Firefighting UAVs for Wildland Fire

1Scope and framing

What this review covers, and how the literature actually clusters.

Research on unmanned aircraft for wildland fire is large but lopsided. Roughly two-thirds of the review literature is about seeing the fire — detection, segmentation, mapping, deep learning on aerial imagery. A much thinner body treats flying near the fire as a dynamics and control problem, and thinner still is work on acting on the fire from the air. This review is written against that imbalance: it summarizes the platform and mission landscape compactly, then spends most of its length on the modeling layer — the equations you would actually implement to simulate, plan for, or control a firefighting UAV.

The anchor surveys are Yuan, Zhang and Liu^[1], which founded the field and is still the most-cited entry point; Akhloufi, Couturier and Castro^[2], the standard modern reference, organizing the field into sensing, perception, cooperation and assistance; Bailon-Ruiz and Lacroix^[3], which is the one survey that frames the problem through autonomy and decision-making rather than sensing; Alsammak et al.^[4], a PRISMA systematic review of 51 included papers on swarm-based extinguishing specifically; Lattimer et al.^[5] for the operational and regulatory view; and Boroujeni et al.^[6], which analyzed over seven hundred articles and organizes them by fire lifecycle phase (pre-fire, active-fire, post-fire). Danish et al. (2025)^[7] is the most recent broad survey, covering 2018–2025 with 200+ references.

2Mission taxonomy and operational reality

What agencies actually fly, versus what the literature proposes.

It is worth separating the two. The US Forest Service's own program documentation names aerial ignition as "the largest applicable mission set to reduce exposure," and states the agency has made large strides toward replacing helicopters for that task^[8]. Detection and suppression get the research attention; ignition is what is operationally mature.

Adoption scale (verified agency figures)

USFS FY2022: approximately 120 UAS platforms, 202 trained pilots, more than 3,800 flights, roughly 1,300 flight hours (45% of them training), ~140 hours of aerial ignition (up 58% year-over-year) and ~360 hours of fire imagery acquisition (up 22%)^[10]. The national UAS program manager has stated that roughly 80% of UAS-conducted burns would otherwise have required a helicopter^[11]. CAL FIRE reported 19 aircraft and 180 statewide pilots as of late 2025^[12].

Constraints that shape the engineering

• One pilot, one aircraft. 14 CFR 107.35 forbids a remote PIC from operating more than one aircraft simultaneously. NASA's prescribed-fire CONOPS study identifies this explicitly as the blocker for multi-drone burn operations, requiring a 90-day waiver process^[13]. Every multi-UAV coordination paper in Section 8 assumes this constraint away.
• BVLOS is still unsettled. The FAA's Part 108 NPRM published 7 August 2025 and the comment period was reopened 28 January 2026^[14]. Claims circulating that the rule was finalized are not supported by the Federal Register record; wildfire BVLOS currently runs on Part 107 waivers or public-aircraft COAs.
• Airspace deconfliction. Wildfire TFRs under 14 CFR 91.137 and the USFS "If You Fly, We Can't" campaign address intruder drones grounding crewed aircraft — the operational reason agency UAS integration is conservative.

3Platform taxonomy and specifications

Configurations, verified specifications, and what each class can physically do.

Five configuration classes appear in wildland fire work, separated less by aerodynamics than by what the mission demands of endurance, payload mass, and hover capability:

• Small/medium multirotor — hover-capable, 30–60 min, sub-kilogram payload. Dominates ISR.
• Heavy-lift multirotor — 10–35 kg MTOW, tens of kilograms of payload, hybrid gas-electric propulsion when endurance matters. Ignition and cargo.
• Fixed-wing and VTOL hybrid — 1.5–3 h, wide-area coverage, no hover. Mapping and persistent watch.
• Unmanned / optionally-piloted helicopter — turbine-powered, thousands of kilograms of external load. The only class with airtanker-adjacent suppression capability.
• Tethered and lighter-than-air — endurance decoupled from onboard energy. Overwatch and comms.

3.1 ISR multirotors and thermal payloads

† 7,000 m with high-altitude propellers. ‡ With H30T, no wind. All figures from manufacturer specification pages.

The jump from 640×512 to 1280×1024 on the H30T, with 32× thermal digital zoom and a 3,000 m laser rangefinder, is the single most consequential recent change for perimeter mapping: it roughly doubles the standoff at which a given ground sample distance is achievable, which matters directly for plume avoidance (Section 5.5).

3.2 Aerial ignition systems

Aerial ignition is the mature suppression-adjacent mission, and the hardware is effectively single-vendor. Drone Amplified's IGNIS line dispenses potassium-permanganate spheres injected with antifreeze; ignition follows the injection by a controlled delay after the sphere lands.

Note the modeling relevance: at 120 spheres/min the IGNIS II sheds 450 × ~5 g in under four minutes — a slow, roughly linear mass ramp that is benign for control, in sharp contrast to a liquid dump (Section 5.1). The mass-flow rate, not the total mass, is what determines whether the variable-mass terms matter.

