Stochastic Nonlinear Elliptical-Growth Fire Models

• Stochastic nonlinear elliptical-growth fire front models are a class of mathematically rigorous simulation tools that incorporate randomness and nonlinearity to model wildfire spread.
• They leverage elliptical firelets, cell-based paradigms, and vertex-based tracking to capture the effects of fluctuating wind, fuel, and moisture on fire perimeters.
• These models enable predictive control and adaptive monitoring through scalable algorithms and Monte Carlo sampling, informing risk analysis and decision support.

Stochastic nonlinear elliptical-growth fire front models provide a physically rooted and mathematically rigorous framework for simulating, predicting, and adaptively monitoring the evolution of wildfire perimeters under complex, uncertain environmental conditions. These models formalize fire spread as an inherently stochastic, spatially and temporally nonlinear process, where the local propagation is governed by the evolution of elliptical firelets subject to randomly fluctuating wind, fuel, and moisture parameters. Two notable research threads—the cell-based simulation paradigm (Cell2Fire) (Pais et al., 2019) and probabilistic perimeter evolution with closed-loop monitoring (Papaioannou et al., 16 Jan 2026)—demonstrate the breadth and applicability of this modeling approach.

1. Mathematical Foundations of Nonlinear Elliptical Growth

In stochastic nonlinear elliptical-growth models, fire progression is parameterized via a dynamically evolving family of ellipses that locally approximate the fireline's advance. In the cell-based paradigm, each burning cell generates an “elliptical firelet” whose axes and orientation result from empirical rate-of-spread (ROS) values supplied by fire behavior prediction systems (e.g., Canadian FBP). Given time step $\Delta t$, the semi-axes are

$$a = \tfrac{1}{2}(R_h + R_b)\Delta t,\qquad b = R_f\Delta t,\qquad R_f = \frac{R_h + R_b}{2\,LB}$$

where $R_h, R_b$ denote head and back ROS, and $LB$ is the length-to-breadth ratio (Pais et al., 2019). Ellipse orientation aligns with the local wind, so the spread toward a neighbor at angle $\phi$ from the head axis is given by the Richards’ ellipse models, e.g.,

$$\mathrm{ROS}(\phi) = \frac{a b}{\sqrt{(b \cos\phi)^2 + (a \sin\phi)^2}}$$

In vertex-based formulations, the fire front is explicitly tracked as a sequence of $N$ perimeter points $x^i_t \in \mathbb{R}^2$, each evolving according to stochastic discrete-time nonlinear dynamics with locally randomized wind and fuel influence (Papaioannou et al., 16 Jan 2026):

$x^i_{t} = x^i_{t-1} + \Delta t\,\dot{x}^i_{t-1}$

with

$$\dot{x}^i = f_{\text{Richards}}(\alpha_1(i),\alpha_2(i),\alpha_3(i),\theta(i), w_s(i), r_f(i),\text{geometry})$$

where $\{\theta(i), w_s(i), r_f(i)\}$ are random fields over wind direction, wind speed, and local spread rates, typically modeled as stationary independent processes.

Nonlinearities arise due to the empirical dependence of ROS on wind, moisture, fuel, and slope, as well as through the interaction geometry as overlapping ellipses merge and drive ignition probability.

2. Stochastic Components and Their Implementation

Stochasticity enters at both the parameter and process levels, representing environmental uncertainty and intrinsic randomness. In cell-based models, three primary mechanisms are implemented (Pais et al., 2019):

1. Random ignition location: Draws from user-specified spatial probability distributions.
2. Empirical ROS perturbation: For each time step and active cell, $R_h', R_b', R_f'$ are sampled from $\mathcal{N}(R,\sigma^2)$, with user-determined variance, modeling uncertainties in ROS estimation.
3. Weather scenario sampling: Monte Carlo over finite sets of probable weather streams, with burn probability across the landscape estimated as the empirical frequency of cell ignition over scenarios.

In perimeter-tracking models (Papaioannou et al., 16 Jan 2026), stochasticity is encoded via random field draws at each vertex and time:

• $$\theta(i) \sim \mathrm{vonMises}(\mu_{\theta(i)},\kappa)$$ for wind direction,
• $$w_s(i) \sim \mathcal{N}_R(\mu_{w_s(i)}, \sigma^2_{w_s(i)})$$ for wind speed,
• $$r_f(i) \sim \mathcal{N}_R(\mu_{r_f(i)}, \sigma^2_{r_f(i)})$$ for spread rate,

with the state transition probability formalized as a one-step Markov kernel integrating over environmental randomness. These stochastic models enable the computation of fire spread probability distributions and risk-weighted fire perimeter forecasts.

3. Algorithmic Frameworks and Numerical Implementation

Cell-based frameworks partition the domain into a rectangular grid of $n\times m$ square cells, each with attributes encoding local slope, fuel, moisture, and state $\in \{\text{Available, Burning, Burned}\}$. The algorithm proceeds in discrete time steps, applying a parallelized message-passing scheme where burning cells send ignition “messages” to neighbors if their ellipses would envelop the neighbor center within $\Delta t$. Cells that receive a message and satisfy a threshold on ROS become burning at the next step (Pais et al., 2019).

