Abstract
Since settlement of São Miguel Island in the middle of the fifteenth century, several destructive landslides have occurred. These have been triggered by various factors, of which rainfall is the most common. Between 1900 and 2008, based on extensive documentary evidence, 193 landslide events with socio-economic impact have been identified. These events have been responsible for 67 deaths, about 20 people have been seriously injured, and dozens of houses destroyed and people rendered homeless. In this chapter, the Information Value Method is used to produce a landslide rupture susceptibility map for São Miguel. Success and prediction rate curves (SRC and PRC) are computed and areas under the curve (AUC) calculated in order to not only support the results of modelling, but also assess the robustness of the suggested susceptibility algorithm. Twelve predisposing factors are used as independent variables and a total of 9890 landslide depletion areas are used as the dependent variable. The AUC for both the SRC and the PRC was 0.90. The methodology used for the construction of the susceptibility map was shown to be both accurate and reliable. The map is an important tool for land-use/emergency planning and for landslide risk mitigation.
Landslides occur worldwide and involve a broad range of predisposing conditions and triggering factors (e.g. intense rainfall, earthquakes, volcanic eruptions, water level changes, storm waves, rapid stream erosion and human activity) that cause an increase in shear stress or decrease in shear strength of slope-forming materials.
Landslides are among the most significant geomorphological processes in mountain environments. They are one of the primary causes of property damage, loss of life and injuries. Since the beginning of the twenty-first century, landslides have affected about 1.5 million people worldwide and have cost more than US$875 million (EM-DAT 2010). These losses are currently increasing as the pressure of population causes urban areas to expand into more unstable areas (NRC 2004).
Since the settlement of São Miguel in the middle of the fifteenth century, several landslide events with important socio-economic impact have been reported in historical accounts. The most deadly and destructive landslide took place on 22 October 1522, when an earthquake with a maximum intensity X (EMS-98) (Silveira et al. 2003; Gaspar et al. 2015; Silva et al. 2015) triggered several landslides over the island, one of which was a debris-flow of about 6.75×106 m3 that covered an area of 4.5 km2 (Marques 2004; Marques et al. 2009b), killing 5000 people and destroying the entire village of Vila Franca do Campo, the island's capital at that time (Frutuoso [1522–91] 1981). This earthquake–landslide event is the most destructive and deadliest natural catastrophe to have affected the Azores and the second largest in Portugal, after the 1 November 1755 earthquake.
The occurrence of such a high-magnitude low-frequency event underlines the potential significance of the landslide threat and poses severe problems of hazard management. To focus discussion solely on these very rare failures, however, serves to divert attention away from the overall landslide magnitude/frequency distribution, because there are many relatively small to medium-sized landslides that cumulatively impose a threat to society at least as great, if not a greater (Jones 1992). The costs of these smaller-scale events are more widely distributed in space and time, with total losses attributed to them growing rapidly during the past few decades.
Smaller landslides have been responsible for many deaths and have had a major socio-economic impact on São Miguel (Marques 2013). Since 1900 hundreds of damaging landslide events (days with confirmed landslide activity) have been triggered, the most catastrophic occurring on 31 October 1997, when nearly 1000 shallow individual landslides were triggered by extreme rainfall (Gaspar et al. 1997; Marques et al. 2008; Marques 2013). This episode claimed 29 human lives in the village of Ribeira Quente and, in addition, seven people were injured and 69 made homeless. The total financial loss was estimated at €21.3 million, a figure that includes both direct and indirect damage (Cunha 2003).
Scientific knowledge allows non-extreme failures to be predicted with a reasonable degree of accuracy by means of the construction of landslide susceptibility/hazard maps. These allow mitigation to be carried out by means of stabilization measures, monitoring and the development/application of better land-use management policies.
In the literature confusion exists between the terms ‘landslide susceptibility’ and ‘landslide hazard’. According to Soeters & Van Westen (1996), landslide susceptibility is the likelihood of a landslide occurring in an area conditioned by local terrain conditions. It expresses the predisposition of a terrain to slope movement and it is an estimate of the likely locations where future landslides may occur. Susceptibility does not consider the temporal probability of failure (i.e. ‘when’ and ‘how frequently’ landslides might occur), nor the magnitude of the expected landslide (i.e. ‘how large’ or ‘how destructive’ the failure will be). Mathematically landslide susceptibility is the probability of the spatial occurrence of a landslide, based upon a combination of a well-known set of geo-environmental conditions.
Landslide hazard, in contrast, requires a far wider knowledge than landslide susceptibility. It involves calculating the probability that a landslide of a given magnitude will occur in a given time period and in a given area. Besides predicting ‘where’ a slope failure will occur, landslide hazard forecasts ‘when’ and/or ‘how frequently’ a given landslide will occur and ‘how large’ it will be (Guzzetti et al. 2005a).