3.3 Fixed-wing and VTOL hybrids

The endurance spread here — 59 min to 18 h — is essentially a propulsion-chemistry spread. Every sub-2-hour platform in the table is battery-electric; ScanEagle's 18 hours comes from a heavy-fuel piston engine. This is the practical reason persistent fire watch is not a solved problem with COTS electric aircraft.

3.4 Heavy-lift and direct-suppression platforms

3.5 Tethered and stratospheric platforms

Tethering decouples endurance from onboard energy entirely. The Elistair Orion 2 sustains 24 hours at 100 m on a micro-tether carrying 2 kg of georeferenced EO+IR; Hoverfly's Spectre runs effectively continuously at up to 60 m on <2 kW, passing all video and data down the tether so there is no RF signature to jam or deconflict. Both are overwatch-and-relay assets for the incident base rather than fire-front sensors. At the other extreme, Sceye's stratospheric airship targets ignition detection "within minutes" from 55,000–65,000 ft with pre-commercial flights planned for 2026 — no operational fire-agency deployment is verified.

4Flight dynamics models

The standard rigid-body layer, stated in the form used in the UAV literature, before the firefighting-specific terms are added in Section 5.

4.1 Six-degree-of-freedom rigid body

Notation: inertial NED frame, body frame, mapping body to inertial, the skew map. The 12-state vector standard to Stevens & Lewis^[15] and Beard & McLain^[16] is

with kinematics and the Euler-rate relation

Newton–Euler, constant mass (the variable-mass case is Section 5.1):

4.2 Multirotor specifics

Per-rotor thrust and drag torque, following Mahony, Kumar and Corke^[17]:

Control allocation for an X-configuration quadrotor with arm length inverts the mixer

Secondary aerodynamics that matter near a fire

• Blade flapping. Advance ratio tilts the tip-path plane by , producing a force opposing motion and a hub moment. This is the dominant pitch-up/speed-stability coupling.
• Rotor drag, lumped as — the term that makes multirotor velocity observable from accelerometers alone.
• Ground effect, Cheeseman–Bennett: , exceeding 10% thrust augmentation below . Directly relevant to low hovers over a fire front and to bucket dipping over water.

Control structure and controller families

The standard cascade exploits the ~10× timescale separation between attitude (200–1000 Hz) and position (50–100 Hz) loops. For firefighting the choice among controller families is not arbitrary:

Trajectory generation rests on differential flatness with flat outputs ^[20], reducing planning to a QP in piecewise-polynomial coefficients minimizing snap — the basis for fast drop-run and perimeter-following trajectories.

4.3 Fixed-wing and VTOL hybrid

Fixed-wing fire-monitoring work almost universally abstracts the vehicle to a Dubins car at constant altitude and airspeed with bounded turn rate, , giving minimum turn radius . This is what makes perimeter-tracking and coverage planning tractable, and it is the abstraction underlying most of Section 8. The full model retains the coefficient build-up and its drag and moment counterparts^[16].

VTOL hybrids (quad-planes, tail-sitters) add a transition phase in which the aircraft passes through deep stall at from roughly 15° to 90°, where quasi-steady coefficient models fail and a blended or flat-plate post-stall model is used. Transition is the highest-risk flight phase for these aircraft and is where the formulation stops being optional.

5Firefighting-specific dynamics

The terms that distinguish a firefighting UAV from generic aerial robotics. This is the least-surveyed part of the literature and the most consequential for anyone building a simulator or controller.

5.1 Variable-mass dynamics during agent release

For a vehicle ejecting mass at rate with exhaust velocity relative to the airframe, the body-frame translational equation gains two families of terms the constant-mass model omits:

and the rotational equation must retain :

With airframe and liquid , the composite CG and inertia are

Three consequences dominate practice:

• Thrust-to-weight doubles. A 10 kg multirotor carrying 5 L changes by a factor of two through a full dump. A PID integrator wound up for the loaded condition produces a balloon-up transient at release — the most commonly reported failure mode.
• Loop gain doubles. Effective plant gain doubles, so nominal gains become 2× hotter and can limit-cycle. Mass-scheduled gains, adaptive control, or INDI (inherently gain-normalizing) are the fixes.
• Asymmetric release produces a constant roll torque for a lateral CG offset — a bias a pure P/D attitude loop cannot null.

Sloshing

The standard reduction replaces the liquid with a rigid mass plus equivalent pendulums. For a rectangular tank of width and fill depth , the first antisymmetric mode is

5.2 Water-jet reaction and hose-tethered architectures

Nozzle recoil is the defining disturbance of active-suppression UAVs. From Bernoulli with discharge coefficient and momentum flux:

The magnitude is the point. A 6 mm nozzle at 0.8 MPa gives and N — comparable to the entire weight of a 4.5 kg aircraft. Recoil is not a perturbation. Two further non-ideal effects are documented: water-hammer transients on valve actuation reaching 2–3× steady state (Joukowsky, ), and persistent broadband two-phase flow fluctuation^[21].

