Forest fire susceptibility prediction using machine learning models with resampling algorithms, Northern part of Eastern Ghat Mountain range (India)

Periodic forest fires destruct to biodiversity, ecosystem productivity and multiple ecosystem services. Forest fires are currently turning a leading cause of forest degradation. The principal objective of this research is to predict forest fire vulnerable zones over Similipal biosphere reserve (SBR; Odisha) using different machine learning (ML) models, such as support vector machine (SVM), random forest (RF) and multivariate adaptive regression splines (MARS). Different resampling methods (CV and bootstrap) have also been applied for optimizing the result and better accuracy. Results show that 10-fold cross validation (CV) technique performed best on SVM model (AUC = 0.83) whereas bootstrap performed best on RF (AUC = 0.80) and MARS model (AUC= 0.84). The main advantage of MARS model is that it only uses input variable and significantly increases the performance of the model. The novelty of this research is application of various ML algorithms through resampling techniques to reduce the biasness and improves the reliability of the models.