Keywords

1 Introduction

Landslides are unpredictably destructive catastrophes that cause severe risks to construction, people's lives, economic growth, and business. These are common catastrophic disasters, resulting in extensive damage and putting individuals and societies at risk (Zhao et al. 2021; Chang et al. 2023). Also, landslides make up roughly twelve percent of any natural catastrophe, causing more than eleven thousand mortality yearly, with a financial loss of $4.5 billion (Froude and Petley 2018; WHO 2022). Japan is particularly susceptible to landslides because it has exceptional topographical and geological circumstances, which include unstable ground and granite structures, high rains, as well as tectonic activity. Arguably, the most prevalent landslide-prone region in Japan is the southern region, which is located near the Pacific Ocean. Also, there are several landslides and flash floods there every year.

The Chugoku Mountain in Western Japan is well-known for its hilly environment and sloping topography. These locations have been recognized as particularly vulnerable to landslides and have been categorized as high-risk areas of these types of natural disasters by many studies on landslide modeling. In addition, the global incidence of landslides is now much affected by changing climates. The susceptibility and occurrence of landslides are significantly impacted by climate change-induced components, like intense precipitation, warmer temperatures, and rising sea levels (Pei et al. 2023; Qiu et al. 2019). To offer reliable and trustworthy estimates, it is crucial to integrate climate variables when developing Landslide Susceptibility Mapping (LSM) models.

According to previous studies, landslides are susceptible to occurring in many geographic regions. Using several landslide model predictions could lessen the harm resulting from landslides (Pradhan and Lee 2010). The LSM has been simpler to create in the last 20 years due to significant advancements in computational capacity, remote sensing (RS), and geographic information systems (GIS) (Mirus et al. 2020; Moragues et al. 2023). Such developments are transforming the field of landslide research.

The disastrous landslides in Hofu City in 2009 that claimed over 14 lives underscored the need to analyze landslide susceptibility in Western Japan precisely (ICHARM 2020). This analysis is essential for reducing the destruction of structures and fatalities. As a result, several academics have used mapping landslide susceptibility as an essential initial phase in the strategies, assessment, and mitigation of landslides. This technique has been used to create landslide vulnerability maps in several publications, including (Yu et al. 2021; Cheng et al. 2022; Kohno and Higuchi 2023). Nevertheless, the landslide susceptibility assessments developed in Japan have mainly concentrated on significant transportation lines and particular sites in extremely susceptible provinces. In order to address landslide threats effectively, establishing and creating the LSM at various spatial resolutions is crucial (Persichillo et al. 2016). This modeling could help with the deployment of early warning and surveillance systems, strategic planning for risk reduction, and the development of feasible and effective land development strategies for policymakers (Roccati et al. 2021).

In general, the LSM was generated into two main approaches: knowledge-driven and data-driven. The knowledge drive was heavily dependent on expert judgment (e.g., the Analytical Hierarchy Process (AHP)), whereas this kind of approach is useful when the inventory data are limited. The AHP, for example, is a widely used knowledge-driven approach for the LSM worldwide (e.g., Kohno and Higuchi 2023; Zangmene et al. 2023; Panchal and Srivastava 2022). Given the rapid advancement of computational and GIS approaches, various complexities of data-driven models for performing LSM have recently been proposed. There are two kinds of data-driven models that exist: machine learning and statistical models (Yu et al. 2021; Cheng et al. 2022). Recently, machine learning has been used so vast and generally has high predictive accuracy (Cheng et al. 2022). However, their fundamental principles are complex and hard to describe directly. As a result, these methods are unsuitable for studying the interaction among landslides and variables (Cheng et al. 2021). Therefore, in this chapter, we proposed a statistical model approach that could overcome those challenges.

Numerous GIS and RS approaches have been used to evaluate the LSM over the past decade. Quantitative and qualitative techniques were employed extensively with such approaches (Cheng et al. 2021; Kohno and Higuchi 2023). The LSM using logistic regression (LR) represents a primarily quantitative approach that provides a practical and useful means of mapping landslide susceptibility groups through numerous proxies. Investigators worldwide have frequently used the LR approach to examine the link between historical landslide disasters and various categories of possibly significant elements. The LR model is especially well-suited for evaluating landslide vulnerability in hilly settings, claim Ali Mohammadi et al. (2014). Even though a severe landslide hit Hofu City in 2009, a lack of research assesses the LSM. This assessment is crucial for the study area's local strategic disaster risk reduction planning. Therefore, this chapter aims to use an LR model to develop the LSM and use GIS and RS approaches to identify the physical elements that influence the likelihood of an incident using the pilot study of Hofu City, Japan.

