classify, and segment wildfire scenes while providing real-time data and decision support for suppression.The reviewed approaches aim to limit fire spread, protect lives and property, and support ecological restoration through firefighting teams and AI-enabled UAV systems.Source: §5 · p. 25
  • The algorithms are categorized as supervised, unsupervised, and agent-based learning, supporting wildfire detection, prediction, assessment, modeling, mapping, monitoring, and control.Figure 9 summarizes these algorithm classes and their potential wildfire-management applications.Source: §5 · p. 25, Figure 9 · p. 26

5.2 Wildfire Monitoring

  • Section 5.2 examines RL-based wildfire monitoring techniques, including agent-based learning algorithms that learn through environmental interaction and reward feedback.Such algorithms support sequential decision-making in complex wildfire systems.Source: §5 · pp. 25–26

5.1 Wildfire Detection

  • Section 5.1 reviews ML- and DL-based wildfire detection, including supervised and unsupervised methods alongside classification, segmentation, and object-detection algorithms.These methods address wildfire science and management through both conventional machine learning and deep learning approaches.Source: §5 · p. 25

5.3 Wildfire Control

  • Section 5.3 focuses on wildfire-control methods that facilitate effective active-fire management and support suppression decisions.Control strategies provide operational support for preventing further spread and reducing wildfire damage.Source: §5 · pp. 25–26

5.1. Wildfire Detection

This section reviews AI-enabled wildfire detection methods for UAV and image-based scenarios, spanning machine learning and deep learning approaches to classification, localization, segmentation, and recognition. It covers architectures and techniques including lightweight CNNs, encoder–decoder and attention mechanisms, transformers, multi-scale and YOLO-based detectors, and hybrid pipelines targeting accurate, interpretable, real-time, and resource-efficient analysis.

5.1. Wildfire Detection

  • AI-based detection deployed on UAVs can process image and video streams for earlier, more reliable, robust, and efficient wildfire monitoring.ML learns wildfire indicators such as smoke, flames, and vegetation change, while DL processes large datasets for rapid real-time detection.Source: §5.1 · pp. 26–27

5.1.1. ML-Based Techniques

  • ML supports wildfire management across stages by using satellite and drone data for detection, prediction, mitigation, resource allocation, and response planning.Supervised classification identifies fire-prone vegetation and terrain, whereas regression predicts fire spread and intensity from temperature, humidity, and wind speed.Source: §5.1.1 · p. 27
  • Unsupervised ML identifies wildfire hotspots, fire perimeters, and behavioral factors through clustering and reduces environmental-data complexity through dimension reduction [102, 103, 104, 105, 106].Clustering requires no prior data knowledge, while dimension reduction removes irrelevant or redundant features that increase computational complexity and can impair learning performance.Source: §5.1.1 · pp. 27, 29

5.1.2. DL-Based Techniques

  • DL methods review wildfire classification, segmentation, and object detection using satellite, aerial, UAV, and real-time video data.These models automatically learn patterns and representations for active-fire detection, monitoring, and post-fire analysis.Source: §5.1.2 · p. 29

5.1.2.1. Wildfire Classification Approaches

  • Deep-learning wildfire classification approaches categorize wildfire types and severity by analyzing flame color, smoke density, and temperature patterns, supporting risk assessment, response selection, and efficient resource allocation.Deep CNN frameworks use convolutional and pooling layers for feature learning, then a flattening layer and fully connected layers for feature classification.Source: §5.1.2.1 · pp. 29, 31

