Environmental Geology

, Volume 47, Issue 7, pp 956–966

Comparative evaluation of landslide susceptibility in Minamata area, Japan

Authors

    • Disaster Prevention Research InstituteKyoto University
    • Geotechnical InstituteHohai University
  • K. Sassa
    • Disaster Prevention Research InstituteKyoto University
Original Article

DOI: 10.1007/s00254-005-1225-2

Cite this article as:
Wang, H.B. & Sassa, K. Environ Geol (2005) 47: 956. doi:10.1007/s00254-005-1225-2

Abstract

Landslides are unpredictable; however, the susceptibility of landslide occurrence can be assessed using qualitative and quantitative methods based on the technology of the Geographic Information Systems (GIS). A map of landslide inventory was obtained from the previous work in the Minamata area, the interpretation from aerial photographs taken in 1999 and 2002. A total of 160 landslides was identified in four periods. Following the construction of geospatial databases, including lithology, topography, soil deposits, land use, etc., the study documents the relationship between landslide hazard and the factors that affect the occurrence of landslides. Different methods, namely the logistic regression analysis and the information value model, were then adopted to produce susceptibility maps of landslide occurrence. After the application of each method, two resultant maps categorize the four classes of susceptibility as high, medium, low and very low. Both of them generated acceptable results as both classify the majority of the cells with landslide occurrence in high or medium susceptibility classes, which could be believed to be a success. By combining the hazard maps generated from both methods, the susceptibility was classified as high–medium and low–very low levels, in which the classification of high susceptibility level covers 6.5% of the area, while the areas predicted to be unstable, which are 50.5% of the total area, are classified as the low susceptibility level. However, comparing the results from both the approaches, 43% of the areas were misclassified, either from high–medium to low–very low or low–very low to high–medium classes. Due to the misclassification, 8% and 3.28% of all the areas, which should be stable or free of landsliding, were evaluated as high–medium susceptibility using the logistic regression analysis and the information value model, respectively. Moreover, in the case of the class rank change from high–medium susceptibility to low–very low, 35% and 39.72% of all mapping areas were predicted as stable using both the approaches, respectively, but in these areas landslides were likely to occur or were actually recognized.

Keywords

LandslidesSusceptibilityLogistic regression analysisInformation value modelGeographic Information Systems

Introduction

Landslide is a general term used to describe the down-slope movement of soil, rock and organic materials under the influence of gravity. Landslides, ranging from shallow debris flows to deep-seated slumps, commonly occur with other major natural disasters, such as earthquakes and floods, which exacerbate relief and reconstruction efforts. Expanded development and other land uses have also increased the incidence of landslide disasters. As one of the major natural hazards, landslides could cause significant life losses and property damage: loss of lives, destruction of homes, businesses, and public buildings, undermining of bridges, derailment of railroad cars, coverage of clam and oyster beds, interruption of transportation infrastructure, and damage to utilities, etc. Recent disasters in the Kyushu Island of Japan, especially during the months of July and August of the rainy season, include torrential downpours around Kumamoto and Nagasaki in 1972, torrential downpours in Nagasaki in 1982, heavy rainfall in Izumi in 1997 and in Minamata in 2003, and many others. Human casualties due to rainfall-induced landslides of these disasters include 543 deaths in the 1972 event, 299 deaths in the 1982 event, and 21 deaths in 1997. Debris flows occurred at Hougawachi and Fukugawa of Minamata area due to localized rainfall which was caused by a seasonal rain front from 20th July 2003, leaving 19 persons dead, 7 persons injured and 15 houses damaged . Landslides triggered by heavy rainfall are the most common throughout Kyushu Island, but few attempts have been made to predict them effectively to remedy the damage.