2 Database and Methodology

2.1 Study Area

This assessment was carried out in Hofu City, Yamaguchi Prefecture, Chugoku region, the westernmost region of Honshu, Japan's largest island. The entire area is 188.59 km2 and has an elevation ranges from 0 to 630 m. Hofu is south of Yamaguchi and faces the Sea of Suo in the Seto Inland Sea. From the rocky slopes of Migita in the north to Mount Ohira in the east and an island mountain on the southern seaside, the city is nearly encircled by mountains and rice fields. The Saba River flows from Tokuji town (legally part of Yamaguchi City) to the Seto Sea. Hofu City faced severe landslides due to extreme rainfall in July 2009. Moreover, this area also faced depopulation from 2010 to 2020 by 2.3%. Figure 2.1 depicts the research site.

Fig. 2.1
A landslide inventory map of Hofu City in Japan. Landslide occurrence is concentrated at the central region along a southwest northeast diagonal of medium to high elevation.

Landslide inventory map in Hofu City, Japan

Full size image

Furthermore, Hofu City has an annual average rainfall of 1652 mm, with the highest precipitation in summer (μ = 222 mm/ month) and the lowest during the winter season (59 mm/month) (JMA, 2022). Throughout 1990–2022, July was recorded as the highest monthly average rainfall (i.e., 300mm/ month), while December was the lowest (52 mm/month) (JMA, 2022). Moreover, the highest monthly rainfall in Hofu City was 812 mm/month in July 2009; therefore, a severe landslide occurred during this time (JMA, 2022). Also, the study area has a dominant granitic rock, followed by marine and marine sediment and pelitic schist (GSI 2014), whereas the inventory landslide dataset is mainly located along the granitic rock. Moreover, according to a previous study by Chigira et al. 2011 and Yamashita et al. 2017, the incidents in the study area mainly occurred in weathered granite, whereas this type of rock is prone to sliding due to extreme rainfall. Yamashita et al. 2017 also revealed that the granite areas had thinner soil and more considerable shear strength, which could cause a slip plane.

2.2 Dataset

2.2.1 Map of Landslide Inventory

This dataset was obtained from the Ministry of Land, Infrastructure, Transport and Tourism (MLIT), Chugoku District, Japan. The data consists of landslide locations and areas. This inventory map recorded landslide events on 21 July 2009 (Fig. 2.1). We utilized this dataset for validation of the susceptibility map.

2.2.2 Remote Sensing Data

In this chapter, we used the ALOS Advanced Visible and Near Infrared Radiometer type 2 (AVNIR-2) satellite dataset to get the land cover information within the study area. This collection contains data from a visible and near-infrared (NIR) radiometer applied to observe terrestrial and coastal regions., which has 10m spatial resolution images. Table 2.1 explains the attribute details of the ALOS AVNIR-2 images.

Table 2.1 AVNIR-2 details (JAXA 2008)
Full size table

In this chapter, ALOS AVNIR-2 was utilized to build a land cover map within the pilot sites. The image of ALOS AVNIR-2 acquired on July 14, 2009, was processed using ArcGIS 10.8.2.

2.2.3 Rainfall Data

Rainfall data used in this chapter were obtained from the Japanese rain-gauged dataset called the Automated Meteorological Data Acquisition System (AMEDAS). This dataset is freely available and could be downloaded at http://www.jma.go.jp/jma/index.html. Rainfall data on a specific date (21 July 2009), both daily and hourly, were collected from rain gauges in Hofu City and surrounding cities located near the study area. The rainfall data in each station was downloaded and interpolated to create an isohyet map.

2.2.4 Geology and Elevation Data

Geology and elevation dataset was collected from the Geospatial Information Authority of Japan (GSI), and they were downloaded from the GSI homepage at www.gsi.go.jp. Geology data were downloaded in raster data format, while the elevation data were downloaded in 10 mesh of point data. Ten mesh has a spatial resolution of 0.4 arc-second (1 arc-second = 30.86 m). The elevation data was interpolated to build a Digital Elevation Model (DEM) and then derived from the slope map and aspect map.

2.3 Method

This chapter applied three main data sources: satellite remote sensing data from ALOS AVNIR-2, rainfall data from AMEDAS, and elevation and geology data from GSI. These data were derived to generate a landslide parameter map using the Weight of Evidence (WOE) approach to demonstrate the relevance of environmental factors or variable groups and the LR modeling to determine the likelihood of incidence or absence of landslides. Figure 2.2 describes the overview of this chapter.

Fig. 2.2
A flowchart of generating a landslide susceptibility map. It involves A M E D A S, A L O S A V N I R - 2, and G S I to extract data such as rainfall, elevation, and geology to generate respective maps followed by weight of evidence, logistic regression to give the desired result.

