Forest structure, roads and soil moisture provide realistic predictions of fire spread in modern Swedish landscape

Abstract

Recent increases in fire activity in Sweden call for the quantification of forest fire susceptibility, in order to develop management strategies to mitigate fire risk. Using the data from 100 large Swedish forest fires (>10 ha), mapped from sentinel-2 images from 2016 to 2022, we explored the predictive power of vegetation properties in estimating relative likelihood of fires within a landscape using logistic regression. To model spatially explicit fire susceptibility within a given landscape, we used the outcome of logistic regression as an input into a cellular automata model (CA model), which simulates fire spread in a 2D grid.

The CA was model calibrated on three fires that occurred between 2016 and 2022, then verified on six 2023 fires and featured a mean sensitivity of 0.74 and specificity of 0.79. The logistic regression model had an accuracy of 54 %, showing increased fire susceptibility from high Scots pine volume (p-value = 0.02), and decreased fire susceptibility from high volumes of deciduous trees and wet soil. Realistic outcomes of the CA model and reliance of our approach on publicly available data with nation-wide coverage of vegetation cover in Sweden allows for the development of an automated protocol of fire susceptibility assessment at the operational level and its integration in existing decision support systems. This would allow forest owners to obtain estimates of forest fire susceptibility for different forest management strategies.