A comparative Bayesian optimization-based machine learning and artificial neural networks approach for burned area prediction in forest fires: an application in Turkey

Abstract

This study presents a prediction methodology to assist in designing an effective resource planning for wildfire fighting. The presented methodology uses artificial neural networks, bagging and boosting in ensemble learning algorithms, and traditional machine learning algorithms decision tree regression, Gaussian process regression and support vector regression to determine the size of the area to be burned in a forest fire that will start. The Bayesian optimization algorithm, which is used in the learning process of the methods, provides the optimum hyperparameter values of the methods to obtain the minimum error value. The methodology, which is first used to predict the size of the fires that occurred in different regions of Turkey between 2015 and 2019, yielded successful results. Second, it is applied to the Montesinho Natural Park forest fire dataset in Portugal to validate its robustness in different geographical regions. Finally, the results are compared with different studies in the literature. Compared with the literature, it is seen that the presented methodology has high accuracy and high speed in the prediction of the burned area. The results of the study are significant as the proposed methodology provides valuable information to the authorized units regarding resource planning during the forest fire response phase. Furthermore, the findings show that the presented methodology is reliable and can be used as an additional tool to predict the burned area for different countries.