Keywords

1 Introduction

According to Thapa (2015), from 1971 to 2015, landslides killed approximately 200 people in Nepal each year, with one of the most severe landslide events in July 1993, killing 1336 people in central Nepal. The risk of slope failure has been further exacerbated due to the extreme rainfall events caused by climate change. We studied landslide hazards owing to extreme rainfall events in Bagmati and Madhesh provinces under several climate change scenarios. We developed high-resolution 1 km × 1 km downscaled extreme rainfall projections under selected extreme Global Climate Models (GCMs) from the Coupled Model Inter-comparison Project Phase 5 (CMIP5), GCMs, with representative concentration pathways (RCP4.5 and RCP8.5) scenarios for 1976–2005 (baseline) as well as the 2030s, 2050s, and 2080s. Based on these extreme rainfall scenarios, we modelled and mapped rainfall-triggered regional landslide hazards by merging the Frequency Ratio (FR) and the Analytical Hierarchy Process (AHP) techniques. The developed approach applies Geographic Information System (GIS) as a tool along with open data to develop the regional landslide hazard models. The study aims to understand better the spatial distribution of landslide hazard levels for current climate conditions and future time horizons and the potential exposure levels of the road transport infrastructure to the landslide hazard scenarios.

Exposure to climate-related extremes and hazards is the first step in assessing the vulnerability and risk of the system. In the context of a transport system, if an asset or system is located in a region directly affected by climate change and climate-induced hazards, it is referred to as “exposure” (Filosa et al. 2017).

2 Description of the Study Area

Bagmati and Madhesh Provinces are situated in the central region of Nepal, bordering Provinces 1 to the east, India to the south, Gandaki Province to the west, and China to the north. Bagmati and Madhesh Provinces are located in between latitudes of 27° 4′ and 28° 19′ N, and longitudes of 83° 56′ and 86° 59′ E (Fig. 1). Bagmati Province has thirteen districts and encompasses a total area of 20,300 km2. Madhesh Province has eight districts and covers an area of 9661 km2.

Fig. 1
figure 1

Digital map shows the location, topography, and distribution of strategic road network (SRN) of the study area

Full size image

The road network in Nepal comprises the Strategic Road Network (SRN), which is maintained by the Department of Roads (DoR), and the Local Road Network (LRN), which is maintained by local governments (GoN/World Bank 2013). The spatial distribution of SRN is shown in Fig. 1. The total lengths of SRN roads in Bagmati and Madhesh Provinces are 2544 km and 1466 km, respectively (GoN/DoR 2018). The road density (km/100 km2) of SRN roads in Bagmati and Madhesh Provinces are 12.5 and 15.21, respectively (GoN/DoR 2018).

The study area is divided into five physiographic regions: the High Himalaya, High Mountain, Middle Mountain, Siwalik (Chure), and the Terai zones (Upreti and Dhital 1996). Land area in Bagmati Province spans from Siwalik region characterized by highly dissected sloping terrain, Middle Mountain with steep terrain, and High Mountain and High Himalayas with very steep mountainous terrain and very high prevalence of landslides. The majority of land in Madhesh Province is located in the flat Terai Region with northern areas in the Siwalik Hills, where the majority of landslides occur. Tectonically, those physiographical zones are separated by three thrust faults, viz. the Main Boundary, Main Central, and Main Frontal Thrust. The regional geology of the study area is divided into four tectonic zones, viz. Terai, Sub-Himalayan, Lesser Himalayan, and Higher Himalayan zones (Dhital 2015).

According to Dahal (2015), small and large-scale landslides occur in Nepal. The majority of small-scale landslides are discovered to develop on large-scale landslides. In many cases, the shallow rotational and translational landslides merge into the mainstream valley, resulting in devastating debris flows at nearby highways or communities.

3 Analysis Input Data

Three important datasets were compiled to map and analyze the landslide hazard viz the landslide inventory data, landslide controlling factors dataset, and landslide triggering factor (rainfall) dataset.

3.1 Landslide Inventory Map

The landslide inventory data was generated by Durham University and the National Society for Earthquake Technology of Nepal (NSET) under a project funded by the Department for International Development of the United Kingdom (DFID) - DFID/NERC SHEAR project number: 201844, in collaboration with the Durable Solutions 2 project. In this study, the data is on polygon shapefile, and we used the post-monsoon data of 2020 as landslide datasets. The landslides in the landslide inventory were randomly divided into two subsets, i.e., the training set area (80%) and the test/validation set area (20%), as suggested by Getachew and Meten (2021). The landslide size in the study area varied from 0.01 to 1.3 km2.