Flowchart of the landslide susceptibility map

Full size image

2.3.1 The Pre-Processing of ALOS AVNIR-2

This satellite image has eight-bit values denoting brightness metrics, generally known as the Digital Numbers (DN). The DN value is varied according to the byte type of the satellite RS dataset; thus, converting it into common features such as radiance value is necessary. The procedure for converting DN to radiance was described by Eq. 1 (Bouvet et al. 2007).

$${L}_{\uplambda }= {GAIN}_{rescale}\times QCAL+{BIAS}_{rescale}$$
(2.1)

where Lλ is the spectrum radiance (W/m2/sr/m), GAINrescale indicates the rescaled gain, QCAL represents the DN, and BIASrescale is referred to as the rescaled bias. Table 2.2 summarizes the rescaling gains and biases utilized for the DN-to-radiance conversion.

Table 2.2 GAIN and BIAS rescaling for conversion of DN to radiance in ALOS AVNIR-2
Full size table

Normalization for solar irradiance could reduce between-scene variability in comparatively "clear" images by translating the spectral radiance into a reflectance. The algorithm for converting radiation to reflectance is described in Eq. (2.2) (Sah et al. 2012).

$${\rho }_{\uplambda }=\uppi \times {L}_{\uplambda }\times \frac{{d}^{2}}{{ESUN}_{\uplambda }}\times {{\text{cos}}\,\,\theta }_{s}$$
(2.2)

The ρ_λ defines as the reflectance (no unit), L_λ represents the spectral radiance, d2 represents the Earth-Sun distance in units of astronomy according to a navigational guidebook, ESUN_λ represents the average solar irradiances, while the θ_s defined as the solar zenith angle (degrees).

2.3.2 Land Cover Classification Using Object-Based Image Analysis OBIA

This is a technique of image analysis that derives information from pictures by using objects in the context of an image pixel-by-pixel (Burnett and Blaschke 2003). This is a two-step technique that begins with the segmentation of images and ends with image categorization. Throughout the segmentation stages, the image is initially separated into homogeneous and neighboring regions that consider region context, spectral data, texture, and shape into account. After, the image is grouped by employing the K Nearest Neighbor classification approach, a process that considers the object's Euclidean distance in the n-dimensional area to the parts of the training sample, in which n is set by the number of attributes of the object applied during the categorization. The length of the distance of two planar locations having coordinates (x, y) and (a, b) is obtained using the Euclidean distance method as shown below:

$$dist\left(\left(x,y\right),\left(a,b\right)\right)=\sqrt{{(x-a)}^{2}+{(y-b)}^{2}}$$
(2.3)

2.3.3 Land Cover Accuracy Assessment

This chapter employed error matrices and Cohen's kappa (k) as an accuracy assessment. This accuracy assessment can be beneficial in developing algorithms that anticipate different groupings and classify images. Moreover, the assessment aids in determining how precise and beneficial the model is. In order to generate an error matrix, a database of locations by which the true categorization can be determined is required. Since the location(s) were examined on the ground, the columns include reference data or established categorization. Rows represent projected categories. Kappa can be applied to provide a metric for assessing the degree of coherence of simulation results and actual (Congalton 1991) or as a proxy for whether the numbers in an error matrix reflect an outcome that is considerably more accurate (Chen et al. 2010). Equation 4 shows the Kappa formula.

$$k = \frac{{N\sum\limits_{i = 1}^{r} {x_{ii} - \sum\limits_{i = 1}^{r} {\left( {x_{i + } \times x_{ + i} } \right)} } }}{{N^{2} - \sum\limits_{i = 1}^{r} {\left( {x_{i + } \times x_{ + i} } \right)} }}$$
(2.4)

2.3.4 Rainfall Isohyet Map

We defined isohyet as a line derived from a weather map to connect dots that get equivalent rainfall quantities throughout a specific period. This map illustrates rainfall statistics. The contoured lines link regions that receive equivalent precipitation, while a color scale is often applied to make the distinction among them. Further, isohyet maps were generated by interpolating the precipitation data from the in-situ rain gauge dataset (Chow et al. 1988).

The arithmetic-mean approach is the most basic way of calculating the average amount of rainfall in a particular area. This entails averaging the rainfall volumes measured at various gauges. This strategy works well if the rain gauge is evenly distributed across the region and each gauge measurement stays within the average value.

2.3.5 Digital Elevation Model (DEM)

The initial GSI elevation points were transformed into a DEM of the raster dataset using an interpolation method of Inverse Distance Weighted (IDW). This interpolation approach used a linearly weighted ensemble of a series of features from the sample to calculate the value of each cell. Moreover, the weight was determined by the inverse distance.

