Wildfire-Aware Modeling Framework

• Wildfire-aware modeling frameworks explicitly integrate wildfire-specific drivers such as fuel, wind, and ignition history into simulation and operational decision-making loops.
• They combine techniques like coupled fire–atmosphere simulation, graph-based neural ODEs, and event-centered data harmonization to optimize both prediction accuracy and response strategies.
• These frameworks incorporate uncertainty quantification and infrastructure resilience analysis to support dynamic risk management and guide emergency intervention.

to=arxiv_search.search 天天中彩票中 彩神争霸快ებრივjson {"query":"wildfire-aware modeling framework arXiv wildfire edge UAV graph ODE uncertainty FireDataForge SFIRE", "max_results": 10} to=arxiv_search.search 】【。】【”】【json {"query":"(Huang et al., 23 Feb 2026) OR (Xu et al., 4 Jan 2026) OR (Xia et al., 19 Jun 2026) OR (Mandel et al., 2011) wildfire", "max_results": 10} to=arxiv_search.search 】【。】【”】【json {"query":"wildfire spread uncertainty framework arXiv 2025 2026", "max_results": 10} A wildfire-aware modeling framework is a class of modeling systems in which wildfire-specific processes are represented explicitly rather than treated as generic geospatial prediction problems. In the literature surveyed here, the term spans coupled fire–atmosphere simulators that exchange wind, heat, and moisture fluxes; multi-scale temporal and graph models for wildfire activity and danger; event-centered data harmonization pipelines; boundary-aware uncertainty protocols; and operational decision frameworks for UAV monitoring, suppression, and power-system resilience (Mandel et al., 2011, Wang et al., 2022, Xu et al., 4 Jan 2026, Xia et al., 19 Jun 2026, Funk, 4 May 2026, Campos et al., 5 Mar 2026). Taken together, these works describe wildfire-aware modeling as an integration problem across hazard dynamics, sensing, computation, uncertainty, and action.

1. Conceptual scope and representative forms

The literature does not use “wildfire-aware modeling framework” to denote a single algorithmic family. Instead, it denotes frameworks that encode wildfire-specific drivers such as fuel, terrain, wind, smoke, ignition history, operational deadlines, infrastructure exposure, and community consequences inside the modeling loop. In some works, the emphasis is physical fidelity; in others, it is geospatial ETL, spatiotemporal learning, uncertainty quantification, or decision optimization.

Framework	Primary target	Wildfire-aware mechanism
SFIRE / WRF-Fire (Mandel et al., 2011)	Coupled fire–atmosphere simulation	Level-set fire spread with wind input and latent/sensible heat-flux feedback
"A Risk-Aware UAV-Edge Service Framework for Wildfire Monitoring and Emergency Response" (Huang et al., 23 Feb 2026)	Monitoring and emergency response	Fire-history-weighted clustering, QoS-aware edge assignment, dynamic rerouting
"Advanced Global Wildfire Activity Modeling with Hierarchical Graph ODE" (Xu et al., 4 Jan 2026)	Global wildfire activity forecasting	Multi-level graph hierarchy, adaptive message passing, Neural ODE dynamics
"FireDataForge" (Xia et al., 19 Jun 2026)	Event-centered data preparation	Retrieval and harmonization of 11 wildfire-related sources to a common grid
"Boundary-Aware Uncertainty Quantification for Wildfire Spread Prediction" (Funk, 4 May 2026)	UQ evaluation	Fire-Centered Evaluation Region around the fire boundary
"Electrical Power Network Modeling Framework for Wildfire Risk and Resilience Analysis" (Campos et al., 5 Mar 2026)	Grid resilience analysis	Joint modeling of wildfire propagation, grid interaction, operations, and community impact

A recurrent misconception is that wildfire-aware modeling is synonymous with wildfire spread simulation alone. The surveyed literature contradicts that view. Some frameworks focus on spread and plume physics; others on risk prediction, event reconstruction, edge computing, multi-source data fusion, or infrastructure resilience. This suggests that “wildfire-aware” is best understood as a systems property: wildfire-specific mechanisms are embedded wherever they materially affect prediction, control, or evaluation.

