Manifestation of SVM-Based Rectified Linear Unit (ReLU) Kernel Function in Landslide Modelling

Support vector machines (SVM) are the most popular machine learning methods currently used to model landslides. To model the complex nature of landslides, the SVM model parameters (kernel function, penalty parameter) should be fine-tuned to achieve the best possible accuracy. The main objective of this paper is to evaluate the commonly used rectified linear unit (ReLU) activation function in deep learning for the SVM model as a kernel function. A case study of the Cameron Highlands, located in the Peninsular Malaysia, was selected and a dataset was acquired through the airborne LiDAR system, topographical databases, and SPOT satellite images. The SVM modelling with ReLU kernel was implemented in a Matlab environment. Overall, 11 landslide factors and 81 landslide locations (inventory map) were used. Experimental results showed that the ReLU kernel function could achieve a higher accuracy (0.81) than other kernels when using a lower number of landslide factors. It was found that the ReLU kernel function is more accurate (0.73) than RBF kernel (0.71) when using very limited factors (such as altitude, slope, and curvature). The kernel ReLU could improve the performance of landslide susceptibility modelling with SVM while reducing the need to use large datasets.