Keywords

1 Introduction

According to the global landslide dataset published by Froude and Petley (2018), 75% of all landslide events between 2004 to 2016 occurred in Asian countries. Notably, the highest number of non-seismically triggered landslides, i.e. more than 600 records (16% of the whole catalogue), was recorded in India. Most of these landslides were triggered by rainfall, particularly during the monsoon seasons. According to the EM-DAT disaster database (Guha-Sapir et al. 2022), 4739 people lost their life due to landslides in India and more than 400,000 people have been affected by this hazard since 1950. A significant upward trend in the occurrence of fatal landslides was observed in India in the period 1995–2014 (Haque et al. 2019). Due to the significant toll in terms of human lives and damages, the adoption of reliable tools to predict the occurrence of rainfall-induced landslides in India is urgent.

The LANDSLIP (LANDSLIde Multi-Hazard Risk Assessment, Preparedness and Early Warning in South Asia: Integrating Meteorology, Landscape and Society) project (www.landslip.org) aimed at reducing the impact of hydrologically related landslide multi-hazards and increasing the resilience to landslides in vulnerable, hazard-prone areas in India. One of the operational goals of the LANDSLIP project was the development of a landslide early warning system (LEWS) to forecast the occurrence of landslides induced by rainfall in two pilot areas. Recently, several regional LEWSs have been designed and implemented in various countries (Piciullo et al. 2018; Guzzetti et al. 2020), and many of them rely on rainfall thresholds, often combined with a susceptibility zonation. Rainfall thresholds are the most widely used tool to define critical probability levels for the possible occurrence of landslides in wide areas. Empirical rainfall thresholds are derived from the statistical analysis of rainfall conditions that have resulted in past landslides (e.g. Guzzetti et al. 2007, 2008; Segoni et al. 2018). Establishing reliable thresholds in large areas requires the availability of a large amount of information on the occurrence of past landslides (a statistically significant sample size) and of rainfall data.

To date, only a few works have studied the relationship between rainfall and landslides in some areas or states in India, and mostly in the Indian Himalayan Region (Dikshit et al. 2020b). Analyses of rainfall conditions responsible for slope failures have been conducted for single phenomena (e.g. Sengupta et al. 2010) and for population of landslides (e.g. Sajinkumar et al. 2020) and using both rain gauges and satellite data. Examples of rainfall thresholds defined using ground-based rainfall data are found in Uttarakhand (Kanungo and Sharma 2014), Sikkim (Sengupta et al. 2010; Harilal et al. 2019), West Bengal, particularly in Darjeeling (Dikshit and Satyam 2019; Dikshit et al. 2020a) and Kalimpong (Dikshit and Satyam 2018; Teja et al. 2019; Abraham et al. 2020), Tamil Nadu (Jaiswal and van Westen 2009; Thennavan et al. 2020), and Kerala (Naidu et al. 2018; Abraham et al. 2019, 2021). Some authors have calculated rainfall thresholds over large areas in India using satellite rainfall products: e.g. in Uttarakhand, northern India, using TRMM precipitation estimates (Mathew et al. 2014) and over the Western Ghats, south-western India, using TMPA and IMERG-derived rainfall data (Thakur et al. 2020). Brunetti et al. (2021) defined rainfall thresholds for the whole Indian Peninsula using GPM and SM2RAIN-ASCAT satellite rainfall products, and their merging, at daily and hourly temporal resolution, and daily rain gauge measurements from Indian Meteorological Department. Finally, a few works have analyzed landslides affecting highways and other infrastructures (Jaiswal and van Westen 2009; Sengupta et al. 2010; Kanungo and Sharma 2014).

In this work, we describe data, procedures, analysis, and challenges faced in calculating empirical rainfall thresholds for the possible initiation of shallow landslides in two pilot areas in India, using a consolidated approach.

2 Study Area

The Indian Peninsula is bounded on the north by the Himalayas and includes the Western Ghats, a N-S mountain chain on the western edge, and the Eastern Ghats, running mostly in a NE–SW direction. The steep topography and geological complexity produce high landslide activity, especially during the monsoon season, which dominates the climate for over one-third of the year. Neal et al. (2020) identified seven broad-scale weather regimes in India. The summer monsoon usually happens between May and September, being mostly active during July and August.

Two pilot areas were selected in the Indian territory: the Darjeeling and the Nilgiris districts (Fig. 1). Both areas have population densities up to over 600 inhabitants per km2 and have experienced rainfall-induced landslides that caused fatalities and heavy socio-economic impacts. Both areas are subject to intense rainfall in the monsoon period, which affects the two regions for several months triggering numerous landslides.