2. Coupled physical and geometric representations of fire behavior

Physics-grounded wildfire-aware frameworks typically represent fire spread as a moving interface or as a coupled fluid–combustion process. In SFIRE, the fire front is advanced by a level-set equation,

$$\frac{\partial \phi(\mathbf{x},t)}{\partial t} + S(\mathbf{x},t)\,\lvert\nabla \phi(\mathbf{x},t)\rvert = 0,$$

where $\phi$ is the level-set field and $S$ is the normal spread speed. The model is tightly coupled to WRF: “the fire module takes the wind as input and returns the latent and sensible heat fluxes,” so the atmosphere drives the fire and the fire feeds heat and moisture back into the atmosphere (Mandel et al., 2011). The same environment adds perimeter ignition with atmosphere–fire spin-up, interpolation from an ideal logarithmic wind profile for nonhomogeneous fuels, and diagnostic quantities such as Byram’s fireline intensity and the speed-of-burning-aware intensity

$J = \frac{H\,R\,w_0}{2\,T_f}.$

SWIRL-FIRE extends this physical lineage toward high-resolution large-eddy simulation. It solves Favre-filtered conservation equations for mass, momentum, oxygen mass fraction, and potential temperature under a low-Mach formulation, with vegetation drag, oxygen consumption, convective and radiative heat exchange, moisture evaporation, and buoyancy all represented as source terms (Wang et al., 2022). In FireFlux II validation, the framework showed that global quantities such as volumetric heat release and fire-spread rate were insensitive to horizontal mesh resolution between $0.5\,\mathrm{m}$ and $2\,\mathrm{m}$, while finer grids better captured intermittency and dynamic fire properties associated with fine-scale turbulent structures in the atmospheric boundary layer (Wang et al., 2022). FireBench then scaled this LES paradigm into an ensemble setting by integrating SWIRL-FIRE with Vizier and TPU execution; it produced 117 simulations, each with 1.35 billion mesh points, and identified a critical convective Froude number near $0.5$ separating plume-dominated from wind-dominated regimes (Wang et al., 2024).

A different but still wildfire-aware physical abstraction appears in the satellite-first geometric framework based on generalized elliptical frames. That work advances fire fronts either by a Huygens envelope of locally anisotropic frames or by a frame-only enclosure strategy, using thermal detections, FRP, wind, and LFMC proxies rather than detailed local fuel models (Dehkordi, 11 Dec 2025). Its directional spread geometry is encoded through a generalized polar form $r(\theta,x,y,t)$, while LFMC is estimated by

$$\mathrm{LFMC}(x,y,t)=A+B\cdot \mathrm{NDVI}(x,y,t)+C\cdot \mathrm{LST}(x,y,t)+D_1\sin(2\pi \mathrm{DOY}/365)+D_2\cos(2\pi \mathrm{DOY}/365).$$

This provides a data-limited alternative to coupled CFD-style models.

These physical and geometric formulations differ in fidelity and computational burden, but they share a common wildfire-aware trait: wind, terrain, fuel state, or anisotropic spread are not post hoc covariates but structural components of the state evolution.