A landslide susceptibility map is the first step in risk assessment/management and, therefore, makes an important contribution to public safety (Guzzetti et al. 1999; Glade et al. 2005). Assessment of landslide susceptibility has shown significant improvements in recent years owing to the implementation of statistically based techniques using Geographical Information Systems (GIS) methodologies. At a regional scale, landslide susceptibility using statistical techniques is based on the evaluation of terrain conditions of land that has been subject to previous landslides (Carrara et al. 1995). Analysis is based on the principle that future landslides have a greater likelihood of occurrence under conditions that are similar to those that caused past slope instability (Varnes et al. 1984; Soeters & Van Westen 1996; Guzzetti et al. 1999; Zêzere et al. 2004). In this way the spatial relationships of past landslides and the multiple predisposing factors that cause them may be efficiently investigated and provide useful information so as to identify the most probable locations of future landslides.
The first task is to construct an inventory of landslides that have occurred in the study area so that the characteristics of the affected slopes may be identified. Traditionally, landslides have been mapped through the interpretation of aerial photographs, fieldwork and, more recently, by exploring digital ortho-rectified photographs and satellite images.
There are a variety of methods used to produce landslide susceptibility maps and these methods may be classified into two main categories: qualitative and quantitative (Soeters & Van Westen 1996; Aleotti & Chowdhury 1999; Guzzetti et al. 1999; Ercanoglu et al. 2008; Bai et al. 2010). Qualitative methods were used extensively during the 1970s and 1980s (Aleotti & Chowdhury 1999) and include field geomorphological analysis (e.g. Rupke et al. 1988; Concha-Dimas et al. 2007; Kouli et al. 2010) and the use of index or parameter maps (heuristic) (e.g. Soeters & Van Westen 1996; Van Westen & Lulie 2003; Gomes et al. 2005). Quantitative methods include geotechnical engineering approaches, statistical/probabilistic analysis and soft computing techniques (Aleotti & Chowdhury 1999). Geotechnical approaches use deterministic methods (e.g. Gökceoglu & Aksoy 1996; Van Westen & Terlien 1996; Zhou et al. 2003; Wang & Lin 2010). Statistical/probabilistic methods include bivariate and multivariate analysis.
In the bivariate analysis (e.g. Remondo et al. 2003; Süzen & Doyuran 2004a; Zêzere et al. 2004, 2007, 2008; Saha et al. 2005; Marques et al. 2009a; Marques 2013) each individual class of each predisposing factor is individually combined with the landslide distribution map. Weighted values based on landslide distribution are calculated for each class. In the literature, the most common bivariate statistical techniques are Information Value (Yin & Yan 1988) and Weights of Evidence (Bonham-Carter et al. 1989). In multivariate analysis many predisposing factors are sampled and, for each of the terrain units, the presence or absence of landslides is determined. This allows an estimation of the relative weights of each contributing factor by means of statistical procedures (Baeza & Corominas 2001). In the literature there is a tendency to use multivariate statistical analysis techniques, including Discriminant Analysis (e.g. Carrara et al. 1991, 2003; Santacana et al. 2003; Guzzetti et al. 2005b, 2006a, 2006b; Baeza et al. 2010a; Marques 2013), Factor Analysis (e.g. Ercanoglu & Gökceoglu 2002; Ercanoglu et al. 2004), Logistic Regression Analysis (e.g. Atkinson & Massari 1998; Dai et al. 2001; Dai & Lee 2002, 2003; Süzen & Doyuran 2004a, b; Ayalew & Yamagishi 2005; Mathew et al. 2009; Marques et al. 2009a; Baeza et al. 2010b; Marques 2013) and Conditional Analysis (e.g. Clerici et al. 2002; Duman et al. 2005; Özdemir 2009).
In recent years researchers have used soft computing approaches to assess landslide susceptibility, characterized by the use of inexact solutions to computationally hard tasks for which there is no known algorithm that can compute an exact solution in polynomial time, such as Fuzzy Logic, Artificial Neural Networks and Neuro-Fuzzy methods (e.g. Ercanoglu & Gökceoglu 2002, 2004; Lee et al. 2003, 2004, 2006; Remondo et al. 2003; Ercanoglu et al. 2004; Lee & Evangelista 2006; Miles & Keefer 2009; Marques 2013).
In spite of the vast number of methods used to produce landslide-susceptibility maps, recent studies comparing different statistic/probabilistic approaches have shown that the quality of landslide susceptibility maps and their capacity for predicting where future landslides will occur lies in the quality of the inputs (e.g. good statistical representation of the landslide database, cartographic quality and the fine detail of predisposing factors), rather than the complexity or computational details of the statistical methods adopted (e.g. Süzen & Doyuran 2004b; Marques et al. 2009a; Cruz de Oliveira 2012; Marques 2013).
