Towards Sustainable Heritage Conservation: A Hybrid Landslide Susceptibility Mapping Framework in Japan’s UNESCO Mountain Villages

Abstract

Sustainable management of cultural heritage in mountainous regions requires effective strategies to mitigate natural hazards such as landslides. Landslide susceptibility mapping (LSM) provides a critical tool to support these conservation efforts. This study presents a hybrid framework that integrates probabilistic slope stability modeling with ensemble learning for LSM in the UNESCO World Heritage sites of Shirakawa-gō and Gokayama, Japan. The framework uses probabilities of failure from Bishop’s simplified method combined with Monte Carlo simulations to guide non-landslide sample selection. An enhanced tri-parametric optimization was applied to refine the slope unit segmentation process. SHAP analysis revealed that the hybrid framework emphasizes physically meaningful features such as rainfall. The proposed method results in AUC gains of 0.072 for XGBoost, 0.066 CatBoost for, and 0.063 for LightGBM compared to their buffer-based counterparts. Future landslide susceptibility was mapped based on the 2035 precipitation projections from ARIMA time-series modeling. By enhancing accuracy, interpretability, and geotechnical consistency, the proposed approach delivers a robust tool for sustainable risk management. The study further evaluates the exposure of Gasshō-style houses and other historic buildings to varying levels of landslide susceptibility, offering actionable insights for local planning and heritage conservation.

1. Introduction

1.1. Background and Recent Advances in LSM

Landslides are one of the most dangerous and common geohazards, and they often cause a lot of damage to property and people in mountainous terrains [1]. Landslides, also referred to as slope failures, may occur as isolated events or in association with other natural hazards such as earthquakes, floods, avalanches, and forest fires, and can also be triggered by improper construction practices or other anthropogenic activities [2]. Global annual losses attributed to landslides are estimated at around 18 billion Euros, accounting for approximately 17% of the average yearly global losses from natural disasters, which total about 110 billion Euros [3]. Global landslide fatalities have been poorly quantified, but a dataset covering 2004–2010 recorded 2620 non-seismic fatal landslides worldwide, resulting in 32,322 deaths. These figures may still slightly underestimate total impacts, with the majority of fatalities occurring in Asia [4]. Moreover, the number of landslides has been going up in recent decades [5]. Therefore, Landslide Susceptibility Mapping (LSM) has become an essential tool used to identify regions prone to future slope collapses. The most commonly used susceptibility analysis approaches include geomorphological hazard mapping, landslide inventory analysis, heuristic methods, statistical models, as well as geotechnical or physically based models [6].

Recently, LSM has moved away from qualitative expert-based methods toward quantitative data-driven methods [7]. Machine learning (ML) techniques have transformed LSM by identifying complex correlations between conditioning factors and landslide events. These patterns may not be detected by conventional statistical methods. Boosting methods, such as XGBoost, CatBoost, and LightGBM, are a sort of ensemble learning that have constantly shown improved accuracy, generalization, and robustness when making LSMs [8,9]. This improved performance makes ensemble ML the current standard for assessing landslide susceptibility [10].

1.2. Challenges in ML-Based LSM

Despite recent advancements, some important issues remain in ML-based LSM. Traditional ML models may result in physically inconsistent or unrealistic predictions and weak generalization capability in new regions [11]. Also, due to “black-box” nature of ML algorithms, it is difficult to interpret how variables affect model predictions, which reduces transparency and practical applicability [12].

Although grid cells are generally adopted as the basic mapping unit in LSM studies because of their simplicity, they often fail to properly represent the complex geometry of the terrain and the geomorphological processes that lead to landslides. On the other hand, slope units, defined by natural hydrological boundaries such as ridges and valleys, define terrain in a more geomorphologically meaningful and physically coherent way [13]. The development of specialized tools, such as r.slopeunits, has lately made their automated and optimized delineation computationally feasible, enhancing LSM spatial accuracy and interpretability [14,15].

One of the most difficult tasks in LSM development involves identifying reliable non-landslide (negative) samples. While positive landslide locations are typically well documented in landslide inventories, determining areas that can be confidently labeled as stable is more difficult. A frequently utilized method includes the exclusion of buffer zones surrounding identified landslides. While buffer-based methods are commonly employed [16,17,18,19], they do not have a solid physical foundation and might incorporate unstable terrain as negative samples.

To overcome such limitations, recent studies have increasingly turned to hybrid models that integrate physically based and data-driven approaches. These hybrid models incorporate fundamental landslide mechanics to ensure that predictions are not only data-driven but also physically consistent. However, most existing hybrid applications rely on deterministic landslide hazard assessments, thereby neglecting the uncertainty associated with geotechnical parameters [20,21].

A well-established approach for addressing geotechnical uncertainty is the use of Monte Carlo simulations (MC), which are widely applied in probabilistic physics-based frameworks to translate deterministic factors of safety into probabilities of failure (PoF). For example, Zhang et al. [22] proposed a physics-based probabilistic framework for rainfall-induced shallow landslides in Sichuan Province, China, where MC simulations were used to sample uncertain soil parameters and convert safety factors into landslide probability estimates. Similarly, Li et al. [23] developed a probabilistic seismic landslide hazard mapping approach for Anchorage, Alaska, incorporating MC simulations to account for uncertainties in geotechnical properties and ground-motion variability, and defining hazard categories based on critical slope angles. Marin and Mattos [24] also applied a probabilistic physically based approach using MC simulations to address uncertainties in landslide susceptibility assessments derived from geotechnical parameters, emphasizing the limitations of assuming fixed parameter values.

While these studies clearly demonstrate the value of MC-based probabilistic formulations in improving the realism of physically based landslide assessments, they remain largely disconnected from machine-learning-based susceptibility modeling and do not exploit probabilistic outputs to enhance hybrid ML frameworks.

The key contribution of this study is the integration of MC with physically based models to account for uncertainties in geotechnical parameters at the slope-unit scale, using the resulting probabilities of failure to select negative samples for hybrid ML models. Negative samples are chosen from slope units exhibiting the lowest probabilities of failure derived from MC coupled with Bishop’s simplified method. These reliable negative samples enhance the training of hybrid ML models, leading to improved robustness and more reliable landslide susceptibility predictions.

The proposed strategy ensures the physical consistency of non-landslide (negative) samples, in contrast to commonly used buffer-based sampling approaches that may include unstable or potentially failing areas [16,17,18,19]. In addition, it improves deterministic hybrid models [20,21] by incorporating uncertainty in geotechnical parameters, such as cohesion, internal friction angle, and unit weight, that strongly control slope stability. Furthermore, compared with probabilistic physics-based landslide susceptibility models [22,23,24], the proposed method exploits probabilistic outputs to inform negative sample selection for hybrid ML models, thereby directly enhancing data-driven susceptibility prediction rather than relying solely on probabilistic hazard indices.

Given the incorporation of both physical principles and probabilistic filtering, the proposed strategy significantly decreases ML models classification errors and improves model performance. Three ensemble learning methods, including XGBoost, LightGBM, and CatBoost, were employed and compared with Area Under the ROC Curve (AUC) as the primary metric. A spatial cross-validation strategy was adopted to examine the generalization of the model across geographically distinct folds. Besides, the proposed approach uses SHapley Additive exPlanations (SHAP) to interpret how input features drive the model outcome. The results showed that the proposed framework outperformed traditional buffer-based approaches by offering superior predictive accuracy, interpretability, and physical coherence.

