Prediction ability of machine learning algorithms in Himalaya region of Pakistan for landslide susceptibility mapping

Machine learning has proven most effective in mapping landslide susceptibility. We carry out experiments with two machine learning algorithms, SVM and MaxENT to study their effectiveness for some mountaneous areas in Pakistan. A data set of 112 historic landslides are used in the study with 70% of the landslides are used for training and the rest for validation. 15 landslide casuative factors are used initially and ineffective ones are eliminated based on information Gain Ratio and Multicollinearity test techniques.  The perfromances of the landslides susceptibility maps generated are assessed using receiver operating curves (ROC), confusion matrix (CM) (Kappa, root mean square error, mean absolute error and balanced accuracy), landslide density (LD), R-index and Pearson’s Chi-squared tests. The result show that both of the models work well in this area. However, the lowest significant value ‘p’ (<0.05) during Chi-square test, showed that both the landslide models have statistical significant difference.