Modeling forest fire behavior is very important for the effective control of forest fires and the setting up of necessary precautions before fires start. However, studies of forest fire behavior are complex studies that depend on many variables and usually involve large data sets. For this reason, the predictive power and speed of classical forecasting models are lower than of artificial intelligence models in cases involving big data and many variables. Moreover, classical forecasting models must satisfy certain statistical assumptions, unlike artificial intelligence methods. Thus, in this study, predictions were made of surface fire behavior, especially the rate of fire spread and the fire intensity, at the location at which fires started using two artificial intelligence methods, an artificial neural network and a decision tree. The accuracy of the developed models was fitted and tested. Finally, the classical regression model for predicting surface fire behavior was compared with the two artificial intelligence methods. The accuracy measures of the artificial intelligence models were found to be better than those of the classical model.
Forests are some of the most important natural resources in the world and play a key role in maintaining ecological balance, and forest ecosystems provide many ecological and economic services for human life. Forest fires are considered some of the most detrimental events that interrupt these services. Extreme meteorological conditions greatly increase the destructive effects of forest fires. The fires have complicated causes and are often very difficult to fight. Therefore, the prediction of fire behavior is essential for the successful management of fires, the effective planning of resources for fighting them, and the mitigation of the damage they cause (Mitsopoulos et al. 2017; Yavuz et al. 2018; Sevinc et al. 2020; Abid 2021). Various classical regression models have been developed to predict forest fire behavior (Fryer and Johnson 1988; Alexander and Cruz 2006; Sullivan 2007; Yassemi et al. 2008; Fernandes 2009; Matthews et al. 2012; Kucuk et al. 2012; Cruz et al. 2017; Kucuk et al. 2018; Bilgili et al. 2019; Alhaj-Khalaf et al. 2021; Cruz et al. 2022). Other models of forest fires based on the machine learning method of artificial neural networks (Pham et al. 2020) have predicted flame characteristics and fire spread (Chetehouna et al. 2015).
Fire detection and mapping, fire weather and climate change (Li et al. 2009; San-Miguel-Ayanz et al. 2012), fire probability and risk, fire hazard assessment, and fire behavior prediction have become very popular in recent years, driven by advances in fire sciences, digital and statistical information, the remote sensing technologies, including GIS, and the growing climate crisis (Vakalis et al 2004; Finney et al. 2011; Aricak et al. 2014; Rodrigues and Riva, 2014; Preisler et al. 2014; Goldarag et al. 2016; Lary et al. 2016; Huiling at al. 2016; Zhang et al 2018; Sivrikaya and Küçük, 2022). Probabilistic methods such as logistic regression, neural networks, and fuzzy logic regression are commonly used for forest fire studies (Jaafari et al. 2019). Traditional models for predicting fire risk and behavior include generalized linear models based on logistic, Poisson, and negative binomial distributions. However, these models cannot process multidimensional big data. Researchers have stated that artificial intelligence outperforms traditional statistical methods in solving the big data problem encountered in modeling forest fires. In addition, traditional statistical models must satisfy certain statistical assumptions, unlike artificial intelligence methods.
In this study, predictions were made of the surface forest fire behavior based on the rate of fire spread and the fire intensity using two artificial intelligence methods, an artificial neural network and a decision tree. Additionally, the prediction performances of the two methods were compared with the performance of the conventional regression model. The artificial neural network performed better than the other models in predicting both the fire intensity and the fire spread rate. The decision tree model had a considerably more successful performance than the regression model in predicting the fire intensity. However, when it came to predicting the fire spread, the regression model performed slightly better than the decision tree model. Overall, the artificial neural network was the most powerful model in predicting the fire intensity and the fire spread rate, and this demonstrated that artificial intelligence models can be used quite successfully in predicting fire behavior. To expand on this study, other studies of the prediction of the fire intensity and fire spread rate can be performed with other artificial intelligence methods and using additional or different variables. This method can also be used for Pinus brutia in Mediterranean region.