Validation of the robustness and accuracy of landslide susceptibility models is possible through the computation of both success-rate and prediction-rate curves (SRCs and PRCs; Chung & Fabbri 1999, 2003; Zêzere et al. 2004; Beguería 2006; Guzzetti et al. 2006a, b). The SRCs check the goodness of fit of the susceptibility model by assessing whether the landslides that are used to build the model are correctly classified. The PRCs are more effective in the process of validation because they check the predictive power of the susceptibility model by verifying if landslides independent of those used to build the model are correctly classified (Chung & Fabbri 2003).
This chapter focuses the temporal and spatial distributions of landslide events and their associated casualties on São Miguel between 1900 and 2008. Susceptibility analysis of landslide rupture was performed for the complete island at a 1:25 000 scale using a GIS-based bivariate statistical technique developed by Yin & Yan (1988), termed Information Value, and using 12 landslide predisposing factors. The final landslide susceptibility map was divided into four susceptibility classes, that is, very high, high, moderate and low. The performance and robustness of the models were evaluated and the predictive power validated using both SRC and PRC and computing their areas under the curve (AUC). Finally the percentages of buildings, main and secondary roads that are included in each landslide rupture susceptibility class are defined for each concelho (i.e. the first-order administrative division of Portugal) of the island.
Geological and geomorphological setting
The Azores archipelago is located in the triple junction between the Eurasian, African and North American tectonic plates (Fig. 13.1). The complex geodynamic setting of the Azores archipelago (Miranda et al. 2015) is reflected in both seismic and volcanic activity (Gaspar et al. 2015), which are frequently responsible for important landslide events. Nevertheless, landslides are the most common type of geological hazard, being conditioned by the volcanic materials' physical, mechanical and hydrological properties and by the geomorphology. Among instability causes, rainfall is of major importance in landslide occurrence both in the Azores archipelago generally and specifically on São Miguel (Marques et al. 2008; Amaral et al. 2009; Amaral 2010; Marques 2013).
The location of São Miguel within the Azores Archipelago setting.
São Miguel has an area of approximately 745 km2, presenting an east–west elongated shape and an altitude ranging between 0 and 1143 m. It is the most densely populated island in the Azores with a total of 137 856 inhabitants (INE 2012), and encompasses six concelhos (i.e. Ponta Delgada, Ribeira Grande, Lagoa, Vila Franca do Campo, Povoação and Nordeste).
The island comprises three quiescent central volcanoes with summit calderas (Fig. 13.2): Sete Cidades Volcano (Queiroz 1997; Queiroz et al. 2008, 2015); Fogo Volcano (Wallenstein 1999; Wallenstein et al. 2015); and Furnas Volcano (Gaspar et al. 1996; Guest et al. 1999, 2015). Non-cohesive materials (mostly pumice fallout deposits) are found on the steep slopes of the volcanoes, particularly on inner caldera walls and sea cliffs, and also on the slopes of the deep narrow valleys cut by intense fluvial erosion. These valleys show a radial distribution around the calderas. These factors together make the island subject to very prominent landslide activity.
Morphostructural units of São Miguel Island: 1, Sete Cidades Volcano; 2, Picos Fissural Volcanic System; 3, Fogo Volcano; 4, Congro Fissural Volcanic System; 5, Furnas Volcano; 6, Povoação Volcano; 7, Nordeste Volcanic Complex (based on Zbyszewski et al. 1958, 1959; Wallenstein 1999; Ferreira 2000; Carmo 2013).
The central volcanoes are linked by zones of fissural volcanism (Fig. 13.2): Picos Fissural Volcanic System (Ferreira 2000; Ferreira et al. 2015) and Congro Fissural Volcanic System (Booth et al. 1978; Guest et al. 1999; Ferreira et al. 2015). Picos Fissural Volcanic System is dominated by the presence of numerous cinder cones and associated young lava flows that extend north and south from the central volcanic ridges to the coastline, developing very gentle slopes towards the sea. This region is characterized by a poorly developed drainage network. Landslides in this area do not have a wide geographical distribution and very frequently are concentrated along sea cliffs. The Congro Fissural Volcanic System is also dominated by several basaltic cinder cones and associated lava flows. This region is mantled by thick trachytic deposits originating from eruptions centred at Fogo and Furnas volcanoes. Gentle slopes are developed to the north and south towards the coastline. Landslide activity is frequent on sea cliffs but, unlike Picos Fissural Volcanic System, landslides on the Congro Fissural Volcanic System also occur on the slopes of the deep narrow valleys cut by intense fluvial erosion.
