Introduction

On 8 November 2013, typhoon Haiyan (Typhoon “Yolanda” in the Philippines) reached its peak intensity and swept through the central Philippines with 1-min sustained winds of 315 km/h at 1800 UTC (JTWC 2013). Powerful winds, heavy rains, and storm surges triggered widespread flooding and landslides, particularly affecting the provinces of Leyte (Fig. 1). Typhoon Haiyan left 6300 fatalities, 28,689 injured, and 1061 missing, as well as damage amounting to over USD 2 billion (NDRRMC 2014).

Fig. 1
figure 1

Track of super typhoon Haiyan across the central Philippines. The location of the study area (Leyte province) is enclosed in red. Typhoon track from IBTrACS (Knapp et al. 2010)

Full size image

Although strong winds and storm surges of up to 5–6 m (Lagmay et al. 2015) were the primary causes of damage in Leyte, landslides were also a significant factor. Shallow landslides are rapid downhill flows of rocks and mud—typically induced by heavy rainfall—with speeds exceeding 55 km/h (Hong et al. 2007). They constitute one of the most significant hazards in complex terrain exposed to frequent heavy rainfall (Liao et al. 2010). Relevant shallow landslide-controlling factors include local conditions such as slope materials and slope surface morphology, soil properties, topography, groundwater, lithology, bedrock permeability, and land cover (D’Amato Avanzi et al. 2004; Hong et al. 2007; Liao et al 2010; Wieczorek and Sarmiento 1988). Usually, the triggering mechanism for shallow landslides on destabilized slopes usually begins with excessive rainfall. It is characterized by infiltration into the soil leading to the buildup of pore pressure, reduced soil suction value, reduced shear strength, and increased shear stress, weakening slope stability (Antronico and Gullà 2000; Campus et al. 2001; D’Amato Avanzi et al. 2003; Giannecchini 2006; Van Asch et al. 1999). This could cause surface sliding due to the textural-granulometric discontinuity within the soil-bedrock interface, causing drastic changes in the infiltration rate (Wieczorek 1987). Slope instability brought about by the complex interaction of interrelated factors becomes a major concern when they lie in close proximity to people and property (Soeters and Van Westen 1996).

Conventionally, landslide susceptibility assessments are conducted through field surveys to account for multiple relevant controlling factors (Hong et al. 2007). However, the zonation of shallow landslide potential for provinces with large areas may be difficult in the Philippines, especially given the difficulty of accessing mountainous areas and the cost of ground monitoring networks. Thus, remote sensing may be the only practical means to obtain data for landslide susceptibility mapping (Catani et al. 2005; Hong et al. 2007; Nadim et al. 2006). High-resolution satellite remote sensing products can provide detailed provincial-wide shallow landslide susceptibility maps. Studies conducted by Fabbri et al. (2003) show that the layers extracted solely from a digital elevation model (DEM) generated more accurate results than layers derived from geological data, surficial materials, and land use, indicating that topography is the primary determinant of landslide location (Coe et al. 2004; Hong et al. 2007).

It is also important to develop accurate hazard maps pinpointing the locations of areas vulnerable to landslides for rehabilitation work. The identification of potential landslides and the delineation of unstable and suitable areas for development can help ensure public safety through improved land use and disaster management planning, management of residential and recreational areas, and infrastructural development (AGS 2007; Sarkar and Kanugo 2004). In the case of areas heavily devastated by Haiyan, reliable, detailed, and readily accessible landslide susceptibility maps are critical for well-planned rehabilitation and reconstruction.

Current approach to landslide susceptibility mapping in the Philippines

Prior to this work, government agencies in the Philippines implement a qualitative method for landslide susceptibility mapping. This involves the use of landslide inventories, geologic maps, and geomorphologic information to identify possible sites of failure.

However, weights of contributing factors depend on expert opinion of specialists on the subject. The experts use a checklist to give subjective weights for the factors (e.g., slope angle, slope weight, material type, nearness to fault, degree of gully erosion, vegetative cover, presence of road cuts on hillsides, etc.) and then derive a cumulative influence factor representing the degree of hazard (Table 1). Other than the slope gradient, there are no other quantitative bases for the hazard level assigned to an area (MGB 2013). Thus, the designating weights are highly subjective.

