Citation

Fernández-García V, Beltrán-Marcos D, Fernández-Guisuraga JM, Marcos E, Calvo L. Sci. Total Environ. 2022; ePub(ePub): ePub.

Copyright

(Copyright © 2022, Elsevier Publishing)

DOI

10.1016/j.scitotenv.2022.154729

PMID

35331756

Abstract

The large environmental and socioeconomic impacts of wildfires in Southern Europe require the development of efficient generalizable tools for fire danger analysis and proactive environmental management. With this premise, we aimed to study the influence of different environmental variables on burn severity, as well as to develop accurate and generalizable models to predict burn severity. To address these objectives, we selected 23 wildfires (131,490 ha) across Southern Europe. Using satellite imagery and geospatial data available at the planetary scale, we spatialized burn severity as well as 20 pre-burn environmental variables, which were grouped into climatic, topographic, fuel load-type, fuel load-moisture and fuel continuity predictors. We sampled all variables and divided the data into three independent datasets: a training dataset, used to perform univariant regression models, random forest (RF) models by groups of variables, and RF models including all predictors (full and parsimonious models); a second dataset to analyze interpolation capacity within the training wildfires; and the third dataset to study extrapolation capacity to independent wildfires.

RESULTS showed that all environmental variables determined burn severity, which increased towards the mildest climatic conditions, sloping terrain, high fuel loads, and coniferous vegetation. In general, the highest predictive and generalization capacities were found for fuel load proxies obtained though multispectral imagery, both in the individual analysis and by groups of variables. The full and parsimonious models outperformed all, the individual models, models by groups, and formerly developed predictive models of burn severity, as they were able to explain up to 95%, 59% and 25% of variance when applied to the training, interpolation and extrapolation datasets respectively. Our study is a benchmark for progress in the prediction of fire danger, provides operational tools for the identification of areas at risk, and sets the basis for the design of pre-burn management actions.

Language: en

Keywords

Climate; Topography; Copernicus; Fire severity; Fuel continuity; Fuel load; Fuel moisture