Comparative analysis of multiple conventional neural networks for landslide susceptibility mapping

In landslide susceptible mountainous regions, the precondition for avoiding and alleviating perilous dangers is the susceptibility mapping of the landslide. In northern Pakistan, landslides due to vigorous seismic zones, monsoon rainfall, extremely sheer slopes, and unfavorable geological conditions present a considerable threat to the mountain areas. This study targets and advances the research in mapping landslide susceptibility in northern Pakistan (Mansehra and Muzaffarabad districts). The central objective of the analysis is to analyze different convolutional neural network (CNN) frameworks and residual network (ResNet) that were constructed by developing distinct data representation algorithms for landslide susceptibility assessment and compare the results. This study considers sixteen landslide conditioning factors related to the incident of landslides centered on the literature review and geologic attributes of the pondered area. The marked historical landslide positions in the deliberated area were arbitrarily split into training and testing datasets, with the earlier containing 70% and the former having 30% of the total datasets. Several commonly exploited measures were used to validate the CNN architectures and ResNet by comparing them with the most prevalent machine learning (ML) and deep learning (DL) techniques. The outcomes of this study revealed that the proportions of regions having very high susceptibility in all the landslide susceptibility maps of the ResNet model and CNN models are considerably alike and less than 20%, which implies that the CNN models are significantly helpful in managing and preventing landslides as to the orthodox techniques. Moreover, the suggested CNN architectures and ResNet attained greater or similar prediction accuracy than other orthodox ML and DL techniques. The values of OA (overall accuracy) and MCC (Matthew’s correlation coefficient) of proposed CNNs and ResNet were greater than those of the optimized SVM (support vector machine) and DNN (deep neural network).