Comparative analysis and landslide susceptibility mapping of Hunza and Nagar Districts, Pakistan

Landslides are the most common and catastrophic activities in the mountainous topography which are responsible for extensive economic and human losses. A regional-scale area susceptible to landslides, located in Hunza and Nagar Districts in the Northern part of Pakistan, was selected for landslide susceptibility mapping (LSM). The objective of the work aims to evaluate the reliability of the models by a comparative analysis of Bi-variate statistical and deep machine learning models (DMLTs). The statistical models such as “weight of evidence (WofE), frequency ratio (FR), information valve (IV), Shannon entropy (SE) and machine learning technique (MLTs), deep machine neural network (DM-NN), and random forest (RF)” were applied for landslide susceptibility mapping (LSM). The major scope of the work was to prepare a reliable landslide susceptibility map by the comparative analysis of six aforementioned statistical and deep machine learning models. It also seeks to analyze the influence of geo-environmental landslide conditional factors contributing to landslides. Information about landslide inventory and 12 pre-defined geo-environmental landslide causative factors were selected for LSM of the study area. The locations of 148 landslides were identified and mapped from detail field survey using Garmin GPS. The results were validated using area under cure (AUC) for prediction accuracy and seed cell area index (SCAI) tests were applied for the classification ability of LSM models. The results revealed that the prediction accuracy of the models for WofE, FR, SE, IV, DM-NN, and RF are 83.70%, 82.26%, 75%, 70.7%, 80.5%, and 80.6% respectively. From the differential value of seed cell area index the (SCAI) represented by ( D _value), the results shown for the LSM are (IV=10.82) (WofE=12.99) (FR=8.47) (SE=5.32) (RF=11.39) (DM-NN=12.19). The results revealed that WofE and DM-NN have similar accuracy and far better classification ability than other models. In terms of the prediction accuracy rate, WofE model has the highest prediction accuracy and the best classification ability. As such, the landslide susceptibility map produced from WofE is proposed to be more useful for this study area. The susceptibility map of this area can be useful for land use planning and engineering purposes.