Fig. 1
figure 1

Location of the two pilot areas of Darjeeling and Nilgiris in India, with indication of the boundaries of the West Bengal and Tamil Nadu states

Full size image

The Darjeeling District is located in the Eastern Himalayas, NE India, and belongs to the state of West Bengal. It covers 3232 km2 with elevations ranging 100–4500 m a.s.l. and slope inclinations ranging 0–60° (average≈18°). The population of this district is approximately 1.9 million people. The area is underlain by a sequence of metamorphic rocks from the Himalayan Fold Thrust Belt (FTB). Landslide failure mechanisms encountered within this pilot area are controlled by FTB deformation of the rocks. The average annual rainfall varies between 3000 and 5000 mm∙yr−1, of which approximately 90% is within the active monsoon period (June to September).

The Nilgiris District, located within the state of Tamil Nadu, in the southern part of the Indian Peninsula, covers 2566 km2 and hosts approximately 0.8 million persons. It includes the Nilgiris hills (1900 km2), a plateau with elevation reaching 2400 m a.s.l. that is one of the most landslide-prone regions in the state. The plateau includes high peaks and rocky escarpments composed by gneiss overlain by lateritic soils. The Nilgiris is impacted by the northeast monsoon (October through December), which accounts for nearly all of the annual precipitation in the area and is responsible for the triggering of most landslides (predominantly debris/earth slides or flows).

3 Data

We used information on the occurrence of 684 landslides in the Darjeeling district from October 1968 to July 2015 and of 392 landslides in the Nilgiris district from December 1987 to January 2017. Rainfall measurements were obtained from regional rain gauge networks in the two pilot areas. In particular, we used daily rainfall data recorded by 36 rain gauges located in Darjeeling district in the period 01/Jan/1959–31/Dec/2017 and by 24 stations from 01/Jan/1987 to 31/Jul/2017 in Nilgiris district (Fig. 2).

Fig. 2
figure 2

Location of rain gauges and landslides selected for threshold calculation in the a Darjeeling and b Nilgiris pilot areas

Full size image

Rain gauge selection was based on the number of time series available, and on data quality and sensor location, being these characteristics crucial to model the spatio-temporal variation of precipitations.

Using the entire rainfall data set, we estimated the distribution of monthly rainfall in the areas. In Darjeeling (Fig. 3a) the monthly rainfall has a peak in July and high values from June to September. As a result, all the 684 landslides are concentrated in the monsoon period, May through October, with nearly half of the events (327 out of 684, 48%) recorded in July (Fig. 3b). Hence, the monthly distributions of rainfall and landslides overlap, except in June and August, when the number of recorded landslides is very low if compared to the less rainy May. In Nilgiris, the distribution of monthly rainfall shows two peaks, with a significant increase, in October and November (Fig. 3c). The 392 landslides are heterogeneously distributed during the year, with more than 80% of the events (326 out of 392) between October and December (Fig. 3d), closely replicating the rainfall distribution.

Fig. 3
figure 3

Top: annual distribution of monthly rainfall in a Darjeeling (1959–2017) and c Nilgiris (1987–2017). Bottom: monthly distribution of landslides in b Darjeeling and d Nilgiris; lighter bars indicate the total numbers of catalogued landslides in each pilot area; darker bars indicate the landslides used to calculate the thresholds (numbers in brackets)

Full size image

4 Method

For the reconstruction of the rainfall conditions responsible for landslide initiation and for the calculation of rainfall thresholds, we adopted the procedure and the automatic tool named CTRL-T (Calculation of Thresholds for Rainfall-induced Landslides—Tool) described in Melillo et al. (2018) and successfully used in different contexts (e.g. Gariano et al. 2019; Teja et al. 2019; Jordanova et al. 2020; Melillo et al. 2020; Abraham et al. 2021).

CTRL-T exploits continuous rainfall measurements, and landslide information, to (1) reconstruct rainfall events; (2) select automatically the representative rain gauges; (3) identify multiple rainfall conditions responsible for the failure; (4) attribute a probability to each rainfall condition; and (5) calculate probabilistic rainfall thresholds and their associated uncertainties. Input data consists of: (i) rainfall event setting parameters, (ii) rainfall data, (iii) rain gauge locations, (iv) landslide locations, and (v) landslide occurrence times.