The eastern part of the island is older and considered inactive. It includes the Povoação Volcano and the Nordeste Volcanic Complex (Fig. 13.2; Duncan et al. 2015). The Nordeste Volcanic Complex is the oldest unit on the island and corresponds to a basaltic (s.l.) shield volcano incised by marine and fluvial erosion. The SE sector of this complex is cut by the Povoação Caldera, a circular collapse structure, which is open to the sea to the south. The non-cohesive materials (essentially pumice fallout deposits) that mantle the Povoação Volcano, are mainly derived from Furnas Volcano, and are associated with the steep slopes on inner caldera walls, sea cliffs and the deep narrow valleys that are cut by intense fluvial erosion inside the caldera. These steep slopes promote very frequent landslide activity. The Nordeste Volcanic Complex, unlike all the other sectors of São Miguel where landslides normally occur, is mainly composed of thick basaltic lava flows (s.l.). The existence of clay-rich levels resulting from weathered lava flows and some soil horizons that mantle the slopes promotes landslide occurrence. In addition the extreme slopes of very deep and narrow valleys oversteepened by intense fluvial erosion also favour landslide activity.
Landslide events in the period 1900–2008
Work is currently being carried out on developing the Azores Natural Hazard Documental Database and, so far, 193 damaging landslide events (days with confirmed landslide activity) have been defined on São Miguel for the period 1900–2008 (Marques 2013). These have been directly responsible for 67 casualties, with more than 20 people being seriously injured, dozens of houses being destroyed and many people being made homeless. This information was collected from historical accounts, theses, technical-scientific reports and newspapers and periodicals, both national and regional. Lack of occurrence in any given period may be due either to the catalogue's incompleteness or to variations in the conditions that led to slope failure. The latter could be related to climatic anomalies, including rainfall events, earthquakes, land-use changes and other human actions.
Temporal analysis
Figures 13.3 and 13.4 portray the temporal pattern of damaging landslide events on São Miguel and their associated casualties. Analysis of these figures allows two general conclusions to be made. First, although important landslide activity occurred in the first half of the 1930s, the number of events per year has been increasing over time, especially since the second half of the 1970s and particularly after 1995. Secondly, the severity of recorded events, measured by the number of casualties, varies with time, but was rarely higher than five per event. On 14 October 1942 and on 31 October 1997, in the concelho da Povoação, two landslide events caused 7 and 29 casualties, respectively, the deadliest events known on São Miguel since 1900. Clearly there is a correlation between periods when more damaging landslide events are identified and periods when more casualties are recorded. From Figure 13.5a it is clear that a high percentage of damaging landslide events are triggered by rainfall (75.1%) with 22.8% of the total landslide events having associated flash floods. Some 76% of the casualties were also caused by rainfall-triggered landslide events (Fig. 13.5b). This highlights the importance of rainfall as a factor in landslide generation and the need to promote detailed research aiming to better understand the complex relationships between rainfall, slope instability and landslide risk mitigation.
Temporal distribution of damaging landslide events in São Miguel between 1900 and 2008, taking into account the landslide trigger event: (a) annual frequency; (b) cumulative frequency. LS, landslide; FF, flash flood.
Temporal distribution of landslide casualties in São Miguel between 1900 and 2008 taking into account the landslide trigger event: (a) annual frequency; (b) cumulative frequency.
(a) Percentage of damaging landslide events recognized to have occurred between 1900 and 2008 in São Miguel triggered by different factors and (b) percentage of casualties by triggering factor. LS, landslide; FF, flash flood.
Spatial analysis
The spatial distribution of damaging landslide events and the fatalities caused by them is presented by concelho for the time period 1900–2008 (Figs 13.6 & 13.7). Povoação is by far the most threatened concelho. This concelho was affected by 105 landslides events (i.e. 54.4% of the total), which caused 44 casualties (65.7% of the total). The least-affected concelhos were, in constrast, Lagoa (43 events and 1 casualty) and Nordeste (46 events and no casualties).
(a) Spatial distribution of damaging landslide events on São Miguel and (b) their incidence by triggering factor, considering the time period 1900–2008. PDL, concelho de Ponta Delgada; LAG, concelho da Lagoa; RGD, concelho Ribeira Grande; VFC, concelho de Vila Franca do Campol; NOR, concelho do Nordeste; POV, concelho da Povoação; LS, landslide; FF, flash flood.
(a) Spatial distribution of landslide casualties on São Miguel and (b) their incidence by triggering factor, considering the time period between 1900 and 2008. Abbreviations as Figure 13.6.
In all concelhos, rainfall has been the most important triggering factor for landslides and has been responsible for the highest number of casualties (Figs 13.6 & 13.7). Only Lagoa does not follow this trend, because the only casualty identified was caused by a landslide triggered by anthropogenic activity.