5.1.2.2. Wildfire Segmentation Approaches

  • DL segmentation performs pixel-wise fire, no-fire, smoke, and no-smoke classification, using encoder-decoder CNNs with upsampling to recover boundaries and softmax to assign class probabilities.U-Net is especially prevalent because encoder-decoder skip connections support precise semantic segmentation and wildfire mapping.Source: §5.1.2.2 · pp. 35–36
  • Segmentation architectures increasingly combine multi-scale context, attention, temporal information, and lightweight computation through models including FuseNet, DeepLabV3+, FlameTransNet, FFS-UNet, and SqueezeNet derivatives.These designs address spatial–temporal dependencies, class imbalance, fine details, UAV-video constraints, and resource-limited real-time deployment.Source: §5.1.2.2 · pp. 38–41
  • 94.46% F1-score: Deep-RegSeg outperforms recent state-of-the-art techniques while segmenting small fire areas under smoke, changing luminosity, weather, and brightness conditions.The method supports multiple RegNet backbones and Dice-based loss options for adapting wildfire segmentation performance.Source: §5.1.2.2 · p. 41
  • U-Net with a ResNet50 backbone achieves the highest segmentation accuracy among FCN, U-Net, PSPNet, and DeepLabV3+ models, while DeepLabV3+ with ResNet50 runs faster with satisfactory performance.The comparison evaluates four fully convolutional models across VGG16 and ResNet50 backbones.Source: §5.1.2.2 · p. 42

5.1.2.3. Wildfire Object Detection Approaches

  • YOLO variants dominate real-time wildfire object detection, while hybrid systems combine YOLO localization with U-Net segmentation, pruning, knowledge distillation, or EfficientDet ensembles.A coarse-to-fine YoloV3 framework targets sparse, small, and irregular fires in high-resolution UAS imagery through adaptive sub-region selection and detailed scrutiny.Source: §5.1.2.3 · pp. 43, 45
  • 96.8% improvement over the gold-standard detection approach, mAP of 50, and 122 FPS: one YOLO study reports stronger wildfire detection, especially for smoke under challenging conditions.Fire-Yolo separately achieves 0.04 seconds per frame at 416 × 416 resolution for very small target detection, outperforming Faster R-CNN and unimproved YoloV3 in efficiency.Source: §5.1.2.3 · p. 46
  • FireNet combined saliency detection with a 15-layer self-learning DCNN and achieved 98% overall accuracy for rapid wildfire identification and segmentation.Its practical utility was demonstrated on sampled news-report images for real-time inspection.Source: §5.1.2.3 · p. 47
  • FireXnet uses a lightweight architecture and SHAP explainability for efficient drone deployment and interpretable wildfire segmentation, with comparisons against five pretrained models [146].The compared models were VGG16, InceptionResNetV2, InceptionV3, DenseNet201, and MobileNetV2.Source: §5.1.2.3 · p. 47
  • Detection systems use specialized architectures, including SqueezeNet-inspired compact CNNs, Mask R-CNN-based Detectron2, and staged ensemble CNN-Yolo pipelines, to reduce computation or improve localization [168, 141].Detectron2 outlines fire regions at pixel level, while the staged pipeline first identifies abnormalities before localizing smoke or fire.Source: §5.1.2.3 · p. 47
  • The dual-channel CNN with attention and feature fusion surpassed its single-channel network in fire recognition, while transfer learning reduced overfitting and training time.The model combines distinct input sizes to handle different-sized fire scenes and uses two feature-fusion approaches to improve feature characterization.Source: §5.1.2.3 · pp. 47–48
  • Spatio-Temporal Self-Attention Networks fuse frame-sequence and single-frame features for fire detection, using segmentation maps as region proposals before classifying fire presence.TemporalNet, SpatioNet, and FuseNet model temporal, spatial, and joint spatial-temporal dependencies through multi-stage training.Source: §5.1.2.3 · p. 48
  • Transformer-based detectors improve contextual and small-target wildfire detection through multi-head attention, multi-scale pooling, Swin Transformers, PAFPN or BiFPN fusion, and attention-enhanced losses [211, 216, 215].FireDetn targets real-time complex scenarios, STPM SAHI addresses small fires, and FCDM combines SIoU Loss with CBAM and BiFPN.Source: §5.1.2.3 · pp. 48–49

5.2. Wildfire Monitoring

Wildfire monitoring is framed as active exploration in which detected ignited areas guide aerial path planning, balancing coverage, computation, communication, and travel distance. The surveyed approaches span trajectory optimization, control-based methods, reinforcement learning, and coordinated aerial fleets under physical, safety, and connectivity constraints.

