Evaluation of landslide susceptibility using multivariate statistical methods: a case study in the Prahova subcarpathians, Romania Z. Chitu, I. Sandric, B. Mihai & I. Savulescu Faculty of Geography, University of Bucharest, Bucharest, Romania. ABSTRACT: The shallow landslides affect most of the hillslopes in the Prahova Subcarpathians and are mainly caused by the agriculture practices. Though, these processes do not imply immediately losses for human activity, the landslide susceptibility assessment is necessary to prevent terrain degradation. The evaluation of landslide susceptibility requires understanding of spatial distribution of the factors that control slope instability. It is known that the behavior of landslides is difficult to evaluate because of the various factors that trigger the mass movements. One of the most efficient methods to establish the importance of the factors in landslides susceptibility is the multivariate analysis, like Principal Component Analysis (PCA) and logistic regression. For this study, the PCA was applied to remove the redundant information from the elements that control slope instability like lithology, faults, slope gradient, plan curvature, profile curvature, land use and land cover, soils etc. The PCA analysis had reduced the factors at six principal components and the first three components comprise over 85% of the information. The logistic regression analysis was used to correlate the presence of the past and actual landslides with the PCA components and derive a landslides susceptibility map. To validate the model we used the Receiver Operating Characteristic (ROC) analysis. ROC shows a good correlation between the observed and predicted values of the validation dataset. 1 INTRODUCTION Landslide susceptibility is defined as the likelihood of a landslide occurring in an area on the basis of local terrain conditions (Brabb, 1984). The landslide susceptibility map must provide information about the areas affected by landslides and that may be affected by future landslides. Landslide susceptibility mapping have received special attention in the latest 25 years from various fields of science and several methods have been proposed. Though, there is no agreement for unique methodology, the landslide susceptibility mapping can be performed using one of the available methodologies: geomorhological mapping, heuristic approach, statistical methods or process based models (Guzzetti, 2005). Due to the large area of our study, the high number of variables and especially due to the redundancy in the information, we choose to apply multivariate statistical methods. This approach offers the possibility to handle a high number of layers, to establish the importance of factors in slope instability and the relation between this factors and the past and present distribution of landslides (Carrara 1995). The development of these quantitative techniques with GIS tools, make possible to obtain, in short time, landslide susceptibility maps for large areas (Santacana 2003). Complete overview of the use of GIS for landslide susceptibility assessment can be found in Carrara (1991, 1995), van Westen (1994) and Guzzetti (2005). In Romania, landslide susceptibility assessment is a recent field of research, some contributions were given by Sandric 2003, 2005, Mihai 2005; Micu, in prep. In this study the methodology used to assess shallow landslides susceptibility is principal component analysis and logistic regression, using the procedure implemented in IDRISI Kilimanjaro 2 STUDY AREA Landslides and especially shallow landslides are a common feature for the Romanian hills and plateaus landscapes (Ielenicz et al., 1999; 2005). The study area is located in the Prahova Subcarpathians, along the Prahova River and is the administrative area of Breaza (Fig. 1). 2.1 General presentation The area is characterized by high slope gradient values between 15-30, specific to most of the hill- -265-