Four architectural responses appear in the literature:

1. Treat recoil as an unbounded disturbance and estimate it online. Ni et al.^[21] model it as in with unknown a-priori bounds, using adaptive robust constraint-following (Udwadia–Kalaba style) on a vector-rotor UAV.
2. Full actuation. Tilting rotors generate direct lateral force to cancel recoil without an attitude excursion^[22].
3. Turn recoil into lift. Viegas et al.^[23] combine multi-rotor and water-jet propulsion on a tethered forest-firefighting UAV so the jet contributes to lift, extending endurance rather than fighting it.
4. Levitate the hose itself. The Tohoku "Dragon Firefighter" line^[24],[25] dispenses with the airframe: a hose is levitated and steered by its own jets. Yamauchi et al.^[25] give the complete hydraulic-network model — six coupled Bernoulli equations for head and middle unit pressures driven by a root pressure ≤0.9 MPa through resistance elements, with 6 mm nozzles at 0.8 MPa reaching an 80 N total reaction target, and a 100 Hz quadratic program distributing the desired wrench across four steerable nozzles under angle and rate constraints. That QP is structurally identical to rotor control allocation (Section 4.2), with jets substituted for rotors.

For a hose-fed aircraft, the hose is not a boundary condition but a distributed dynamic element. Hament and Oh^[26] import the pipes-conveying-fluid formulation,

5.3 Slung load and suspended bucket

The helicopter bucket, scaled down. For quadrotor mass , load , cable length and unit vector , the taut-phase dynamics are

When the cable goes slack, the bodies decouple, and the load free-falls: the system is hybrid, with impulsive taut↔slack transitions. Sreenath, Michael and Kumar^[27] showed the load position and yaw form a flat output set for this hybrid system, enabling trajectory generation that plans through mode switches.

The design-level result is the small-angle reduction , giving . A 2 m sling yields 2.2 rad/s ≈ 0.35 Hz — inside the position-loop bandwidth of a typical multirotor, so the outer loop excites the pendulum unless it is notched, input-shaped, or explicitly damped. A full water bucket gives mass ratio , at which the load appears in the attitude loop as well. And a water bucket nests slosh inside the swing, producing a double-pendulum cascade with , terminated by a variable-mass event at dump.

5.4 Drop ballistics and ground pattern

For a compact payload — fire-extinguishing ball, gel capsule — released with vehicle velocity :

with vacuum baseline , horizontal miss , and crosswind miss where is the drag relaxation time. The error budget — release-timing jitter , wind-estimate error , position error , attitude error — scales with or in every term, so low release altitude dominates accuracy. That is in direct tension with plume-updraft and thermal safety (Section 5.5), and this tension is the central design trade for autonomous suppression.

A bulk liquid drop is not a ballistic particle. The sequence is column formation → aerodynamic breakup → droplet dispersion → deposition, governed by

The operational output is the ground pattern: coverage level in gal/100 ft² or L/m², and the length of contiguous line built at or above a threshold (typically 1–3 gal/100 ft² by fuel model). Amorim's two-part model^[28],[29] is the standard numerical treatment.

5.5 Wind, plume updraft, and hot-air thrust loss

Decompose the wind field as : steady, turbulent, and fire-induced. Only the first two appear in the classical UAV literature. Background turbulence uses Dryden or von Kármán shaping filters (MIL-HDBK-1797), e.g.

For the plume term, Byram fireline intensity sets a buoyancy flux, and the Morton–Taylor–Turner similarity solution gives centerline velocity and radius

The empirical anchor: Shawon et al.^[30] flew a fixed-wing through prescribed-grass-fire plumes and estimated vertical wind of 6–10 m/s at ~115 m AGL using 2-state and 9-state EKFs on inertial angle-of-attack and sideslip — comparable to or exceeding the climb-rate authority of many small multirotors and enough to saturate a fixed-wing altitude loop. Brewer and Clements^[31] profiled the fire environment with a DJI M200 carrying a TriSonica Mini at 5 Hz, documenting super-adiabatic layers and surface inversions with ~±1 m/s wind accuracy.

A tractable disturbance model for control design is a Gaussian-cored, position-dependent mean field plus scaled turbulence:

Hot air, thrust, and battery

Since and , at constant RPM . Going from ISA 288 K to 333 K (60 °C, plausible near a front) costs 13.5% of density; 288 K to 400 K (inside a plume) costs 28%. Holding thrust requires , and shaft power rises as — required power increases precisely when battery capability is degrading. Compounding it: copper winding resistance rises as , roughly +30% at 100 °C, NdFeB remanence falls ~0.12%/K reducing , and convective cooling is worse in hot air. Thrust margin should be evaluated at fire-environment density altitude, not sea-level ISA.

6Energy and endurance models

Endurance is the binding constraint on almost every wildfire UAV mission, and the energy model is what couples the dynamics layer to the planning layer.

6.1 Momentum-theory formulation

Induced power in hover, with induced-power factor :

and in forward flight, with Glauert induced velocity ,

The equivalent form used throughout the communications and path-planning literature, from Zeng, Xu and Zhang, collapses the same physics into four fitted constants:

6.2 Payload penalty — why multirotor suppression is hard

In hover and , so endurance scales as . Adding 5 kg of water to a 10 kg airframe gives — a 46% endurance loss on the induced-power term alone, before accounting for the battery mass that must also be lifted.