3.2 Landslide Controlling Factors Map

Pawluszek and Borkowski (2017) mentioned that there are no specific rules for defining the number of controlling factors sufficient for the landslide hazard analysis. With the combination of the works of literature mentioned above and the nature of the study area, nine different thematic layers together with terrain slope (°), distance from lineaments (m), distance from streams (m), elevation (m), lithology, soil, rainfall, land use, and terrain aspect were taken into consideration. The landslide controlling factor maps were rasterized with 12.5 m × 12.5 m pixels, and each raster was reclassified into appropriate thematic classes.

3.3 Rainfall Datasets and Projected Scenarios

Climate projections are widely used to understand climate extremes and their probability of occurrence in the future. Development of climate change projections, including regional and national level projections for climate extremes, could be accessed from four sources such as the Global Climate Models (GCMs), the Regional Climate Models (RCMs), from an understanding of the regional and national climate influences, and from the historical climate change recently observed (Seneviratne et al. 2012; Knutti et al. 2010; Christensen et al. 2007). As illustrated in Fig. 2, the suggested climate impact modelling methodology for identifying extreme events in the study area comprises six methodological steps.

Fig. 2
figure 2

Climate impact modeling for evaluating risks from severe landslides utilizing downscaled GCMs

Full size image

As the current study focuses on identifying future extreme occurrences to determine the maximum hazardous levels for the study area, it is essential to choose those GCMs that give extreme conditions in the targeted areas in the future. The future climate scenarios for 2030 (averaging 2016–2045), 2050 (averaging 2036–2065), and 2080 (averaging 2066–2095) are developed in the study area based on current climate (rainfall and mean temperature from 1976 to 2005) across the same study region during the wet season. Because landslides are more common during the rainy (monsoon) season, it is thought to be the best time period to choose suitable GCMs. Then a skills test is suggested for the identification of the models. The Skill Test entails calculating the change in annual mean temperature (∆T) and percent change in yearly precipitation (∆P%) from each of the models (CMIP5 models).

4 Analysis Methods

The integrated weighted index method was used in this study. The integrated weighted index method was applied by merging the analytical hierarchy process (AHP) and the Frequency Ratio (FR) method. The integration between these two methods has been peer-reviewed by Yi et al. (2019), Mondal and Maiti (2013), and El Jazouli et al. (2019). There are four important processes in the integrated weighted index method. The first step is to determine the relative importance of landslide controlling and triggering parameters using the AHP technique (Fig. 3). The second step is to describe a mutual relationship between the location of landslides and the controlling & triggering parameters by using the FR method. In the third step, landslide hazard maps were created using the weighted overlay tool in a GIS environment. In the fourth and last step, the operating characteristics curve (ROC) technique was used to validate the results of landslide hazard model predictions (Fig. 3).

Fig. 3
figure 3

Flow chart with study approaches

Full size image

4.1 Frequency Ratio Method

Frequency ratio (FR) values represent the connection between the landslide occurrence and the classes of every single controlling and triggering parameter. In terms of landslide evaluation and prediction, the frequency ratio approach is one of the most widely used statistical methods (Mandal and Mondal 2019). The FR relates the likelihood of the presence to the absence of landslide occurrence (Lee and Pradhan 2007). The landslide hazard can be evaluated from the spatial relationship between the controlling factor and landslides occurrences. The greater the FR ratio is the stronger is the link among the landslide controlling and triggering parameters and the incidence of landslides (Lee and Pradhan 2007).

The frequency ratio value is computed as described in the following formula:

$$ FRi = \frac{{Ncell \left( {Si} \right)/Ncell \left( {Ni} \right)}}{{\sum Ncell \left( {Si} \right)/\sum Ncell \left( {Ni} \right)}} $$
(1)

where, Ncell (Si) is the number of grid cells in class I that have been identified as landslides, and Ncell (Ni) is the total number of grid cells in class I in the entire area. The total number of grid cells that have been identified as landslides in the entire area is ∑Ncell (Si), while the total number of grid cells in the entire area is ∑Ncell (Ni).

4.2 Analytical Hierarchy Process (AHP) Method

The analytical hierarchy process (AHP) method was first proposed by Saaty (1982, 1991, 2000) and is widely applied in the decision-making paradigm. AHP avoids the difficulties associated with random weights and rating systems by allowing judgment to influence relative importance features or weights rather than randomly assigning those features (Yalcin 2008).