Landslides are not currently amenable to risk assessment as there is no basis on which the probability of landslide occurrence within a given time period can be determined. Hazard assessment is estimation of an area’s susceptibility to landslide occurrence based on a few key factors. These are each capable of being mapped and allow land areas to be evaluated on their relative susceptibility to landslides. Landslide hazard and risk zoning and mapping for urban and rural areas is widely performed around the world, as the GIS technique was applied to deal with this problem decades ago (Carrara et al. 1991, 1992; van Westen 1993; Hansen et al. 1995; Cruden and Fell 1997; Aleotti and Chowdhury 1999; Gupta and Anbalagan 1997; Dai et al. 2001; Lee et al. 2002; Alcantara-Ayala 2004; Chau et al. 2004). A landslide susceptibility map can rank slope stability of an area into categories that range from stable to unstable, based on an estimated significance of causative factors in inducing instability. Engineers, earth scientists, and planners are interested in assessing landslide susceptibility due to the following considerations: (1) landslide hazard maps identify and delineate unstable hazard-prone areas, so that environmental regeneration programs can be initiated to adopt suitable mitigation measures; (2) These maps help planners to choose favorable locations for site development schemes, such as building and road construction. Even if hazardous areas cannot be avoided altogether, their recognition in the initial stages of planning may help to adopt suitable precautionary measures. Presently, tools for handling and analyzing geospatial data (i.e., the Geographic Information Systems (GIS)) may facilitate the application of quantitative techniques in landslide hazard assessment and mapping. They also allow the management of several themes concerning instability factors. Hence, GIS-based techniques are best suited to the study of landslide susceptibility, instead of traditional manual handling and processing of the data in landslide susceptibility assessment (Chung and Fabbri 1995; Guzzetti et al. 1999). Most of the proposed methods are based upon a few widely accepted principles or assumptions (Varnes 1984; Yin and Yan 1988; Carrara et al. 1991; Hutchinson 1995; Turner and Schuster 1996), particularly the well-known and widely applied principle, “the past and the present are keys to the future”. The detailed classification of assessment methodologies is somewhat subjective and depends upon the choice of the aspects to be emphasized.

This study aims to evaluate the landslide incidence and produce susceptibility maps portraying their spatial distribution at a scale of 1:50,000 in Minamata area of Japan, employing the multiple logistic regression method and the information value model based on GIS techniques as shown in the flowchart (Fig.1). This study also aims to compare two methods predicting landslide susceptibility with respect to their accuracy and reliability in the classifications of areas prone to landsliding.
Fig. 1

The procedure of landslide susceptibility assessment in the study area

Minamata is a city of about 32,000 persons, on the coast of the southern Kyushu Island. The site lies between the latitudes 32°6′30′′N and 32°14′10′′N, and longitudes 130°21′40′′E and 130°36′E, and covers an area of 162.6 km2 . The bedrock geology of the study area consists mainly of andesite from Pliocene to Early Pleistocene, such as lava, tuff breccia, etc. The topography of the area of interest is characterized by mountainous areas with few hilly lands in northwest Minamata. There are various types of destructive earthquakes in the Kyushu region, including earthquakes that occurred at sea, and at shallow locations on land; however, there were few reports about landslides triggered by earthquakes in the study area. With its average annual temperature of 16.8°C and its average precipitation at 1,711 mm, it is characterized by a temperate oceanic climate with abundant rainfall.

Spatial databases

The evidence of past landslides is among the most important factors in the prediction of future failures. In this study, landslide factors involving many characteristics, both natural and man-induced, of the environment that are directly or indirectly related to the causes of landsliding in a given region, such as elevation, slope, aspect, lithology, land cover, soil deposit, and a vast collection of geospatial data, were observed.

The databases consist of products ranging from vector and raster GIS layers, to 1:30,000-scale aerial photographs (Table 1). The vector data consist of point, line, and polygon vector layers, such as lithological units, soil deposits, land use and landslide inventory, while the raster GIS is comprised of cell-based layers, such as digital elevation models (DEM). The sets of raster data are composed of collected air photos available at Geographical Survey Institute, Japan. Many of the vector data within the database of the study area were digitized from 1:25,000 to 1:50,000 scale thematic maps, as well as other government-issued maps. From these products, contour lines, soil deposits, land use and lithology were digitized and stored as ArcView shape files. In addition to the air photographs taken in 1999 and 2002 used to detect landslides, other raster datasets deal with the terrain and topography of the landscape. Included in the terrain/topography data is the DEM, at 25-m grids. Additionally, a host of information is derived from the 25-m grid DEM, such as elevation, aspect, and slope. As digitizing all the maps is highly laborious and time intensive, data already in a digital format were sought whenever possible.
Table 1

The description of datasets in the spatial database constructed

 

