Comparison of multiple conventional and unconventional machine learning models for landslide susceptibility mapping of Northern part of Pakistan

Landslide susceptibility study is a critically important topic throughout the globe owing to the social and financial catastrophes of landslides. The present research aims to evaluate as well as compare the proficiencies of six advanced machine learning techniques (MLTs) for mapping the landslide susceptibility of northern parts of Pakistan. The considered MLTs include Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis, Artificial Neural Network (ANN), Naive Bayes, Multivariate Adaptive Regression Spline (MARS), along with Random Forest. The present research was performed applying GIS and R programming language (an open-source software). Primarily, the landslide inventory map was formulated with the help of an overall 3251 historical landslide events obtained through a variety of data sources. All the historical landslide locations were arbitrarily split into two groups with a proportion of 70% for training plus 30% for validating purposes. In total, sixteen landslide influencing factors were considered for modeling landslide susceptibility. These factors comprise aspect, elevation, slope, lithology, fault density, land cover classification system, topographic wetness index, earthquake, sediment transport index, normalized difference vegetation index, rainfall, soil, stream power index, road density, profile curvature, and plan curvature. The receiver operating characteristic, the area under curve (AUC), and root mean square error approaches were employed to appraise, validate, and relate the performance of the practiced MLTs. The outcomes demonstrated that AUC for six MLTs vary from 88.5% for LDA to 92.3% for ANN. The results reveal that among the six practiced MLTs, ANN (AUC = 92.3%) and MARS (AUC = 91.7%) have shown outstanding performances. Policymakers can use the findings of the present research and the produced landslide susceptibility maps for devising mitigation measures to curb the damages.