Machine-learning modelling of fire susceptibility in a forest-agriculture mosaic landscape of southern India.

The recurrent forest fires have been a serious management concern in southern Western Ghats, India. This study investigates the applicability of various geospatial data, machine learning techniques (MLTs) and spatial statistical tools to demarcate the forest fire susceptible regions of the forested landscape of the Wayanad district in the southern Western Ghats (Kerala, India). The inventory map of 279 forest fire locations (period = 2001–2018) was developed via Sentinel 2A satellite images, NASA fire archives, and field visits. The forest fire susceptibility modelling involves twelve influencing factors, such as ambient air temperature, wind speed, rainfall, relative humidity, atmospheric water vapor pressure (WVP), elevation, slope angle, topographical wetness index (TWI), slope aspect, land use/land cover (LU/LC), distance from the road and distance from the villages. Considering the varying level of performances (i.e., receiver operating characteristics-area under curve (ROC-AUC) values ranging from 0.869 to 0.924 in the testing phase) of the MLTs, viz., artificial neural network (ANN), generalized linear model (GLM), multivariate adaptive regression splines (MARS), Naïve Bayesian classifier (NBC), K-nearest neighbour (KNN), support vector machine (SVM), random forest (RF), gradient boosting machine (GBM), adaptive boosting (AdaBoost) and maximum entropy (MaxEnt), we propose a weighted approach to characterize the forest fire susceptibility of the region using the outputs of the different MLTs. The proposed method demonstrates improvement in accuracy (AUC = 0.890) for mapping the forest fire susceptibility of the region compared to the individual MLTs (AUC = 0.715 to 0.869) while validating with the recent forest fire data (i.e., 2019–2021). This study suggests that roughly one-third of the study area is highly susceptible to the occurrence of forest fires, implying the severity of the disturbance regime. The analysis also indicates the role of anthropogenic factors in the occurrence of forest fires in the region. It is expected that the demarcation and prioritization of the forest fire susceptibility zones in the region, which is a part of one of the global biodiversity hotspots, have significant implications on biodiversity conservation at a regional scale. • We compare the efficiency of ten machine learning techniques for forest fire susceptibility modelling in a tropical region • Moran's I Spatial autocorrelation, and Getis-Ord Gi* hotspot analysis help identify forest fire pattern of the region • We propose a weighted approach to integrate the forest fire susceptibility outputs of different machine learning techniques • One-third of the study area is highly susceptible to forest fires implying the severity of the fire disturbances • The proposed approach has better accuracy for mapping forest fire susceptibility zones compared to the individual techniques [ABSTRACT FROM AUTHOR]