Attribute

Description

Code of subtypes

Attribute

Description

Code of subtypes

Attribute

Description

Code of subtypes

Lithology

Unconsolidated sediment

Reclaimed land

1

Soil deposit

Lithosol

1

Slope

0–10

1

Gravel sand, and mud (lowland sediments)

2

Ando soil

2

  

Consolidated sediment

Alternating beds of sandstone and mudstone, sandstone, and mudstone

3

Dry brown forested soil

3

10–15

2

Sandstone and conglomerate

4

Brown forested soil

4

15–20

3

Mainly sandstone with subordinate slump deposit

5

Red soil

5

20–30

4

Mainly slump deposits

6

Yellow soil

6

30–40

5

Limestone

7

Dark red soil

7

40–50

6

Volcanic rock

Pyroclastics

8

Gray lowland soil

8

50–60

7

Andesitic rock (lava, tuff breccia)

9

Other soil

9

 

≥60

8

Land use

 

Irrigated field

1

Elevation

0–100

1

Aspect

Flat

1

 

Non-irrigated farmland

2

  

N

2

 

Garden

3

100–200

2

NE

3

 

Forested land

4

200–300

3

E

4

 

Grass land/barren land

5

300–400

4

SE

5

 

Residential area

6

400–500

5

S

6

 

Water

7

500–600

6

SW

7

Aerial photograph

 

Photos taken in 1999

Raster

≥600

7

W

8

 

Photos taken in 2002

NW

9

Shallow landslides and debris flows that occurred during the days of high intensity rainfall, or shortly after were dominant. A landslide inventory defining the type and activity of all landslides, as well as their spatial distribution, is essential before any analysis of the occurrence of landslides and their relationship to environmental conditions is undertaken. Pre-1993 landslide inventory was primarily compiled on the basis of work from the Asikita Civil Engineering. Aerial photographs taken in 1999 and 2002 at a scale of 1:30,000 were interpreted for landslide occurrence, and then the locations of landslide source areas were plotted on the map of landslide inventory. Landslides were identified in Hougawachi and Fukugawa areas using aerial photographs taken after the disaster of July 2003. All the landslides of different periods were merged to create a cumulative landslide distribution map. A total of 160 landslides was identified in four periods, in which 11 landslides were detected from the aerial photographs taken in 1999 and 39 landslides from the ones taken in 2002 (Fig. 2). As the landslide inventory was displayed as point features, the buffer distance was determined based on the range of width of landslide sources to create a raster map of landslide distribution.
Fig. 2

Landslide inventory in Minamata area, Japan

Models of susceptibility mapping

Approaches of landslide susceptibility assessment can be qualitative or quantitative, and advantages and disadvantages of different methods are discussed in detail by Aleotti and Chowdhury (1999). Landslide susceptibility assessment by means of the GIS is based on the analysis of relationship between past events (response variable) and various local ground parameters (predictor variable) at a medium scale (1:25,000 to 1:50,000). In the study, the logistic regression analysis and the information model were adopted to produce the susceptibility maps, and then the comparison was carried out to assess the capacity of two models to predict future landslide occurrence.

Logistic regression analysis

The logistic regression is useful for predicting the presence or absence of a characteristic or outcome based on values of a set of predictor variables. It is similar to a linear regression model and is suited to models where the dependent variable is dichotomous. Binomial (or binary) logistic regression is a form of regression which is used when the dependent is a dichotomy and the independents are of any type, while multinomial logistic regression exists to handle the case of dependents with more classes than two. The logistic model assumes a dichotomous dependent variable Y with probability πi,
$$ \pi _i = \frac{{\exp \left( {\eta _i } \right)}} {{1 + \exp \left( {\eta _i } \right)}} $$
(1)
$$ \ln \left( {\frac{{\pi _i }} {{1 - \pi _i }}} \right) = \eta _i = {\rm X}_i \beta $$
(2)
where β and Xi are the coefficient and the observed value of the ith case, respectively. In this study, the dependent variable is the absence or presence of a landslide, and the independent variables are the factors which affect the landslide occurrence, including slope gradient, elevation, aspect, lithology, soil deposit and land use. Then, the binomial logistic regression is adopted to predict the probability of landslide occurrence. A landslide as a dependent variable (Y) can be represented through a binary variable, in which Y=1 means landslide occurrence or Y=0 indicates no landslide occurrence. The logistic model can be written in the simplest way as
$$ P\left( {Y = 1} \right) = {1 \mathord{\left/ {\vphantom {1 {\left( {1 + {\text{e}}^{ - z} } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {1 + {\text{e}}^{ - z} } \right)}} $$
(3)
where P (Y=1) is the probability of landslide occurrence given different independent variables, and z is defined as:
$$ z = \alpha + \beta _1 X_1 + \beta _2 X_2 + \cdots + \beta _n X_n $$
(4)
where βi (i=1,2,... ... n) is the coefficient estimated from the sample data and α is the intercept, Xi (i=1,2, ... ...n) represents n independent variables.