5.2. Wildfire Monitoring

  • Monitoring uses detected ignited areas as inputs to path planning, with objectives including fire-front coverage, computational and communication efficiency, and short local paths.The monitoring cycle comprises sensing, preprocessing and detection, modeling and planning, and control.Source: §5.2 · pp. 49–50, Figure 13 · p. 49

5.2.1. Trajectory Optimization and Path Planning

  • Trajectory optimization and path planning generate efficient 2D or 3D routes or waypoints while accounting for time, power, speed, angle, and other UAV constraints.Waypoint selection affects flight-path cost, while optimized trajectories support data collection, situational awareness, and timely firefighting decisions.Source: §5.2.1 · p. 51, Figure 15 · p. 51

5.2.2. Control-based Optimization

  • Control-based methods guide UAVs toward detected fires through safe waypoint and trajectory generation, while cooperative low-altitude fleets align fire perimeters in their fields of view and reduce image-transfer latency.One framework uses a 360° thermal camera detecting fires within 3.5 km, followed by collision-free safety points and orbit-based sub-trajectories; another uses thermal-image classification and load balancing.Source: §5.2.2 · p. 52

5.2.3. Reinforcement Learning (RL) Methods

  • Reinforcement learning models wildfire monitoring as reward-driven sequential decision-making, but many surveyed algorithms remain untested for wildfire path planning, marking an algorithmic research gap.Applications include Q-learning and deep Q-learning for safe fire-front tracking, cooperative multi-agent learning, and POMDP-based planning under wind, fuel, uncertainty, and partial observability.Source: §5.2.3 · pp. 52–55

5.2.4. Aerial Fleets for Wildfire Monitoring

  • Aerial-fleet approaches coordinate centralized or decentralized UAVs—and, in some systems, ground vehicles—to optimize coverage, tracking, communication, collision avoidance, energy, latency, and mission completion.Methods include hierarchical UAV–UGV control, importance-based decentralized perimeter reconstruction that improves accuracy in 2-, 3-, and 4-UAV scenarios, leader-follower coalitions, consensus, communication trees, and cooperative model-predictive control.Source: §5.2.4 · pp. 55–56

5.3. Wildfire Control

Wildfire control uses UAVs with control theory, optimization, and reinforcement learning to support suppression, containment, and decision-making. Research spans multi-agent coordination, aerial-drop evaluation, and constrained control of suppressant delivery under environmental and vehicle constraints.

Wildfire Control

  • UAVs support active wildfire suppression by delivering retardants or water and informing aircraft-deployment decisions through containment-time comparisons.Their mobility and low-altitude operation make them suitable for early-stage control algorithms and preventing fire spread toward industrial or residential borders.Source: §5.3 · pp. 56–57
  • A distributed multi-agent RL framework models forests as locally interacting MDPs and trains MADQN agents to move and dump retardant using local information.Agents communicate with nearest neighbors, receive rewards tied to fire boundaries and threatened healthy trees, and use a shared replay buffer for exploration and generalization under partial observability.Source: §5.3 · p. 57
  • Drone suppression systems [268] launch from firetrucks, encircle or cover fires, and repeatedly crash-deploy retardant, while evaluates drop placement, coverage, and fire-behavior effects.Ortho-rectified airborne infrared imagery measures drop dimensions, perimeter proximity, and effects on fire spread to compare tactics, suppressants, and delivery systems.Source: §5.3 · pp. 57–58
  • The UAV suppressant-delivery problem is a constrained optimal control problem balancing altitude, heat exposure, speed, payload, and aircraft dynamics.The formulation incorporates wind, terrain gradients, and payload-release mechanisms, and is verified through simulations and real-world experiments with evaluated drop-off locations.Source: §5.3 · p. 58

5.4. Challenge, Discussion, and Future Directions

AI-enabled UAV systems support active-fire wildfire detection, classification, segmentation, monitoring, and control, but practical deployment remains limited by real-time processing constraints and reliability challenges across diverse environments. Future progress depends on optimized algorithms, hardware improvements, cross-disciplinary collaboration, and rigorous validation.