slopes, medium drainage density, with values between 0.3-1 km/km 2 (Popp 1939). The natural vegetation was replaced during the last two centuries by the secondary vegetation. The human activity has a relevant impact for this area, with a population density of about 50-75 inhabitants/km 2. The local economy is dominated by agricultural activities, based mainly on grazing and orchards. The main factors that control the spatial distribution of landslides are: lithology, slope gradient, land-use and land-cover. (Surdeanu 1998). After heavy rainfalls, the Subcarpathian area is affected by landslides, usually superposing on older landslide bodies. Balteanu (1986, 1997) estimated an average denudation rate generated by landslides to 0.5-10 mm/year. The shallow landslides are a frequent feature with an average recurrence of 5-7 years (Balteanu 1986). landslides have an important significance for the region, because of the losses caused annually in agriculture. The landslide inventory includes 156 landslides, mostly rotational slides (ca. 92%) and few translational slides (3%), complex landslides (3%) and earth flows (2%). The shallow landslides are spread especially in areas with small vegetation. The deep seated landslides are old landslides which extend over the entire hillslope (Fig. 2). They are either in a dormant state or have been partially reactivated during the rainfall events of 1982, 1997 and 2005. The methodology used to create the landslide inventory, is a combination between field observations and aerial image interpretation. The field observations were carried out for about three years (2006-2008), using a Garmin GPS receiver and topographical maps at a scale 1:5000. The aerial images are color orthophotos (acquisition time: 2005), grayscale air-photos from 1972, 1978, 1986 and aerial photos obtained from two flights taken on May 7th and 15th 2008 from a small aircraft, model Cessna F172H, with a Canon 400D camera with 10Megapixels resolution and Tamron AF 18-250mm Di-II lens Figure 1. Location of the study area. 2.2 Geology From the geological point of view, the study area is composed from a nappe system. These nappes belong to the Cretaceous and the Paleogene flysch areas, as well as to the Upper Cretaceous and Lower Miocene post tectonic sedimentary covers. The faults, as an effect of the tectonic transformations, show a system of parallel synclines and anticlines, oriented from the East to the West (Damian 2003). The most affected lithostratigraphic units are: Gura Beliei marls (Upper Cretaceous post tectonic sedimentary covers) which produce rotational and translational slides and earthflow; molasses deposits of Doftana (marls, clays, sandstones of Lower Miocene post tectonic sedimentary cover), affected by rotational and translational slides; Pucioasa deposits (marls, clays of Paleogene Flysch) affected by earthflows, rotational and translational slides. 2.3 Landslide inventory We use the term shallow landslides to define the small size rotational and translational slides and earth flows (Santacana et al., 2003). The shallow Figure 2. Shallow and deep-seated landslides on the hillslopes between Prahova Valley and Târsa Valley, reactivated in 2007 by human activity; oblique aerial picture, May 15 th 2008. 3 METHODOLOGY When many factors are available an issue of information redundancy is arising. To reduce the number of variables and to limit their interdependence we choose to use the PCA analysis (Carrara, 1995). It is well known that morphometrical parameters derived from DEM present a high interdependence, and this, often, leads to underestimation or overestimation of the results. The principal components analysis was used in order to reduce the redundant information from the variables and transform them from correlated variables into uncorrelated variables, named principal components (Gorsevski 2001). To model the susceptibility for -266-
landslides we used the logistic regression analysis, which calculates the probability that an individual pixel will contain a landslide. 3.1 Materials The data used to assess the landslide susceptibility is composed from: a DEM with a cell size of 10m, lithology, faults density, drainage density, land-use and land-cover, landslide inventory and several terrain parameters derived from DEM: slope gradient, aspect, profile curvature, plan curvature, energy of the relief. The digital elevation model was obtained by interpolation of contour lines and point elevation. The contour lines and elevation points were digitized from the topographical maps at a scale of 1:5000. The lithology and the faults were digitized from the geological maps at scales ranging from 1:25000 up to 1:50000. Land-use and land-cover was mapped by image interpretation from the color orthophotos (acquisition time: 2005) and from field surveys. The morphometrical parameters like slope gradient, plan curvature etc were derived from the DEM using Idrisi Kilimanjaro GIS software. 