This study focuses on the UNESCO World Heritage villages of Shirakawa-go (Ogimachi) and Gokayama (Suganuma and Ainokura) in Japan, known for their cultural Gasshō-style houses. These villages are located in steep mountainous terrain with complex geomorphology, making them susceptible to landslides. This study applies hybrid models to produce medium-scale landslide susceptibility maps under present and projected future rainfall scenarios. The expected outcomes include identification of high-risk areas, assessment of exposure of historic buildings, and spatial information to support targeted hazard mitigation and early warning systems. To achieve this, a comprehensive framework is implemented, including probabilistic slope stability modeling for negative sampling, optimized slope unit segmentation, SHAP-based interpretability, and ensemble ML for robust LSM.

2. Materials and Methods

2.1. Study Area

The study area includes the historic villages of Shirakawa-go and Gokayama, which have been UNESCO World Heritage Sites since 1995. The property includes the villages of Ainokura and Suganuma in the Gokayama region of Toyama prefecture and the village of Ogimachi in the Shirakawa-go region of Gifu prefecture. Figure 1 shows that the whole study area, which includes the villages and their buffer zone, is about 546.77 km² and lies between 36°5′41″ N and 36°30′1″ N latitude and 136°45′52″ E and 137°0′41″ E longitude.

The high, rugged mountains in this area, which rise to 2700 m, make landslides very likely to happen there. This is because the gravitational forces are stronger on these slopes. The study area is also situated in the Shō River valley, which cuts through the mountainous landscape, thereby increasing the likelihood of slope instability resulting from erosion processes.

Figure 2 shows the distinctive gasshō-zukuri architectural style with the steeply pitched thatched roofs, which the villages are famous for. This layout was engineered to shed the heavy snowfall, demonstrating a direct cultural adaptation to the harsh natural environment. These traditional structures have been preserved for a long time, showing interaction between human settlement and the environment [25,26]. Figure 3 shows the layout of Gasshō-style houses and other historic buildings in the three villages of Ogimachi, Ainokura, and Suganuma, according to UNESCO mapping. These villages contain most of the Gasshō-style architecture in the Shirakawa-go and Gokayama region. The UNESCO maps indicate that Ogimachi has 110 Gasshō-style houses, Ainokura has 20 Gasshō-style houses and 41 other historic buildings, and Suganuma has 9 Gasshō-style houses and 17 other historic buildings.

Its geographical location near a converging plate boundary makes Japan highly susceptible to frequent landslides, often triggered by earthquakes. The area is highly vulnerable to geological hazards due to the presence of multiple active faults. Data from the Geological Survey of Japan indicates that the Shokawa, Atotsugawa, Kokufu, and Ushikubi fault zones are important line faults in the region. Larger area faults, such as the eastern section of the Tonami-heiya/Kurehayama fault zone and the Morimoto-Togashi fault zone, form complex tectonic structures of this area [28]. These are significant contributors to regional seismicity, raising the possibility of landslide occurrences.

The Japanese archipelago features extensive fragile ground and steep terrain, resulting from high volcanic activity, crustal changes, and weathering processes. These geological weaknesses greatly contribute to slope instabilities [29].

This study area has a humid climate, with average yearly rainfall ranging from about 2500 mm to 3100 mm. This high moisture level can significantly lead to landslide initiation. The combination of the previous factors, including active tectonics, fragile geological structures, and extreme weather events, creates a high risk of landslides. For instance, the historic Kaerikumo castle and its surrounding were ruined by a massive landslide during the Great Tensho Earthquake of 1586 [30]. Accordingly, the historic villages of Shirakawa-go and Gokayama were classified as having high exposure to landslide hazards, as indicated in Resilient Cultural Heritage: Learning from the Japanese Experience [31]. This classification highlights the need to robust disaster risk management plans to safeguard these culturally significant sites. This directly supports Sustainable Development Goal (SDG) Target 11.4 on protecting the world’s cultural and natural heritage, as well as Targets 11.5 and 13.1, which emphasize reducing disaster-related losses and strengthening resilience to climate-related hazards.

The scale of LSM depends on the extent of the study area, the purpose of the assessment, and the availability of data and resources. Different work scales influence the choice of methodology; for instance, statistical or machine learning approaches may not be suitable for site-specific studies of individual slopes, but they are appropriate for analyzing the relationship between landslides and contributing factors at a medium scale. The medium scale is appropriate when evaluating the susceptibility of an area and conducting hazard zonation [6,32]. In this study, covering approximately 546.77 km² with a 12.5 m DEM, using slope unit as a mapping unit, and aimed at targeted hazard mitigation, the work can be considered as a medium-scale study. ML landslide susceptibility models are particularly useful at this scale to provide an overview of slopes prone to landslides, enabling the identification of hot spots where more detailed slope-stability analyses can subsequently be conducted, and targeted mitigation measures can be applied [33].

The framework for conducting LSM in Shirakawa-go and Gokayama integrates advanced geospatial data processing, physically based slope stability analysis, MC, and ensemble ML algorithms, as illustrated in Figure 4. This study adopts a hybrid scheme that builds upon well-established physically based and ML-based landslide assessment approaches while introducing a probabilistic integration strategy. In many existing hybrid LSM frameworks, physically based analyses are combined with ML using single value estimates of geotechnical parameters as inputs, thereby neglecting parameter uncertainty [20,21]. Physically based slope stability modeling and MC have been widely applied to address geotechnical uncertainty in probabilistic hazard assessments, independent of ML approaches (e.g., Zhang et al. [22]; Li et al. [23]; Marin and Mattos [24]) Unlike these deterministic hybrid models and purely probabilistic physics-based studies, the proposed framework uses probabilities of failure of slope units to guide the selection of reliable negative samples for hybrid ML models. Thus, while the individual methodological components are grounded in established practices, their integration into a unified probabilistic–physics-ML framework represents the main emphasis and contribution of this study.

2.2. Landslide Inventory

In LSM studies, future susceptibility can be predicted by analyzing statistical relations between past landslides and predisposing factors. Therefore, a robust landslide inventory is fundamental for generating accurate landslide susceptibility maps [34]. Here, we employed the national landslide inventory maintained by the National Research Institute for Earth Science and Disaster Resilience, which has also been successfully utilized in previous studies to generate landslide hazard maps in various regions of Japan [17,29,35]. Landslides in this inventory are predominantly rotational in this study area; therefore, Bishop’s simplified method was used for stability analysis. This historical landslide inventory was created from monochrome aerial photographs at the scale of 1:40,000 to identify historical landslides. This approach allowed the detection of relatively large landslide topography, wider than 150 m, since small-scale topographies could not be included. Applying such historical records for different conditions and triggering events provides more comprehensive and representative data, enhancing model reliability and enabling stronger predictions regarding the probable landslide areas in the future [36,37,38,39]. Figure 5a shows the landslide inventory throughout the investigation area and its surrounding regions, where it is obvious that the study area is one of the most susceptible zones in this part of Japan. Figure 5b shows the landslide inventory within the research area in detail.

2.3. Data Acquisition

In this research, twenty-two critical conditioning factors have been selected and classified into topographic, soil-related, geological and seismic, hydrological, and land cover and anthropogenic groups. The selection was informed by examination of relevant literature [41,42,43,44], where common key variables were included for landslide susceptibility modeling. These conditioning variables are derived from commonly used sources as in

Table 1. Input conditioning factors for landslide susceptibility assessment.

Feature Name	Resolution	Data Sources
Elevation	12.5 m	https://search.asf.alaska.edu/ (accessed on 4 March 2025)
Slope, Aspect, Slope height, Plan curvature, Profile curvature, TPI, TRI, and TWI	12.5 m	Processing of DEM
Sand, Silt, Clay contents, and Density	250 m	https://soilgrids.org/ (accessed on 4 March 2025)
Lithology and Distance to faults	Vectors	https://gbank.gsj.jp/ (accessed on 4 March 2025)
Peak Ground Velocity (PGV) and Shear Wave Velocity (VS30)	250 m	https://www.j-shis.bosai.go.jp/ (accessed on 4 March 2025)
Precipitation	Stations	https://www.data.jma.go.jp/ (accessed on 4 March 2025)
Distance to water	Vectors	https://data.humdata.org/ (accessed on 4 March 2025)
LULC	10 m	https://www.eorc.jaxa.jp/ALOS/en/index_e.htm (accessed on 4 March 2025)
NDVI	30 m	https://glovis.usgs.gov/app (accessed on 4 March 2025)
Distance to roads	Vectors	https://data.humdata.org/ (accessed on 4 March 2025)

. Further sections will be explaining these factors in detail.