6.3 Empirical and learned energy models supplied papers

Analytic momentum-theory models are structurally correct but poorly calibrated for sub-2 kg aircraft. Cabuk, Tosun, Dagdeviren and Ozturk^[33] address this directly, building four energy models for drones under 2 kg and 1 m diameter from an extensive flight-test campaign and comparing them:

The physically interesting result is an energy-efficient "valley" in airspeed: power is minimized at an intermediate speed and rises on both sides. The valley's location moves with mass — measured minima at roughly 6 m/s for 1102 g, 10 m/s for 1418 g, and 12.5 m/s for 1734 g, at approximately 130 W, 170 W and 200 W respectively, against ~250–330 W at a constant 20 m/s. The paper also derives a swarm topology-control cost function for evaluating connectivity restoration and formation change in energy terms.

Two implications for fire work. First, the optimal cruise speed is payload-dependent — a UAV that has just dropped its water should fly home slower than it flew out, and a fixed cruise setpoint is leaving endurance on the table. Second, this is the empirical validation of the U-curve that path planners assume; for the polynomial form means the analytic structure is right and only the coefficients need fitting per airframe.

6.4 Energy-aware planning with adaptive speed supplied paper

Diller and Han^[32] close the loop from the energy model to the route. Rather than treating speed as fixed, they exploit the fact that total distance achievable is

They formalize the resulting problem — Minimum-Time while On-The-Move — as a multiple-depot, multiple-terminal Hamiltonian path problem with a non-stopping ground vehicle, and solve it two ways: a MINLP tailored to optimize UAV speed, and a tractable k-means-plus-integer-program heuristic. Reported completion-time reductions against baseline are 23.8% (MINLP) and 14.5% (k-IP), validated on a physical testbed.

The transfer to wildfire is direct and, as far as this review found, unexploited: a fire-monitoring UAV rendezvousing with a moving engine or crew buggy for battery swap is exactly the MT-OTM structure, and the observation that speed is a planning decision variable, not a constant is the piece most wildfire coverage planners omit.

7Fire propagation models and their coupling to UAV planning

The fire is a moving, uncertain, partially-observed boundary. How it is modeled determines what the planner and the estimator can do.

7.1 The Rothermel kernel and what is built on it

Rothermel's 1972 surface-spread model^[34] is the kernel inside BEHAVE, FARSITE, FlamMap, FSim, ELMFIRE, Cell2Fire, and most cellular-automaton simulators used in UAV studies. It computes a quasi-steady one-dimensional rate of spread of the flaming front of a surface fire through a homogeneous, dead-fuel-dominated fuel bed, from an energy balance of heat source over heat sink:

It does not model crown fire, spotting, transient acceleration, or fire–atmosphere feedback. Everything a UAV planner might want from it — where the front will be in ten minutes — is downstream of assumptions the model does not itself check.

7.2 Level-set form and fire-front estimation

For estimation the level-set formulation is the right one, because it is a PDE on a scalar field rather than a list of markers, and it handles merging and topological change automatically:

Three filtering formulations appear in the literature:

• Parameterized-perimeter EKF/KF. Discretize the perimeter into nodes (or a Fourier parameterization ), state , measurement from UAV IR/visual front-crossing detections. Lin and Liu^[42] run this with a fleet of UAVs and use it to schedule coverage. Observability is the crux: a single UAV sees only a local arc, so covariance grows on unvisited segments — and that growth is exactly what drives information-driven path planning.
• Particle filter / SMC. Preferred when the posterior is multimodal (spotting, barrier breaching) and latent fuel-moisture and wind-aloft parameters must be estimated jointly. The natural home for data assimilation of UAV imagery into a coupled fire–atmosphere model.
• Bayesian occupancy grid over burning/burned/unburnt cells with per-cell IR detection likelihood — a hidden Markov field, well matched to camera footprints.

The perception front-end geo-registers fire/smoke segmentations by projecting detections through the camera model onto a DEM using UAV pose, so attitude-solution error propagates directly into perimeter measurement noise. The estimation quality of Section 4.1 is upstream of the fire estimate.

7.3 Closing the loop

7.4 Simulators and RL environments

MITRE's SimFire implements Rothermel spread with a PyGame front-end and a historical data layer; SimHarness^[43] is the RL harness on top of it, training agents to place mitigations. PyroRL^[44] is a Gymnasium environment but targets evacuation routing rather than UAV sensing. On the planning side, Julian and Kochenderfer^[45] applied deep RL to multi-aircraft wildfire surveillance over a stochastic gridded fire model, and Seraj, Silva and Gombolay^[37] give the cleanest example of a genuine simulator-in-the-loop planner: an adaptive extended Kalman filter infers latent propagation dynamics, FARSITE supplies the spread, and a TSP formulation produces the multi-UAV paths, with analytical temporal and tracking-error bounds and reported tracking error 7.5–9.0× smaller than baselines.