3.2 PCA analysis The PCA analysis was used to reduce de number of variables and to eliminate de redundant information from these variables. A number of 15 variables, represented by various topographical attributes, land-use and land-cover, lithology etc were introduced in the PCA analysis and only six principal components were obtained. The variables were previously classified and the classes were created to maximize the information and minimize the variance inside a class. Component 1 has 35.31%, component 2 has 28.33%, component 3 has 21.27% of the information, summarizing together ~85% of the information. The next three principal components have 8.73%, 5.08% and 1.28% of the information, summarizing together ~15% from the total information. The first four components have together 93.64% of the information and if we add the fifth component it reaches 98.72%. 3.3 Logistic regression analysis Logistic regression was applied for three different combination of the principal components obtained from the PCA analysis. The principal components were used as independent variables and the past and present shallow landslides as dependent variable. We have combined the first three components, then the first four components and the first fifth components. The result was the landslides susceptibility map for each of the combination, indicating the probability that a pixel will contain a landslide. 3.4 ROC analysis ROC analysis is suitable to assess the validity of a model that predicts the location of the occurrence of a class by comparing the predicted image with the actual presence of the class (Pontius, 2001). The ROC analysis was used to validate the results obtained from the logistic regression analysis (Gorsevski, 2001). The susceptibility map was introduced as the image to be validated and the past and present landslides were used as the reference image (Pontius, 2001). The ROC analysis was used for each susceptibility map obtained from the three logistic regression equations using three, four and five components Figure 3. ROC graph showing the validation of the three susceptibility maps 4 RESULTS Upon examination of the normalized eigenvectors, the elements with the absolute largest value indicated that the first principal component has a strong relationship with the slope gradient and the relief energy, the second principal component has a strong relationship with land-use, land-cover and lithology, the third principal component with width drainage density and plan curvature, the fourth principal component with aspect and the fifth principal component with the profile curvature and fault density. Logistic regression analysis has produced a series of three landslide susceptibility. The logistic regression for the combinations of the first three, four and five principal components with the past and present shallow landslides are: -267-
Figure 4. Susceptibility map for shallow landslides. Administrative area of Breaza, Prahova Subcarpathians, Romania. -268-
logit(sl)= -8.2836+0.643115*PC1+0.383955*PC2-0.104961*PC3 (1) logit(sl)= -7.3973+0.573523*PC1 + 0.299697*PC2-0.120301*PC3-0.267935*PC4 (2) logit(sl)= 7.8027+0.611648*PC1+0.330487*PC2-0.110811*PC3-0.310245*PC4+0.352099*PC5 (3) From all the equations from above we can observe that the susceptibility is increasing by a value of ~0.6 when the PC1 is present and by a value of ~0.3 when the PC2 is present and is decreasing by a value of ~-0.1 when the PC3 is present. By adding PC4 (Equation 2) the susceptibility is decreasing by a value of ~-0.3, but PC5 is increasing the susceptibility value by ~0.35. All three equations have produced susceptibility maps with a ROC value very similar (0.739-0.75). Figure 3 plots the ROC values for each susceptibility map. The best results were obtained by combining the first four principal components, having the highest ROC value of 0.75 (Fig. 3). The highest susceptibility values are present on the hillslopes of Tarsa River and on the left hand side of Prahova River. These area correspond marls and clay deposits, small vegetation and slope gradients between 10⁰ and 25⁰(Fig. 4) Low susceptibility areas correspond to flat areas on top of the terraces with deposits of sand and gravel and under the forest vegetation from the hillsslopes of Tarsa River. The cutting threshold between low susceptibility and high susceptibility is set by the logistic regression at 0.06 (Fig. 4). 