The tool is divided into three main logical blocks. The first block executes the reconstruction of rainfall events. A rainfall event is a period of continuous rainfall separated from consecutive events by a dry (no-rain) period of two days. The second block combines rainfall events and landslide information, selects the rainfall events that have resulted in landslides and determines the rainfall duration, D, and the cumulated event rainfall, E, responsible for the landslides. Specifically, rainfall conditions were reconstructed from rainfall measurements recorded by rain gauges located within a 12-km radius circular buffer centered on the landslide. To select the representative rain gauges and the rainfall condition likely responsible for each failure, a weight function was used that considers D, E, and the distance between the rain gauge and the landslide. In the case of two or more landslides occurring in the same day and associated with the same rainfall condition, we considered the first one in order to avoid duplicates that would invalidate the frequentist method adopted to define the thresholds. The third block calculates rainfall thresholds at different non-exceedance probabilities adopting a frequentist statistical method (Brunetti et al. 2010; Peruccacci et al. 2012). The approach assumes the threshold curve is a power law that links the cumulated event rainfall E (in mm) to the rainfall duration D (in days);

$$ E = \left( {\alpha \pm \Delta \alpha } \right) \cdot D^{{\left( {\gamma \, \pm \,\Delta \gamma } \right)}} $$
(1)

where, α is a scaling constant (the intercept), γ is the shape parameter (that defines the slope of the power law curve), and Δα and Δγ represent the uncertainties of α and γ, respectively. The uncertainties Δα and Δγ are calculated using a “bootstrap” non-parametric statistical technique (Peruccacci et al. 2012). The uncertainties associated with the thresholds depend on the number and the distribution of the empirical data points, and reduce as the number of the empirical data increases.

For the reconstruction of the rainfall conditions likely responsible for the failures, CTRL-T outputs were checked and eventually revised by expert investigators from CNR IRPI and GSI to consider (i) the difference in local rainfall regimes at the two pilot areas (presence of monsoons), (ii) the coarse (daily) temporal resolution of rainfall measurements and (iii) the possible inaccuracy of the landslide occurrence date. For a few uncertain events, multiple investigators analyzed all available landslide and rainfall information and agreed to revise the rainfall conditions automatically reconstructed by CTRL-T. We acknowledge that this procedure has introduced a hint of subjectivity into the reconstruction of the rainfall events responsible for failures, but we maintain that this additional check was necessary to reduce the uncertainties listed above.

5 Results and Discussion

5.1 Rainfall Conditions

Applying the above-mentioned procedure, we reduced the number of rainfall conditions associated with landslides in the two areas to 96 and 114, respectively. In addition, in the Darjeeling pilot area, the subset was further reduced to 84 landslides by discarding 12 more records for which a reliable triggering rainfall condition could not have been associated. These two landslide subsets have the same monthly distribution as the initial ones.

In the Darjeeling pilot area, the 84 rainfall conditions associated with the landslides have durations ranging between one and eight days (24–192 h) and cumulated rainfall varying between 74.0 and 10,775.0 mm (average value = 320.5 mm). In the Nilgiris pilot area, the rainfall conditions have durations ranging between one and six days (24–144 h) and cumulated rainfall varying between 45.0 and 675.5 mm (average value = 260.3 mm).

5.2 Magnitude Analysis

For civil protection purposes, it is worth mentioning that the number of landslides triggered by each rainfall condition can be assumed as a proxy for the magnitude of the event. For the purpose, we divided the ensemble of rainfall conditions responsible for the failures in the two pilot areas into two classes: those that triggered (i) from one to ten, and (ii) more than ten landslides.

We analyzed separately the D and E distributions of the rainfall conditions associated with landslides for the two magnitude classes. Figure 4 shows the box-and-whisker plots of D and E for each class for the two pilot areas (in shades of blue and green, respectively). For Darjeeling, the number of landslides associated to a single rainfall condition varies from one to 182 (event occurred on 26th May 2009). For Nilgiris, the number of landslides associated with a single rainfall event ranges from one to 138 (event occurred on 14th November 2006). Overall, median and maximum values of D remain stable for the two classes, whereas median values of E increase with the magnitude for both pilot areas.

Fig. 4
figure 4

Box-and-whisker plots of the duration, D (top) and cumulated event rainfall E (bottom) of the 84 and 114 rainfall conditions used to calculate rainfall thresholds in Darjeeling (in shades of blue) and Nilgiris (in shades of green) pilot areas, respectively, categorized into two classes of magnitude: (i) 1–10, and (ii) >10 landslides

Full size image

5.3 Rainfall Thresholds

Using the 84 and 114 rainfall conditions, we defined frequentist rainfall thresholds for the two areas. Table 1 lists the parameter values and uncertainties for thresholds at different non-exceedance probability levels. Figure 5 shows, in a log–log plane, the rainfall conditions responsible for the failures and the rainfall thresholds at 5% non-exceedance probability levels with the associated uncertainties for Darjeeling (Fig. 5a) and Nilgiris (Fig. 5b).