Active-fire management

  • AI-enabled UAVs use ML, DL, and RL computer-vision algorithms for wildfire detection, classification, segmentation, monitoring, and control.These technologies are positioned as central to effective active-fire management.Source: §5.4 · p. 58

Challenges

  • Limited real-time data-processing capabilities and computational constraints hinder rapid decision-making in dynamic wildfire scenarios.Efficient and timely analysis is particularly important during active-fire operations.Source: §5.4 · p. 58
  • Algorithm reliability and accuracy must be maintained across varying weather, terrain, and other diverse environmental conditions.Robustness under changing scenarios is necessary for practical implementation.Source: §5.4 · p. 58

Future directions

  • Future research should combine optimized algorithms, hardware enhancements, interdisciplinary collaboration, and rigorous testing to develop robust active-fire management systems.Collaboration among computer scientists, wildfire experts, and UAV engineers can address wildfire-behavior complexities and support more resilient, adaptive, and efficient systems.Source: §5.4 · p. 58

6. Post-Fire Management

UAV-assisted post-fire management supports forest recovery monitoring, damage assessment, evacuation planning, and workforce training through high-resolution imagery, optimized operations, and immersive AR/VR systems. The surveyed literature identifies computer vision, machine learning, safe autonomy, and AR-based training as priorities for future development.

6. Post-Fire Management

  • UAVs extend post-fire management from damage assessment to forest recovery, evacuation planning, rehabilitation, and workforce training using aerial imagery and immersive technologies.The section surveys UAV-assisted approaches across forest recovery monitoring, damage assessment, evacuation planning, and AR/VR for safe operations.Source: §6 · p. 60

6.1. Forest Recovery Monitoring

  • High-resolution RGB, multispectral, and low-cost UAV imagery supports monitoring vegetation regrowth and ecosystem dynamics, including settings where field campaigns and satellite data are constrained.UAV imagery is also presented as a viable alternative to satellite remote sensing for evaluating post-fire forest vegetation.Source: §6.1 · p. 60

6.2. Damage Assessment

  • UAV-SfM, LiDAR-derived variables, multispectral imagery, satellite comparisons, and deep learning segmentation enable mapping fire severity, land cover, and burnt areas.These methods provide rapid, high-resolution information for evaluating fire-affected areas and supporting recovery planning.Source: §6.2 · p. 60

6.3. Evacuation Planning

  • UAV imagery supplies road, traffic, hazard, access-point, and burnt-area information for evacuation planning, while swarm coverage and path optimization improve disaster-response operations.In, particle swarm optimization surpassed the tested greedy, intra-route, inter-route, and tabu methods in runtime and cost.Source: §6.3 · p. 61

6.4. AR/VR for Workforce Training

  • AR overlays and VR simulations support wildfire training and safe operations, with display format affecting situational awareness, immersion, portability, and workload.HMDs produced greater situational awareness and presence and lower workload than SimPit in one study, while SimPit provided the best environment for the actual role; experts also managed stress better under communication disruptions.Source: §6.4 · pp. 61–62

6.5. Challenge, Discussion, and Future Directions

  • Future work should automate recovery monitoring and damage assessment, develop safe-autonomy evacuation guidance, and expand AR-based wildfire-fighting training.The section’s taxonomy is summarized in Table 9, while MAR-CPS and AR/VR interfaces demonstrate controlled wildfire simulation, perception, and drone-swarm coordination.Source: §6.5 · p. 62, §6.4 · p. 62, Table 9 · p. 63

7. Wildfire Modeling

Wildfire modeling supports understanding observed mechanisms, exploring hypothetical scenarios, inferring unobserved processes, and predicting future conditions. The section examines how UAV observations and AI contribute across statistical, physically based, dynamic, and machine-learning fire models.