CR stands for the likelihood that the matrix judgments were created at random, as follows:

$$ CR = \frac{{{\text{CI}}}}{{{\text{RI}}}} $$
(2)

Based on the arrangement of Malczewski’s matrix (1999), RI is the average of the resultant consistency index, and CI is the consistency index and may be represented as:

$$ CI = \frac{{\left( {\lambda_{\max } - n} \right)}}{{\left( {n - 1} \right)}} $$
(3)

where λmax is the matrix’s largest value, which may be easily computed from the matrix, and n is the matrix’s order. R is a value that ranges from 0 to 1. A consistency ratio near one implies that the matrix’s rating was most likely generated randomly. A decent level of consistency is one with a consistency ratio of 0.10 or less (Malczewski 1999).

4.3 Integrated Weighted Index Method

FR is a statistical model based on data that can determine the spatial correlation between landslide locations and the controlling and triggering parameters (Yi et al. 2019). The FR technique, on the other hand, does not take into account the interdependence of controlling and triggering parameters. The AHP technique is a critical multiple-criteria decision-making method to address this problem. To a certain degree, the integrated approach retains the benefits of the FR and AHP methods while reducing their limitations.

By combining the FR and AHP techniques, the integrated weighted model examined the connection among the controlling and triggering parameters as well as the effect of each landslide controlling and triggering parameters on landslide frequency. The following formula can be used to obtain the integrated weighted index:

$$ I = \mathop \sum \limits_{i}^{m} \left( {{\text{W}}i{\text{ x FR}}i} \right) $$
(4)

where m is the number of controlling and triggering factors, Wi denotes the weight assigned to each controlling and triggering factor using the AHP method, and FRi denotes the FR value assigned to the ith controlling and triggering factor using the FR method.

4.4 Validation Method

Validation of the model predictions is critical for landslide hazard mapping. The receiver operating characteristics curve (ROC) technique (Brenning 2005) was used to assess the effectiveness of the integrated weighted index model. The ROC curve is a plot of the model prediction’s sensitivity (percent of true positives) against the complement of its specificity (proportion of false positives, Jaiswal 2011). The true positive rate (TPR) was plotted in contradiction of the false positive rate (FPR) on the ROC curve, with TPR on the y-axis and FPR on the x-axis.

Specificity is the probability that a mapping unit with no landslide is correctly classified 1-specificity is the false positive rate measured along the curve’s x-axis (Jaiswal 2011). The area under the curve (AUC) ranged between 0.5 and 1.0, where a value of 1.0 suggested that the model performed perfectly, but a value near 0.5 indicated that the model performed poorly.

4.5 Exposure Assessment Method

The exposure of roads under the SRN is assessed as the physical exposure which is spatially exposed to the landslide hazard areas. The physical exposure of the road network is assessed as “the ratio or the proportion of road section length potentially affected by landslides in the current climate scenario and future climate change scenarios to the total length of the road,” the landslide being represented by the landslide hazard and the severity represented by the hazard classes. The exposure index for a road segment is calculated as

$$ E_{i} = \frac{{{\text{L}}_{r} }}{{{\text{L}}_{i} }} \times 100 $$
(5)

and the road exposure index for entire region (province) is calculated as

$$ E = \frac{{\mathop \sum \nolimits_{i = 1}^{n} {\text{L}}_{ri} }}{{\mathop \sum \nolimits_{i = 1}^{n} {\text{L}}_{i} }} \times 100 $$
(6)

where, Ei is the exposure index of road segment i with the total length of Li and landslide exposed length of Lr. E is the exposure index of the road in the entire geographic region calculated as the proportion of the total length.

The landslide exposed length Lr for each road section is calculated using spatial overlay analysis of the existing SRN road network data (GoN/DoR 2018) with landslide hazard scenario maps of the current baseline period (1976–2005) and future climate change scenarios (2030, 2050 and 2080) in a GIS environment.

5 Results

5.1 Regional Landslide Hazard Maps

The weight for each controlling & triggering parameter was assigned using the analytic hierarchy process approach. The larger the weight is, the more significant the impact on landslide occurrences may be. The comparison quality is described by the consistency ratio (CR). Some works of literature related to landslides hazard mapping were investigated to find the appropriate correlation between the controlling and triggering factors (Mandal and Mondal 2019; Sonker et al. 2021; Yi et al. 2019).

Four factors considered to be the most important landslide controlling and triggering factors were slope, lithology, rainfall, and distance from lineaments. Thus, in this study, those four factors were given a higher weight than the other controlling factors. Some studies found no significant influence of aspect on slope stability (Ayalew and Yamagishi 2005), while some others found its effect on landslide initiation (Dai and Lee 2002). Thus, low weight was given to the aspect factor.