Table 1 Parameters of the ED rainfall thresholds at different non-exceedance for the possible initiation of landslides in Darjeeling (T1,Dar, T5,Dar, T10,Dar) and Nilgiris (T1,Nil, T5,Nil, T10,Nil) pilot areas
Full size table
Fig. 5
figure 5

DE conditions that have resulted in landslides in a Darjeeling and b Nilgiris, and corresponding threshold at 5% non-exceedance probability level. Shaded area shows uncertainty of the thresholds. Data shown in logarithmic coordinates

Full size image

The thresholds have ranges of validity equal to the range of rainfall durations D, and are defined at discrete (daily) rainfall durations D. Therefore, thresholds can be used only for the prediction of landslides for D equal to one day or multiples.

Figure 6 portrays the comparison of T5,Dar (blue curve) and T5,Nil (green curve) in logarithmic coordinates, with the shaded areas showing the uncertainty associated with the thresholds. Inspection of the figure reveals that T5,Dar is higher than T5,Nil, meaning that, regardless of the rainfall duration, a larger amount of rainfall is necessary to trigger landslides in the Darjeeling pilot area. Further inspection of the figure reveals that T5,Nil is slightly steeper than T5,Dar. This result agrees with the different rainfall regimes of the two pilot areas.

Fig. 6
figure 6

Comparison between the 5% ED thresholds for Darjeeling (T5,Dar) and for Nilgiris (T5,Nil) pilot areas. Data shown in logarithmic coordinates

Full size image

The relative uncertainties in the γ parameter, Δγ/γ, are equal to 14.5% and 11.5% in Darjeeling and Nilgiris, respectively. Moreover, considering the 5% non-exceedance probability level, the relative uncertainties in the α parameter, Δα/α, are equal to 8.9% and 4.9%, respectively. The two values are lower than 10%, considered an acceptable value for reliable thresholds, as proposed by Peruccacci et al. (2012).

The thresholds defined for the two pilot areas have high values of the parameter γ, as a result of the distribution of the empirical data points in the DE plane. The uncertainties associated with thresholds are high and depend primarily on the number and distribution of the (D, E) pairs.

The daily temporal resolution of landslide and rainfall data affects the calculation of thresholds for two main reasons. First, as mentioned above, thresholds are valid and can be used only for durations ranging 1–8 days and 1–6 days in Darjeeling and Nilgiris, respectively. Second, the daily rainfall resolution results in a large dispersion of E values, especially for the shortest events (lasting one or two days). Presumably, the cumulated rainfall recorded at one day (especially the lower cumulated values) fell over a shorter time period. According to the data point distribution, thresholds are very steep (high values of γ parameter) and the associated uncertainties are high. Marra (2019) and Gariano et al. (2020) noted that the use of daily data leads to a worse threshold predicting capability than using hourly data.

6 Conclusions

An effective landslide forecasting model is a key component of a LEWS. In this work, we described the procedures and the challenges for calculating frequentist rainfall thresholds for the possible occurrence of shallow landslides in two Indian pilot areas. Thresholds defined for Darjeeling and Nilgiris areas were based on 84 and 114 landslide-triggering rainfall conditions, respectively. Despite some limitations, mostly due to the daily temporal resolution of rainfall data and to the spatial and temporal distribution of the reported landslides, the thresholds defined for the two pilot areas have acceptable values of uncertainties.

Frequentist rainfall thresholds obtained for different non-exceedance probabilities are particularly suitable for designing probabilistic schemes to be implemented in LEWS for the prediction of rainfall-induced landslides. In particular, the thresholds here defined were implemented in the LANDSLIP short-range prototype LEWS. Every day, the system compares measured and forecasted rainfall amounts against the rainfall thresholds, and assigns, for a given combination of cumulated rainfall and rainfall duration (i.e. for a given rainfall scenario), a possibility level of landslide occurrence, also based on the susceptibility maps of the two pilot areas. The landslide possibility level is categorized into four classes: low, moderate, high, and very high possibility of landslide occurrence—based on the number of landslides that might occur in one or many locations. The warning message is issued to each administrative area within the two pilot areas: four areas in Darjeeling and six areas in Nilgiris. The short-range landslide forecast only covers shallow landslides caused by rainfall and is valid for the successive 48 h, with different values for each day. The forecast is included in a bulletin of information for district authorities to support them in their decision-making.

The thresholds and the whole forecasting model require ongoing evaluation and refinement by assessing both forecasting and operational capabilities. The evaluation relies upon standardized landslide and rainfall data collection and management. Additional landslide data and different rainfall data inputs (e.g. hourly rainfall measurements, and/or radar and satellite rainfall estimates) will be useful to validate the thresholds and improve the forecasting model.