2.3.1. Topographic Factors

Topographic factors are fundamental in slope stability, as they shape the morphology of terrain and control the hydrological and gravitational processes. In the subject study, nine topographic factors have been generated mainly from a Digital Elevation Model (DEM) of 12.5 m resolution acquired from ALOS PALSAR data. Figure 6 shows these factors, described below.

• Elevation allows for the calculation of potential energy along the slope. The increase in elevation correlates generally with mass movement potential. The elevations of the study area range from 223 to 2719 m above sea level (Figure 6a).
• Slope Angle is a major factor in landslide occurrence. Sharper slopes tend to be less stable. Slope layer was derived from the DEM and has always been one of the most important factors in LSM studies [41] (Figure 6b).
• Aspect shows how the slope faces are oriented. It affects how much sunlight gets to the ground, how much moisture stays in the soil, how the wind blows, and how much vegetation covers the slope. All of these factors affect slope stability [45] (Figure 6c).
• Slope Height is estimated as the difference between the maximum and minimum elevations in each slope unit (${\mathrm{Z}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{Z}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$). This is a metric that is particularly useful in assessing landslide magnitude [46].
• Plan Curvature refers to the curvature of contour lines on a horizontal plane. It reflects the lateral flow accumulation of surface water and related erosion processes [47] (Figure 6d).
• Profile Curvature represents the curvature along a vertical slice of the terrain, as it influences the acceleration of flowing surface water. The upwardly concave slope is reflected by positive values, while negative values relate to convex surfaces and zero relates to flat terrain [48] (Figure 6e).
• Topographic Position Index (TPI) defines the relative position of a particular cell based on the comparison of its elevation with the average elevation of neighboring cells.

$$\mathrm{T}\mathrm{P}\mathrm{I}={\mathrm{Z}}_0-\overline{\mathrm{Z}}$$

(1)

where ${\mathrm{Z}}_0$ is the elevation of the cell and $\overline{\mathrm{Z}}$ is the mean elevation of its surroundings within a predefined neighborhood. It helps in distinguishing between geomorphological features for representing positional influence on landslide occurrences (Figure 6f).

• Terrain Ruggedness Index (TRI) is a measure of surface heterogeneity, which affects soil stability (Figure 6g). The TRI formula was proposed by Riley et al. [49] and is calculated as follows:

$$\mathrm{T}\mathrm{R}\mathrm{I}={\sum }_{\mathrm{i}=1}^8{({\mathrm{Z}}_0-{\mathrm{Z}}_{\mathrm{i}})}^2$$

(2)

where ${\mathrm{Z}}_0$ is the elevation of the cell, and ${\mathrm{Z}}_{\mathrm{i}}$ is the elevation of each of its 8 surrounding neighboring cells.

• Topographic Wetness Index (TWI) is a measure of the effect of topography on soil moisture accumulation, a key factor for landslide initiation (Figure 6h). It is calculated as follows:

$$\mathrm{T}\mathrm{W}\mathrm{I}=\mathrm{l}\mathrm{n}({\mathrm{A}}_{\mathrm{S}}/\mathrm{t}\mathrm{a}\mathrm{n}\mathsf{\beta })$$

(3)

where ${\mathrm{A}}_S$ is the specific catchment area and $\mathsf{\beta }$ is the slope angle.

These nine factors were selected because together they capture the key controls on landslide susceptibility in mountainous terrain, including slope shape, water movement, and surface variability, providing a complete representation of terrain conditions affecting stability.

2.3.2. Soil-Related Factors

Sand, silt, and clay content along with density were extracted from SoilGrids [50], which is a global dataset that contains detailed estimates for various soil attributes at 250 m resolutions. The attributes are made available at six standard depth intervals up to 200 cm, derived by applying ML models on global soil profile data. For each cell, means of soil parameters along the given depth were calculated (Figure 7). These parameters are vital in the hybrid approach for selecting negative samples, as they directly influence slope stability assessments. Previous research has also utilized SoilGrids data for this purpose [11,51].

2.3.3. Geological and Seismic Factors

• Lithology is a key element affecting landslide susceptibility by Reflecting the strength and weathering behavior of rocks. Variations in lithological and structural composition result in significant differences in the strength of rocks and soils [52]. In this research, a seamless geological map from the Geological Survey of Japan (Figure 8a) was rasterized at the same resolution as DEM (12.5 m). The dominant lithological units include the igneous rocks listed in Table 2. These geological formations control the mechanical properties of the ground, which control the spatial pattern and frequency of landslides.

  Table 2. Main lithologies in the study area and their formation era [53].

  Symbol	Formation Era	Lithology
  L1	Mesozoic Era Late Cretaceous Campanian to Maastrichtian	Dacite–rhyolite (Large-scale pyroclastic flow)
  L2	Cenozoic Era Paleogene Paleocene Danian to Eocene Yaplesian	Dacite–rhyolite (Lava and pyroclastic rock)
  L3	Cenozoic Era Neogene Miocene Burdigalian to Early Rangian	Andesite–basaltic andesite (lava and pyroclastic rock)
  L4	Cenozoic Era Paleogene Paleocene Danian to Eocene Yaplesian	Granite (Massive)

  Table 2. Main lithologies in the study area and their formation era [53].

  Symbol	Formation Era	Lithology
  L1	Mesozoic Era Late Cretaceous Campanian to Maastrichtian	Dacite–rhyolite (Large-scale pyroclastic flow)
  L2	Cenozoic Era Paleogene Paleocene Danian to Eocene Yaplesian	Dacite–rhyolite (Lava and pyroclastic rock)
  L3	Cenozoic Era Neogene Miocene Burdigalian to Early Rangian	Andesite–basaltic andesite (lava and pyroclastic rock)
  L4	Cenozoic Era Paleogene Paleocene Danian to Eocene Yaplesian	Granite (Massive)

• Distance to Faults is another important geological factor that controls landslide susceptibility. Areas close to active faults generally have rock strength considerably weakened, which increases the possibility of landslide events. Therefore, distance to lines of fault become a very significant factor in LSM [54]. Fault data were obtained from the Geological Survey of Japan. Line faults in the area include the Atotsugawa, Kokufu, Ushikubi, and Shokawa fault zones, and area faults include the eastern part of the Tonami-heiya/Kurehayama fault zone and the Morimoto-Togashi fault zone [28]. A distance raster with a resolution of 12.5 m was generated to quantify proximity to these faults (Figure 8b).
• Peak Ground Velocity (PGV) is an important seismic parameter quantifying the intensity of an earthquake and has been illustrated to be closely related to landslide initiation. In the study of Liu et al. [55], PGV was among the most important predictors in LSM. The PGV dataset was obtained from the Japan Seismic Hazard Information Station (Figure 8c).
• Shear-Wave Velocity (Vs30) is used to estimate the local site amplification for the top 30 m of material during seismic shaking and is one of the most widely accepted parameters in seismic hazard analysis. It represents the shear modulus and elastic properties of near-surface materials [56]. Vs30 data was obtained from the Japan Seismic Hazard Information Station in a 250 m resolution raster format (Figure 8d).