Aspect and slope were generated from DEMs and determined by comparing the point's elevation to its neighbors. The slope was expressed as "a horizontal tangent to a topographic surface" as an equation of a DEM for a given position (Burrough 1986). Moreover, slope and aspect were calculated by comparing the elevations of points in a 3 × 3 neighborhood. Further, the slope and aspect at one point were calculated using an elevation dataset in that spot location and eight other points. Equation (2.5) describes the initial procedure employed to compute the slope.

$$\mathrm{Slope radians}=\mathrm{ATAN }\left(\surd \left({\left[\frac{dz}{dx}\right]}^{2}+{\left[\frac{dz}{dy}\right]}^{2}\right)\right)$$
(2.5)

The slope is typically expressed in degrees, where employs the equation as follows (6):

$$\mathrm{Slope degree}=\mathrm{ATAN }(\sqrt{{\left[\frac{dz}{dx}\right]}^{2}+{\left[\frac{dz}{dy}\right]}^{2})}*57.29578$$
(2.6)

The aspect indicates the downslope direction of the greatest extent of the variation in magnitude of one grid to the next. In other words, the slope direction could be referred to as the aspect. Also, the compass-derived vector of the aspect will be represented by the values in the finalized raster. The aspect of a raster surface is the direction of the highest alteration rate in the z-value from each cell. Aspect is stated as a positive number with values from 0 to 359.9, identified clockwise from north. A typical aspect of computation is shown in Eq. 2.7:

$${\text{tan}}\left(slope\right)= \frac{{\text{dz}}/{\text{dx}}}{{\text{dz}}/{\text{dy}}}$$
(2.7)

where (dz/dx) denotes the horizontal surface while (dz/dy) is the vertical surface. The arithmetic-mean method is the most accessible approach to measuring the average rainfall in a specific area. It entails averaging the rainfall depths measured at various gauges. This approach works well if the gauges are spread evenly across the region and each gauge measurement stays within the average value.

2.3.6 WOE and LR Model

WOE is a statistical analysis according to the Bayesian probability model in which the distribution is defined based on evidence (Bonham-Carter 1994). The relevance of variable classes is expressed by W + (positive weight) and W- (negative weight), correspondingly to the formulas (2.8) and (2.9). When a parameter class has a positive W + , landslides have a greater likelihood of taking place. Meanwhile, the landslide susceptibility decreases when the negative W-value indicates that the class is absent. Likewise, a negative W + suggests that the category is not conducive to landslides occurring, yet a positive W − suggests that the variable category's absence increases landslide susceptibility. Weights nearly “0” show no association between the parameter class and the frequency of landslides. Table 2.3 (Bonham-Carter 1994) indicated that the weight computation was according to a cross-tabulation of areas of the parameters class and the landslide area.

Table 2.3 Area cross-tabulation
Full size table
$${W}^{+}={\text{ln}}\left[\frac{\frac{{T}_{11}}{{T}_{21}}}{\frac{{T}_{1}}{{T}_{2}}}\right]={\text{ln}}[\frac{{T}_{11}{T}_{2}}{{T}_{21}{T}_{1}}]$$
(2.8)
$${W}^{-}={\text{ln}}\left[\frac{\frac{{T}_{12}}{{T}_{22}}}{\frac{{T}_{1}}{{T}_{2}}}\right]={\text{ln}}[\frac{{T}_{12}{T}_{2}}{{T}_{22}{T}_{1}}]$$
(2.9)

The multivariate statistical method known as logistic regression (LR) can determine the likelihood of a categorical regressand, including the likelihood (i.e., Yes) and “No” of landslides, according to the association among the response and predictors variables, such as land cover, lithology, slope gradient, isohyet, etc. Then, the probabilistic values are between zero to one, representing the susceptibility of landslides. Mining resource estimations (Agterberg et al. 1993) and the likelihood of landslide analysis (Feby et al. 2020) have both benefited from the use of LR. The benefits of LR include the fact that it can be employed when variables have conditional dependency and when variables have several classes or are continuous. The logistic equations f(z) underpins LR was described as follows:

$$f\left(z\right)=\frac{1}{1+{e}^{-z}}$$
(2.10)

where,

$$z={\alpha }_{0}+{\beta }_{1}{X}_{1}+{\beta }_{2}{X}_{2}+\dots +{\beta }_{x}{X}_{x}$$
(2.11)

LR involves calculating the coefficients β1, β2,…βk, as well as the constant or intercept, α0. The parameters of X1, X2, and Xk represent the predictor factors that are associated with landslides.

2.3.7 Process of WOE and LR Model

First, all the predictor factors (i.e., elevation, slope, aspect, geology) are converted into the same grid size of 10 × 10 m and overlaid with the evidence point. Figure 2.3 explains the procedure of the weight calculation. The figure describes 6 × 6 pixels for each environmental factor, representing the presence and absence of the incident likelihood in Hofu City.

