Extreme Fire Severity Classification using Clustering and Decision Tree
Nov 2022
With climate change, large, unpredictable, and difficult to suppress forest fires are increasingly frequent. To increase
the ability to anticipate and respond to these extreme events it is necessary to characterize the meteorological conditions
associated with the risk levels of these events. The main objective of this work is to automatically identify those severity
conditions and extract classification rules to characterize extreme forest fires with at least 100ha of burned area (90%
percentile) in mainland Portugal for the period 2001-2020.
The conditions characterizing the extreme fires are elicited by applying fuzzy clustering and predictive methods to forest
fire data and corresponding fire risk indices, namely the Canadian Forest Fire Risk Index (FWI), and subindices, as well
as the Continuous Haines Index (CHI), provided by the Portuguese Institute of Sea and Atmosphere (IPMA). The dates
and localization of fires are obtained from the shapefiles provided by the Portuguese Institute for Nature Conservation
and Forests (ICNF), and complemented with data from the MODIS Global Burned Area Product MCD64A1 downloaded
from the University of Maryland repository.
The popular fuzzy c-means (FCM) algorithm is applied to group fires into five and seven clusters, with no pre-specified
ground-truth severity. Then each cluster is labelled with the fire risk scale class assigned to the cluster’s prototype
considering the EEFIS scale (European-Forest-Fire Information System) for five clusters and IPMA fire risk scale for
seven clusters, respectively. Fuzzy Sammon mapping has been used to visualize and validate the fuzzy partitions.
Using the data from 2001-2018, decision trees (DT) were induced in order to obtain the conditions and thresholds that
characterize the obtained clusters, and tested with the data from 2019 and 2020. To ensure the quality of the classification
results robust validation techniques such as cross-validation and bootstrapping as well as evaluation metrics are applied.
The DT rules described by conjunctions of the fire risk indices and thresholds, were not always in agreement with the
reference forest fire risk prediction scales, revealing the importance of adapting the indices values according to the
region in question and taking into account several factors (forest fire risk indices) in the analysis of the conditions
associated with the level of risk of an extreme forest fire. The proposed approach shown to be a proof of concept to
derive an empirical fire severity risk scale for the collection of used indices and to compare the results with the two fire
risk scales used by IPMA and EEFIS