2.3.4. Hydrological Factors

• Precipitation (Annual) is a primary contributing factor to landslides because it increases the level of soil moisture, which decreases shear strength and increases slope instability [57]. This study adopted average annual rainfall data from 2000 to 2024 from 13 meteorological stations operated by the Japan Meteorological Agency, following [57], which used average 10-year precipitation for LSM. The utilization of ground-based observation overcomes the limitation raised by Allam et al. [58], who asserted that the basis of hazard studies on remote sensing data solely without considering gauge station records may limit the accuracy of the precipitation data used in hazard assessments. The presented station records (Table 3) were used to form a precipitation map in a 12.5 m resolution through the application of the Inverse Distance Weighted (IDW) interpolation technique. The annual rainfall in the investigation area falls in a range between 2466 and 3135 mm (Figure 9a).
  Future rainfall was modeled using a seasonal AutoRegressive Integrated Moving Average (ARIMA) approach to enable temporal transfer of landslide susceptibility. Monthly precipitation records from 2000–2024 were analyzed for each meteorological station using the pmdarima python library [59]. Seasonality was represented with a 12-month period, while the orders of non-seasonal and seasonal differencing were determined automatically using the KPSS and OCSB tests. Model orders were identified through a grid search over non-seasonal (p,q ≤ 4) and seasonal terms (P,Q ≤ 3) [60]. The last two years (24 months) were reserved for validation, and the optimal model for generating future rainfall projections was selected based on parameters configuration that minimizes the mean absolute error (MAE) of the validation set.
  Following model selection process, the final ARIMA configuration was applied to the complete 2000–2024 rainfall dataset to generate future rainfall projections. The monthly rainfall estimates for the target year (2035) were summed to calculate the total annual precipitation, then they were spatially interpolated using IDW, and the mean projected rainfall was extracted for each slope unit. These future rainfall values (instead of historical precipitation values) with other conditioning factors were used to estimate future landslide susceptibility [61,62].
• Distance to Water is another important factor that affects landslide occurrence [19]. Proximity to water bodies tends to lead to increased surface erosion and soil moisture, reduction in soil cohesion, and thereby increased likelihood of slope instability. Water bodies, including rivers, streams, canals, and ditches, were mapped in vector format, and a corresponding 12.5 m resolution distance-to-water raster was derived (Figure 9b).

Figure 9. Hydrological factors: (a) Average annual precipitation (2000–2024) and (b) Distance to water bodies.

Figure 9. Hydrological factors: (a) Average annual precipitation (2000–2024) and (b) Distance to water bodies.

Table 3. Rainfall stations and observed (2000–2024) and projected (2035) precipitation.

Station	Latitude	Longitude	Average Annual Precipitation (2000–2024) “mm/Year”	Projected Precipitation (2035)
Shirakawa	36.27333	136.8967	2466.78	2458.5
Miboro	36.14667	136.9083	3135.06	3143.9
Kawai	36.305	137.1	2038.06	2045.4
Kiyomi	36.18	137.045	2355.16	2337.6
Hakusan Kawachi	36.39667	136.62	2934.34	2968.1
Mount Hakusan	36.18	136.625	2956.04	2965.7
Mount Io	36.52	136.745	2235.68	2719.4
Hirugano	36.01	136.8933	3398.96	3455.2
Rokumaya	36.06	137.035	2555.80	2599.3
Gokayama	36.43	136.9417	2837.80	2974.1
Nanto Takamiya	36.545	136.8717	2645.92	2650.5
Tonami	36.61	136.955	2218.16	2280.2
Yao	36.57	137.1583	2611.56	2609.2

Table 3. Rainfall stations and observed (2000–2024) and projected (2035) precipitation.

Station	Latitude	Longitude	Average Annual Precipitation (2000–2024) “mm/Year”	Projected Precipitation (2035)
Shirakawa	36.27333	136.8967	2466.78	2458.5
Miboro	36.14667	136.9083	3135.06	3143.9
Kawai	36.305	137.1	2038.06	2045.4
Kiyomi	36.18	137.045	2355.16	2337.6
Hakusan Kawachi	36.39667	136.62	2934.34	2968.1
Mount Hakusan	36.18	136.625	2956.04	2965.7
Mount Io	36.52	136.745	2235.68	2719.4
Hirugano	36.01	136.8933	3398.96	3455.2
Rokumaya	36.06	137.035	2555.80	2599.3
Gokayama	36.43	136.9417	2837.80	2974.1
Nanto Takamiya	36.545	136.8717	2645.92	2650.5
Tonami	36.61	136.955	2218.16	2280.2
Yao	36.57	137.1583	2611.56	2609.2

2.3.5. Land Cover and Anthropogenic Factors

• Land Use/Land Cover (LULC) maps classify the Earth’s surface based on physical features such as forests, croplands, lakes, and wetlands. Land cover is considered as a tool to evaluate the influence of human activities on landslide initiation [63]. In this study, land cover classification was derived from ALOS-JAXA data, revealing that the majority of the investigated area is occupied by deciduous broadleaf forest (DBF) then by grassland and evergreen needleleaf forest (ENF) (Figure 10a).
• Normalized Difference Vegetation Index (NDVI) is a frequently utilized parameter to assess vegetation density (Figure 10b). High vegetation would decrease erosion, hold soil by means of rooting systems, and increase stability on slopes [63,64]. NDVI was estimated from Landsat 8 satellite imagery with the formula: NDVI = (IR − R)/(IR + R), where IR is the infrared band and R is the red band [65].
• Distance to Roads is one of the primary anthropogenic factors contributing to landslides [19,66]. Road network data were obtained in vector format, and a 12.5 m resolution distance-to-road raster was generated (Figure 10c).

Each of the above factors represents a different aspect of either terrain, geology, hydrology, or human activity. The datasets applied have been sourced from ALOS PALSAR, SoilGrids, J-SHIS, JMA, USGS, and JAXA. All these datasets are supported by recent literature, ensuring their applicability and coherence in modern landslide susceptibility modeling.

2.4. Enhanced Slope Units Delineation

The selection of an appropriate mapping unit is crucial to perform a correct and meaningful LSM. Even though grid (pixel) units are commonly used in LSM, they often fail to capture the natural boundaries and hydrological processes that control slope stability. Slope units are hydrologically and geomorphologically homogeneous terrain subdivisions bounded by drainage and ridge lines, representing individual slopes or terrain segments. This subdivision allows for a more physically meaningful analysis of slope processes and landslide susceptibility [67].

The r.slopeunits tool (v1.0) is a free, open-source GRASS GIS (v8.4) plugin developed to automatically delineate and optimize slope units [14,15]. This tool generates multiscale slope units from elevation data in a way that the resulting units best match natural terrain features.

The generation process within r.slopeunits is controlled by a set of parameters, including the initial flow accumulation threshold (t), minimum circular variance (c), minimum surface area (a), cleaning size, and the iterative reduction factor (r). By systematically testing various combinations of these parameters, the optimal configuration that yields the most geomorphologically appropriate slope units is determined [14,15].

Developers of r.slopeunits suggested that the iterative process should be done using different sets of the (a, c) variables to get the set that maximizes the objective function. Numerous studies have followed the same technique [68,69]; however, a study carried out by Yang et al. [70] obtained 30 slope unit segmentation layers using different sets of the (t, c) parameters while keeping minimum surface area (a) at a constant value of 300,000 m². In this study, we have generated 48 slope unit layers from multiple combinations of the (t, a, c) parameters to examine the combined impact of the three parameters on the slope unit generation process. Adopted parameter values are listed in

Table 4. Parameters and corresponding values used in r.slopeunits for generating slope units in the study area.

Parameter	Value
Initial flow accumulation threshold (t)	(50, 100, 300, 500) × 104 m2
Minimum circular variance (c)	(0.1, 0.2, 0.3, 0.4)
Minimum surface area (a)	(10, 20, 30) × 104 m2
Reduction factor (r)	10
Cleaning size	30,000 m2

.