Table 1 Landslide susceptibility parameters currently used MGB-DENR—Philippines to determine the landslide hazard susceptibility of a given area (MGB 2013)
Full size table

To date, quantitative and deterministic methods for landslide susceptibility assessment has not been used on a nationwide scale in the Philippines. The recent availability of high-resolution topographic models for the country has made this possible. Quantitative procedures are advantageous in rehabilitation purposes wherein rapid assessment is highly critical to conduct well-planned reconstruction and rehabilitation.

A sample of maps made by Mines and Geosciences Bureau - Department of Environment and Natural Resources (MGB-DENR) Philippines is shown in Fig. 2. These belong to the current national maps for landslide susceptibility. In the context of disaster management, hazard maps of 1:100,000 to 1:50,000 are not suitable for localized emergency response. Current enhancement efforts by the MGB include producing hazard maps at a finer scale of 1:10,000 (Fig. 2b).

Fig. 2
figure 2

a MGB’s landslide susceptibility map for Tacloban City, Leyte, Philippines, 1:50000 scale (modified from MGB 2006). b MGB’s detailed landslide and flood hazard map of Tacloban City, Leyte, Philippines, 1:10000 scale (modified from MGB 2014)

Full size image

The same scenario occurs in all the available susceptibility maps in the country. The major goal is to improve these existing government maps of the Philippines to enhance post-disaster rehabilitation using newly available high-resolution terrain data and landslide information. It is also imperative to expand options of evacuation and relocation to prevent repeat of disasters within the area in the future.

Here, we implement a widely used method to characterize shallow landslides that can be employed in the nationwide mapping effort as a rapid shallow landslide assessment method. Using the typhoon- and landslide-prone Leyte province as a case study, available high-resolution digital elevation data is utilized to generate shallow landslide susceptibility maps. Unlike delineating hazard zones using historical data, maps generated using a deterministic slope stability geographic information system (GIS) model incorporates pixel-specific soil strength parameters and hydraulic characteristics. Also, it takes full advantage of the fact that landslide source areas are, in general, strongly controlled by shallow groundwater flow convergence accompanied by the reduction of shear strength and increased pore water pressure (Pack et al. 2001, 2005).

The most recent approaches to deterministic slope stability research involves diverse hydrology models using geotechnics at different levels of complexity (Connell et al. 2001; Dietrich et al. 1992, 1993, 1995, 2001; Grayson et al. 1992a, b; Lan et al. 2004; Montgomery and Dietrich 1994; Pack et al. 1998, 2001, 2005; Wu and Sidle 1995). Successful physically based stability model examples are SHALSTAB (shallow slope stability model; Montgomery and Dietrich 1994), SINMAP (Stability Index Mapping; Pack et al. 2005; Tarboton 1997), TRIGRS (Transient Rainfall Infiltration and Grid-Based Regional Slope-Stability Model; Baum et al. 2008), and SLIP (Shallow Landslide Instability Prediction; Montrasio 2000; Montrasio and Valentino 2008). Other examples of advanced hydrology models that involve modelling of groundwater and vadose zone flows are GEOtop-FS (Rigon et al. 2006), TOPMODEL (Beven and Kirkby 1979), dSLAM (Wu and Sidle 1995), and SHETRAN (Burton and Bathurst 1998). SINMAP has been widely used and tested in several landslide-prone areas because of its applicability, minimum requirement for input while allowing multi-calibration of the studied area, and model modification in smaller areas (Calcaterra et al. 2004; Deb and El-Kadi 2009; Lan et al. 2004; Meisina and Scarabelli 2007; Morrissey et al. 2001; Zaitchik and van Es 2003; Zaitchik et al. 2003). It also adds the capability of representing results into a probabilistic framework.

