Prediction of forest fire susceptibility applying machine and deep learning algorithms for conservation priorities of forest resources

Abstract

Periodic forest fires occur in sub-tropical deciduous forest of Ayodhya hill region which is extended part of Eastern Ghats Mountain Range (India). This forest provides diverse ecosystem services (mostly non timber forest products)to tribal and non-tribal people throughout the year. The main objective of this study is to demarcate the accurate forest fire susceptibility zones applying three relevant machine learning models i.e., Random Forest (RF), Multivariate Adaptive Regression Splines (MARS) and Deep Learning Neural Network (DLNN). In this research, more than 300 hundreds historical forest fire events along with 14 forest fire predictors have been considered. DLNN model has the highest efficacy (AUC = 0.925). Features of DLNN are instinctively presumed and optimally tuned for required outcomes. Applied models explicitly examined that the north western and eastern segments of Ayodhya hill are highly susceptible for forest fire due to its physical and social complexity. DLNN model investigates that total 18% areas of Ayodhya hill have come under very high susceptible zone. The forest fire incidents are taken place during March to May because of the effect of high temperature, poor moisture content and existence of high fuel woods. Both the machine and deep learning techniques have been applied first time in this dry deciduous forest belt and models indicated high accuracy and precision which is harmonized with ground reality. This study will definitely help the local government to take proper management and conservation strategies for sustainable forest resources conservation along with improvement of livelihood of forest and forest fringe dwellers.