8Multi-UAV coordination, coverage, and search

Where the fire-specific literature meets the general autonomy literature — and where transferable results sit.

8.1 The fire-specific lineage

Three groups define the canon. Casbeer, Kingston, Beard and McLain^[46] established cooperative perimeter tracking with a team of small fixed-wing UAVs — still the most-cited multi-UAV wildfire paper and the origin of the "regulate the fire/no-fire boundary in the image" control law. The Seville group's COMETS project^[47] built the heterogeneous multi-UAV system (helicopters plus airships) with cooperative perception, culminating in Merino et al.'s automatic fire monitoring and measurement system^[48] validated in real field fires, measuring front position, height and width. LAAS-CNRS (Bailon-Ruiz, Bit-Monnot, Lacroix) contributed the planning-centric line, from the autonomy-lens survey^[3] to real-time fleet monitoring^[49]. Pham, La, Feil-Seifer and Deans^[50] added distributed formation control for dynamic fire-front tracking, and Seraj and Gombolay^[37] the QoS-guaranteed coverage-and-tracking formulation.

8.2 Transferable results from the general autonomy literature supplied papers

Two of the supplied papers are not wildfire papers, and that is precisely their value — each supplies a mechanism the wildfire literature is missing.

Coverage under a connectivity constraint

Devaraju, Ihler and Kumar^[51] address the fundamental tension between area coverage and network connectivity in decentralized fixed-wing UAV networks. Pure repulsion-pheromone (stigmergic) models achieve fast coverage by pushing UAVs apart — and thereby break the network. Their connectivity-aware pheromone (CAP) model weights each candidate next-waypoint cell by both pheromone value and estimated local connectivity. The pheromone map evolves with evaporation and repulsion-deposit ,

with a distance-weighted pairwise connectivity that saturates inside 60% of transmission range and falls linearly beyond it:

Evaluated over a 6 km × 6 km area with 20–40 UAVs at 1 km transmission range and speeds of 20–40 m/s, averaged over 30 runs, CAP improves coverage time and fairness over both the repel-pheromone and CACOC2 baselines while maintaining higher average neighbour count. The wildfire relevance is direct: a fire-monitoring fleet without a connectivity term will spread out to cover the perimeter and lose the mesh exactly when relaying observations matters most — and no fire-specific paper in this review treats coverage and connectivity jointly.

Complete search in cluttered unknown space

Luo et al.'s Star-Searcher^[52] targets autonomous target search in complex unknown environments — structurally the mop-up and search-and-rescue problem. Its two mechanisms are worth importing. Visibility-based viewpoint clustering groups viewpoints with mutual collision-free rays into clusters, decomposing planning into a global path over clusters and a local path over individual viewpoints, so real-time replanning stays tractable as the viewpoint count grows. A history-aware global planner penalizes motion inconsistency from frequent map changes, suppressing the indecisive back-and-forth that frontier-based planners exhibit when new area keeps appearing. Against FUEL baselines across four environments, Star-Searcher shortened path length and flight time while achieving 100% search completeness in every trial, where FUEL-4m dropped to 68.8–95.0%.

Viewpoint scoring uses a weighted information gain , separating "unknown space to explore" from "known surface to inspect." That separation maps cleanly onto wildfire mop-up, where the interior is mapped but not yet inspected for residual heat — a distinction current wildfire coverage planners, which treat the problem as pure area coverage, do not make.

8.3 Coverage and informative path planning

Beyond these, the relevant families are coverage path planning with energy-efficient and cooperative strategies, informative path planning for terrain monitoring, and decentralized importance-based multi-UAS planning for wildfire monitoring, where cells are weighted by expected information rather than treated uniformly. The common thread with Section 7.2 is that the planner should be driven by the estimator's covariance, not by a geometric coverage criterion — the fire moves, so uniform coverage is provably wasteful.

9Datasets and benchmarks

What exists, what it supports, and the one thing none of it supports.

FLAME 3's radiometric TIFFs are the qualitative advance: preserving temperature rather than a rendered palette makes threshold-based auto-labeling possible and, more importantly, makes it possible to measure spread rate and energy release from UAV imagery rather than merely detect fire.

10Gaps and research directions

Where the modeling literature is thin, ordered by how load-bearing the gap is.