Modeling purposes and approaches

  • Wildfire models support mechanism analysis, what-if exploration, inference about unobserved processes, and prediction of future fire behavior and effects.Approaches range from statistical correlations and semi-empirical formulas to computational fluid dynamics, weather and climate models, and machine learning.Source: §7 · p. 64

Modeling paradigms

  • Historically, statistical and AI approaches modeled fire behavior and fire effects separately, while physically based systems represent time-dependent fire processes and interactions with the surrounding fluid medium.These dynamic simulations capture impacts on the atmosphere, vegetation, and soil.Source: §7 · p. 64

UAV and AI integration

  • UAVs contribute to wildfire modeling through observations and other system components, while AI methods are examined across modeling levels using UAVs or UAV imagery.The section also identifies unresolved areas where combining UAV and AI technologies could advance wildfire modeling.Source: §7 · p. 64

7.1. Physics-aware Approaches to Fire Behavior and Effects

Physics-aware wildfire modeling has progressed from separate kinematic calculations toward computational systems that couple fluid dynamics, fire behavior, observations, and fire effects. These approaches improve prediction of complex fire evolution while remaining costly and requiring stronger reconciliation between modeled effects and coarse-scale observations.

Physics-aware fire behavior modeling

  • Coupled weather-fire models integrate fluid dynamics with fire behavior to forecast growth and anticipate bifurcation, merging, direction changes, and large fire whirls.The CAWFE model simulates three-dimensional atmospheric interactions with fire behavior and can be combined with VIIRS active-fire detections [296].Source: §7.1 · p. 64

Modeling evolution

  • Modeling has evolved from separate FARSITE, FOFEM, and CONSUME calculations toward computational fluid dynamics, physics-based, data-driven, and ML approaches.Physical models commonly predict heat release rate, fuel mass loss, and smoke concentration, including PM2.5.Source: §7.1 · p. 64

Physics-aware fire behavior modeling

  • 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 [296].The visualization shows modeled conditions during the fire alongside contemporaneous active-fire detections and post-fire imagery.Source: §7.1 · p. 64, Figure 19 · p. 65

Fire effects and limitations

  • Fire modeling increasingly addresses effects including vegetation and soil burn severity, mortality, fuel consumption, and emissions, but modeled variables remain only loosely compared with analogous observations.The field calls for joint fire-behavior and fire-effects calculations and reconciliation with observations typically available at much coarser landscape scales.Source: §7.1 · p. 65

7.2. Data Driven Fire modeling

Data-driven wildfire modeling infers spread from physical and environmental features, framing prediction as classification, segmentation, or regression across multiple future time steps. The survey covers cellular automata, deep learning, Bayesian methods, and particle-filter approaches, including UAV, satellite, and probabilistic modeling strategies.

7.2. Data Driven Fire modeling

  • Data-driven spread models depend critically on rich, diverse data and meaningful features, using static factors such as vegetation and slope alongside dynamic fuel, humidity, wind, and temperature inputs.Targets include fire intensity, rate of spread, fire presence, remaining fuel, or spatially segmented ignition probabilities, with predictions generated for one or multiple future time steps.Source: §7.2 · pp. 65–66

7.2.1. Cellular Automata for Spread Modeling

  • Cellular automata simulate wildfire propagation through recursive spatial rules, but accurate modeling requires addressing simultaneous updates, combustion properties, ignition locations, spread directions, and heterogeneous landscapes.GIS integration provides a quick, efficient, and versatile way to represent fire-break evolution and support environmental safeguarding and fire management.Source: §7.2.1 · p. 66

7.2.2. Deep-Learning based Spread Modeling

  • Deep-learning models extract complex spatio-temporal features from satellite and UAS data, while hybrid physics-aware approaches combine observations with physical-model biases for wildfire prediction.Satellite models address large-area forecasting but face low spatial resolution and smoke occlusion; UAS methods combine thermal perimeter detection with Rothermel spread estimation and elliptical expansion.Source: §7.2.2 · pp. 67–69