Fig. 2.3
5 pixel samples of 5 environmental factors with 3 landslide spots each. a to e present pixelated grids with multiple shades for elevation, slope, geology, land cover, and aspect in order. 2 spots occur in the upper half while the third occurs in the lower half.

The pixel sampling of each environmental factor

Full size image

In order to determine the weight of all parameter classes, the cross-tabulation table was created and computed by pixel-based analysis, as shown in Table 2.4. For example, the elevation parameter for 150 m is shown in black (8 pixels), 200 m is shown in green (28 pixels), and the landslide evidence point is shown in a white dot (Fig. 2.3a). The W + and W- values are calculated to identify where the landslide is likely to occur.

Table 2.4 Area cross-tabulation of elevation
Full size table
$${W}^{+}={\text{ln}}\left[\frac{{T}_{11}{T}_{2}}{{T}_{21}{T}_{1}}\right]={\text{ln}}\left[\frac{1 x 33}{7 x 3}\right]=0.452$$
$${W}^{+}={\text{ln}}\left[\frac{{T}_{11}{T}_{2}}{{T}_{21}{T}_{1}}\right]={\text{ln}}\left[\frac{2 x 33}{26 x 3}\right]=0.053$$
$${W}^{-}={\text{ln}}\left[\frac{{T}_{12}{T}_{2}}{{T}_{22}{T}_{1}}\right]={\text{ln}}\left[\frac{2 x 33}{26 x 3}\right]=0.017$$
$${W}^{-}={\text{ln}}\left[\frac{{T}_{12}{T}_{2}}{{T}_{22}{T}_{1}}\right]={\text{ln}}\left[\frac{1 x 33}{7 x 3}\right]=0.165$$

The weight calculation process for other environmental factors was also applied in the same steps. The weight raster of each environmental factor was generated from the weight calculation. The weight raster of elevation, slope, geology, land cover, and aspect are described in Fig. 2.4.

Fig. 2.4
5 tables, a to e, with 6 rows and 6 columns each present the weight map of elevation, slope, geology, land cover, and aspect in order. a has 2 values, 0.452 and 0.053, b ranges from 0 to 2.4, c and d have the entry 1 alone, while e has 2 entries of value 0 and the rest with a value of 1.06.

The weight raster of elevation, slope, geology, land cover, and aspect

Full size image

The process of the logistic regression model performed in GIS was done by overlaying all environmental factors weight raster (Fig. 2.5a) to the landslide evidence point and boundary of the pilot sites (Fig. 2.5b) for calculating the LSM value.

Fig. 2.5
2 illustrations. a. A set of 5 weight raster tables for 5 categories. b. A 6 by 6 grid with 3 landslide spots, 2 in column 2 at row 1 and 5, and 1 in column 5 at row 2.

Logistic regression process

Full size image

The value for landslide susceptibility ranges from 0 to 1. The number of 0 indicates that there is no chance of a landslide likelihood in this region. A higher value means a higher likelihood of landslides. As per the calculation in this particular sampling area, the landslide susceptibility value for all pixels is 0.008 (Fig. 2.6).

Fig. 2.6
A 6 by 6 grid with the same entry, 0.008 for all cells.

The simulation of landslide susceptibility value 3

Full size image

3 Result

3.1 Dem

The original dataset of elevation points obtained from GSI has a spatial resolution of 0.4 arc-second. The elevation point was used in creating the DEM map, slope map, and aspect map, which will be used as the environmental variables to estimate the LSM. Data processing started from the interpolation of elevation points through the IDW method. The interpolation result was raster data in a spatial resolution of 10 m. The spatial reference of the data has been converted into World Geodetic System (WGS) 1984 with the coordinate system of Universal Transverse Mercator (UTM) zone 52 north.

The elevation, slope, as well as aspect map that was produced from the GSI's elevation point, is displayed in Fig. 2.7. The slope varied from 0 to 64 degrees with an average of 15.7 degrees, while the altitude ranged from 0 to 630 m with an average of 141 m. Furthermore, most of the lowland area was surrounded by elevations greater than 300 m, with nearly half (48.7%) of the research area having an elevation of less than 100 m. Additionally, approximately 41.6% of the Hofu City has a slope of less than 10 degrees, while around 17% and 41.4% have moderate to high slopes. Also, the aspect was concentrated in the south and southeast, where the slope direction might be conceptualized as this physical component.

Fig. 2.7
3 spatial distribution maps. a. Elevation. It increases from the southwest to the northeast ranging from 50 to 630 meters. b. Slope. It increases from 5 to 25 S W to N E, with scattered stretches of 25 to 54 degrees. c. Aspect. Predominantly south and southeast.