Data and methods used in the analysis of landslide susceptibility
Owing to the socio-economic impact of landslides over past decades, government, environmental and research organizations worldwide have mobilized resources to assess landslide susceptibility. Existing methods are based upon a few, widely accepted assumptions, namely:
slope failures leave discernible features that can be recognized, classified and mapped by field work or through techniques of remote sensing;
landslides are controlled by physical and mechanical laws that may be identified using empirical, statistical or deterministic approaches, so that conditions that cause slope failures (landslide predisposing factors) may be collected and used to build predictive models of landslide occurrence; and
that the past and present are keys to the future, this uniformitarian principle implying that future landslides will be more likely to occur under conditions that led to past instability.
Susceptibility assessment of landslide ruptures on São Miguel has included three main steps: (1) preparation of a landslide rupture inventory map and predisposing factors maps; (2) landslide susceptibility analysis and mapping; and (3) validation of the resulting susceptibility map.
The landslide inventory map
The landslide inventory for São Miguel Island was constructed using detailed geo-referenced digital orthophoto maps (pixel=0.4 m). These orthophoto maps, dating from 2005, were combined with the topographic information in the form of contour lines spaced every 10 m height. Landslides were identified and located at a 1:1000 scale based on interpretation of distinctive changes in the morphology, vegetation and draining conditions of slopes. In addition, photographic characteristics such as tone and texture, pattern, shape and lineaments were related to expected ground conditions or landforms associated with processes of slope instability as summarized in Table 13.1.
Morphologic, vegetation and drainage characteristic of slope-instability processes and their photographic characteristics (from Soeters & Van Westen 1996)
A total of 9890 individual landslide rupture areas were identified and digitized on to a GIS as polygons with a total planimetric area of about 13.16 km2 (Fig. 13.8) and then converted to a raster data file (131 405 10×10 m pixels). The final landslide inventory is statistically representative of the spatial incidence of landslides on the island, although it is not complete. Indeed, the smaller landslides (i.e. <250 m2) were very difficult to identify using only the orthophoto interpretation because of dense forest.
Landslide rupture area distribution in São Miguel based on aerial-photo interpretation.
Typically, landslides on São Miguel are shallow translational slides and soil slips that evolve downslope into debris-flows under favourable geomorphological, hydrological and geological conditions (Fig. 13.9). Most landslides are narrow – with widths normally not exceeding 30–40 m – and the depth of the slip surface does not exceed 2–3 m. The vast majority of landslides originated near to the crest of very steep slopes and spread over complete slope lengths that normally control the total area depleted. Usually failures occur on steep slopes, comprising unsaturated cohesion-less pyroclastic deposits. Under these conditions matric suction plays a crucial role in the shear strength, owing to the effect of apparent cohesion, and contributes to maintaining slopes in physical equilibrium (Amaral et al. 2009; Amaral 2010).
Examples of landslide occurrences in São Miguel: (a, b) landslides on Furnas Volcano; (c, d) landslides on Fogo Volcano.
It is clear from Figure 13.8 that landslides are mostly concentrated on the slopes of young central volcanoes and sea cliffs. Vertical caldera walls, steep slopes, deep incised valleys and several fault scarps characterize these central volcanoes, which are covered by unconsolidated pyroclastic deposits from more recent explosive eruptions. These geological and geomorphological circumstances make such areas very prone to the occurrence of landslides. The old basaltic volcanic complex of Nordeste is marked by deep gullies developed on thick sequences of weathered lava flows and also has a large number of landslides. On the other hand, the Picos Volcanic Fissure System shows the lowest concentration of damaging landslide events. This region is built mainly from recent basaltic (s.l.) lava flows that are very cohesive, with minor interbedded scoria deposits, and presents gentle slopes and a few very immature valleys. Landslides with larger areas are concentrated on slopes characterized by higher slope lengths, normally associated with extremely deep narrow valleys (Fig. 13.8).
For methodological purposes the 9890 landslide rupture areas (LS-TOT) were subdivided randomly into two subsets (each containing 50% of the total number of landslide rupture areas), and then converted into two raster data files, henceforth referred to, respectively, as the ‘modelling subset’ (LS-MOD – 65 102 10×10 m pixels) and the ‘validation subset’ (LS-VAL – 66 311 10×10 m pixels) (Fig. 13.10). The modelling subset was used to calibrate the landslide susceptibility model, whereas the validation subset was used to validate the model results. Ideally the model should be constructed using an older landslide inventory and more recent landslides should then be used for the testing the prediction. This rule could not, however, be followed because the age of most landslides is unknown.
Landslide rupture areas modelling (LS-MOD) and validation (LS-VAL) subsets.
The predisposing factors
Landslide rupture susceptibility assessment on São Miguel was carried out using 12 predisposing factors that are widely used in susceptibility analysis (Fig. 13.11, Table 13.2). The landslide predisposing factors were used as independent variables for the elaboration of the landslide susceptibility map. Their choice relies on both their physical characteristics and the availability of data. The predisposing factors were separated into two: (1) digital elevation model- (DEM-) derived factors; and (2) other geo-environmental factors.