As mentioned above, in the AHP analysis, the consistency ratio characterizes the comparison’s reliability. In this study, the consistency ratio (CR) of the matrix of paired comparisons between the nine controlling and triggering factors in the hazard map is 0.012. A consistency ratio of 0.10 or less is considered a satisfactory degree of consistency (Malczewski 1999). Even though the consistency ratio of this study is consistent (0.012), the pair-wise comparison is subjective, and the precision is heavily reliant on the expert’s perspective. In order to reduce the subjectivity, the integrated weighted index model by combining FR and AHP was applied.

Based on the pairwise comparison matrix of AHP, the weight of slope was the highest (0.205 suggesting that slope has the greatest effect on the frequency of landslides). The weight of distance from lineaments and rainfall were the second and third highest (0.198 and 0.174, respectively), followed by lithology (0.134), soil (0.067), and distance from streams (0.061). Finally, the weights of aspect, elevation, and land use were the lowest (0.054). This suggested that these three parameters had the lowest impact on the prevalence of the landslide. The application of the integrated weighted index model demonstrated that landslide hazard levels were essentially compatible with landslide occurrence.

A total of seven maps of landslide hazard were developed from two RCP4.5 and RCP8.5 for time horizons 2030s, 2050s, and 2080s and baseline (1976–2005). The landslide hazard maps were divided into five categories: very low, low, moderate, high, and very high (Fig. 7). The total area of landslide hazard was quantified by calculating the area geometry in the GIS environment.

Based on the acquired landslide hazard maps, the area for each hazard level in scenarios shows different values. In the RCP8.5 scenario, for time horizon 2030, 2050, and 2080s, the hazardous zones, including zones with extremely high and high hazard levels, cover about 24%, 25%, and 30% of the study area, respectively.

In the case of the RCP4.5 scenario for time horizon 2030, 2050, and 2080, the very high and high hazard level zones covered approximately 22%, 23%, and 24% of the study area, respectively. The highest area value (8994 km2) of the high to very high hazard level was reached by the hazard map of RCP8.5 for time horizon 2080, where most of the high-very high hazard area is located in Bagmati province (8862 km2) and a small part of Madhesh province (132 km2, Tables 1 and 2). Concerning the hazard maps of RCP4.5 and RCP8.5 for time horizon 2030, 2050, and 2080, most parts of high to very high-level hazard zones are mainly distributed in Bagmati Province, especially in Rasuwa, Niwakot, Sindhupalohok, Dolahak, Dhading, Nuwakot, Ramechhap, the southern part of Lalitpur, north and northwestern part of Chitawan, Makawanpur, and Sindhuli district (Fig. 7).

Table 1 Total area of landslide hazards in Bagmati Provinces for current climate (baseline), future scenarios and periods
Full size table
Table 2 Total area of landslide hazards in Madhesh Provinces for current climate (baseline), future scenarios and periods
Full size table

5.2 Validation of Landslide Hazard Maps

Validation of modelled outputs is critical to evaluate the landslide hazard map results. A sample dataset of 8813 landslides was used as a validation dataset from the total (43,519 landslides sample dataset) to validate the reliability of the landslide hazard model.

The AUC values for the landslide hazard map (current climate and future projected scenarios) varied from 81 to 86%, which showed that the integrated weighted index model used in this analysis has a high degree of predictability. Figure 4 showed the ROC of the landslide hazard map for time horizon 2080s with RCP8.5.

Fig. 4
figure 4

The integrated weighted index model's forecast effectiveness. An example of ROC of landslide hazard map for time horizon 2080s with RCP8.5

Full size image

5.3 Exposure Assessment Results

Exposure assessments of SRN roads in the study provinces were done for seven landslide hazard scenarios. In Bagmati Province, the majority of roads have low (35.5%) and medium (40.9%) risk to landslides in the baseline period, with 16 and 2.5% lengths of the roads exposed to high and very high landslide hazards. The proportion of road length exposed to very low and low landslide hazards decreases in all the time horizons for RCP4.5 and RCP8.5. It consequently increases in moderate, high, and very high hazardous categories compared to the baseline years (Fig. 5). Exposure to moderate, high, and very high hazard increases significantly for the RCP8.5 scenario in 2080.

