Using efficient parallelization in Graphic Processing Units to parameterize stochastic fire propagation models

Abstract

Wildfires are a major concern in Argentinian northwestern Patagonia and in many ecosystems and human societies around the world. We developed an efficient cellular automata model in Graphic Processing Units (GPUs) to simulate fire propagation. The graphical advantages of GPUs were exploited by overlapping wind direction, as well as vegetation, slope, and aspect maps, taking into account relevant landscape characteristics for fire propagation. Stochastic propagation was performed with a probability model that depends on aspect, slope, wind direction and vegetation type. Implementing a genetic algorithm search strategy we show, using simulated fires, that we recover the five parameter values that characterize fire propagation. The efficiency of the fire simulation procedure allowed us to also estimate the fire ignition point when it is unknown as well as its associated uncertainty, making this approach suitable for the analysis of fire spread based on maps of burnt areas without knowing the point of origin of the fires or how they spread.

Fire propagation visualization in real time

“…(1) Where I f , ψ, ω and σ , are related with the vegetation type, aspect, wind direction and slope respectively, as explained in [8,12]. Fuel type coefficient β 0 is the baseline fire propagation probability for shrubland cells and β 1 is the difference between shrubland and forest.…”

“…In order to determine if a target cell is reached by fire, the ignition probability is calculated taking into account the state of the 8-cells neighborhood using Equation 1 as explained in [8]. The following pseudocodes illustrate the main characteristics of algorithms (in order to improve clarity these algorithms does not show visual interface details): fire spread flag setting ⊲ If fire stops 7: end while Algorithm 1 presents the main fire spread simulation loop, which is executed to perform a complete simulation.…”

“…In previous works [8], a High Performance Computing forest fire simulator was developed as a parallel cellular automata implemented in CUDA-C ( [9,10,11]). Our simulator spreads the fire over a landscape taking into account several layers of topography, vegetation type and wind ( [12,8]). The model for fire propagation is a semi-physical model and takes into account as input the actual vegetation, slope, average wind intensity and direction.…”

“…The search in the parameter space was done using a Genetic Algorithm (GA), a previously implemented methodology [13,14,15] that we programmed in parallel using CUDA-C. A Monte Carlo method was also used to find the best parameters but GA was proved to be more efficient. The results of this calibration step are exposed in [8].…”

“…Once the calibration was performed and the propagation parameters were estimated, we can simulate and visualize 10 stochastic simulations in parallel in the order of fraction of seconds which is acceptable for the time needed to make decisions in real time [18]. In this contribution we will not discuss in detail the model fitting stage that was previously explained elsewhere [8]. The main contribution of this work is the design and implementation of a powerful visualization tool for fire propagation, based on the needs of firefighters in Patagonia, Argentina.…”