1. No validated UAV-scale drop model. Airtanker ground-pattern work^[28],[29] operates 2–4 orders of magnitude away in volume and in Weber number. There is no published coverage-level model for a 1–20 L release, so no principled way to answer whether a given small-UAV drop does anything useful to a fire.
2. Plume dynamics are not in anyone's controller. The plume field is a state-dependent, feedforward-compensable disturbance with measured magnitudes of 6–10 m/s vertical at 115 m AGL^[30]. Yet the fire estimate and the flight controller remain separate subsystems in essentially all reviewed work. Coupling them — MPC preview from , or a plume term in an INDI disturbance observer — is the most concrete unexploited opportunity in this domain.
3. Variable-mass and slosh dynamics are borrowed, not derived. The best available treatments come from agricultural spraying and spacecraft propellant slosh. The firefighting case is worse than both: faster discharge, higher mass ratio, higher control bandwidth. The guaranteed downward sweep of slosh frequency through attitude-loop crossover during partial discharge (Section 5.1) has, as far as this review found, never been analyzed for a firefighting UAV.
4. No standard multi-UAV wildfire planning benchmark. Bailon-Ruiz, Pham/La and Seraj/Gombolay each built a different simulator with a different fire model (custom propagation, custom CA, FARSITE respectively). Cross-paper comparison is effectively impossible. SimFire/SimHarness is the most plausible base for a shared benchmark but is mitigation-oriented rather than sensing-oriented.
5. Coverage and connectivity are optimized separately. The wildfire literature optimizes coverage; the networking literature optimizes connectivity^[51]. A fire fleet needs both simultaneously, and the joint problem is unaddressed in the fire-specific work.
6. Speed is treated as a constant. The energy models^[33] show a mass-dependent optimal-speed valley, and the planning literature^[32] shows that making speed a decision variable buys double-digit percentage improvements. Wildfire coverage planners almost universally assume fixed airspeed, and the payload-dependence of the optimum is exactly the regime an ignition or suppression UAV lives in.
7. Regulatory structure shapes the achievable architecture. One-pilot-one-aircraft^[13] and unsettled BVLOS^[14] mean that every multi-UAV result in Section 8 currently requires a waiver to fly. Research that assumes swarms should say so; research that works within the constraint — one supervised aircraft with high onboard autonomy — is closer to deployable.