5 DISCUSSION The map presented in the Figure 4 represents the landslides susceptibility for shallow landslides in the administrative area of Breaza. The map accurately describes areas that are prone to shallow landslides, but it also presents areas that are not prone to shallow landslides or are affected by deep-seated landslides. The terrace scarp of Prahova, on the east part of Breaza de Sus city, are partially identified with high susceptibility for shallow landslides, but these areas are covered by old deep-seated landslides, with a latent state of activity. The areas correctly identified with high susceptibility for shallow landslides are on the hillslopes of Tarsa River, between Irimesti village in the south and Valea Tarsei village in the north and at east of Surdesti village. The values of ROC are higher than 0.5, which indicate a random fir. The ROC value of 0.75 is comparable with ROC values reported in other fields are ranging from 0.71 to 0.89 for weather forecasting, from 0.75 to 0.97 for library information retrieval, 0.81 to 0.93 for medical imaging diagnosis, from 0.68 to 0.93 for material strength testing and from 0.55 to 0.93 for polygraph lie detection (Pontius, 2001). Special attention must be given to the criteria used in variables classification, so the maximum of information is retained in the class with minimum variance between classes. 6 CONCLUSION The PCA analysis can be used to identify the shallow landslides susceptibility with satisfactory results. ROC analysis enables scientist to validate a model ability to identify a specific location. The ROC values obtained are comparable with ROC values reported by scientist from other research fields. REFERENCES Balteanu, D. 1986. The importance of mass movements in the Romanian Subcarpathians. Zeitschrift fur Geomorphologie. Suppl. Bd.58: 173-190. Balteanu, D. 1997. Romania. In C. Embleton & C. Embleton- Hamann (eds.), Geomorphological hazards of Europe: 409-427. Amsterdam: Elsevier. Brabb, E. 1984. Innovative approaches for landslide hazard evaluation in IV International Symposium on Landslides, 1: 307-323, Toronto. Carrara, A., Cardinali, M., Detti, R., Guzzetti, F., Pasqui, V., Reichenbach, P. 1991. GIS techniques and statistical models in evaluating landslide hazard. Earth Surface Processes and Landforms 16: 427-445. Carrara, A., Cardinali, M., Guzzetti, F., Reichenbach, P. 1995. GIS technology in mapping landslide hazard. In Carrara, A. & Guzzetti, F. (eds.). Geographic Information Systems in Assessing Natural Hazards: 135-175. Dordrecht: Kluwer Academic Publishers. Damian, R. 2003. In Armas et al. Vulnerabilitatea versantilor la alunecari de teren in sectorul subcarpatic al Vaii Prahova, Bucuresti, Fundatia Romania de Maine. Ielenicz, M.,Patru, Ileana, Mihai, B.1999. Some geomorphic types of landslides in Romania, Transactions.Japanese Geomorphological Union., 20(3) : 287-296. Ielenicz, M., Ghincea (Clius) M., Patru, I.G. 2005. Subcarpatii Romaniei. Bucuresti: Universitara. Gorsevski, P. 2001. Statistical modeling of landslide hazard using GIS, Proceedings of the Seventh Federal Interagency Sedimentation Conference, March 25 to 29, 2001, Reno, Nevada.: 103-109 Guzzetti, F. 2005. Landslide hazard and risk assessment. PhD dissertation. Bonn, Germany Mihai, B. 2005. Muntii Timisului (Carpatii Curburii). Potential geomorfologic si amenajarea spatiului montan, Bucuresti: Universitatii din Bucuresti. Pontius, R., Schneider, L.C. 2001. Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA. Agriculture, Ecosystems and Environment 85: 239-248 Popp, N. 1939. Subcarpatii dintre Prahova si Dambovita. Studiu geomorfologic, Bucuresti: Societatea Regala Romana de Geografie -269-
Sandric, I. in Armas et al. Vulnerabilitatea versantilor la alunecari de teren in sectorul subcarpatic al Vaii Prahova, Bucuresti, Fundatia Romania de Maine. Sandric, I., 2005, Aplicatii ale teoriei probabilitatilor conditionate in geomorfologie, Analele Universitatii Bucuresti, Anul LIV-2005 Santacana, N., Baeza, B., Corominas, J., De Paz, A., Marturia, J. 2003. A GIS-based multivariate statistical analysis for shallow landslide susceptibility mapping in la Pobla de Lillet area (Easter Pyrenees, Spain). Natural Hazards 30: 281-295. Surdeanu, V. 1998. Geografia terenurilor degradate. I. Alunecarile de teren: Cluj-Napoca: Presa Universitara Clujeana van Westen, C.J. 1994. GIS in landslide hazard zonation: A review with examples from the Andes of Colombia. In Price, M.F. & Heywood, D.I. (eds.), Mountain Environments and Geographic Information Systems: 135-165. Taylor and Francis Publishers. -270-