The spatial distribution of a Elevation, b Slope, and c Aspect within the study area

Full size image

3.2 Geological Map

Geological data were downloaded from GSI websites in Tiff format. Similar to the elevation data, geological data were also converted into the spatial reference of the WGS 984 with the UTM zone 52 north coordinate system. Japan's 1:200,000 scale geological map acted as the source of the lithology unit keys included in the geological map. Lithology was classified into 14 classes: fan deposits, granite, granodiorite, lower terrace, non-alkaline, marine and non-marine sediments, middle terrace, non-alkaline mafic volcanic rocks, dune deposits, politic gneiss, politic schist, reclaimed land, mafic schist, and water. Figure 2.8 depicts the spatial variation of the lithology map within Hofu City.

Fig. 2.8
A lithology map of the study area highlights 14 features. It includes Peltic Schist concentrated along the eastern borders, Mafic Schist along the central S W-N E diagonal, reclaimed land on the south with scattered sand dune deposits above them.

Lithology map

Full size image

3.3 The Spatial Distribution of Land Cover

The spatial distribution of land cover has been generated using an ALOS AVNIR-2 image acquired on July 14, 2009. Image processing started with converting the DN image into absolute radiance value, followed by radiance to reflectance conversion. Moreover, we used the OBIA approach to cluster the land cover into several categories within the region. The land cover map is shown in Fig. 2.9.

Fig. 2.9
A land cover map of the study area. Most of the western and northeastern regions are forests while most of the central and southeastern regions are urban areas with a small patch of bare soil in it. Waterbody flows diagonally along the center. Southwestern region has agricultural lands.

Land cover map

Full size image

As shown in the figure, forest land accounts for 57.3% of the pilot sites, followed by urban area (26.4%), agricultural land (12.8%), bare soil (2.2%), and water (1.4%).

An accuracy assessment is needed to verify the result of the classification. Typically, it was determined by comparing it to reference data, which is deemed to represent its actual condition accurately. Moreover, a confusion matrix is typically used to describe the findings of an accuracy assessment. The Kappa statistic was applied as a consensus on predicted models and the actual condition. As the reference, the high-resolution image provided by Google Maps (Copyright of Google MapIT, Zenrin) with the time acquisition of 3 February 2006 was used. The accuracy assessment result is shown in Fig. 2.10. From this analysis, the overall accuracy was 90.83%, with a Kappa coefficient of 0.89. Solid/ good classification performance was indicated by Kappa values greater than 0.80. Meanwhile, classification performance was considered medium once the Kappa values fell between 0.40 and 0.80, while it is called low if below 0.40 (Jensen 2005).

Fig. 2.10
A copy of land cover accuracy assessment data has 2 tables. 1. 6 columns and 8 rows give the corresponding entries including ground truth and 4 layers for 8 classes such as bare soil settlement, and water. 2. 3 columns and 8 rows give ground truth layer and percent total values for 8 classes.

Land cover accuracy assessment

Full size image

3.4 Rainfall Map

In late July 2009, a stationary rain front (Baiu Front) brought significant precipitation across Chugoku and northern Kyushu districts. An intense rainstorm was recorded in the pilot study on July 21, 2009. An enormous amount of precipitation was reported on that particular day, more than 270 mm, and the city's most significant hourly rainfall was 63.5 mm/hour. These rains caused numerous small landslides on the mountain slopes around Hofu. Figure 2.11a shows the daily rainfall rate in the landslide event, and Fig. 2.11b depicts the hourly rate in the landslide event on July 21, 2009.

Fig. 2.11
2 fluctuating line graphs of daily and hourly precipitation in Hofu City during the landslide event in July 2009. Former peaks on day 21 while the latter has ascending peaks from hour 1 to 9 and declining peaks from hour 10 to 23. Values are approximated.

Daily a and hourly b precipitation (mm) in Hofu City during the landslide event in July 2009

Full size image

The precipitation dataset was obtained from rain gauges at the pilot study and surrounding city. Moreover, the precipitation dataset was taken on July 21, 2009, which triggered many shallow landslides. The data was stored in point-type data, which was interpolated through the IDW method to create the rainfall map. Figure 2.12 shows the interpolation results in the raster data set, which describes rainfall distribution. During that day, the highest precipitation occurred in the lowland area, which is illustrated by the purple color.

Fig. 2.12
A spatial distribution map of the study area for rainfall. Precipitation decreases from 274 millimeters in the southwest to 194 millimeters in the north.

Spatial distribution of rainfall in the study area on July 21, 2009

Full size image

3.5 The LSM

The LSM was conducted after we defined the landslide's physical factors. All factors were then reclassified based on the WOE approach using the Arc-Spatial Data Modeller (Arc-SDM) tool in ArcGIS 10.8.2 to determine the relationship and the influence of each environmental variable during the incident. The weights of each environmental factor were calculated through the WOE tool in the Spatial Data Modeller. Weights and contrast indicate the relationship and the effect of landslide occurrence. WOE of all environmental factors is described in Table 2.5.