3. Data-driven temporal, graph, and segmentation models

Recent wildfire-aware learning frameworks have moved beyond short-window classifiers toward explicit multi-scale temporal modeling. HiGO represents the Earth system as a three-level graph hierarchy with 4-neighbor intra-level edges and learnable parent–child inter-level mappings. At each level, hidden states evolve according to a GNN-parameterized Neural ODE,

$$\frac{d h^{(l)}(t)}{d t} = f_{\theta}^{(l)}(h^{(l)}(t), x^{(l)}(t), t; G^{(l)}),$$

while adaptive filtering message passing performs context-aware aggregation and inter-level pooling and unpooling (Xu et al., 4 Jan 2026). On SeasFire Cube, HiGO was reported as best across all evaluated horizons; for example, at 8 days it achieved AUPRC/M-F1 of $\phi$0 versus $\phi$1 for GraphCast, and at 48 days $\phi$2 versus $\phi$3 (Xu et al., 4 Jan 2026). The same work reported “strong observational consistency” and showed that removing land variables harmed performance most.

At event scale, deep segmentation models have been used to predict final burned extent from ignition-centered spatiotemporal cubes. In the Mediterranean framework for final burned-area prediction, the task is binary semantic segmentation on $\phi$4 km patches at $\phi$5 resolution, using a 10-day window from four days before ignition to five days after ignition (Anastasiou et al., 23 May 2025). The best-performing model was a 3D U-Net with input shape $\phi$6, trained with a composite BCEDice loss,

$\phi$7

On the 2022 test set it reached Dice $\phi$8 and IoU $\phi$9, improving over the ignition-day-only U-Net2D baseline at Dice $S$0 and IoU $S$1 (Anastasiou et al., 23 May 2025). The ablation over post-ignition window length showed systematic degradation as fewer post-ignition days were included.

For next-day danger forecasting, the uncertainty-aware LSTM framework models both epistemic and aleatoric uncertainty over 45-day per-pixel sequences. Its predictive decomposition follows

$S$2

and its best-performing model, BBB plus aleatoric heteroscedastic logits, achieved Precision $S$3, Recall $S$4, F1 $S$5, AUPRC $S$6, and ECE $S$7 for next-day prediction, compared with a deterministic baseline at F1 $S$8 and ECE $S$9 (Kondylatos et al., 29 Sep 2025). The same work reported that aleatoric uncertainty increases with forecast horizon up to 10 days, whereas epistemic uncertainty remains stable.

A complementary trustworthy risk framework for western Canada uses a compact WaveletMixer encoder with three levels of Haar low-pass decomposition, patch-wise time–channel mixing, and checkpoint-ensemble uncertainty estimation. Over the 2023 and 2024 fire seasons, it reported F1 $J = \frac{H\,R\,w_0}{2\,T_f}.$0, PR-AUC $J = \frac{H\,R\,w_0}{2\,T_f}.$1, parameter count $J = \frac{H\,R\,w_0}{2\,T_f}.$2, and FLOPs $J = \frac{H\,R\,w_0}{2\,T_f}.$3 (Xu et al., 4 Jan 2026). SHAP analysis in that work showed temperature-related drivers as dominant in both years, with moisture-related constraints more influential in shaping spatial and land-cover-specific contrasts in 2024.

A persistent methodological point emerges from these studies: wildfire-aware learning is not merely about replacing physics with neural networks. It more often means making temporal scale separation, land-cover dependence, hazard-specific uncertainty, or continuous-time structure explicit inside the model.

4. Data integration, event-centered pipelines, and digital twins

Wildfire-aware modeling depends on geospatial harmonization because wildfire-relevant inputs arrive in incompatible spatial resolutions, cadences, and coordinate systems. FireDataForge addresses this directly by retrieving 11 wildfire-related datasets for a single MTBS Event ID, reprojecting vectors and rasters to a target CRS, clipping them to the event extent, resampling them by data type, temporally aligning dynamic sources, and writing analysis-ready arrays with CRS, affine transform, resolution, extent, units, native resolution, source attribution, and checksums embedded in the metadata (Xia et al., 19 Jun 2026). Its common grid is defined in EPSG:5070 by default, with a GDAL-style affine transform, and its validation across eight events reported coordinate round-trip error $J = \frac{H\,R\,w_0}{2\,T_f}.$4, FRP conservation error $J = \frac{H\,R\,w_0}{2\,T_f}.$5, DEM RMSE $J = \frac{H\,R\,w_0}{2\,T_f}.$6, landcover accuracy $J = \frac{H\,R\,w_0}{2\,T_f}.$7, and WUI accuracy $J = \frac{H\,R\,w_0}{2\,T_f}.$8 (Xia et al., 19 Jun 2026).