References

1. Yuan, C., Zhang, Y., Liu, Z. (2015). A survey on technologies for automatic forest fire monitoring, detection, and fighting using unmanned aerial vehicles and remote sensing techniques. Canadian Journal of Forest Research 45(7):783–792. doi:10.1139/cjfr-2014-0347
2. Akhloufi, M.A., Couturier, A., Castro, N.A. (2021). Unmanned Aerial Vehicles for Wildland Fires: Sensing, Perception, Cooperation and Assistance. Drones 5(1):15. doi:10.3390/drones5010015
3. Bailon-Ruiz, R., Lacroix, S. (2020). Wildfire remote sensing with UAVs: A review from the autonomy point of view. ICUAS 2020, 412–420. hal-02614975
4. Alsammak, I.L.H., Mahmoud, M.A., Aris, H., AlKilabi, M., Mahdi, M.N. (2022). The Use of Swarms of Unmanned Aerial Vehicles in Mitigating Area Coverage Challenges of Forest-Fire-Extinguishing Activities: A Systematic Literature Review. Forests 13(5):811. doi:10.3390/f13050811
5. Lattimer, B.Y., Huang, X., Delichatsios, M.A., et al. (2023). Use of Unmanned Aerial Systems in Outdoor Firefighting. Fire Technology 59(6):2961–2988. doi:10.1007/s10694-023-01437-0
6. Hosseini Boroujeni, S.P., Razi, A., Khoshdel, S., Afghah, F., Coen, J.L., et al. (2024). A comprehensive survey of research towards AI-enabled unmanned aerial systems in pre-, active-, and post-wildfire management. Information Fusion 108:102369. arXiv:2401.02456
7. Danish, S., Piran, M.J., Khan, S.U., et al. (2025). Vision-based fire management system using autonomous unmanned aerial vehicles: a comprehensive survey. Artificial Intelligence Review 59(1):16. doi:10.1007/s10462-025-11415-3
8. USDA Forest Service. How we use unmanned aircraft systems. fs.usda.gov
9. USDA Forest Service. Mechanized birds: Wildland firefighting with drone intelligence. fs.usda.gov
10. USDA Forest Service (2022). Forest Service UAS Report FY22. PDF
11. Commercial UAV News. Success and scale with drone programs: how the U.S. Forest Service saves taxpayer dollars and reduces wildfire response risk. link
12. Firehouse (Dec 2025). CAL FIRE's drones that nearly didn't launch now soaring. link
13. Buchmann, et al. (2024). Autonomous Drone Integration in Prescribed Fire Operations. NASA/TM–20240010585. NTRS
14. FAA. Normalizing Unmanned Aircraft Systems Beyond Visual Line of Sight Operations. NPRM 90 FR (7 Aug 2025); comment period reopened 28 Jan 2026
15. Stevens, B.L., Lewis, F.L., Johnson, E.N. (2015). Aircraft Control and Simulation, 3rd ed. Wiley. doi:10.1002/9781119174882
16. Beard, R.W., McLain, T.W. (2012). Small Unmanned Aircraft: Theory and Practice. Princeton University Press.
17. Mahony, R., Kumar, V., Corke, P. (2012). Multirotor Aerial Vehicles: Modeling, Estimation, and Control of Quadrotor. IEEE Robotics & Automation Magazine 19(3):20–32. doi:10.1109/MRA.2012.2206474
18. Lee, T., Leok, M., McClamroch, N.H. (2010). Geometric tracking control of a quadrotor UAV on SE(3). 49th IEEE CDC, 5420–5425. doi:10.1109/CDC.2010.5717652
19. Smeur, E.J.J., Chu, Q., de Croon, G.C.H.E. (2016). Adaptive Incremental Nonlinear Dynamic Inversion for Attitude Control of Micro Air Vehicles. J. Guidance, Control, and Dynamics 39(3):450–461. doi:10.2514/1.G001490
20. Mellinger, D., Kumar, V. (2011). Minimum snap trajectory generation and control for quadrotors. IEEE ICRA, 2520–2525. doi:10.1109/ICRA.2011.5980409
21. Ni, Z., Zhang, X., Bai, J., et al. (2026). Adaptive robust constraint-following control of vector–rotor UAVs subject to high-intensity time-varying water-jet disturbances. Dynamics 6(2):19. doi:10.3390/dynamics6020019
22. A novel firefighter quadrotor UAV with tilting rotors: modeling and control (2024). Aerospace Science and Technology 151:109248. doi:10.1016/j.ast.2024.109248
23. Viegas, C., Chehreh, B., Andrade, J., Lourenço, J. (2022). Tethered UAV with Combined Multi-rotor and Water Jet Propulsion for Forest Fire Fighting. J. Intelligent & Robotic Systems 104:21. doi:10.1007/s10846-021-01532-w
24. Ando, H., Ambe, Y., Ishii, A., Konyo, M., Tadakuma, K., Maruyama, S., Tadokoro, S. (2018). Aerial Hose Type Robot by Water Jet for Fire Fighting. IEEE Robotics and Automation Letters 3(2):1128–1135. doi:10.1109/LRA.2018.2792701
25. Yamauchi, Y., Maezawa, Y., Ambe, Y., Konyo, M., Tadakuma, K., Tadokoro, S. (2023). Development of a remotely controllable 4 m long aerial-hose-type firefighting robot. Frontiers in Robotics and AI 10:1273676. doi:10.3389/frobt.2023.1273676
26. Hament, B., Oh, P. (2023). High Pressure Hosing-Drone Dynamics and Controls. J. Intelligent & Robotic Systems 109(4):90. doi:10.1007/s10846-023-01954-8
27. Sreenath, K., Michael, N., Kumar, V. (2013). Trajectory generation and control of a quadrotor with a cable-suspended load — a differentially-flat hybrid system. IEEE ICRA. IEEE 6631275
28. Amorim, J.H. (2011). Numerical modelling of the aerial drop of firefighting agents by fixed-wing aircraft. Part I: model development. Int. J. Wildland Fire 20(3):384–393. CSIRO
29. Amorim, J.H. (2011). … Part II: model validation. Int. J. Wildland Fire 20(3):394–406. CSIRO
30. Shawon, M.H., Chao, H., Rhudy, M., Johansen, T.A., Tian, P., Flanagan, H.P., Goyer, J. (2025). Vertical Wind Velocity Estimation during UAS Fire Plume Encounters. AIAA SciTech Forum. doi:10.2514/6.2025-1624
31. Brewer, M.J., Clements, C.B. (2020). Meteorological Profiling in the Fire Environment Using UAS. Fire 3(3):36. doi:10.3390/fire3030036
32. Diller, J., Han, Q. (2023). Energy-aware UAV Path Planning with Adaptive Speed. AAMAS 2023, London, 922–930. supplied
33. Cabuk, U.C., Tosun, M., Dagdeviren, O., Ozturk, Y. (2024). Modeling Energy Consumption of Small Drones for Swarm Missions. IEEE Transactions on Intelligent Transportation Systems 25(8):10176–10189. doi:10.1109/TITS.2024.3350042 supplied