Table 2.5 WOE of all environmental factors: (A) Elevation, (B) Slope, (C) Aspect, (D) Lithology, (E) Land Cover
Full size table

The reclassification of elevation is shown in Table 2.5A. The weights and contrast show that classes with elevations ranging from 50 to 350 m (positive W + ) were more likely to have landslides than the other classes. Table 2.5B describes slope reclassification and WOE. It shows W + was 0 for slope gradient classes < 5° and > 50°, which means that the possibility of disaster likelihood within the slope gradient was zero. Meanwhile, the W + was positive for slope range > 5° and < 50°; this implies that gradients of slopes higher than 5° and lower than 45° indicate that this range of slope gradient was vulnerable to the occurrences of landslide. Moreover, the aspect was classified by numbers 1 to 9, respectively, referring to flat, north, northeast, east, southeast, south, southwest, west, and northwest (Table 2.5C). The value and contrast of W + revealed that slope directions of north and northeast were suitable for this type of natural disaster. Table 2.5D defines the reclassification and WOE for the lithology parameter. The result stated that the lithology class numbers 10, 11, and 12 (respectively referring to granodiorite, middle terrace, and fan deposits) were conducive to the likelihood of landslides. Moreover, land cover was classified by numbers 1 to 6 (Table 2.5E), referring to water, agriculture, urban area, bare soil, forest, and cloud. A positive contrast suggests that the category is favorable for the threat of landslides, whereas a negative contrast indicates that the class is not. Based on the WOE calculation in Table 2.5E, agriculture, urban area, bare soil, and forest were suitable conditions for landslide incidents. In this assessment, the WOE was calculated for all environmental factors except for rainfall.

The logistic regression tool from SDM combined the point landslide inventory data, evidence maps, and weight of evidence table. The tool generates tables with the environmental factor coefficients and computes the posterior probability. The coefficients represent the degree of significance of the different environmental component types. A positive coefficient suggests that the category increases susceptibility level, whereas a negative lessens the likelihood of landslide occurrence. The results of LR coefficients are listed in Table 2.6.

Table 2.6 LR Coefficient
Full size table

The LR calculation shows that the susceptibility value ranged from 0 to 0.00895. A value of 0 indicates very low susceptibility or low likelihood that a landslide will hit in this area; the higher the value, the higher the likelihood of the disaster. Figure 2.13 describes the classification of susceptibility distribution of landslide occurrences. We classified the susceptibility map into five classes, where red indicates the highest susceptibility and blue indicates no susceptibility.

Fig. 2.13
A landslide susceptibility map of the study area. It includes Hofu-Shi in the southeast with a low susceptibility and most of the northern regions including Shunan-Shi and southwestern ones with a very high susceptibility.

Landslide susceptibility map

Full size image

4 Discussion

In this chapter, the LSM was created utilizing LR techniques and five physical parameters comprising land cover, slope, elevation, aspect, lithology, and excessive precipitation as the trigger factors. These physical factors gave a good overview and effectiveness in the landslide mapping (Hong 2023). In terms of DEM preparation, the elevation point provided by GSI Japan made the availability of a resolution elevation map possible. Moreover, the application of the IDW method in elevation point interpolation performed good results in creating a raster image of DEM.

Hofu City has a variety of elevations, from 0 to 630 m (Table 2.5A), with slope gradients varying between zero to sixty-four degrees (Table 2.5B). According to the inventory map, the disaster arises at elevations and slopes ranging from 50 to 550 m and 5° to 50°, with the most significant frequent incidence occurring at 200 m and 25° with the most likely land cover type of forest land (Figs. 2.7, 2.9, 2.13). Meanwhile, based on the weight computation, landslides tend to get increasingly prevalent between 50 to 350 m of elevation and in the slope angle zone of 5° to 50°(Table 2.5 A, B). Further research revealed that soil and colluvium is an additional element that leads to shallow landslides; thus, landslides are less common in areas with slope angles greater than 35 degrees (Chen et al. 2015, 2018). As a result, a highly steep slope is the most vulnerable zone for landslide incidents all of the time. In addition, the slope directions north, northeast, and northwest were conducive to the incidence of landslides (Fig. 2.7c and Table 2.5C). Also, territories such as forests, bare soil, urban areas, and agricultural land were conducive to the emergence of landslides (Table 2.5E). Even having identical slope angles and lithology qualities, a landslide can be more potent in wooded highland locations due to the reduction in root strength caused by vegetation. Moreover, the granite was the most lithologically favorable for the likelihood of a landslide (Fig. 2.8 and Table 2.5D). This result has the same agreement as the previous research conducted in Japan (Chigira et al. 2011; Yamashita et al. 2017).

