Assessing fire risk homogeneity and patterns with GIS and machine learning: A study from the Czech Republic

Abstract

Forest fires have historically been a concern in fire-prone regions, where climatic and environmental conditions naturally foster fire occurrences. However, in Central Europe, the increasing frequency of wildfires is driven by a complex interplay of factors, reflecting both environmental changes and human influences. This study uses machine learning to address key challenges in understanding fire risk in low-risk regions, specifically: the limited applicability of models developed for high-risk areas and the struggle to incorporate dynamic spatiotemporal factors. By analyzing eight years of fire data alongside variables such as weather, geography, and population density, our objective is to develop a machine learning and GIS-based procedure for identifying and understanding the underlying complexity of fire risk in Czechia. K-Medoids clustering revealed challenges in forming distinct clusters, suggesting the need for additional or more refined predictors. Classification models (e.g. Random Forest, Gradient Boosting, k-Nearest Neighbors) were used to predict periods of increased fire risk. Soil temperature and population density emerged as the most significant predictors. This study proposed two procedures: one that combines clustering and bivariate analysis to identify clusters with homogeneous factors, and another as a classification-bivariate visualization framework for assessing fire risk. The results indicate that classification performance can serve as a proxy for spatial complexity and highlight regions where multiple factors drive fire occurrence. To enhance model accuracy and applicability, incorporating high-resolution weather data, population mobility patterns, vegetation types, and advanced modeling techniques is recommended. Introducing advanced predictors could help differentiate between fires driven by natural factors and those influenced by human activities.