4.2 Relationship between Landslide Causative Factors and Landslide Occurrence
LSM maps of the study area were prepared using three bivariate statistical models and were further analysed for landslide susceptibility change due to road construction. This was done by preparing two types of LSM maps. The first map, termed as Landslide Susceptibility Map Natural (LSMN), was formulated, taking into account only ten landslide causative factors excluding distance from road factor. Similarly, the second map, termed as Landslide Susceptibility Map Road (LSMR), was prepared, taking into account all the eleven landslide causative factors, including distance from the road factor, using Equation (4). The LSMN and LSMR maps were classified into five landslide susceptibility zones: very low, low, moderate, high, and very high susceptibility, as shown in Figure 5.
The FR values of each class of eleven landslide causative factors based on their correlation with the landslide occurrences are shown in Table 3. The LSMN and LSMR maps analysis indicates that drainage density, TWI, NDVI and distance from road (for LSMR only) were the critical factors that affect the study area’s landslide susceptibility. While analysing the hydrological parameters, the highest FR values were obtained for areas with very high drainage density (14.7) and very high TWI (49.7). Such areas were found to be more susceptible to landslides. Also, it was found that areas near the vicinity of roads are generally more prone to landslides. Further, the FR values of the distance from the road classes of 0-100 m (6.4) and 100-200 m (5.7) were found to be highest therefore, such areas were found to be more landslide prone. The study area was categorised into five NDVI classes: waterbodies, urban area, barren land, shrubs and grasslands, and sparse and dense vegetation. The areas closer to waterbodies had the highest FR value (7.7), followed by urban areas where human interference with natural slopes was observed.
The analysis of CF values indicated a similar trend with classes of very high drainage density (0.93), very high TWI (0.98), NDVI waterbodies (0.87) and 0-100 m (0.84) distance from road indicating the highest correlation with landslide occurrence. Further, the slope gradient directly relates to landslides occurrence as steeper slopes tend to be more unstable than flatter terrains. It was interpreted from the data that slope gradient classes, namely; steep (35°-45°) and very steep (>45°), had the highest CF values and were more prone to landslides, whereas no landslides were reported for flat (<15°) slope gradient class. Similarly, it was observed that the probability of landslides was moderate at lower elevations (400m-1000m) due to modest terrain characteristics. The highest probability of landslides was observed at high elevations (2000m-2500m). At very high elevations (2500m-3500m), the probability of landslides again decreases. This might be attributed to lesser reporting of landslides due to rugged terrain at areas with higher elevations. Regarding the geological aspects, the Middle Siwalik Group was found to have the highest CF values (0.74). This group predominantly consists of medium to coarse grained sandstone and red clay alternation, soft pebble with subordinate claystone and a locally thick prism of the conglomerate which might be attributed to its higher landslide susceptibility. During the analysis of soil classes, the highest CF was obtained for the lesser Himalayan soils of fluvial valleys (0.75), followed by Siwalik soils of side and reposed slopes (0.62). Both these soil types are described as loamy to loamy-skeletal soils, facilitating moderate to severe erosion.
The analysis of the Shanon Entropy model indicated that the highest Wij values were obtained for drainage density, TWI, NDVI and distance from road factors and highest Pij values were attributed to very high drainage density (0.58), very high TWI (0.42), NDVI waterbodies (0.57) and 0-100 m distance from roads (0.35). Hence these factors had the highest influence on landslide occurrence. Along with these, the areas with steep (35°-45°) and very steep (>45°) slopes and moderately high elevation class was found to have the highest Pij values and had a moderate influence on landslides occurrence. The geology map analysis again confirms that the Middle Siwalik Group, with the highest Pij value (0.15), was highly prone to landslides, followed by the Dharmasala Group. Similarly, soil classes analysis confirms that the fluvial valley soils of lesser Himalayas with the highest Pij value (0.70) were highly prone to landslides because of their shallow depth and excessive drainage characteristics. All other landslide causative factors such as aspect, curvature, lineament density etc. and their classes with the highest FR, CF and SE values had low to moderate influence on landslides occurrence.

4.3 Accuracy Assessment and Validation of Models

In this study, the LSMN and LSMR maps prepared using FR, CF and SE models were evaluated for accuracy of prediction and validation using ROC curves and AUC technique. These are well known techniques to determine the quality of a statistical model by plotting the fraction of true positives values out of total positives values and false positives values out of total negatives values by determining Sensitivity and Specificity (Devkota et al., 2013; Nohani et al., 2019). Each model’s prediction rate curve was plotted using training data set of 1199 landslides (70%), and the validation rate curve of each model was plotted using a validation data set of 524 landslides (30%). The relative ROC curves of the three models are shown in Figure 6. Based on the ROC results and AUC evaluation, all three models offer the satisfactory prediction and validation accuracy. However, the Shanon Entropy model was found to have the most accurate prediction and validation for landslide susceptibility mapping of LSMN and LSMR maps.