Fire modelling in an uncertain world

By Rachael Quill

This article appeared in the Winter 2016 edition of Fire Australia magazine. A version of this article first appeared in the March–April issue of Wildfire magazine. Article reproduced with permission.

Uncertainty is all around us. We account for it in all aspects of everyday life: “It takes 10 minutes to get there but I’ll allow 15, just in case.” So why is it that when it comes to the wicked problem of bushfire, we have put uncertainty to one side for so long?

Traditionally, operational fire prediction has been based on deterministic methods—for each set of input conditions there is a single output, with no allowance for uncertainty in the process. After almost 30 years of discussion in the literature, probabilistic approaches to fire-spread and behaviour modelling are now beginning to emerge. These probabilistic approaches account for uncertainty in fire-spread by allowing for random fluctuations in the input variables and predicting a range of fire propagation scenarios. When considered together, these multiple predictions are overlaid to form what is known as an ensemble. Ensemble-based predictions allow fire-spread across the landscape to be defined in terms of probabilities, such as likelihood of burning and risk to assets.

However, within these developing frameworks we still rely on deterministic models and simplified probabilistic inputs. That is to say, we are trying to understand the variability of fire-spread without capturing the true variability of the driving factors. As input errors are a major source of prediction error in fire modelling (see Cruz and Alexander 2013), it is imperative that we seek to acknowledge the uncertainties of these inputs. 

Wind, in particular, is known to account for much of the variability displayed in the spread of bushfires (see Cruz and Alexander 2013). Because of the constraints of operational requirements (i.e. real-time or near real-time prediction), the current physics-based deterministic wind models for fire prediction do not well capture the variability of wind flow in key areas, particularly across complex terrain. In the worst case, on leeward slopes, errors in wind direction of up to 180° have been recorded. Even in the new ensemble-based approaches, wind direction is only characterised as random whereas analysis of data collected across complex terrain has shown that wind direction takes a highly structured form. To better capture the uncertainty of fire-spread across the landscape, we must characterise the structured nature of wind direction within fire prediction frameworks.

Without capturing the true variability of wind flow across complex terrain, the curse of error accumulation through the modelling process leaves us to question the uncertainty of fire-spread predictions in these key regions. It has been suggested that traditional fire-modelling techniques are failing to capture dynamic processes such as vorticity-driven lateral spread (dynamic spread of the fire front on a leeward slope in a direction perpendicular to that of the wind) in areas of flow separation (see Simpson, Sharples, Evans and McCabe, 2013). Lee-slope eddies are not accounted for by the wind models used today.

Figure 1: wind direction in a valley

The first step in handling the wicked problem of uncertainty in wind modelling is to understand the effects of the physical environment, such as vegetation or topography, on the statistical representation of wind fields. A statistical representation, rather than the traditional physics-based approach, allows discussion of probability— leading to analysis of scenarios with quantified likelihoods. 

Statistical analyses of the effects of topographical aspect on wind direction clearly indicate thresholds for dynamic behaviour. This can of course be understood using detailed physical analysis and has been studied using sophisticated mathematical models (see Simpson, Sharples, Evans and McCabe, 2013). However, from the firefighting perspective, we must look to understand the uncertainty around this behaviour and capture it within our operational models under the constraints of real-time prediction.

When considering the effects of vegetation on wind direction across complex terrain, the story becomes less clear—and the role of uncertainty becomes yet more important. Changing vegetation structures have distinct effects on wind direction in some parts of the terrain—and the behaviours are consistent with the current predictions. However, in other areas of the terrain, the impact of vegetation on wind direction is far less obvious—and observed behaviours vary from those currently captured by state-of-the-art models. 

Better statistical understanding of the variations in wind fields across the landscape will improve on current physics-based methods by better capturing wind dynamics in complex terrain. Development of hybrid models, which combine probabilistic information with deterministic approaches to wind modelling, will provide better understanding of uncertainty within the fire-modelling process while maintaining operational real-time (or near real-time) prediction. The result of such a hybrid model would ultimately provide more information to fire managers and decisions-makers dealing with the wicked problem on the ground.

Learn more about Rachael’s PhD and watch a three minute thesis here.

REFERENCES

1. Cruz MG and Alexander ME, 2013, ‘Uncertainty associated with model prediction of surface and crown fire rates of spread’, Environmental Modelling & Software, 47, 16–28.

2. Simpson CC, Sharples JJ, Evans JP and McCabe MF, 2013, ‘Large eddy simulation of atypical wildland fire spread on leeward slopes’, International Journal of Wildland Fire, 22(5) 599–614.