Earlier large-scale wildfire data engineering made a similar point from a machine-learning perspective. The dataset-construction pipeline behind the uncertainty-aware wildfire management framework combined VIIRS active-fire detections with LANDFIRE raster covariates over a California grid of $J = \frac{H\,R\,w_0}{2\,T_f}.$9 cells, using zonal statistics over 30 m rasters and a scalable intersections-file strategy instead of raster-vector conversion (Diao et al., 2020). The resulting dataset contained $0.5\,\mathrm{m}$0 cell-day samples, and the geospatial processing complexity was reduced to $0.5\,\mathrm{m}$1 (Diao et al., 2020). The same pipeline illustrates that wildfire-aware modeling often begins well before model training, at the level of spatial joins, neighborhood construction, and covariate provenance.

Perception pipelines also require wildfire-aware adaptation. LADA reformulates wildfire detection as semi-supervised domain adaptation for object detection, using Faster R-CNN with ResNet-50 and FPN, CoordConv-style location channels in FPN and RPN, teacher–student pseudo-labeling, and masked image consistency (Jang et al., 2024). Its revised HPWREN labeling policy merges smoke bounding boxes into fewer, larger targets, and with only 1% target-domain labeled data the SSDA version improved over the source-only baseline by $0.5\,\mathrm{m}$2 mAP on HPWREN (Jang et al., 2024). That improvement is not a generic domain-adaptation result; it relies on wildfire-specific spatial priors such as the asymmetry between sky regions and likely smoke-source regions.

Digital twins provide a further extension of the data layer into embodied simulation. FIRE-VLM constructs a wildfire digital twin from USGS DEM terrain, LANDFIRE fuels, and CAWFE-derived fireline products inside Unreal Engine 5.3, then places a dual-view AirSim multirotor into that evolving environment (Webb et al., 6 Jan 2026). This suggests that event-centered harmonization, perception adaptation, and embodied simulation are increasingly contiguous rather than separate stages.

5. Uncertainty, evaluation, and operational decision support

Wildfire-aware uncertainty quantification is increasingly evaluated where it matters most: near the active boundary. FCER defines a fire-centered region by dilating the fire set,

$0.5\,\mathrm{m}$3

or by using band-limited regions around the signed distance field (Funk, 4 May 2026). This protocol changes what counts as good uncertainty. At the ASD anchor, the ensemble achieved mean AUROC $0.5\,\mathrm{m}$4 and AUPRC $0.5\,\mathrm{m}$5, whereas the distilled student model achieved $0.5\,\mathrm{m}$6 and $0.5\,\mathrm{m}$7; pooled Wilcoxon signed-rank tests gave $0.5\,\mathrm{m}$8 for both AUROC and AUPRC in favor of the student (Funk, 4 May 2026). The same study found comparable calibration at the ASD anchor, with Brier $0.5\,\mathrm{m}$9 for the ensemble and $2\,\mathrm{m}$0 for the student.

A related uncertainty problem arises when wildfire forecasts must be initialized from partial satellite evidence. The cWGAN framework for fire-arrival-time inference treats the fire arrival time as the latent object to be reconstructed from VIIRS active-fire detections, then uses that field to initialize WRF-SFIRE through atmosphere-consistent replay (Shaddy et al., 2023). Across four California fires, it reported average Sorensen’s coefficient approximately $2\,\mathrm{m}$1 for fire perimeters and mean absolute ignition-time error approximately 32 minutes (Shaddy et al., 2023). Here, uncertainty is not auxiliary; it directly determines whether a coupled forecast starts from a plausible fire history.

