Forest wildfire risk mapping, performance comparison of machine learning algorithms

Abstract

Abstract Mapping wildfire risk using proper models and algorithms is one of the top execution priorities for forest managers to prevent wildfires before fires occur. This study evaluates the abilities of the Artificial Neural Network (ANN), Support Vector Machines (SVM), Random Forest (RF), Multivariate Adaptive Regression Splines (MARS) machine learning methods for the prediction and mapping of fire risk across the forests of Golestan Province, Iran. For modeling, the area was first gridded into 1 ha grids, and then pixel values ​​of influential factors were extracted and standardized based on the point shape file of grid centers. The nonparametric algorithms were implemented using 70% of fire points as training samples. The obtained forest fire risk maps were classified into three zones, including low-risk, medium-risk, and high-risk classes. The classification accuracy of the obtained risk maps was evaluated using 30% of the remained fire points. The results showed that the RF algorithm, with an overall accuracy of 75%, had the best performance in fire risk predictions compared to other algorithms. Forest managers can use this methodology to predict areas of most significant fire risk to prevent future fires through land use management, strategic decision-making, and planning. The results enable forest managers to find the best way to monitor, manage, and control fire outbreaks based on fire risk maps of forests in northeastern Iran or other regions with similar conditions.