The reliability and degree of SINMAP model has been tested by Zizioli et al. (2013) by using receiver operating characteristic (ROC) analysis as compared to other widely used physically based models: SHALSTAB, TRIGRS, and SLIP. The values in the ROC curve (sensitivity against one specificity) represent the ability of a model to correctly classify between negative and positive observations. The global accuracy obtained for the SINMAP model expressed by the area under the curve (AUC) was equal to 0.79. The global accuracy for SHALSTAB, TRIGRS, and SLIP was 0.78, 0.79, and 0.78 respectively. The values demonstrate that the four models have similar degrees of success for prediction of shallow landslide source areas. Considering this, SINMAP model was selected for the case study due to limitation on geotechnical data or soil strength parameters of Leyte province. With SINMAP, the spatial variability of soil strength parameters can be expressed by a range of values.

After the SINMAP simulation, the maps were then compared to a landslide inventory derived by interpreting high-resolution satellite imagery dated 2002–2014. We then tested for accuracy and performance to determine the detection percentage (DP) and miss factor (MF) values. The shallow landslide susceptibility maps can be complemented with debris flow and structurally controlled hazard maps to identify safe and unsafe areas for disaster rehabilitation.

Methods

The Stability INdex MAPping approach to stability index modelling

The methodology (Fig. 3) involves the use of SINMAP, an ArcView GIS extension and an objective terrain stability mapping tool that combines steady-state hydrologic concepts with the infinite slope stability model and incorporates grid-based digital terrain model (DTM).

Fig. 3
figure 3

Process flow of shallow landslide susceptibility mapping using SINMAP

Full size image

The infinite slope stability model balances the resisting components of cohesion and the destabilizing components of gravity on a failure plane parallel to the ground surface. It is applicable to shallow translational landslides controlled by shallow groundwater flow convergence. The slope stability theory does not apply to deep-seated instability (Montgomery and Dietrich 1994).

The input data required to implement the methodology include topographic slope, specific catchment area, and parameters quantifying material properties (such as soil strength) and climate (hydrologic wetness parameter). The topographic variables are computed from the DTM data. SINMAP does not require numerically precise input and accepts a range of soil strength and hydrologic values to account for uncertainty. In an approximate sense, this capability predicts landslide zones without additional input of weather data and analysis by specifying a range of values for the hydrologic wetness parameter.

Each of the input parameters is delineated on a numerical grid over the study area. The accuracy of the output is therefore heavily reliant on the accuracy of the input DTM (Pack et al. 2005).

The primary output of this modelling approach is a stability index (SI), the numerical representation of the stability of terrain and hence the probability that a pixel is stable assuming uniform distributions of parameters over the uncertainty ranges. It does not predict if shallow translational slope movements will occur, but it forecasts where they are more likely to initiate given the assumptions and input parameters used in the analysis. The indices are not to be interpreted as numerically precise and are most appropriately interpreted as indications of relative shallow landslide susceptibility. SI values range between 0 (most unstable) and 1 (least unstable) (Pack et al. 2001).

The selection of breakpoints of stability index values is subjective. For mapping purposes in the Philippines, three major susceptibility ratings are used for shallow landslide susceptibility maps: high (1.0 ≥ SI > 0), moderate (1.25 ≥ SI > 1.0), and low (1.5 ≥ SI > 1.25). Stability index values greater than 1.5 are classified as stable zones. Moreover, SI less than 0 classify regions where the slope is held in place by forces not represented in the model (e.g., bedrock outcrops, man-made slope protection).

SINMAP uses the following formula to calculate the stability index based on the infinite slope equation proposed by Hammond et al. (1992).

$$ \mathrm{F}\mathrm{S}=\frac{C+ \cos \theta \left[1- \min \left(\frac{R}{T}\frac{a}{ \sin \theta },1\right)r\right] \tan \phi }{ \sin \theta } $$
(1)

where

FS:

Factor of safety

a :

Topographic catchment area

C :

Dimensionless cohesion = (Cr + Cs) = hρ s g

C r :

Root cohesion

C s :

Soil cohesion

h :

Soil thickness

ρ :

Soil density

g :

Gravity constant

hw:

Height of water

R :

Recharge

r :

Water density (ρ w) to soil density (ρ s) ratio

T :

Soil transmissivity = soil hydraulic conductivity × h

φ :

Soil internal angle of friction

θ :

Slope

The variables θ and a are obtained from the DTM topography. Other parameters, C (cohesion), φ (soil angle of friction), T/R (transmissivity divided by recharge), and r (the ratio of water and soil density) are manually entered into the model. These are the more uncertain parameters and are set in terms of lower and upper bound values. The smallest C and tan phi together with the largest R/T define the most conservative (worst case) scenario within the assumed variability in the input parameters (Pack et al. 2001).

