1. Landslide susceptibility mapping along earthquake affected road-section in Kashmir Himalaya 2. Thirteen causative factors involved in Prediction of landslide susceptibility using six data-driven models 3. Machine learning models outperformed heuristic and statistical models 4. Distribution of predicted landslides was common in the proximity of road/faults and deeply dissected valley slopes

The interaction of seismic events with geo-environmental conditions and anthropogenic activities may exacerbate the risk of landslide hazard in a mountainous region. As an example of this, 2005 Kashmir earthquake triggered a large number of shallow to deep slope failures, which was further intensified in following years by human activities notably along road networks, posing a long-term hazard. Hence, this study was planned to evaluate the effectiveness of landslide susceptibility prediction along earthquake affected road-section of Neelum Highway using six different data-driven models. We applied analytical hierarchy process as heuristic approach, weight of evidence and index of entropy as statistical models and multi-layer perceptron, support vector machine and binary logistic regression (BLR) as machine learning models. Initially, 224 landslides locations were marked through field surveys to prepare landslide inventory which was further randomly divided into training (70%) and testing (30%) datasets. Then, 13 landslide causative factors (LCFs) were extracted from geo-spatial database and analysed by measuring collinearity among factors and assessing their contribution in landslide occurrence using different feature selection methods for inclusion in susceptibility modelling. Thereafter, six employed models were trained to produced landslide susceptibility maps of investigated road-section. Finally, the area under receiver operating characteristics (AU-ROC) curve and various statistical measures were applied to validate and compare the performance of modeled landslide susceptibility. The results revealed that no collinearity issue exists among all 13 LCFs, and all six models exhibited satisfying performance in predicting landslide susceptibility of study area. However, BLR model have produced most promising and optimum results as compared to other models with AU-ROC (0.881), Matthew’s correlation coefficient (0.609), Kappa coefficient (0.604), accuracy (0.797) and F-score (0.787). The outcomes of this study can be used as pertinent guide for preventing and managing the landslide disaster risk along Neelum Highway and beyond.

No abstract