Developers of the r.slopeunits tool proposed that the optimization process within r.slopeunits should be guided by the segmentation metric (F) presented by Espindola et al. [71] that combines the internal terrain aspect variance (V), which quantifies the internal homogeneity of slope direction within units, and external terrain aspect variance (I), which assesses the heterogeneity between adjacent slope units [71].

$$\mathrm{V}={\displaystyle \frac{\sum _{\mathrm{n}}{\mathrm{S}}_{\mathrm{n}}{\mathrm{C}}_{\mathrm{n}}}{\sum _{\mathrm{n}}{\mathrm{S}}_{\mathrm{n}}}}$$

(4)

$$\mathrm{I}={\displaystyle \frac{\mathrm{N}\sum _{\mathrm{n},\mathrm{l}}{\mathrm{w}}_{\mathrm{n}\mathrm{l}}\left({\mathsf{\alpha }}_{\mathrm{n}}-\overline{\mathsf{\alpha }}\right)\left({\mathsf{\alpha }}_{\mathrm{l}}-\overline{\mathsf{\alpha }}\right)}{\left(\sum _{\mathrm{n}}{\left({\mathsf{\alpha }}_{\mathrm{n}}-\overline{\mathsf{\alpha }}\right)}^2\right)\left(\sum _{\mathrm{n},\mathrm{l}}{\mathrm{w}}_{\mathrm{n}\mathrm{l}}\right)}}$$

(5)

where N is the count of slope units (SU), ${\mathrm{C}}_{\mathrm{n}}$ represents the circular variance of aspect for the nth SU, ${\mathrm{S}}_{\mathrm{n}}$ designates the surface area of the nth SU, ${\mathsf{\alpha }}_{\mathrm{n}}$ denotes the mean aspect of the nth SU, $\overline{\mathsf{\alpha }}$ indicates the mean aspect of the whole study area, and ${\mathrm{w}}_{\mathrm{n}\mathrm{l}}$ is equal to 1 if the (n, l) SUs are adjacent, 0 otherwise.

The optimization step aims to identify the group of parameters that generate a slope unit layer that maximizes the segmentation metric (F) value defined by the following equation:

$$\mathrm{F}={\displaystyle \frac{{\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\mathrm{V}}{{\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{V}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}+{\displaystyle \frac{{\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\mathrm{I}}{{\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{I}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}$$

(6)

where ${\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ and ${\mathrm{V}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ are the maximum and minimum internal terrain aspect variance, while ${\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ and ${\mathrm{I}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ represent the maximum and minimum external terrain aspect variance. This will ensure that each slope unit is relatively homogeneous internally and distinguished from its neighbors to enhance the reliability of LSM.

The process of slope unit delineation may generate slope units with unrealistic length–width ratios. In order to overcome this problem in our study, we have adopted the modified objective function (F) used by Li et al. [72] in their new technique for generating SU using the PSO-SLIC algorithm. They recommended including a squareness parameter (S), which considers average length–width ratios in such a way that the generated slope units will have similar length–width ratios as much as possible.

$$\mathrm{S}={\displaystyle \frac{\sum _{\mathrm{n}}\mathrm{M}\mathrm{a}\mathrm{x}(\frac{{\mathrm{D}}_{\mathrm{n}}}{{\mathrm{L}}_{\mathrm{n}}},\frac{{\mathrm{L}}_{\mathrm{n}}}{{\mathrm{D}}_{\mathrm{n}}})}{\mathrm{N}}}$$

(7)

where ${\mathrm{L}}_{\mathrm{n}}$, ${\mathrm{D}}_{\mathrm{n}}$ represent the length and width of the nth SU. Therefore, the value of the squareness parameter (S) relative to its maximum (${\mathrm{S}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$) and minimum (${\mathrm{S}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$) values are used to update the objective function (F) as follows:

$$\mathrm{F}={\displaystyle \frac{{\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\mathrm{V}}{{\mathrm{V}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{V}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}+{\displaystyle \frac{{\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\mathrm{I}}{{\mathrm{I}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{I}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}+{\displaystyle \frac{{\mathrm{S}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-\mathrm{S}}{{\mathrm{S}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{S}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}$$

(8)

In each slope unit, the mean values for continuous variables were used to characterize unit-level properties, and the majority class represented the dominant categorical feature.

2.5. Sample Selection Strategies

The precise identification of both positive and negative samples is a crucial step in LSM, as it directly affects model accuracy and reliability. In this work, slope units containing at least one mapped landslide point were defined as positive samples, totaling 118 positive samples. The number negative samples in LSM is usually much higher than the positive ones, so to ensure sufficient representation of non-landslide slope units, negative samples were chosen at a ratio of 2:1 relative to the positive samples, yielding 236 negative slope units. To address this class imbalance, the models relied on internal weighting of classes during training to reduce bias and improve predictive performance.

This study employed two distinct strategies for negative sample selection to compare their effectiveness and ensure a high-quality dataset for model training.

2.5.1. Buffer Method (500 m Buffer from Landslides)

The buffer method is a widely adopted technique to define non-landslide areas by creating a buffer zone around known landslide locations. For this study, a 500 m buffer is applied around all identified landslide points. Negative samples are then randomly chosen from zones out of this 500 m buffer area, as adopted by Wang et al. [73], to ensure a spatial separation between known landslide occurrences and the designated stable areas, thereby. This method is straightforward and commonly used to minimize the chance of including potentially unstable zones as negative samples and to provide a baseline for comparison with the physically based strategy.

2.5.2. Physically Based Method with Monte Carlo Simulation

This hybrid method integrates a physically based slope stability analysis with probabilistic modeling to select highly reliable non-landslide samples. The goal is to identify areas that are genuinely stable based on their geotechnical properties, rather than relying solely on spatial distance from landslide locations.

Factor of Safety Calculations: The dominant failure mechanism in the study area is rotational; therefore, the Simplified Bishop method was selected [74,75]. This method is used to compute the factor of safety (FoS) for each slope unit (SU), where FoS values greater than 1 indicate stability and values less than 1 indicate potential failure. The method assumes a circular failure surface and considers normal inter-slice forces, while neglecting inter-slice shear forces by assuming horizontal side forces [76]. The FoS is calculated as:

$$\mathrm{F}\mathrm{o}\mathrm{S}={\displaystyle \frac{\sum _{\mathrm{i}=1}^{\mathrm{n}}(\mathrm{C}\ast {\mathrm{b}}_{\mathrm{i}}+{\mathrm{W}}_{\mathrm{i}}\ast \mathrm{t}\mathrm{a}\mathrm{n}\mathrm{\varnothing })\frac{1}{{{\mathrm{m}}_{\mathsf{\alpha }}}_{\mathrm{i}}}}{\sum _{\mathrm{i}=1}^{\mathrm{n}}({\mathrm{W}}_{\mathrm{i}}\ast \mathrm{s}\mathrm{i}\mathrm{n}{\mathsf{\alpha }}_{\mathrm{i}})}}$$

(9)

$${{\mathrm{m}}_{\mathsf{\alpha }}}_{\mathrm{i}}=\mathrm{c}\mathrm{o}\mathrm{s}{\mathsf{\alpha }}_{\mathrm{i}}+{\displaystyle \frac{\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{\varnothing }\ast \mathrm{s}\mathrm{i}\mathrm{n}{\mathsf{\alpha }}_{\mathrm{i}}}{{\mathrm{F}}_{\mathrm{S}}}}$$

(10)

where c is cohesion, γ is the unit weight of the soil, φ is the internal friction angle, ${\mathrm{b}}_{\mathrm{i}}$ is the width of the i-th slice, ${\mathrm{W}}_{\mathrm{i}}$ is the weight of the i-th slice, and ${\mathsf{\alpha }}_{\mathrm{i}}$ is slope angle of the i-th slice.