Fig. 5
figure 5

Proportion SRN Road hazardous to current and future landslide scenarios in Bagmati Province

Full size image

For Madhesh Province, the majority proportion (68%) of the road is exposed to very low landslide hazard, with 31% low and negligible 0.9% to moderate hazard in the baseline period. The very low hazard proportion of the roads decreases significantly in the future horizons, consequently with a significant increase in the proportion of road length exposed to low and moderate hazardous landslides (Figs. 6 and 7).

Fig. 6
figure 6

Proportion SRN Road hazardous to current and future landslide scenarios in Madhesh Province

Full size image
Fig. 7
figure 7

Landslide hazard and exposure maps for a baseline period (1976–2005), b time horizon 2030s for RCP4.5, c time horizon 2030s with RCP8.5, d time horizon 2050s for RCP4.5, e time horizon 2050s for RCP8.5, f time horizon 2080s for RCP4.5 and g time horizon 2080s for RCP8.5

Full size image

6 Discussion

Compared to the landslide-hazardous areas modelled for the current baseline climate conditions (1976–2005), the landslide-hazardous areas increase significantly in extreme rainfall conditions for the future climate change scenarios for both RCPs. Results revealed that the regions with very high landslide hazard increased to 312 km2 (for RCP4.5) and 961 km2 (for RCP8.5) in the 2080s. Thus, it is predicted that landslide hazards will increase in future time horizons (in the 2030s, 2050s, and 2080s). RCP8.5 scenario showed larger hazardous areas than RCP4.5 in future time horizons, as RCP8.5 models predicted heavier rainfall than RCP4.5. In future periods, the Bagmati Province and northern part of Madhesh Province are likely to have higher landslide hazard areas than the current condition. In particular, it is predicted that landslide hazards in Bagmati and the northern area of Madhesh Provinces would increase in the 2030s (2016–2045), 2050s (2036–2065), and 2080s (2066–2095) in both RCP scenarios. In both scenarios, the landslide-hazardous areas for the 2080s were predicted to expand significantly to the area adjacent to Rasuwa, Sindhupalchok, and Dolakha districts.

The hazard map for time horizon 2080 with RCP8.5 represents the worst-case scenario for landslides. Although the majority of the study area (31%) was predicted to be on moderately hazardous to future landslides, 30% was indeed in the high and very high hazardous categories. The unsafe areas mentioned in the previous section require immediate mitigation actions. Reactivation of existing landslide sites and new landslides may occur, particularly along lineaments. The findings revealed the dependability (indicated by the AUC value that showed a satisfactory result of 81–86% accuracy) and practicality of the integrated weighted index model in regional landslide hazard mapping.

In Bagmati Province, the proportion of road length hazardous to very low and low landslides decreases for RCP4.5 scenarios. Consequently, the proportion increases for moderate, high, and very high landslide hazards for 2030, 2050 and 2080 horizons. Similarly, in Madhesh Province, the proportion of road length in low hazardous areas decreases in the future horizons, consequently with a significant increase in low and moderate hazards.

7 Conclusions

The landslide hazard modelling and mapping spatially locate the regions hazardous to landslides due to multi-variate factors under the extreme rainfall conditions for the current and projected future time horizons. Seven maps of landslide hazard were created, representing scenarios for RCP4.5 and RCP8.5 for future time horizons 2030s, 2050s, and 2080s and the current (1976–2005) baseline period. The significant increase in very high landslide hazardous areas between the baseline period and the long-term horizon in 2080 (for RCP8.5 scenario) indicates that the extreme rainfall due to climate change will substantially increase landslides at Bagmati and Madhesh provinces. These landslide hazard models and maps enable decision-making in the planning of road infrastructure through relatively safer areas indicated by very low, low, and/or moderate hazardous areas. Planning, designing, and construction of road infrastructure in the moderate hazardous areas may require road design parameters to adapt and/or mitigate the impacts of landslide hazards. Moderate hazardous areas may also be managed through climate adaptive measures, including green-grey infrastructures such as the incorporation of Nature-based Solutions (NbS) along with engineering solutions. These landslide hazard maps will further support vulnerability assessment of strategic roads in the provinces and identify critical road segments for planning and implementing climate adaptation measures and activities to enhance the climate resilience of the road infrastructure in the provinces.

The exposure analysis showed that the exposure levels of the roads to landslides would significantly increase in the future due to climate change impacts. The identification and mapping of road sections exposed to higher levels of landslide hazard enable road sector authorities to prioritize adaptation actions in these stretches of the road. Further, the exposure assessment provides an opportunity to target a more detailed study to model and quantify the hazard risk to the road transport network due to specific landslides that could occur in hazardous areas.