5. Discussion

The process of generation of LSM maps is complex and requires multistep in depth analysis. This study analyses three main issues: (a) the mapping of landslide susceptibility of Mandi district based on relevant landslide causative factors (b) the comparison of three statistical models, namely Frequency Ratio (FR), Certainty Factor (CF) and Shanon Entropy (SE) for their accuracies in predicting landslides and (c) the assessment of landslide susceptibility change due to road construction in Mandi district.
An appropriate selection of landslide causative factors requires extensive knowledge of geographical and topographical aspects of the study area and triggering mechanisms associated with them (Guzzetti et al., 1999; Costanzo et al., 2012). Generally, the optimal approach is manual selection based on expert opinion, but there are no universal guidelines for identification and selection of landslide causative factors (Dou et al., 2015). Based on this criterion, eleven landslide causative factors were identified, all of whom had a strong association with landslide occurrence. Further, the interdependence of these factors was investigated using a multicollinearity test, which indicated that the selected factors were independent and credible. Analysis of the three statistical models results revealed that drainage density, distance from the road, TWI and NDVI were the most influential factors for landslide occurrence. At the same time, the slope curvature and aspect were the least influential factors on landslide occurrence. The rest of the parameters, such as elevation, lineament density, geology and soil, had a moderate impact on landslide occurrence. A high density of drainage networks and steeper slopes, and sedimentary rocks like medium to coarse grained sandstone and conglomerate can be attributed as principle factors of landslide occurrence in the study area. These factors, especially when combined with excessively drained soils, high lineament density and improper road construction activities, tend to increase the study area’s landslide susceptibility. Further, the NDVI map analysis suggested that the areas near waterbodies and the areas interfered with by human settlements, road construction or any other infrastructure development activities tend to be more prone to landslides. On the contrary, regions having very high elevations, a higher percentage of vegetation and flatter slope gradients have minimum susceptibility to landslides. These results conform with the findings of similar research reports. (Dou et al., 2015; Hong and Bui, 2015; Roy, 2019).
Statistical modelling is an essential component in determining the landslide susceptibility of an area. The accuracy of statistical models is primarily dependent upon the data quality and model structure. In this study, three statistical models, namely: FR, CF and SE, were used to determine the Mandi district’s landslide susceptibility. The validation of these models was done using ROC curves and the AUC technique, assuming that landslides were dependent only on the given spatial parameters with rainfall as the common triggering factor. The results indicate that all three models have satisfactory values for prediction and accuracy. Still, the relative contribution of the landslide causative factors varied with the models, as shown in Figure 6. The highest accuracy of prediction and validation was demonstrated by Shanon Entropy (83-86%). The SE model is an entropy based data driven model which directly stores information of variables and correlates with the probability of landslide occurrence. This might be the reason for its higher accuracy in comparison to the other two models. The SE model indicated that drainage density and distance from roads as the two major contributing factors towards landslide susceptibility. The SE model also suggested a strong correlation between higher slope gradient and TWI and landslide occurrence. Some other factors like NDVI, soil, geology and elevation also indicated significant contribution. The geology Dharmasala Group, Dagshai and Kasauli Formations combined with soils of fluvial valleys at moderate to high elevations showed the highest correlation with landslide occurrence. FR model being on the observation model had (75-79%) accuracy of prediction and validation. As a rule-based model, the CF model had a relatively good accuracy of predicting and validating (75-82%). These models also suggested drainage density and road construction as two fundamental factors with maximum impact on landslide susceptibility in the study area. The comparison of these models was found to be in accordance with recent studies of landslide susceptibility analysis. (Devkota et al., 2013; Lee and Pradhan, 2007; Nohani et.al, 2019; Wang et al., 2015).
In this study, the LSMN and LSMR maps were prepared using FR, CF and SE models. The results of all the models indicated that the road construction activities in the Mandi district appear to be a primary factor responsible for an increase in landslide susceptibility of the study area. For comparison, ten common landslide causative factors were considered for preparing two susceptibility maps. The additional factor of the distance from the road was only considered for the LSMR map. These susceptibility maps are further classified into five zones of susceptibility: very low, low, moderate, high and very high susceptibility. The analysis of change in susceptibility was done by comparing each class’s percentages in both LSMN and LSMR maps, as shown in Table 4. The analysis of LSMN and LSMR maps of FR model indicates that the percentage of area in high susceptibility zone increases from 22.9% in LSMN map to 24.5% in LSMR map and the percentage of area in very high susceptibility zone increases from 13.2% in LSMN map to 15.8% in LSMR map.
Similarly, for the CF model, the high susceptibility zone area increases from 23% in the LSMN map to 24% in the LSMR map. The area is very high susceptibility zone increases from 7% in LSMN map to 8% in LSMR map. Likewise, in the SE model having the highest prediction accuracy, it was observed that the percentage of area in high susceptibility zone increases from 19.3% in LSMN map to 21.6% in LSMR map, and the percentage of area in very high susceptibility zone increases from 10.8% in LSMNmap to 12.5% in LSMR map. It can be observed from the LSMR maps in Figure 5. that the areas in the vicinity of the road, particularly in classes (0-100m) and (100-200m), witnessed an increase in landslide susceptibility. This can be attributed to the fact that the cutting and tempering of natural bed slope for road construction increases the risk of slope failure in that area. Additionally, road construction may change or block the natural drainage networks operating in mountainous terrains. This might further increase the probability of landslide occurrence in that area.