Wild3R: Advanced Wildland Fire Forecasting

• Wild3R is an advanced framework for real-time wildland fire propagation forecasting, coupling semi-empirical spread dynamics with atmospheric modeling.
• It integrates high-resolution fire spread, level-set front tracking, and a novel morphing ensemble Kalman filter to correct both amplitude and positional errors.
• The system leverages continuous remote sensing and scalable parallel computation to enable operational forecasting for emergency management and ecological modeling.

Wild3R denotes a family of advanced frameworks for real-time, data-driven modeling of wildland fire propagation, integrating semi-empirical surface fire spread dynamics, atmospheric interactions, and data assimilation using morphing ensemble Kalman filtering. The system enables accurate, scalable, and operational forecasting of fire behavior by tightly coupling high-resolution physical simulation with continuous remote sensing assimilation, particularly from thermal infrared imagery. Key algorithmic pillars include the level-set method for front tracking, coupled fire-atmosphere computation using the Weather Research and Forecasting (WRF) model, and a novel morphing EnKF that directly corrects both amplitude and positional errors in the state estimate. The reference implementation and methodology are detailed in (0801.3875).

1. System Architecture and Coupling Paradigm

Wild3R consists of three primary computational modules operating in a coupled loop:

1. The fire–atmosphere model integrates a two-dimensional semi-empirical fire spread solver (based on the Rothermel–Clark model) with WRF, which supplies dynamic atmospheric fields to the fire grid and receives in turn fire-generated heat/mass fluxes.
2. A synthetic-scene rendering engine (e.g., DIRSIG) generates timeresolved thermal infrared radiance fields from the model state, providing direct comparability with airborne or satellite imagery.
3. The morphing ensemble Kalman filter (EnKF), which assimilates observed thermal images and (optionally) ground measurements, updating both physical variables and spatial alignment of the forecast to correct for discrepancies.

These components communicate asynchronously, with forecast–assimilation cycles typically ranging between 5–15 min of simulation time and a wall-clock turnaround on the order of minutes on 500–2000 cores, enabling near-real-time operational forecasting (0801.3875).

2. Semi-Empirical Fire Spread and Post-Frontal Heat Dynamics

Surface fire spread is parameterized following the Rothermel–Clark semi-empirical paradigm. The normal spread rate $S$ of the fire front (m s⁻¹) is

$$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$

where $R_0$ is the no-wind, flat-terrain rate, $a$, $b$, and $d$ are fuel- and wind-dependent empirical coefficients, $\vec{v}$ is the local wind vector, $\vec{n}$ is the outward unit normal to the interface, and $\nabla z$ is the terrain gradient. $S_{max}$ is a fuel-limited maximum.

Post-frontal fuel consumption assumes exponential decay of mass:

$$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$0

for initial fuel mass $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$1 and time constant $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$2, leading to a sensible heat flux

$$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$3

with $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$4 the fuel heat content. These fluxes are vertically distributed into the atmospheric model by exponential weighting.

3. Level-Set Firefront Tracking and Numerical Methods

The burning region is defined by a level-set function $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$5, with $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$6 inside the fire and $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$7 at the fire front. Fire propagation is governed by

$$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$8

solved with a second-order upwind Godunov scheme for spatial derivatives, and Heun’s (second-order Runge–Kutta) method for time integration. This preserves accurate normal propagation speed and geometric fidelity at the interface, outperforming explicit Euler or lower-order discretizations. Boundary conditions include fixed "no-fire" at the fire grid’s outer edge and Neumann (no-flux) for $$S = R_0 + a (\vec{v} \cdot \vec{n})^b + d (\nabla z \cdot \vec{n}), \quad 0 \leq S \leq S_{max},$$9.

4. Data Assimilation via Morphing Ensemble Kalman Filter

The morphing EnKF corrects both amplitude and positional errors inherent in complex fire propagation. The extended state $R_0$0 includes both a residual field $R_0$1 and deformation (morphing) field $R_0$2 such that the analysis state is

$R_0$3

where $R_0$4 is a candidate forecast and $R_0$5 the reference state. Registration minimizes

$R_0$6

where $R_0$7 is the identity map and regularization terms favor physically plausible deformations.

ΔAfter registration, classical stochastic EnKF analysis is performed:

$R_0$8

where $R_0$9 is the Kalman gain, $a$0 the observation vector (e.g., pixelwise thermal radiances), $a$1 the observation operator mapping model space to observation space, and $a$2 i.i.d. Gaussian noise. State update is completed by reconstructing $a$3.

5. Observation Operator, Synthetic IR Scene Generation, and Innovations

The observation operator $a$4 transforms the model state (fire front, heat flux, and time since ignition) through a physically-based synthetic image pipeline:

• $a$5 → surface heat fluxes ($a$6) + ground cooling → 2-D temperature field → 3-D flame voxelization → full radiative transfer (DIRSIG) → synthetic IR radiance image $a$7.

Assimilation compares $a$8 with the real observation $a$9; the residual field $b$0 (difference in radiance) drives EnKF updates for both field values and front morphology. The system accommodates both mid-wave (3–5 μm) and long-wave (8–12 μm) IR wavelengths, as available from sensor data.

6. Real-Time Scalability, Domain Decomposition, and Operational Use

Typical production domains span 20 km × 20 km with 60 m WRF mesh, 6 m fire grid, and 40–60 vertical atmospheric levels. The natural time base for model coupling is dictated by the smaller of atmosphere/fire CFL constraints, typically $b$1 s, with state assimilation every 5–15 simulation minutes. Ensemble sizes around $b$2 suffice to capture major forecast uncertainty. Parallelization on O(500–2000) cores produces assimilation cycles in minutes, suitable for operational nowcasting and incident response.

7. Synthesis and Extensibility

Wild3R establishes a reproducible, extensible template for coupled, operational wildland fire forecasting that can be adapted to incorporate additional physical processes (e.g., smoke transport), alternative observation types (radar, lidar), or improved fire spread physics. All major algorithmic steps—including the semi-empirical spread equations, level-set and registration solvers, WRF interaction, and ensemble Kalman assimilation—are grounded in established meteorological and computational mathematics literature (0801.3875).

By solving for both front location and fire-atmosphere feedback and tightly integrating with real-world observations, Wild3R advances state estimation fidelity and forecast reliability for emergency management, ecological modeling, and atmospheric science.