Soil Shear Strength Parameters: A common approach in physically based LSM models is to assign fixed values of cohesion (c) and internal friction (φ) over the study area [77]. However, in this study, to better reflect the spatial variability of geotechnical properties, cohesion and friction angle were estimated using the relative proportions of clay, silt, and sand derived from the SoilGrids dataset (Figure 11).

Cohesion (c) was calculated using an empirical equation derived by Wei et al. [78]:

$$\mathrm{c}=3.897\times {\mathrm{e}}^{0.023 \ast \mathrm{c}\mathrm{l}\mathrm{a}\mathrm{y} \mathrm{f}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}} [\mathrm{kPa}, \mathrm{R}2=0.91]$$

(11)

Friction angle (φ) was estimated following the equation introduced by Arvanitidis et al. [79]:

$$\mathsf{\varphi }=32.958\times {\left({\displaystyle \frac{\mathrm{s}\mathrm{a}\mathrm{n}\mathrm{d} \mathrm{f}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}{\mathrm{c}\mathrm{l}\mathrm{a}\mathrm{y} \mathrm{f}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{s}\mathrm{i}\mathrm{l}\mathrm{t} \mathrm{f}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}}}\right)}^{0.5384} [\mathrm{D}\mathrm{e}\mathrm{g}\mathrm{r}\mathrm{e}\mathrm{e},\mathrm{R}2=0.917]$$

(12)

The Simplified Bishop method was implemented using the PySlope Python package (v1.4.0) [80], with MC applied to account for geotechnical parameter uncertainty. For each slope unit, the mean, standard deviation, minimum, and maximum values of geotechnical parameters were calculated from all pixels within the slope unit. Then, 100 realizations of these parameters were generated, and for each set, PySlope performed 2000 iterations over different trial circular slip surfaces to identify the one yielding the lowest factor of safety (FoS). Geotechnical parameters are sampled from truncated normal distributions constrained by the smallest and largest values of the parameter within the slope unit. The truncation ensures that the sampled values remain within realistic physical bounds while capturing both the central tendency and variability of each parameter. The factor of safety (FoS) was computed for each realization, and the physics-based PoF of the slope unit was derived as:

$$\mathrm{P}\mathrm{o}\mathrm{F}={\displaystyle \frac{\mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s} \mathrm{“}\mathrm{F}\mathrm{o}\mathrm{S}<1\mathrm{”}}{\mathrm{t}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{n}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{s}\mathrm{i}\mathrm{m}\mathrm{u}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{s}}}$$

(13)

This probability output gives more information about stability than a single deterministic value of FoS. Non-landslide or negative samples are selected as points with minimum PoF. 236 slope units are selected as negative samples, and a complete dataset is formed.

2.6. Machine Learning Models

In this study, three boosting algorithms were employed: XGBoost, CatBoost, and LightGBM. These models are selected for their proven high performance, efficiency, and suitability for LSM problems.

2.6.1. Data Preprocessing

The dataset contains two subsets: slope units with labeled landslide occurrences (118 slope units) and slope units with non-occurrences (236 slope units). This class imbalance was handled using class weighting during ML models training. Spatial coordinates of each slope unit center (longitude and latitude) were retained temporarily for clustering but excluded from model training features. The remaining features were imputed using mean values to handle missing data and standardized to zero mean and unit variance.

2.6.2. Dataset Splitting

The collected dataset is divided into a training set (70%) and a test set (30%) to provide enough data for model learning while reserving a separate subset for unbiased evaluation [81].

2.6.3. Feature Selection and Importance

A preliminary XGBoost model was trained on the training dataset to compute SHAP values that evaluate the influence of each feature on the predictions. SHAP values were calculated for all 22 input features, and the 19 most influential features were retained for subsequent modeling to reduce model dimensionality while maintaining predictive power. Accordingly, the three least important predictors were excluded from the datasets used in the two proposed approaches. Spearman correlation analysis was used to identify and exclude collinear features, showing very strong Spearman correlations (greater than 0.9) [82]. Additionally, SHAP values were computed for the 19 features used to assess their relative importance in all models [77,83].

2.6.4. Spatial Cross-Validation Strategy

To obtain a realistic and spatially aware assessment of model performance, a spatial cross-validation strategy was implemented in this study. First, the spatial coordinates, longitude and latitude of each slope unit center, were extracted from the dataset and used to perform KMeans clustering, which grouped the samples into five spatial clusters based on their geographical proximity. Longitude and latitude were then excluded from the dataset used for model training.

These clusters were used as group labels in a GroupKFold cross-validation strategy with 5 splits. This ensures that all data points within a spatial cluster are either used entirely for training or for validation in any fold. The grouped folds were applied during hyperparameter tuning for the three utilized algorithms. This provides a strong framework for gauging the generalization capability of the model on unseen geographic regions.

2.6.5. Model Training and Hyperparameter Optimization

Hyperparameters significantly affect the predictive capability of ML models and must be carefully optimized in order to prevent problems such as underfitting or overfitting. Hyperparameter tuning was performed with grid search. The key hyperparameters of the three competing classifiers, XGBoost, LightGBM, and CatBoost: the maximum number of boosting rounds (iterations), the maximum depth of each decision tree, the learning rate, and regularization parameters. These hyperparameters were selected for having a large impact on model behavior, and their value ranges can be found in

Table 5. Grid Search Hyperparameter Ranges for Model Tuning.

Parameter	XGBoost	LightGBM	CatBoost
Iterations	100, 300, 500	100, 300, 500	100, 300, 500
Depth	3, 5, 7	3, 5, 7	3, 5, 7
Learning Rate	0.05, 0.1	0.05, 0.1	0.05, 0.1
Regularization	reg_lambda: 1, 3, 5	reg_lambda: 1, 3, 5	l2_leaf_reg: 1, 3, 5
Model-specific	tree_method: “hist”	boosting_type: ‘dart’	random_strength: 1.25
	gamma: 0.5	skip_drop: 0.15	max_bin: 8

. Early stopping was applied during training to avoid overfitting. 5-fold GroupKFold cross-validation was used with AUC (Area Under the Curve) as the metric to ensure that the models generalized well to new data. The final models with the best cross-validation performance were chosen.

2.6.6. Performance Evaluation

The evaluation of the trained models was done using a set of widely accepted quantitative metrics. These include:

• Area Under the Receiver Operating Characteristic Curve (AUC): This is a basic measure representing the overall classification power of the model. The higher the value of AUC, the greater the reliability and predictive power of the model.
• Accuracy: It is the ratio of correctly identified landslide and non-landslide cases to the total count of instances.
• Precision and Recall (Sensitivity): Precision represents the ratio of true positive predictions to all predicted positives, whereas recall (or sensitivity) indicates the ratio of true positives to the total number of actual positives.
• F1-Score: The harmonic mean of precision and recall, offering a balanced measure of model performance.

Models Performances were compared through these metrics to determine the effect of different negative sample selection methods. SHAP values were computed for interpretability, and confusion matrices were generated for performance assessment on the test data.

2.6.7. Susceptibility Scores Prediction

The trained ML models were applied to the full set of delineated slope units using only the top-ranked features identified through prior feature selection. These features were scaled using the same transformation applied during training to maintain consistency. Subsequently, models that achieved the highest performance metrics were used to generate continuous landslide susceptibility scores for all slope units under both present and projected rainfall conditions. The continuous susceptibility scores were classified into five susceptibility classes, Very Low, Low, Moderate, High, and Very High. Finally, the classified maps were overlaid with the mapped locations of Gasshō-style houses and other historic buildings in the three UNESCO-designated villages, Suganuma, Ainokura, and Ogimachi, to analyze the distribution of structures across susceptibility classes.