Using the LR function in the Arc-SDM tool, the weights and rasters (evidence rasters) of the physical elements and the evidence point of landslides were integrated to assess the landslide susceptibility. In Hofu, the susceptibility score ranged from 0 to 0.00895. Based on the LR output, 916 of the 928 landslide incident spots were located in very high-susceptibility zones, with 98.7% of the findings from the estimations fitting the historical data of the landslide occurrence (Fig. 2.13).

Since the mechanics of the shallow landslides in the pilot study were considered when building the model of physical characteristics, the LR technique was selected. Moreover, the LR is a useful multivariate method of statistical analysis for assessing the LSM (Goyes-Penafiel et al. 2020; Lombardo and Mai, 2018), and SDM is a good tool for LR-based modeling (Neuhaser et al. 2012; Mohamed &Elmahdy 2016). The LR method and Arc-SDM tools effectively create an LSM for Hofu City. Also, this technique is well applied worldwide, such as in Italy (Lombardo and Mai, 2018), India (Feby et al. 2020), Colombia (Goyes-Penafiel et al. 2020), and China (Zhao et al. 2019). On the other hand, SDM had some limitations; for instance, training data on landslides were provided as points (up to 1,000), but the scarp centroid may not be sufficient to describe scarp circumstances for massive landslides. Also, a distinct condition table with up to six thousand circumstances serves as the foundation for the LR, reducing the amount of environmental parameters that can be input.

In this chapter, rainfall-triggered shallow landslides provided the basis for the predicted landslide susceptibility because the landslide spots in the inventory map were triggered by an equitable distribution of rainfall. Moreover, the destabilizing effects of the rainfall intensity may be affecting the soil cohesion; this is mainly in the case of increasing rainfall from its lowland minimum to its highland maximum. Hofu city was hit by intense rainfall in late July 2009, and on 21st July 2009, the rainfall reached the highest intensity, up to 274 mm/day, then triggered many shallow landslides, resulting in 928 landslides scattered in the hilly area. Additionally, previous studies revealed that incorporating rainfall into the analyses improves the landslide susceptibility models' accuracy (Zhao et al. 2020), and the combination of event inventory and spatial rainfall data could help researchers better understand the factors that determine landslide vulnerability (Smith et al. 2023).

Topography and lithology are static terrain factors that influence susceptibility. However, the development of landslides altered the set of circumstances that existed during the triggering events. Moreover, the LR coefficient result indicates that lithology is the physical element that has the most influence on this shallow landslide occurrence. Since lithological and structural variations frequently result in differences in the durability and permeability of rocks and soils, it is well known that geology is one of the physical elements that strongly influences the likelihood of landslides (Pradhan and Lee 2010). Moreover, type-specific lithology, physical and chemical properties, and mineralogy are essential key factors, together with slope aspect, angle, and land cover. Some literature (Chen et al. 2011; Moragues et al. 2023; Machay et al. 2023; Zhao et al. 2023) most commonly consider lithological conditions.

5 Conclusion

This chapter used LR models with 10 m grid sizes to assess landslide susceptibility. The elements selected as controlling factors were aspect, elevation, slope, rainfall, lithology, and land cover. Based on the weight calculation results of LR models, the physical parameters of elevation ranging between 50 and 350 m with a slope range of 5° to 50° were favorable for the likelihood of landslides. Also, north and northeast slope directions were conducive to the incidence of landslides. Moreover, the land coverings conducive to the likelihood of landslides included forests, bare soil, agricultural fields, and urban areas. Meanwhile, granodiorite were mostly lithologically suitable for the landslides. Further, according to the LSM, of the 928 landslide spots, 916 were located in very high susceptibility zones, which revealed that 98.7% of the LSM estimations were close to the inventory/reference dataset. The results of the proposed method show great promise for the spatial prediction of landslides, and their land susceptibility maps could serve as a first evaluation for any planning project involving municipalities to carry out more in-depth research in areas previously determined to be highly susceptible. Ideally, the susceptibility models should consider many physical variables, including curvature, geomorphology, distance to the geological fault, distance to the road, and flow density. Still, our research was limited to six variables due to local site data availability. In the future, it is necessary to add more variables that can be generated from DEM, such as curvature, flow density, and terrain roughness index, and combine the LR models and Principal Component Analysis to obtain the most influential factors among many proxies of the model.

6 Author Contribution Statement

Benita Nathania did data collection and analysis and wrote the initial draft manuscript. Martiwi Diah Setiawati did data curation, supervised, and revised the manuscript. Both authors have equal contributions.