Soil data of the study areas

Leyte province has an extensive cover of clay soil (Fig. 4). Clay loam and sand are also present but cover only relatively smaller portions of Leyte. The soil type at most mountainous areas are undifferentiated.

Fig. 4
figure 4

Leyte soil map (digitized from BSWM 2014)

Full size image

Digital terrain model grid data

The study used an Interferometric Synthetic Aperture Radar (InSAR)-derived DTM of Leyte provided by the National Mapping and Resource Information Authority (NAMRIA).

InSAR-derived products can be obtained from airborne surveys using single-pass systems or from space-borne satellite platforms using single- and repeat-pass systems (Dowman 2004). This technique is used to generate a high-resolution DEM suited for broad survey coverage. There are two types of DEM products: the digital surface model (DSM) and the DTM. The former is derived directly from the survey and represents the actual earth surface which includes vegetation and infrastructure. The DTM is produced by removing the vegetation and buildings from the DSM using filtering algorithms leaving only a representation of the actual terrain (Dowman 2004).

The raw data is acquired using Intermap’s STAR-3 InSAR system that includes an X-band InSAR mounted on Learjet 36A aircraft at a flight height of 4–12 km above mean sea level and a baseline tilt angle of 1.3°. One of the obtained raw data from the survey is the DSMs. These are processed to generate DTMs with a 5-m spatial resolutions and 2- and 1-m horizontal and vertical accuracy, respectively (NAMRIA 2013) The projection datum used is the Philippine Reference System of 1992 (PRS92) which is based on the World Geodetic System of 1984 (WGS84) (Floris et al. 2010; Geospatial 2011).

Geotechnical data

Parameters not determined by the DTM such as soil strength, cohesion, angle of friction, transmissivity, and hydraulic conductivity are considered more uncertain and are specified in terms of upper and lower boundary values (Pack et al. 2001). The calibration parameters used for the simulation encompasses the typical soil parameter values for clay and clay loam, which are the dominant soil types in Leyte province (Table 2).

Table 2 Soil calibration parameters for SINMAP simulation
Full size table

The equation used to determine the dimensionless cohesion combines root and soil cohesion. Theoretically, this is the ratio of the cohesive strength of the roots and soil relative to the weight of a saturated thickness of soil (Pack et al. 2005).

$$ C=\frac{\left({C}_{\mathrm{r}}+{C}_{\mathrm{s}}\right)}{h{\rho}_{\mathrm{s}}g} $$
(2)
C r :

Root cohesion

C s :

Soil cohesion

h :

Soil thickness

ρ s :

Soil density

g :

Acceleration due to gravity

Internal angle of friction is a measure of the shear strength of soil due to friction determined in the laboratory using direct shear strength or triaxial stress test.

The effective recharge rates used in this study have been derived from lower and upper precipitation limits (i.e., 50 and 200 mm/day). A 50-mm/day rain rate was chosen as a minimum rate and 200 mm/day was used as the maximum (i.e., an extreme example of the precipitation that can produce shallow translational movement). This range matches the maximum accumulated rainfall of 440 mm for a 72-h period or 147 mm/day on average (MapAction 2013).

The transmissivity rate (m2/h) was calculated using the equation, T = Kb, where K is the hydraulic conductivity of the soil and b is the soil depth in meters. In the context of shallow landslide movement, soil thickness of 1.5 m is assumed to be uniform in the simulation. Data on hydraulic conductivity is obtained from a database of soils in the Philippines (Hamazaki et al. 1990). Field and laboratory descriptions and test results, along with literature values for soil properties given in Hammond et al. (1992) were also used to constrain reasonable ranges of soil input parameters for the stability index modelling.