Information value model

The information value model can also predict the susceptibility considering contributions of all the factors to landsliding (Yan et al. 1988; Wu et al. 2001), which can be expressed as the information value shown in the following:
$$ I_{A_i \to B} = {\text{ln}}\frac{{P\left( {B/A_i } \right)}} {{P\left( B \right)}} $$
(5)
where P(B/Ai) is the probability for a landslide to occur at the presence of attribute Ai and P(B) the general probability for the landslide occurrence.
Due to the difficulty in assessing the probabilities of both P(B/Ai) and P(B), simple frequencies are used in the above formulation instead of calculating the contribution of different factors to landsliding, and \(I_{A_i \to B} \) can be expressed as follows:
$$ I_{A_i \to B} = {\text{ln}}\frac{{N_i /N}} {{S_i /S}} $$
(6)
where S is the total number of cells, N is the number of cells with landslide occurrences; Si is the number of cells with the presence of attribute Ai and Ni is the number of cells in which the landslide occurred in the presence of attribute Ai Hence, \(I_{A_i \to B} \) means the information value for the contribution of attribute Ai to landslide occurrences. The larger the value, the more the contribution to landslide susceptibility.
Finally, the total information values, as the susceptibility of each cell x, is calculated according to the following equation:
$$ {\text{Susceptibility index (}}x{\text{)}} = \sum\limits_i {I_{A_i \to B} } = \sum\limits_i {\ln \frac{{N_i /N}} {{S_i /S}}} $$
(7)

Results

After the construction of spatial databases, all the thematic maps, including elevation, slope, aspect, lithology, land cover, soil deposit and landslide inventory, were overlaid to generate 239,589 cells with intra-unit homogeneity and inter-unit heterogeneity, in 2,155 of which landslide occurrence can be observed. Each cell is delineated with a variable indicating presence or absence of landslide occurrence, in which the independent variables categorized represent different factors determined above. Thus, the sample datasets for logistic regression consisted of 2,000 cells with landslide occurrence and the 4,000 cells in the stable areas. The datasets of 6,000 cells were then input to the logistic regression algorithm within SPSS, a statistical software package, to obtain coefficients for the logistic regression model (Table 2).
Table 2

Regression coefficients estimated for the logistic regression model

Variable

Categories

Coefficient

Variable

Categories

Coefficient

Lithology

Reclaimed land

19.519

Slope

0–10

−.777

Gravel sand, and mud (lowland sediment)

−.628

10–15

−.006

Alternating beds of sandstones and mudstone, sandstone, and mudstone

−.778

15–20

−.639

Sandstone and conglomerate

5.564

20–30

−.347

Pyroclastic

−.672

30–40

−.332

Andesitic rock (lava, tuff breccia)

 

40–50

−.845

Soil deposit

Lithosol

−.397

50–60

−.894

Andosoil

−.218

≥60

 

Dry brown forested soil

−1.526

Elevation

0–100

6.167

Brown forested soil

−1.759

100–200

5.154

Red soil

−.725

200–300

4.010

Yellow soil

−1.714

300–400

2.904

Dark red soil

−.920

400–500

2.029

Gray lowland soil

 

500–600

−.213

Other soil

 

≥600

 

Land use

Irrigated field

−.282

Aspect

N

.241

Non-irrigated farmland

−.554

NE

−.253

Garden

−.291

E

.032

Forested land

−.514

SE

−.115

Grassed land/barren land

−.092

S

.835

Residential area

 

SW

.863

Water

 

W

.574

Constant term

 