34. Rothermel, R.C. (1972). A mathematical model for predicting fire spread in wildland fuels. USDA Forest Service Research Paper INT-115. Treesearch 32533
35. Andrews, P.L. (2018). The Rothermel surface fire spread model and associated developments: A comprehensive explanation. RMRS-GTR-371. PDF
36. Finney, M.A. (1998). FARSITE: Fire Area Simulator — model development and evaluation. RMRS-RP-4. Treesearch 4617
37. Seraj, E., Silva, A., Gombolay, M. (2022). Multi-UAV planning for cooperative wildfire coverage and tracking with quality-of-service guarantees. Autonomous Agents and Multi-Agent Systems 36:39. doi:10.1007/s10458-022-09566-6
38. Lautenberger, C. (2013). Wildland fire modeling with an Eulerian level set method and automated calibration. Fire Safety Journal 62:289–298. doi:10.1016/j.firesaf.2013.08.014
39. Pais, C., Carrasco, J., Martell, D.L., Weintraub, A., Woodruff, D.L. (2021). Cell2Fire: A Cell-Based Forest Fire Growth Model to Support Strategic Landscape Management Planning. Frontiers in Forests and Global Change 4:692706. doi:10.3389/ffgc.2021.692706
40. Linn, R.R., et al. (2020). QUIC-fire: A fast-running simulation tool for prescribed fire planning. Environmental Modelling & Software 125:104616. doi:10.1016/j.envsoft.2019.104616
41. Alexandridis, A., Vakalis, D., Siettos, C.I., Bafas, G.V. (2008). A cellular automata model for forest fire spread prediction: the case of the wildfire that swept through Spetses Island in 1990. Applied Mathematics and Computation 204(1):191–201. doi:10.1016/j.amc.2008.06.046
42. Lin, Z., Liu, H.H.T. (2019). Kalman Filter-Based Large-Scale Wildfire Monitoring With a System of UAVs. IEEE Trans. Industrial Electronics 66(1):606–615. IEEE 8331879
43. Tapley, A., et al. (2023). Reinforcement Learning for Wildfire Mitigation in Simulated Disaster Environments. NeurIPS 2023 Workshop on Tackling Climate Change with ML. arXiv:2311.15925
44. Guman, J., O'Brien, J.C., Pondoc, C., Kochenderfer, M.J. (2024). PyroRL: A Reinforcement Learning Environment for Wildfire Evacuation. JOSS 9(101):6739. doi:10.21105/joss.06739
45. Julian, K.D., Kochenderfer, M.J. (2019). Distributed Wildfire Surveillance with Autonomous Aircraft Using Deep Reinforcement Learning. J. Guidance, Control, and Dynamics 42(8):1768–1778. doi:10.2514/1.G004106
46. Casbeer, D.W., Kingston, D.B., Beard, R.W., McLain, T.W. (2006). Cooperative forest fire surveillance using a team of small unmanned air vehicles. Int. J. Systems Science 37(6):351–360. doi:10.1080/00207720500438480
47. Ollero, A., Lacroix, S., Merino, L., et al. (2005). Multiple eyes in the skies: architecture and perception issues in the COMETS unmanned air vehicles project. IEEE Robotics & Automation Magazine 12(2):46–57. doi:10.1109/MRA.2005.1458323
48. Merino, L., Caballero, F., Martínez-de-Dios, J.R., Maza, I., Ollero, A. (2012). An Unmanned Aircraft System for Automatic Forest Fire Monitoring and Measurement. J. Intelligent & Robotic Systems 65(1–4):533–548. doi:10.1007/s10846-011-9560-x
49. Bailon-Ruiz, R., Bit-Monnot, A., Lacroix, S. (2022). Real-time wildfire monitoring with a fleet of UAVs. Robotics and Autonomous Systems 152:104071. doi:10.1016/j.robot.2022.104071
50. Pham, H.X., La, H.M., Feil-Seifer, D., Deans, M.C. (2020). A Distributed Control Framework of Multiple Unmanned Aerial Vehicles for Dynamic Wildfire Tracking. IEEE Trans. Systems, Man, and Cybernetics: Systems 50(4):1537–1548. arXiv:1803.07926
51. Devaraju, S., Ihler, A., Kumar, S. (2022). Connectivity-Aware Pheromone Mobility Model for Autonomous UAV Networks. arXiv:2210.06684 supplied
52. Luo, Y., Zhuang, Z., Pan, N., Feng, C., Shen, S., Gao, F., Cheng, H., Zhou, B. (2024). Star-Searcher: A Complete and Efficient Aerial System for Autonomous Target Search in Complex Unknown Environments. arXiv:2402.16348 supplied
53. Shamsoshoara, A., Afghah, F., Razi, A., Zheng, L., Fulé, P.Z., Blasch, E. (2021). Aerial imagery pile burn detection using deep learning: The FLAME dataset. Computer Networks 193:108001. doi:10.1016/j.comnet.2021.108001
54. Hopkins, B., O'Neill, L., Afghah, F., et al. (2022). FLAME 2: Fire Detection and Modeling — Aerial Multi-spectral Image Dataset. IEEE DataPort. doi:10.21227/swyw-6j78
55. Hopkins, B., O'Neill, L., Marinaccio, M., et al. (2024). FLAME 3: Radiometric Thermal UAV Imagery for Wildfire Management. arXiv:2412.02831; doi:10.21227/w0mz-aq48
56. Toulouse, T., Rossi, L., Campana, A., Celik, T., Akhloufi, M.A. (2017). Computer vision for wildfire research: An evolving image dataset for processing and analysis. Fire Safety Journal 92:188–194. doi:10.1016/j.firesaf.2017.06.012
57. Wang, M., Jiang, L., Yue, P., Yu, D., Tuo, T. (2024). An open flame and smoke detection dataset for deep learning in remote sensing based fire detection. Geo-spatial Information Science. doi:10.1080/10095020.2024.2347922
58. Pesonen, J., Raita-Hakola, A.-M., Joutsalainen, J., et al. (2025). Boreal forest fire dataset. Scientific Data. doi:10.1038/s41597-025-05634-0
59. Zhang, Y., Rui, X., Song, W. (2025). A UAV-Based Multi-Scenario RGB-Thermal Dataset and Fusion Model for Enhanced Forest Fire Detection. Remote Sensing 17(15):2593. doi:10.3390/rs17152593

Platform specifications are drawn from manufacturer specification pages and agency documents; figures marked n/p are not published by the manufacturer. Citation details were checked against Crossref where a DOI exists. Items that could not be verified against a primary source are flagged in the text rather than presented as established.