Validation

The validation process of the generated SINMAP model is summarized on Fig. 5. Landslide inventory map consisting of 363 landslides was used to validate the SINMAP model for Leyte province. The inventory utilized remote sensing techniques using various satellite images that have imagery dates from 2002 to 2014. Satellite imagery includes SPOT of 5 to 10 m spatial resolution, QuickBird of 0.6 to 2.4 m, and WorldView of 0.5 m. These images were inspected using manual detection method to identify landslide events in Google Earth, freeware program, depending on its availability.

Fig. 5
figure 5

Process flow of validation of SINMAP model

Full size image

Satellite images were loaded and inspected in Google Earth. Detected landslide events were delineated with polygons by outlining the visible disturbed body. Each identified landslide was classified based on Cruden and Varnes classification (Varnes 1958, 1978). The types of landslides indentified were mainly slides and flows. The assigned polygons were saved as vector files and analyzed in ArcGIS. The landslide inventory statistics including number of landslide, total surface area damage, and density per locality were computed using the satellite images of Leyte province.

Visual comparison was done with the SINMAP model and the landslide inventory. Accuracy assessment is performed by calculation of detection percentage and miss factor.

Results

Visual comparison of model and landslide inventory

The area distribution for low, medium, and high susceptibility to shallow landslides generated by SINMAP is shown in Fig. 6a. Unsafe areas from shallow landslides were found to be 31 % of the total land area of Leyte. Safe areas from shallow landslides comprise 69 %. The previous 1:10,000 landslide susceptibility map by MGB (Fig. 2b) presents noticeably less stable areas compared to the SINMAP model (54 %) as shown in Fig. 6b.

Fig. 6
figure 6

Distribution of landslide susceptibility by area in Leyte province generated by SINMAP (a) and MGB (b)

Full size image

Initially, visual comparison was done by manual overlaying of the landslide inventory over the SINMAP model and MGB map. The number of landslide polygons that fall within each susceptibility class is tabulated (Tables 3 and 4). Among the 363 landslide polygons in Leyte, 9 were within the stable zones of the SINMAP model (Table 3). Majority of the polygons fell within the medium (66.1 %) susceptibility zones. For the MGB map, most polygons fell within the high and very high susceptibility zones (Table 4).

Table 3 Summary of landslide polygons captured by SINMAP model for Leyte province
Full size table
Table 4 Summary of landslide polygons captured by MGB map for Leyte province
Full size table

Accuracy assessment

The number and extent of mapped landslides are critical to disaster relief and deterministic landslide assessment (Yange et al. 2013). Therefore, in this procedure, the susceptibility zones defined by the SINMAP model were compared with the manually delineated landslide inventory. This accuracy assessment procedure is also done to the MGB map to compare its accuracy with the SINMAP model.

By visual comparison, the landslide polygons were either marked as true positive or false negative using the following category definitions:

  • True positive (TP): a landslide event occurred at the area of instability defined by SINMAP

  • False negative (FN): a landslide event occurred outside the area of instability defined by SINMAP

Once the number of landslide polygons belonging to each susceptibility category is determined, the performance of the SINMAP model for Leyte was evaluated using the following statistical measures:

$$ \mathrm{M}\mathrm{F}=\mathrm{F}\mathrm{N}/\mathrm{T}\mathrm{P} $$
(3)
$$ \mathrm{D}\mathrm{P}=100\times \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{F}\mathrm{N}} $$
(4)

The miss factor (MF) is a measure of omission error where the SINMAP model defined actual landslide areas as stable. The detection percentage (DP) is the percentage of landslide areas correctly classified by the SINMAP model as unstable (Lee et al. 2003; Shufelt 1999).

The 354 polygons captured by the SINMAP model were considered as TP, whereas the 9 landslide polygons outside the model were considered as FN. For the MGB map, 351 polygons are captured (TP) and 12 are outside the susceptibility zones (FN).