18.238

NW

−.089

Table 3 demonstrates a high degree of accuracy for the logistic regression model, and it can be adopted as a predictive model to produce a reliable map of landslide susceptibility in the area of interest. Using the coefficients estimated for the model, the probability of landslides, given all the independent variables, was calculated through all 239,589 cells by Formulas 3 and 4, and a susceptibility map was produced. The resultant hazard zonation map was then classified into four susceptibility levels, namely very low (<0.25), low (0.25–0.5), medium (0.5–0.75), and high (>0.75), as shown in Fig. 3.
Table 3

The description of the logistic regression model to predict the landslide susceptibility

Observed

Predicted

Landslide occurrence

Percentage Correct

.00

1.00

Step 1

Landslide occurrence

.00

3,593

407

89.8

 

1.00

465

1,535

76.8

Overall percentage

79.3

Fig. 3

The susceptibility map of landslide occurrence using the logistic regression approach

Using Eq. 6, the information value of each factor was calculated (Table 4). It shows that the areas with deposits of pyroclastics and andesitic rock (lava, tuff breccia) are susceptible to landsliding, while there is no landslide occurrence in the sites covered by sandstone with subordinate slump deposit, slump deposit and limestone. Forested land and Lithosol are the factors more prone to landsliding, whereas there is no landslide observed in the areas of red soil. With respect to topographical factors, the height of 500–600 m, the slope of 15–20° are susceptible to landslide occurrence, while this aspect has no obvious relation to landsliding as their information values are almost the same.
Table 4

The calculation result of information values

Factor

Landslide occurred

The total number of cells

Information value

Factor

Landslide occurred

The total number of cells

Information value

Count

Count

Lithology

   

Elevation

   

1

16

457

0.78297

0–100

243

38643

1.196171

2

336

18799

0.730195

100–200

707

42484

0.644304

3

392

26492

0.773942

200–300

502

46467

0.888283

4

20

1791

0.955815

300–400

462

42442

0.889758

5

0

1043

0

400–500

201

34897

1.231358

6

0

163

0

500–600

25

23459

1.917944

7

0

72

0

≥600

15

11197

1.621627

8

59

12610

1.221969

Slope

   

9

1332

177179

1.59432

0–10

207

39663

1.302652

Land use

   

10–15

136

26987

1.265309

1

332

22832

0.795662

15–20

185

37298

1.320005

2

91

4841

0.811083

20–30

833

67238

0.748031

3

193

11254

0.788979

30–40

538

44903

0.828757

4

1489

192132

1.67474

40–50

191

16312

0.901843

5

33

3100

0.961246

50–60

41

3567

0.941713

6

17

5310

1.271176

≥60

24

3621

1.07283

Soil deposit

   

Aspect

   

1

10

9908

1.686649

Flat

389

45784

1.034427

2

295

35329

1.038841

N

182

25463

1.10253

3

121

27152

1.322521

NE

154

21422

1.092809

4

881

119124

1.280108

E

186

21198

1.010219

5

0

462

0

SE

276

23672

0.887893

6

221

17462

0.869596

S

367

25830

0.794735

7

232

16559

0.834136

SW

299

28409

0.926313

8

319

10820

0.616735

W

200

30961

1.161779

9

76

2773

0.750127

NW

102

16850

1.149176

The total number of cells is 239,589, and the number of cells with landslide occurrences 2,155

Finally, the information model was adopted to predict the susceptibility of landslides. Frequencies versus information value curves for all the 239,589 units are plotted in Fig. 4 to show frequencies or percentages of different information levels of the constituent units. In this map produced, boundaries for landslide hazard zonation were determined on the basis of distribution of frequencies against the information value, and correspondingly the susceptibility was categorized in four classes shown in Fig. 5, namely very low (<7.15), low (7.15–7.55), medium (7.55–7.88), and high (>7.88), which allows the comparison to the result from logistic regression analysis.
Fig. 4

Distribution of frequency versus information value

Fig. 5

The susceptibility map of landslides using the information value model

Examination of reliability

Two susceptibility maps were generated from the logistic regression analysis and the information value model. It was shown that both approaches classify different susceptibilities of landslide occurrence in the study area, and the misclassification exists inevitably, which could be discussed to validate the combined resultant maps useful for planning, land development, and decision-making. Some work has addressed the issue of reliability of susceptibility maps, and the validation can be implemented to test models used and susceptibility results. It is normally assumed that the future landslides will occur in places similar to the existing landslides of the study area; if the susceptibility maps produced coincide reasonably well with the latter, they are considered as acceptable. In the present study, all the landslides of different periods were considered in order to obtain the appropriate population for statistical analysis and sampled to predict the landslide susceptibility; hence, different models were evaluated to compare the resultant maps with the landslide inventory.