Key pseudocode (paraphrased for clarity):

for t in 1..T_max: for i in BurningCells (parallel): compute R_h, R_b, R_f from FBP form ellipse (a, b, e) for j in neighbors(i): compute φ = bearing(i->j) - θ_wind compute R_dir = ROS_i(φ) if R_dir * Δt >= dist(i, j): send message(i->j) for j in AvailableCells receiving messages: if FBP_ROS_j > δ: mark j as Burning update state

Parallelism is achieved at the burn message passing step, with OpenMP-based shared-memory optimizations. Profiling indicates 80% of runtime is spent in this region, yielding 15×–20× speedups over serial code and up to 30× faster simulation compared to Prometheus on $10^8$ cell domains. Strong and weak scaling efficiencies reach 75–80% with up to 32 threads (Pais et al., 2019).

Vertex-based models enable high-fidelity perimeter evolution and form the basis for closed-loop estimation and control in adaptive monitoring scenarios (Papaioannou et al., 16 Jan 2026).

4. Bayesian Estimation and Predictive Control

Monitoring fire fronts under stochastic, nonlinear propagation necessitates recursive Bayesian estimation and actively informative measurement strategies. The system’s state comprises the agent pose and a multi-object belief $\mathcal{B}_t(X_t)$ over all front vertices. The Bayesian recursion consists of:

• Prediction: Propagate $\mathcal{B}_{t-1}(X_{t-1})$ using the Markov kernel (process model).
• Update: Incorporate new sensor measurements $Z_t$ via a Poisson point-process likelihood, which accounts for detection rates within the agent's sensing disk.

Exact inference is infeasible in the nonlinear, non-Gaussian multi-object context, so practical implementations utilize particle filtering, with the belief approximated by weighted samples $\{w_t^{(j)}, X_t^{(j)}\}_{j=1}^{N_s}$ and Sequential Importance Resampling (SIR) (Papaioannou et al., 16 Jan 2026). For parameterizable scenarios, an Extended Kalman Filter may be used when linearizations are appropriate.

Adaptive trajectory planning for a sensing platform (e.g., UAV) is cast as a finite-horizon Markov Decision Process (MDP) over joint belief-agent-pose state space:

• Action space: Discrete control inputs (e.g., speed, heading).
• Cost function: Posterior uncertainty, quantified as Risk-Weighted Dispersion (RWD):

$$\mathcal{C}(s_{t+1}) = \frac{1}{\omega} \sum_{\text{cell }\varepsilon} \mathcal{R}(\varepsilon)\cdot\det(\Sigma^\varepsilon_{t+1})$$

where $\mathcal{R}(\varepsilon)$ is the risk weight, and $\Sigma^\varepsilon_{t+1}$ is the local sample covariance of particles.

• Policy search: Policies (open-loop over horizon $T$) are evaluated by Monte Carlo rollouts, with an adaptive Lower Confidence Bound (LCB) selection scheme to asymptotically approach the optimal policy under standard multi-armed bandit regret guarantees ($O(\ln n/n)$ decline).

5. Model Validation, Performance, and Adaptive Monitoring Results

The Cell2Fire framework is validated against the Prometheus simulator using metrics such as Mean Squared Error (MSE), Structural Similarity Index (SSIM), and Frobenius norm differences between hourly binary burn grids. Across various real and synthetic fire scenarios (including the Dogrib landscape and five large British Columbia domains), Cell2Fire achieves average $1-\text{MSE}$ of 0.90–0.98 and SSIM of 0.70–0.95, with final simulated fire scars matching remotely sensed burn perimeters within approximately 10–20% (Pais et al., 2019).

In adaptive monitoring regimes, non-myopic (receding horizon $T=3,5$) predictive controllers significantly outperform myopic ($T=1$) strategies, reducing estimation RMSE by factors of 10–50 by actively guiding the agent to high-risk, high-uncertainty locations. As $T$ and the number of LCB rollouts increase, estimation accuracy rises, and observed policy regret matches theoretical $O(\ln n/n)$ rates (Papaioannou et al., 16 Jan 2026).

Baseline scenarios with “infinite” sensing and no agent motion yield the minimum achievable RMSE set by sensor noise alone; realized performance approaches this lower bound as horizon length and sampling increase.

6. Practical Implications and Applications

Stochastic nonlinear elliptical-growth fire front models support strategic wildfire management by producing probabilistic forecasts that account for environmental variability and uncertainty in spread rates and ignitions. Their parallelized, scalable implementation enables timely simulation of very large-scale landscapes (on the order of $10^8$ cells). These models are suitable for integration as subcomponents in coupled landscape management and risk analysis platforms, allowing users to assess probabilistic fire scar outcomes under varying weather and fuel management scenarios. Information-seeking control frameworks enable efficient sensor deployment (e.g., UAV path planning) for situational awareness and adaptive monitoring, with convergence guarantees on policy selection.

A plausible implication is the growing integration of such stochastic nonlinear frameworks within operational wildfire decision support and situational awareness systems, particularly as remote sensing and sensor-driven monitoring technologies mature.

---

References

(Pais et al., 2019) Cell2Fire: A Cell Based Forest Fire Growth Model (Papaioannou et al., 16 Jan 2026) Adaptive Monitoring of Stochastic Fire Front Processes via Information-seeking Predictive Control