The SINMAP results showed that the DP for the study area was 97.5 % with MF of 0.025. These values demonstrate that the SINMAP approach performs well despite the lack of an extensive geotechnical and hydrological database in the study area. This also suggests a consistency in the accuracy of the InSAR-derived DTM and the high-resolution imagery used for the landslide polygon delineation. The MGB map showed a DP of 96.7 % and MF of 0.034. The accuracy of the MGB map is high which is typical for qualitative methods involving historical landslide data and relevant geomorphological information. Both susceptibility mapping methods establish high accuracy when compared to the landslide inventory; however, they differ much in the amount of stable areas they define.

Discussion

The SINMAP simulations for Leyte (Fig. 7) capture very well the landslides that have been mapped by high-resolution imagery over the span of 12 years (2002–2014). Almost all of the 363 landslides in the inventory of Leyte fall within the high (red) and moderately (orange) susceptible landslide areas. Only 2.5 % of the landslide events were not captured by the landslide susceptibility maps generated by SINMAP (Tables 3 and 4).

Fig. 7
figure 7

Shallow landslide susceptibility map of Leyte province with a close up view in an area with identified landslides

Full size image

The stable areas identified by SINMAP within Leyte is 69 % of its total land area, 15 % more than the safe areas identified by previous landslide susceptibility map generated by the MGB (Fig. 2b) which is 54 % of the total Leyte area (Fig. 6). Nonetheless, the accuracy of the SINMAP results remain high, with the 97.5 % DP value and a MF of 0.025.

The idea of using a model that identifies landslide susceptible areas with high accuracy, and at the same time, maximizing safe and habitable areas is beneficial to the development plans of the local government. In post-disaster rehabilitation, it also expands options for siting evacuation and relocation areas. This is critical especially for the rehabilitation work of Haiyan-affected communities because it is imperative that rebuilding be carefully planned to prevent future disasters.

The landslides that were not captured by the SINMAP simulations for Leyte could be due to a different failure mechanism not considered by the infinite slope stability model. For instance, SINMAP does not consider structural controls for slope instability. These landslides occur much deeper and are controlled primarily by fracture planes that pervade in mountainous terrain. Structurally controlled landslide maps are also being prepared under the new Nationwide Hazards Mapping Program called Project NOAH where high-resolution topographic maps derived from Light Detection and Ranging (LiDAR) and InSAR are used. Along with the identification of alluvial fans and simulation of debris flows, the SINMAP and the structurally controlled landslide hazard maps will constitute the detailed hazard maps at 1:10,000 scale for landslides in Haiyan-devastated areas. These maps are available online for free in the Project NOAH portal (http://noah.dost.gov.ph).

It is critical to provide rapid assessment and delineation of areas susceptible to rainfall-induced shallow landslides for well-planned reconstruction and rehabilitation. The method of shallow landslide susceptibility mapping we implement in the Philippines addresses the need for rapidly generated landslide hazard maps. These shallow landslide hazard maps have been made freely available to different relief and rehabilitation agencies in typhoon Haiyan-ravaged areas. The shallow landslide maps that were provided are suited to the requirements of the rehabilitation and reconstruction process. It can still be and should be improved for better accuracy. More detailed geological and geotechnical assessments are necessary in the future. One area of improvement can be the replacement of InSAR-derived DTM data with more accurate LiDAR topographic data.

Conclusion

The landslide map generated using SINMAP are found to be highly accurate with DP value of 97.5 % and MF of 0.025 when compared with the landslide inventory derived from high-resolution satellite imagery for Leyte. In addition, the maps can be further improved with more detailed geological and geotechnical assessment of the study areas. Additional satellite and aerial imagery can provide more landslide inventory data to better determine the spatial extent of landslide occurrences vis-à-vis the results of the SINMAP model. The landslide susceptibility classification found in the landslide hazard maps is useful to identify unsafe and safe areas from shallow landslides. These maps complement the debris flow and structurally controlled landslide hazard maps that are also being prepared for rebuilding Haiyan-devastated areas.