Table 5 exposes the validity of two models employed to assess the areas prone to landsliding based on the GIS technology. In the case of the logistic regression analysis, of the 2,000 observed landslide grid cells, 1,535 were correctly predicted and 465 misclassified with a concordance rate of 76.8% while 0.5 as a classification cutoff value. Of the 4,000 grid cells without observed landslides, 3,593 were correctly predicted and 407 misclassified with a concordance rate of 89.8%. After the application of the logistic regression to the study area, it is found that 14.5% of the total area is classified as high and medium susceptibility classes, 7.1% and 7.4% respectively. On the other hand, 85.5% of the study area is classified as very low and low levels, in which 11.3% was predicted as low susceptibility of landslide occurrence.
Table 5

Comparison of two methods to assess the landslide susceptibility

Logistic regression analysis

Comparison

Information value model

Susceptibility classes

Percentage of cover area

Acceptable classification

Misclassification

Susceptibility classes

Percentage of cover area

High

7.1

6.5%

11.28%a

High

3.1

Medium

7.4

Medium

6.68

Low

11.3

50.5%

74.72%a

Low

11.4

Very low

74.2

Very low

78.82

aIndicates misclassification together by both the logistic regression and the information value model

Compared with the logistic regression analysis, the information value model only predicted 1,530 cells, 71% of all the areas with landslides, as the areas in which landslides occurred, and 585 were misclassified to the cells without landslide occurrences. Of the total area, Fig. 5 illustrates that 9.78% is classified as high and medium susceptibility classes, 3.1% and 6.68%, respectively. In the study area, 90.22% is classified as very low to low level, in which 11.4% was predicted to have low susceptibility of landslide occurrence.

Two approaches produced different susceptibility maps, and both of them generated acceptable results, as both classify the majority of the cells with landslide occurrence in high or medium susceptibility classes, which could be believed to be a success; however, in the combination of the susceptibility maps, the susceptibility was categorized into two classes, of which one included high and medium levels and another was low and very low, correspondingly, herein called high–medium and low–very low susceptibility. As can be seen in Fig. 6, a map, combining the results from different methods, can be observed as acceptable classification in the high–medium susceptibility level of landsliding within the study area. Those areas whose classification can be acceptable as a high–medium susceptibility level, cover 6.5% of the area, in which 66.46% and 44.9% represents both high- and medium-level susceptibility using the information value model and the logistic regression analysis, respectively. Upon the investigation of the accepted pixels of the two methods, hence, it was seen that the logistic regression analysis overestimated the susceptibility to landsliding. As shown in Fig. 3, more pixels were predicted to high and medium susceptibility to landsliding around or near the Minamata City, as most of the landslides were recorded in the northwestern portion of the city, which resulted in the areal concentration of prediction. The information value model, as an approach in assessing the landslide hazard without assumption, can explore the relationship among unstable conditions and factors contributing to the landslide occurrence. Using the information value model, it was found that the prediction of high- and medium-level susceptibility could be accepted, taking into account all the affecting factors. However, Fig. 4 showed the areas of high- and medium-level susceptibility at the elevation of up to 500 m was overestimated, because the information value of the elevation up to 500 m to landsliding was extremely weighted.
Fig. 6

The combined resultant map of high–medium susceptibility (in red color)

Another acceptable classification can also be presented in the low susceptibility class. It was also found that the areas predicted to be unstable cover 50.5% of the total area, corresponding to combination of the units at low and very-low susceptibility to landsliding using two methods. In the study proposed, whichever approach was applied to assess stable areas or landslide free areas, the classification of susceptibility can be acceptable, although nearly 74.72% of the total areas were misclassified taking into account the two resultant maps.