Operational frameworks then convert wildfire-aware estimates into actions. WMaaS defines wildfire monitoring as a service through joint optimization of risk-aware clustering, QoS-aware edge assignment, 2-opt route optimization, adaptive fleet sizing, and dynamic emergency rerouting. Its response-time proxy is

$2\,\mathrm{m}$2

and its fire-history-weighted clustering uses

$2\,\mathrm{m}$3

Against GA, PSO, and greedy baselines, it reported $2\,\mathrm{m}$4–$2\,\mathrm{m}$5 lower average response time, $2\,\mathrm{m}$6–$2\,\mathrm{m}$7 lower energy, and $2\,\mathrm{m}$8–$2\,\mathrm{m}$9 fewer UAVs, while meeting a 300 s emergency deadline with 233 s achieved (Huang et al., 23 Feb 2026).

For active intervention, the hybrid CNN–CA suppression framework treats water and retardant as distinct controls, optimizes aerial drop schedules by backpropagating through a differentiable wildfire model, and evaluates robustness under both aleatoric and epistemic uncertainty (Matei et al., 11 Jun 2026). In the 2020 Bear Fire case study, Monte Carlo analysis at Day 15 reported aleatoric baseline mean $0.5$0 ha versus optimized $0.5$1 ha, and epistemic baseline mean $0.5$2 ha versus optimized $0.5$3 ha, corresponding to reductions of $0.5$4 and $0.5$5 in simulator-conditional area respectively (Matei et al., 11 Jun 2026).

Autonomous monitoring closes the loop between uncertainty and control. FIRE-VLM couples PPO with dual-view sensing and CLIP-style semantic reward shaping inside a physics-grounded wildfire digital twin. The potential-based shaping term is

$0.5$6

and the total reward adds this VLM term and a directional semantic bonus to the physical reward (Webb et al., 6 Jan 2026). On Task 2, time-to-detection fell from $0.5$7 for Base PPO to $0.5$8 for the VLM-guided final policy, while time-in-FOV rose from $0.5$9 to $r(\theta,x,y,t)$0 (Webb et al., 6 Jan 2026).

At a more abstract decision-theoretic level, uncertainty-aware wildfire management formulates suppression as a POMDP over fire status and fuel states, with online planning via a modified POMCPOW belief update (Diao et al., 2020). Across grid sizes from $r(\theta,x,y,t)$1 to $r(\theta,x,y,t)$2, mean negative utility improved over a heuristic baseline in every reported setting, including reductions of $r(\theta,x,y,t)$3 for $r(\theta,x,y,t)$4 and $r(\theta,x,y,t)$5 for $r(\theta,x,y,t)$6 at $r(\theta,x,y,t)$7 (Diao et al., 2020). This line of work emphasizes a point sometimes obscured in predictive studies: wildfire-aware modeling is frequently inseparable from partial observability and resource allocation.

6. Infrastructure coupling, resilience analysis, and open limitations

Wildfire-aware modeling has expanded beyond fire behavior and monitoring into critical-infrastructure analysis. The electrical power network framework organizes the problem into hazard modeling, infrastructure modeling, interaction modeling, operations, community impact modeling, and resilience evaluation (Campos et al., 5 Mar 2026). It explicitly distinguishes grid-to-fire ignition mechanisms from fire-to-grid damage, and it proposes community-facing metrics such as SAIDI, SAIFI, EENS, and functionality-over-time

$r(\theta,x,y,t)$8

The framework’s central claim is that traditional IEEE test cases ignore fuels, topography, time-evolving fire behavior, Tx–Dx coupling, and community impacts (Campos et al., 5 Mar 2026).