7.2.3. Bayesian Spread Modeling

  • Bayesian wildfire models incorporate uncertainty and environmental predictors to produce probabilistic spread, rate-of-spread, and extreme-fire-size predictions for decision-making under varied conditions.Models use observations or historical records together with variables such as wind, humidity, soil moisture, topography, meteorology, and housing data.Source: §7.2.3 · pp. 69–70

7.2.4. Other Data-driven Spread Modeling Techniques

  • Particle filters assimilate evolving fire-front observations into nonlinear spread simulators, improving posterior estimates and supporting more accurate regional-scale forecasting and prevention.Adaptive and spatially partitioned variants adjust particle counts or localize filtering to improve efficiency in uncertain, high-dimensional state spaces.Source: §7.2.4 · pp. 70–71

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

UAVs enable fine-scale investigation of wildfire behavior and burn severity, while multi-instrument learning and cycling simulations improve burned-area evaluation and landscape-scale fire progression modeling.

Fine-scale fire impacts

  • UAVs capture fine-scale burn mosaics and fire impacts by observing combustion modes, fuel consumption, and post-fire fuel status.Their high-resolution observations are well suited to resolving variability in burn severity that remains poorly explained.Source: §7.3 · p. 71

Multi-instrument modeling

  • Multi-scale instrument transfer can refine burned-area deep-learning models and support UAV-based multispectral datasets for fire-severity evaluation.Landsat-8 reference data were refined using PlanetScope CubeSat samples, whose 3–4 m resolution and 30 h median revisit provide finer-grained burned-area information.Source: §7.3 · p. 71

Fire progression modeling

  • Landscape-scale fire progression modeling adapts weather-forecast cycling by repeatedly initializing and validating simulations with the latest weather and fire-mapping data.Fine-scale atmospheric simulations have limited predictability of roughly 1–2 days, motivating sequential updates with VIIRS, airborne, incident, and polar-orbiting satellite data.Source: §7.3 · pp. 71–72

7.4. Challenges, gaps, and future directions

Future wildfire modeling requires better-matched, high-resolution observations of ignition, fire environment, fuels, terrain, and validation extent, alongside complementary models and deployable monitoring systems. Key directions include coordinated multi-UAV perimeter mapping, targeted near-fire sensing, non-satellite spread modeling, and efficient interactive systems for real-time aerial deployment.

Data requirements and limitations

  • Fire behavior models require ignition timing and location, weather, fuel moisture, terrain, and later fire extent for validation, but existing datasets can be incomplete or mismatched to newer models’ spatial and temporal resolutions.The mismatch is especially problematic for higher-dimensional models and model types requiring different spatial or temporal scales.Source: §7.4 · p. 72

UAV perimeter mapping

  • Coordinated multi-UAV systems could address the difficulty of collecting full fire perimeters with small drones and support validation of fire extent, spread mechanisms, hotspots, and fire effects.Mosaicing observations from several drones, potentially acquired at staggered times, remains at the conceptual research stage.Source: §7.4 · p. 72

Model complementarity and sensing

  • Coupled weather-fire models require spin-up and lose skill over time, whereas data-driven models may suit short-term prediction; targeted UAV observations could supply key near-fire environmental conditions.Weather-station observations alone are too sparse to represent all fire-driving conditions, including conditions generated by the fire itself.Source: §7.4 · p. 72

Integrated real-time systems

  • Future systems need non-satellite high-resolution spread modeling, efficient integration with monitoring and tracking, and low-complexity onboard deployment for real-time aerial vehicles.Interactive modeling and monitoring could improve adaptation to dynamic and varied environments.Source: §7.4 · p. 72

8. Conclusion

This survey reviews UAV- and AI-enabled wildfire management across pre-fire, active-fire, and post-fire phases, covering prevention, detection, monitoring, recovery, and damage assessment.

Conclusion

  • The survey spans pre-fire prevention and early warning, active-fire UAV computer vision for detection, classification, and segmentation, and post-fire recovery planning and damage assessment.It also evaluates deep-learning algorithms, explores reinforcement learning for wildfire monitoring, and discusses open problems and future directions.Source: §8 · pp. 72–73

Related papers

Loading 2401.02456v1…