3. Results

3.1. Slope Unit Generation

For slope unit delineation with the widely used r.slopeunits tool, the segmentation metric F should be optimized to achieve the best balance between internal homogeneity, external heterogeneity, and squareness of slope units, providing a more realistic representation of the terrain. In r.slopeunits, this optimization is performed by iterating over combinations of parameters, namely flow accumulation threshold (t), minimum circular variance (c), and minimum surface area (a). While the developers recommend iterating over different sets of (a, c) [14,15], Figure 12 shows fluctuations in F with varying a and c (horizontal axes) and t (colored lines), supporting the need for a tri-parametric optimization over all three parameters (t, a, c) to achieve the optimal slope-unit layer.

For the study area, the analysis identified the optimal configuration as $t=\mathrm{1,000,000}$ m², $a=\mathrm{300,000}$ m², and $c=0.1$, resulting in 1611 slope units with an average size of 339,399 m² (Figure 13). This optimized layer provides a physically consistent and hydrologically meaningful foundation for LSM, preserving natural terrain boundaries.

Figure 14 shows the distribution of the sizes of the slope units, represented as a histogram for the areas of all delineated slope units. The figure indicates a positively skewed distribution: most slope units are concentrated in the lower to mid-size range, which reflects that the segmentation process has caught the localized features of the terrain successfully. A smaller number of units possess significantly larger areas, often corresponding to broad, geomorphologically homogeneous regions. This variation supports the adaptive nature of the r.slopeunits algorithm, which allows slope unit size to reflect underlying terrain complexity rather than enforcing uniform geometries.

3.2. Feature Selection and Importance

SHAP values with a preliminary XGBoost model were used for feature selection to identify the three least important predictors to be excluded from the datasets used in the two proposed approaches. Specifically, silt, aspect, and LULC were removed from the dataset prepared using the buffer-based method, while Vs30, elevation, and LULC were excluded from the physically derived dataset. For both datasets, TRI and plan curvature were excluded due to high Spearman correlation with slope and profile curvature, respectively. Among the correlated factors, slope and profile curvature were retained due to their higher importance scores relative to the excluded variables.

Moreover, SHAP values for XGBoost, LightGBM, and CatBoost models (visualized as bar plots in Figure 15) were used to interpret feature importance for the two negative sampling strategies:

• Buffer-Based Strategy: SHAP plots reveal that the most influential features across all three models are profile curvature, slope height, elevation, and TPI.
• Physics-Based Strategy: Common dominant features across the models include slope angle, slope height, precipitation, lithology, and distance to roads, while the hybrid approach also highlights the enhanced importance of soil-related parameters such as clay, silt, and sand proportions.

These SHAP plots provide a quantitative basis for interpreting landslide mechanisms by showing the impact of each conditioning factor on the model output (horizontal axis) and can offer guidance for future studies on environmental factor selection in similar geomorphological and climatic settings.

3.3. Model Performance

Each model underwent grid search optimization using spatially grouped cross-validation. The tuned hyperparameters are demonstrated in

Table 6. Optimized hyperparameters for landslide susceptibility models.

Strategy	Algorithm	Depth	Iterations	Learning Rate	L2 Regularization
Buffer	XGBoost	7	100	0.1	3
	LightGBM	7	500	0.1	3
	CatBoost	3	100	0.1	1
Hybrid	XGBoost	5	100	0.1	5
	LightGBM	5	100	0.1	5
	CatBoost	7	100	0.1	1

.

Performance was assessed using accuracy, precision, recall, F1-score, and AUC on the test data. The hybrid strategy consistently outperformed the buffer-based method across all models. The most notable AUC improvement was seen in XGBoost, where AUC increased from 0.859 (buffer) to 0.931 (hybrid).

Table 7. Performance metrics of the models on the test dataset.

Strategy	Algorithm	Accuracy	Precision	Recall	F1-Score	AUC
Buffer	XGBoost	0.747664	0.609756	0.694444	0.649351	0.859
	LightGBM	0.785047	0.675676	0.694444	0.684932	0.868
	CatBoost	0.766355	0.622222	0.777778	0.691358	0.863
Hybrid	XGBoost	0.897196	0.857143	0.833333	0.845070	0.931
	LightGBM	0.897196	0.878788	0.805556	0.840580	0.931
	CatBoost	0.925234	0.966667	0.805556	0.878788	0.929

summarizes these results.

Confusion matrices of the hybrid method demonstrate a significant decrease in false negatives and false positives as shown in Figure 16. This indicates better reliability in identifying areas at risk for landslides.

The ROC (Receiver Operating Characteristic) curves in Figure 17 illustrate how effectively each model separates classes. The curves for the hybrid models have a larger area under the curve (AUC), which shows better performance in distinguishing between the classes. The three hybrid models achieved a similar AUC of approximately 0.93, while the buffer-based models showed a lower range of 0.859–0.863. These trends affirm the benefit of the proposed strategy for negative sample selection. The AUC values of the proposed framework (0.93) exceed those reported by Liu et al. [84], which achieved 0.866 based on five-fold cross-validation and identifying negative samples based on deterministic factor of safety estimates.

3.4. Susceptibility Mapping

Hybrid models, trained on physically generated dataset, were employed to predict landslide susceptibility scores for all slope units due to their superior performance over those trained on buffer-based dataset. Figure 18 shows the landslide susceptibility maps generated by the three hybrid models under present conditions (top) and future rainfall projections (bottom).

All three models display broadly similar spatial patterns under present conditions. Under future rainfall projections. some areas, particularly in XGBoost, show an increase in susceptibility class, but overall patterns remain consistent with the present scenario. Increased rainfall patterns have already increased slope unit susceptibility scores, however these changes were not always sufficient to shift the susceptibility class of most slope units to a higher one. This can be explained by the dominance of slope angle and slope height across the three hybrid models, as shown in the SHAP bar plots (Figure 15). Consequently, a larger increase in precipitation is needed to change the model’s susceptibility class, whereas smaller changes in slope angle or height (dominant features) are sufficient to produce the same effect. These maps provide a visual basis for identifying priority areas for hazard mitigation measures and assessing exposure of Gasshō-style houses and other historic buildings to landslide hazard.

3.5. Distribution of Gasshō-Style and Other Historic Buildings Across Susceptibility Classes

The distribution of historic buildings across landslide susceptibility classes was assessed in Ainokura, Suganuma, and Ogimachi using the three hybrid ML models under both current and projected rainfall conditions. Figure 19 shows that buildings in the south and west of Ainokura village consistently fall within the High susceptibility class, while the remaining buildings in the eastern–northern corner range from Very Low to Moderate across all three models. The increase in susceptibility due to future rainfall patterns was not sufficient to shift buildings into higher susceptibility classes.

In Suganuma village, most buildings are classified as Very Low susceptibility. Four buildings in the southeast corner are classified as High by CatBoost and LightGBM, and by XGBoost under historical conditions (Figure 20). Under future rainfall conditions, XGBoost shifts these four buildings to Very High susceptibility, highlighting the influence of future rainfall on these specific locations.

Most buildings in Ogimachi village are classified as Very Low susceptibility across all models (Figure 21); however, 9 buildings in the central area are classified as Low by the LightGBM model, while 26 buildings located in the central and southern parts are classified as Low by the CatBoost model. The historical and future scenarios show similar spatial patterns across all models, as the rainfall-induced increase in susceptibility is insufficient to cause a shift in susceptibility classes.

Overall, Ogimachi exhibits low landslide risk for historic buildings, whereas Ainokura (particularly in the south and west) and Suganuma (in the southeast corner) show elevated susceptibility, highlighting the need for focused hazard management in these areas.