The adversarial wildfire framework for grid disruption then turns the hazard into an optimization variable. It formulates a mixed-integer conic program for minimum time-to-outage and maximum sequential load shed under realistic fire-spread constraints, with test cases based on the Park, Eaton, and Palisades fires (Brun et al., 19 Mar 2026). In SCOPF screening under the LF, $r(\theta,x,y,t)$9 setting, the number of contingencies was reduced from 744 to 148 for Park, 744 to 455 for Palisades, and 744 to 556 for Eaton, with corresponding SCOPF solve-time reductions and essentially unchanged resilience to plausible wildfire contingencies (Brun et al., 19 Mar 2026). This is a distinctly wildfire-aware use of spread modeling: not prediction of the fireline for its own sake, but construction of physically plausible worst-case contingencies.

Decision-dependent uncertainty further changes planning conclusions at the distribution level. In the DRO-DDU framework, the expected line-unavailability bound is modeled as

$$\mathrm{LFMC}(x,y,t)=A+B\cdot \mathrm{NDVI}(x,y,t)+C\cdot \mathrm{LST}(x,y,t)+D_1\sin(2\pi \mathrm{DOY}/365)+D_2\cos(2\pi \mathrm{DOY}/365).$$0

so operating flows through high-threat lines increase the failure bound unless mitigated by hardening (Piancó et al., 2024). In the 54-bus case study, the DDU-aware plan built lines 17 and 34, made lines 5 and 52 switchable, and hardened lines 22 and 55; out-of-sample Monte Carlo analysis then reduced annual loss of load from $$\mathrm{LFMC}(x,y,t)=A+B\cdot \mathrm{NDVI}(x,y,t)+C\cdot \mathrm{LST}(x,y,t)+D_1\sin(2\pi \mathrm{DOY}/365)+D_2\cos(2\pi \mathrm{DOY}/365).$$1 of demand to $$\mathrm{LFMC}(x,y,t)=A+B\cdot \mathrm{NDVI}(x,y,t)+C\cdot \mathrm{LST}(x,y,t)+D_1\sin(2\pi \mathrm{DOY}/365)+D_2\cos(2\pi \mathrm{DOY}/365).$$2, SAIDI from 280 to 18 hours/year, and SAIFI from 59 to 14 interruptions/year, with annual investment cost reported as \$91,049 (Piancó et al., 2024). This is not a marginal adjustment to a conventional planner; it changes the investment portfolio because wildfire risk depends on the decisions themselves.

The literature is also explicit about its limitations. WMaaS assumes offline deterministic planning, stationary fire-history weights within a planning horizon, constant UAV speed, and reliable communications within effective ranges; it omits physics-based fire spread and dynamic risk updates driven by weather or wind (Huang et al., 23 Feb 2026). HiGO does not model uncertainty in its reported experiments and uses a fixed 4-neighbor grid without explicit wind-direction priors (Xu et al., 4 Jan 2026). The generalized elliptical-frame framework omits slope effects in the presented formulation and does not model spotting or crown-fire transitions (Dehkordi, 11 Dec 2025). SWIRL-FIRE assumes homogeneous flat terrain in its FireFlux II study and reports sensitivity of burning-rate statistics to mesh refinement even when spread rate is converged (Wang et al., 2022). The electrical-resilience literature identifies additional gaps in component-specific fragilities, ember transport at corridor scale, community-impact fidelity, and computational scaling of coupled wind–fire–infrastructure simulations (Campos et al., 5 Mar 2026).

These limitations indicate that wildfire-aware modeling remains a federated field rather than a settled stack. The surveyed papers consistently point toward multi-objective optimization, stochastic and chance-constrained planning, coupled weather–fire–infrastructure models, data assimilation, online risk updating, heterogeneous fleets, dynamic clustering, and stronger uncertainty calibration as the next steps (Huang et al., 23 Feb 2026, Xu et al., 4 Jan 2026, Campos et al., 5 Mar 2026, Funk, 4 May 2026). The unifying direction is not toward one universal model, but toward tighter coupling between wildfire physics, geospatial data, uncertainty, and decision-making.