Predicting burn probability: Dimensionality reduction strategies enable accurate and computationally efficient metamodeling

Predicting the probability that a given location will be burnt by a wildfire is an important part of understanding the risk that wildfires pose and how our management actions (e.g., prescribed burning) can reduce this risk. Existing methods to quantify this burn probability involve simulating the spread of many thousands of individual wildfires, making them highly computationally expensive. To reduce this expense, we propose strategies that enable the development of computationally efficient machine learning assisted metamodels for estimating burn probability, which are demonstrated for a case study in South Australia. Artificial neural networks are used as the metamodel to emulate the outputs of a landscape fire simulation model. Development of the metamodel is facilitated by reducing the input and output dimensionality of the simulation model by a factor of 10,000-1,000,000, while still being able to predict burn probabilities with high accuracy (approximately ± 7.4% error, on average) and only requiring 0.6% of the computational time compared with an approach using landscape fire simulation models. This opens the door to obtaining many thousands of spatially distributed estimates of burn probability, as is required when optimising fuel treatment strategies.