Landslide predisposing factors considered for susceptibility analysis: (a) altimetry; (b) slope; (c) aspect; (d) transversal profile; (e) longitudinal profile; (f) insolation; (g) contributing area; (h) inverse of the wetness index; (i) distance to stream lines; (j) drainage density; (k) geology; (l) land use.
Information values (IVi) for classes of each landslide predisposing factor computed using LS-TOT (IVi-TOT) and LS-MOD (IVi-MOD) as dependent variable
For the preparation of the DEM-derived factors, a DEM of the entire island with a 10 m resolution was used. The DEM was prepared using both contour lines (10 m) and spot heights from 1:25 000 topographic maps. The following factors were derived from the DEM using ArcGIS Tools (Table 13.2): altitude (11 classes); slope angle (13 classes); slope aspect (9 classes); slope transversal profile (3 classes); slope longitudinal profile (3 classes); and insolation (12 classes). The stream-line distance (seven classes) was assessed using buffer distances to the stream lines and drainage density (6 classes) calculated using 100×100 m windows. The contributing area (12 classes) and the inverse of the wetness index (10 classes) were calculated using TauDEM 5.0 (Terrain Analysis Using Digital Elevation Models; Tarboton & Mohammed 2010). Both the contribution area and inverse of wetness index were calculated using the D-Infinity algorithms (Tarboton & Mohammed 2010). Geology (10 classes) was recorded by simplifying the Geological Map of São Miguel (Moore 1991) and the land-use (nine classes) used pre-existing data from SRAM (2007).
Each predisposing factor was represented as raster datasets and the selected size of the pixel was set at 10 m, giving a total of 7 445 444 10×10 m pixels for each factor.
Landslide susceptibility analysis
Evaluation of the susceptibility to landslide rupture was carried out using the Information Value method, a bivariate probabilistic/statistical approach. This method was selected taking into account: (1) its low computational requirements; (2) ease of update of the landslide susceptibility map through time; (3) simplicity of underlying principles which are easy to explain to decision makers; and (4) accurate predictions, which are similar to those obtained from more complex bivariate and multivariate methods (e.g. Cruz de Oliveira et al. 2009; Marques et al. 2009a; Marques 2013).
Developed and first used successfully by Yin & Yan (1988) and subsequently by others (e.g. Jade & Sarkar 1993; Zêzere 2002; Marques et al. 2009a), the Information Value method relies on the existence of a robust landslide inventory, the prior definition of terrain units (in this work pixels) and a set of landslide predisposing factors for quantitatively modelling the relationships between landslides and these factors. According to Yin & Yan (1988), the definition of the Information Value for each factor class i is given by equation 13.1:
(13.1)where Si is the number of pixels with landslides within factor class i; Ni is the number of pixels with presence of factor class i; S is the total number of pixels with landslides; and N is the total number of pixels. S/N is the a priori probability. It is the probability for each pixel that a landslide will occur without considering the predisposing factors. Si/Ni is the conditional probability. It represents the probability a landslide will occur in the presence of factor class i.
If IVi<0, then the presence of that class i inhibits landslide occurrence. On the contrary, if IVi>0 then the presence of that class promotes landslide occurrence. The higher the value of IVi, the greater is the importance of factor class for landslide occurrence and vice versa. If IVi = 0, or close to 0, the factor class is not relevant or its influence on landslide occurrence is not clear.
The susceptibility map is obtained by the sum of IVi of each factor class i present in each pixel j, given by equation (13.2):
(13.2)where Xji=1, if factor class i exists in a grid cell j, or Xji=0 if factor class i does not exist in grid cell j; j=1, 2, … , n, where n is the total number of grid cells; and i=1, 2, … , m, where m is the number of factor classes considered.
Validating the landslide susceptibility map
The performance of the susceptibility models may be effectively evaluated and the prediction power of the models assessed using techniques such as SRCs and PRCs, and by calculating the AUCs (Chung & Fabbri 1999, 2003).
The SRCs (Chung & Fabbri 1999) were used to assess the goodness of fit of the models. SRCs are constructed by plotting the cumulative percentage of susceptible areas (starting from the highest probability values) on the x-axis and the cumulative percentage of landslide modelling subset area on the y-axis. The steeper the curve, the better the capability of the model to describe the distribution of landslides.
The last step in the analysis seeks to evaluate the predictive power of the susceptibility model using the PRCs (Chung & Fabbri 2003). PRCs are constructed in the same way as SRCs, but instead of the modelling subset, the validation subset is employed for validation proposes.
