Automatic forest fire danger rating calibration: Exploring clustering techniques for regionally customizable fire danger classification

Abstract

Effective wildfire management begins with fire prevention and the assessment of the local fire danger. Typical approaches to fire danger classification rely heavily on manual analysis for specific regions based on expert knowledge. In this paper, a novel approach is proposed for the automatic calibration of fire danger classes based on the Canadian Forest Fire Weather Index System (CFFWIS) applied for specific regions. The proposed automatic calibration method is based on clustering algorithms, namely k-means, fuzzy c-means, Gaussian mixture models, and data-clouds, which are used to identify clusters in datasets composed of elements from CFFWIS and wildfire historical records. The clusters are associated with fire danger classes which are separated by proposed thresholds based on the fire weather index values contained within each cluster. Exhaustive experiments ensured an accurate comparison of performance with the analysis of our fire danger classes against the classes defined by the European Forest Fire Information System (EFFIS), also based on the CFFWIS. These experiments consider individual information from each of the selected European regions from a total of 769 regions with available data, with validation aimed at the analysis of the large fires in a general context, and a case study for Portuguese regions.