Metaheuristic algorithms in optimizing neural network: a comparative study for forest fire susceptibility mapping in Dak Nong, Vietnam

Abstract

Meta-heuristic algorithms become common approaches in finding sufficiently good solutions for optimization problems. This study proposed and compared three novel hybrid methods, namely Biogeography-based Optimization (BBO), Gravitational Search Algorithm (GSA) and Grey Wolf Optimization (GWO) in combination with the popular Neural Network classifier for forest fire modeling. Dak Nong province was selected as a case study as it had undergone a critical drought season. One thousand three hundred and thirty-eight historic fired locations during the first several months of 2017 were chosen as dependent variables. On the other hand, topological, climatic and socio-economic data were used as independent predictor variables. For accuracy assessment, root mean square error derivable from the neural network was used as an objective function to be optimized by three proposed algorithms. The results showed that the area under Receiver Operating Characteristic curves (AUC) were in BBO (0.9515), GWO (0.9509), (0.9398) outperformed the Regular neural with backpropagation algorithm (AUC = 0.9271). Even though the differences between prediction results were small, but they were significant by using a paired t-test. It could be concluded that three hybrid models are suitable to map forest fire susceptibility in the selected study area and could be considered as alternative methods for studying forest fire in other locations.