In the case of combined classification of high–medium and low–very low susceptibility, hence, all the susceptibility in the research area can be acceptable at the confidence of 57%, including both the areas prone to or free of landsliding, while 43% of the areas were misclassified, in which 11.28% were unacceptably classified into high–medium classes of landslide susceptibility and 74.72% was underestimated to be free of landslide occurrence. Due to the misclassification of the class rank from low–very low to high–medium, 8% and 3.28% of the area were produced, and 35% and 39.72% was underestimated in the landslide susceptibility, using both the logistic regression analysis and the information value model, respectively. In the case of the class rank change from high–medium to low–very low level, it is noted that the prediction indicates either mapping errors or a model that fails to take account of the factors that cause failure in that specific environmental setting. For the sake of misclassification, it could be explained that the logistic regression analysis predicted the susceptibility only from the sample data, in which the data sets without landslide occurrence were obtained randomly. Moreover, the information value model overestimated the contribution of elevation up to 500 m to landslide occurrence.

Whichever method may be adopted to assess the landslide susceptibility, there is no way to compare hazard zones at different sites or to determine the likelihood that a high hazard area, for example, two times or ten times more likely to fail in the future than low hazard areas. It should be stressed that these relative hazard zones are based on the existing landslides and conditions influencing their occurrence in a specific area. This misclassification of susceptibility limits the value of landslide hazard maps and may jeopardize their usefulness for planning, land development, and decision-making. Indeed, where slope failure or the areas prone to landsliding are recognized, actions can be taken and proper regulations can be established before planning or land development takes place.

Conclusion

Landslides are unpredictable, however, the susceptibility assessment of landslide occurrence can be determined using different methods based on the GIS technology. This study presented a landslide inventory map obtained from the previous work in the Minamata area, and from the interpretation of aerial photographs taken in 1999 and 2002 and field investigation, during which a total of 160 landslides waqs identified in four periods, of which 11 landslides were observed from the aerial photographs taken in 1999 and 39 landslides from the ones taken in 2002.

Following the construction of geospatial databases, including lithology, topography, soil deposits, land use, etc., the study documents the relationship between landslide hazard and the factors that affected the occurrence of landslides. The areas with deposits of pyroclastics and andesitic rock (lava, tuff breccia) are susceptible to landsliding, while there is no landslide occurrence in the locations covered by sandstone with subordinate slump deposit, slump deposit and limestone. Forested land and Lithosol are the factors more prone to landsliding, whereas there is no landslide observed in the areas of red soil. With respect to topographical factors, the height of 500–600 m and the slope of 15–20° are susceptible to result in landsliding, while the aspect has no obvious relation to landslide occurrence as their information values are almost the same.

Different methods were adopted to produce the susceptibility maps of landslides: the logistic regression analysis and the information value model. With each method, the resultant maps can be categorized in four classes of susceptibility, such as high, medium, low and very low. By combining the hazard maps generated from both methods, the susceptibility was classified into high–medium and low–very low levels, in which classification a high-susceptibility level can be acceptable, cover 6.5% of the area, while the areas predicted to be unstable are 50.5% of the total area and can be acceptable as the low-susceptibility level. However, comparing the results from both the approaches, 43% of the areas were not acceptably classified, either changed from high–medium to low–very low or low–very low to high–medium susceptibility. As for the misclassification, 8% and 3.28% of all the areas, which should be free of landsliding, were evaluated to high–medium level of landslide susceptibility using the logistic regression analysis and the information value model, respectively. In the case of the class rank change from high–medium to low–very low susceptibility, moreover, 35% and 39.72% of all mapping areas were predicted as stable, corresponding to the logistic regression analysis and the information value model, respectively, but in these areas landslides were likely to occur or were actually recognized . The misclassification of the landslide susceptibility cannot be avoided throughout the study proposed; however, the results, together with the susceptibility maps, would allow planners to determine the actions for mitigating landslide effects, avoid development in susceptible areas, and recommend adjustments to existing land use and restrictions for future land use.

Acknowledgements

This work was performed under the Grant-In-Aid for scientific research from the Ministry of Education, Culture, Sports, Science and Technology (MEXT), Japan. The first author was indebted to the Postdoctoral Fellowship for foreign researchers provided by the Japanese Society for the Promotion of Science (JSPS). Also, the first author would like to gratefully acknowledge the additional support from the China Postdoctoral Science Foundation. They would like to express their sincere appreciation to an anonymous reviewer for his insightful comments and valuable work to both the language and the research.

Copyright information

© Springer-Verlag 2005