4. Discussion

4.1. Impact of Slope Unit Optimization

This study presented an improved methodology for slope unit delineation by systematically analyzing the integrated impact of three parameters, namely flow accumulation threshold (t), minimum surface area (a), and circular variance (c). In contrast to previous studies that have usually considered only two parameters at a time, either (a, c) or (t, c), our approach examines 48 combinations of (t, a, c). Figure 12 indicates large variability in the segmentation metric F within all tested combinations, which proves the relevance of tri-parametric optimization. Larger F values correspond to a slope unit partition that attains a better balance among internal homogeneity, external heterogeneity, and geometric squareness. The optimal configuration was chosen for t = 1,000,000 m², a = 300,000 m², and c = 0.1, which corresponds to the highest F value; it ensures that the generated slope units produce a physically consistent basis for subsequent landslide susceptibility modeling.

4.2. Effect of Negative Sample Selection Strategy

The method of selecting non-landslide (negative) samples had a great impact on the performance of the model. In the buffer-based strategy, the negative samples were chosen from those areas that fall at least 500 m away from mapped landslides. However, this assumption may include unstable zones that have not yet failed.

The physically based strategy used in the hybrid models selected negative samples from slope units with the lowest PoF values. This ensured that training data represented physically stable conditions and led to consistent improvement in all models.

Performance gains in terms of AUC were as follows:

• XGBoost increased from 0.859 to 0.931, which is 0.072 higher.
• LightGBM went from 0.868 to 0.931, a gain of 0.063.
• CatBoost improved from 0.863 to 0.929, an increase of 0.066.

Similar improvements were observed across other classification metrics. For example:

• Accuracy for CatBoost increased from 0.766 to 0.925, precision from 0.622 to 0.967, recall from 0.778 to 0.806, and F1-score from 0.0.691 to 0.879.
• LightGBM had significant improvements under the hybrid strategy across all metrics, especially precision (from 0.676 to 0.879).
• Also, XGBoost had a similar trend: accuracy rose from 0.748 to 0.897 and F1-score from 0.649 to 0.845.

Confusion matrices further illustrate these gains. Under the buffer-based method, a higher number of false positives (stable units incorrectly predicted as slope failures) and false negatives (missed slope failures) were recorded. Conversely, the hybrid method reduced both types of errors substantially, with increases seen in both sensitivity (true positive rate) and specificity (true negative rate).

These results show that the physically based criteria generate cleaner training data. Improvements across all metrics and models indicate that the quality of the training labels is fundamental to effective landslide susceptibility modeling.

4.3. Feature Importance and Interpretability

To identify the most influential conditioning factors in the models, SHAP values were used for all three ML models in both sampling strategies. SHAP values estimate the contribution of each variable to the model output.

As an initial feature selection step, SHAP analysis with XGBoost removed the three features with the lowest impact in each dataset. In the buffer-based strategy, silt content, aspect, and LULC had the least SHAP values and were hence removed. For the hybrid models, Vs30, elevation, and LULC were removed. In both datasets, TRI and plan curvature were removed as they were highly correlated (Spearman) with slope and profile curvature, respectively.

The SHAP summary bar plots in Figure 15 present consistent patterns of feature importance for each strategy:

• Buffer-based strategy: the dominant features across all models include profile curvature, slope height, elevation, and TPI. The strong presence of purely topographic features reflects the nature of this sampling strategy, where physical factors are less considered.
• Physics-Based Strategy: Slope angle, slope height, precipitation, lithology, and distance to roads were the most relevant feature predictors, which are closely related to the physical initiation of landslides. Interestingly, the SHAP plots also highlighted higher levels of importance for soil-related variables (proportions of clay, silt, and sand).

These results clearly indicate that the hybrid approach improves both predictive performance and model interpretability because it provides a closer representation of actual landslide failure mechanisms. This is consistent with Wang et al. [85], who reported that physically informed ML models offer not only improved predictive performance but also greater interpretability. By tying the predictions to physically meaningful features such as precipitation, lithology, and soil properties, models gain more credibility for practical applications in hazard zoning and infrastructural development.

4.4. Comparative Performance of ML Models

All three gradient boosting models, CatBoost, LightGBM, and XGBoost, performed well across both sampling strategies. These findings indicate that, while model selection helps with performance optimization, the quality of input data and sampling strategies have a greater impact on final model performance.

4.5. Implications for Landslide Risk Management in Shirakawa-go and Gokayama

The villages of Shirakawa-go (Ogimachi) and Gokayama (Suganuma and Ainokura) represent an invaluable cultural heritage. The protection of Gasshō-style houses and other heritage structures within these UNESCO World Heritage Sites is therefore of critical importance. This study presents an overall susceptibility assessment, which highlights that buildings in the south and west of Ainokura village have high landslide susceptibility, while those in the eastern–northern corner remain in lower susceptibility classes. In Suganuma, most buildings are classified as Very Low susceptibility, except for four in the southeast corner, which are High under historical conditions and shift to Very High under future rainfall according to XGBoost. In Ogimachi, most buildings remain in the Very Low class, with only a few in the central and southern areas classified as Low.

This spatial information may help provide the basis for the following risk-reduction strategies:

• Targeted mitigation can thus focus on specific slope stabilization, improved drainage, and land-use restrictions in those most vulnerable areas, minimizing impact on the unique cultural landscape.
• Integrating LSM with real-time monitoring of triggering factors such as heavy rainfall in susceptible zones can strengthen early warning systems and disaster preparedness.

4.6. Limitations of This Study and Future Work Proposal

Although the proposed hybrid ML framework has shown high predictive capability and interpretability in the investigated LSM, some limitations have to be recognized in order to direct future improvements:

• Some input layers in this study, such as soil parameters, were available only at a coarser spatial resolution of 250 m that can affect the prediction accuracy.
• Incomplete or imprecise mapping of past landslides could introduce bias in the process of sample selection.

The following are some directions that may help address these limitations and further advance the methodology:

• Acquiring finer-scale soil parameters will help in giving more accurate estimates of PoF.
• Enhancing the physics-based approach by incorporating more sophisticated slope stability models and applying sensitivity analysis to identify which type of parameters contributes the most to PoF.
• Integrating spatial exposure of Gasshō-style houses with physical vulnerability models will allow a more complete and focused landslide risk assessment for the area.

By pursuing the proposed extensions, the framework can evolve into a more adaptive tool for long-term landslide risk management in both heritage and nonheritage landscapes.

5. Conclusions

This study presented a hybrid ML framework for LSM of the culturally significant region of Shirakawa-go and Gokayama, Japan. Unlike traditional buffer based methods, the approach here utilizes MC in order to estimate the Probability of Failure (PoF), considering the uncertainty in geotechnical parameters. Major conclusions derived from the analysis are discussed below:

• The systematic investigation of the three parameters, flow accumulation threshold (t), minimum surface area (a), and circular variance (c) showed how they are inter-related in influencing the segmentation quality metric F. The oscillations of F demonstrated that segmentation quality was sensitive to this tri-parametric optimization.
• The integration of MC gave an enhanced landslide hazard estimation by providing probabilities of failure instead of single deterministic safety factors.
• The selection of negative samples in hybrid models relied on low probability of failure (PoF) values. This provided much cleaner and more reliable training data, with a significant reduction in classification errors, and enhanced interpretability.
• SHAP analyses highlighted physically meaningful variables like precipitation, lithology and soil properties in hybrid models, while the buffer-based models emphasized topographic indicators.
• All three gradient boosting classifiers demonstrated strong and consistent performance across both sampling strategies, with notable improvements observed when trained on physically derived datasets. This close performance across models suggests that factors such as input data quality, and sampling strategy have a greater influence on final outcomes than the specific algorithm used.

With the integration of high-resolution geotechnical and hydrological data, the proposed framework holds strong potential to serve as a physically meaningful tool for LSM and long-term risk management.