The AUC, for both success and prediction rate curves, was computed by the trapezoidal rule using relative frequencies following the proposal of Gorsevski et al. (2000) and Bi & Bennett (2003) (equation 13.3):
(13.3)where (aUi−aLi) is the curve interval i amplitude; oLi is the value of ordinate of abscissa aLi (lower limit of the curve interval i); oUi is the value of ordinate of abscissa aUi (upper limit of the curve interval i); and i=1, 2, … , n, where n is the total number of intervals of the curve. The AUC for the SRCs and PRCs ranges between 0.5 and 1 owing to their cumulative nature and express the percentage of the graph area that lies under the curve, allowing an easy comparison of SRCs and PRCs from different susceptibility models. According to Swets (1988), Guzzetti (2005b) and Thuiller (2009), the AUC results should be interpreted as listed on Table 13.3.
Interval values and their meaning for the areas under the curve (AUC) results
Results and discussion
The probability of landslide rupture area for São Miguel is 0.017, which means that the total area affected by landslide failures corresponds to 1.7% of the total area of the island. The landslide rupture probabilities for the concelhos of São Miguel are summarized in Table 13.4. Nordeste and Povoação have by far the highest landslide rupture probability (0.038 and 0.028, respectively) which is in agreement with the high number of damaging landslide events that affect these concelhos (Fig. 13.6). The concelho with the lower landslide rupture probability is Lagoa (0.009), reflecting a low landslide susceptibility in accordance with the low number of landslide occurrences registered in the area (Fig. 13.6). The total landslide rupture inventory (LS-TOT) and the landslide rupture modelling group (LS-MOD) were cross-tabulated with each predisposing factor class to compute the Information Value scores. These are summarized in Table 13.2. The results obtained support the assertion that there are no significant differences in computed Information Value scores using LS-TOT and LS-MOD, which shows the marked similarity between the two inventories. These findings indicate that the landslide rupture inventory used for the Information Value scores calculation is robust and statistically significant. According to the results obtained (Table 13.2), the most favourable conditions for landslide rupture in São Miguel are: steep slopes (i.e. >30°) composed of pumice deposits, older than 5 ka BP; low insolation zones (i.e. ≤10×105 W m−2 a−1); altitudes above 700 m; denuded areas occupied by low-lying vegetation and high values of the inverse of the wetness index (>0.050).
Landslide rupture probability for each concelho of São Miguel
In contrast, less favourable conditions for landsliding are associated with: gentle slopes (i.e. ≤30°); low wetness index values (i.e. ≤0.001); slopes with planar longitudinal profiles; and fresh lava flows younger than 5 ka BP.
The results obtained emphasize the importance of slope angle in discriminating between low and high landslide areas of susceptibility. This is clearly revealed by the increase in Information Value scores with an increased slope angle. This factor is directly related to the increase in shear stress on slopes, which tends to increase with increases in slope angle.
Geology is another important discriminating predisposing factor, reflecting indirectly the physical, mechanical and hydraulic properties of the materials. This factor is, therefore, a proxy for internal friction angle, cohesion and hydraulic conductivity, which directly control the stability state of slopes. The Information Value scores obtained for geology clearly shows the prevalence of landslides on cohesionless pyroclastic slopes (in agreement with Amaral et al. 2009; Amaral 2010; internal friction angle (°)=[30, 43] and cohesion (kPa)=[0, 9]) and their almost absence on fresh lava flows younger than 5 ka old.
The importance of water content on slopes subject to landslide occurrence is clear on São Miguel. This evidence is given indirectly by the high Information Value scores that are associated with high altitudes, low insolation values and high inverse values of the wetness index. As in most regions in the world, so on São Miguel Island an increase in altitude is reflected by an increase in the mean annual precipitation (Azevedo 1999) and, therefore, an increase in water availability. In addition, lower insolation values reflect lower rates of evaporation and, thus, the preservation of water on the slopes. Furthermore, higher inverse values of the wetness index (i.e. of the proxy that quantifies the moisture retention, soil water content and surface saturation zones) indicate a higher potential for water availability and to the occurrence of saturation conditions on slopes and more favourable conditions for landslide occurrence.
Land-use Information Value scores indicate that denuded slopes with an absence of vegetation and the presence of shallow-rooted low-lying vegetation are the most unstable zones within the study area. Figure 13.12 shows the SRC and PRD of the landslide susceptibility model for São Miguel. The AUCs values are 0.90 in both cases, illustrating the fact that the predictive model is extremely robust and shows excellent predictive accuracy for future landslides (according with Table 13.3).
(a) Success rate curve (SRC) and (b) prediction rate curve (PRC) of the landslide rupture susceptibility map of São Miguel.
The landslide susceptibility map for São Miguel is shown in Figure 13.13. Susceptibility classes were defined as a function of the Information Value final score obtained by integrating the scores corresponding to the classes of 12 predisposing factors, according to the following rules (Table 13.5):
Class I (very high susceptibility), includes an area with Information Values >1;
Class II (high susceptibility), includes an area with Information Values ranging between 1 and 0;
Class III (moderate susceptibility), includes an area with Information Values ranging between 0 and −1; and
Class IV (low susceptibility), includes an area with Information Values lower than −1.
Landslide rupture susceptibility map for São Miguel.
Landslide rupture probability in different landslide susceptibility classes for São Miguel
The respective percentage of landslide rupture area included in each susceptibility class as well as the landslide probability in each one of them is presented in Table 13.5.
The landslide probability for each pixel within very high- and high-susceptibility classes is 0.078 and 0.023, respectively. Furthermore there is a gradual decrease in probability values from very high to low susceptibility classes and this reflects the strong correlation and coherence between the landslide susceptibility map and landslide inventory.
Table 13.6 shows the percentage of each landslide susceptibility class within each concelho of São Miguel. By a wide margin Nordeste and Povoação are the concelhos with the highest percentage of area within the two most susceptible classes: 47.1 and 42.9%, respectively. Ponta Delgada is the concelho with the smallest area included in those two susceptibility classes owing to the geological and geomorphological features of Picos Volcanic Fissure System.
Percentage of area within each landslide susceptibility class for the concelhos of São Miguel
The buildings, main roads and secondary roads were cross-tabulated for each concelho with the landslide susceptibility classes (Tables 13.7, 13.8, 13.9). Povoação and Nordeste are the concelhos with the highest percentage of buildings (1.9 and 1.4%, respectively), main roads (19.7 and 24.1%, respectively) and secondary roads (19.1 and 20.8%, respectively) contained in the two highest landslide-susceptibility classes. Ponta Delgada is the concelho where exposure to landslides is lower: 0.2% of buildings, 3.8% of main roads and 4.0% secondary roads are within high or very high landslide-susceptibility classes.
Percentage of buildings within each landslide susceptibility class for the concelhos of São Miguel
Percentage of main roads within each landslide susceptibility class for the concelhos of São Miguel
Percentage of secondary roads within each landslide susceptibility class for the concelhos of São Miguel
Concluding remarks
The island of São Miguel has a high rate of landslide activity triggered by several mechanisms of which rainfall is the most significant. For the time span 1900–2008, 193 damaging landslide events were registered and 67 casualties were recorded. Since the second half of the decade of the 1970s and, in particular, after 1995, the frequency of landslide events has increased. There is a heterogeneous distribution of the landslides and associated casualties on São Miguel. The highest number of landslide events and casualties have occurred in the concelho of Povoação, owing to its landslide-prone geo-environmental characteristics and because of the presence of buildings and roads located in areas that have a high susceptibility to landslides. Such a situation implies less than optimal land-use planning with respect to this hazard.
A total of 9890 landslide depletion areas were identified and used as the dependent variable in the landslide susceptibility analysis. Twelve geo-enviromental variables were used as independent variables. Slope, altitude, insolation, inverse of the wetness index, geology and land use are revealed by this study to be the most important in discriminating between landslide-prone areas and stable areas. The calculated AUCs for SRC and PRC were extremely strong (i.e. 0.90 for both) and showed the robustness of the method used in this study and the quality of the landslide susceptibility map.
The susceptibility map shows that 23% of the area of the island has a very high to high susceptibility to landslides. Nordeste and Povoação are the concelhos with the largest areas classified within the very high- and high-susceptibility classes (i.e. 47.1 and 42.9%, for each concelho, respectively). Cross-tabulating landslide susceptibility with buildings, main roads and secondary roads at a concelho level reveals that Povoação and Nordeste have the most risk-exposed infrastructure.
The correct delimitation of areas susceptible to landslide rupture is the first step in hazard analysis. Indeed, the study of runout distances of landslides must be a future priority owing to the lack of this information in the Azores. The percentage area of each concelho that could be affected by the propagation of landslides will significantly increase the proportion falling into the higher susceptibility classes (Marques 2013). In addition, there is much more infrastructure that may be affected by landslides.
The landslide-susceptibility map for São Miguel is an important and useful tool for land-use and emergency planning. Areas with very high and high landslide susceptibility may be prioritized for detailed site-specific geotechnical investigations. New land-use planning policies need to be defined and implemented in order to mitigate landslide risk.
Acknowledgments
This work was supported by the Azores Regional Government through the project ‘Evaluation of geological hazards and vulnerabilities for risk assessment and land-use planning in the Azores Islands’ contract Ref. CC17/2009. The authors are grateful to Maria João Poim who helped to compile and process the documentary information for the Azores Natural Hazards Database (NATHA). We also are grateful to the reviewers Neil Macdonald and Fernando Marques, and to the editor David Chester, whose comments and suggestions helped to improve this manuscript.
- © 2015 The Author(s). Published by The Geological Society of London. All rights reserved
References
Chapter 13 Landslides on São Miguel Island (Azores): susceptibility analysis and validation of rupture zones using a bivariate GIS-based statistical approach
