Building Damage Potential Assessment Based on Landslide Susceptibility Analysis in Major South Korean Cities
Article information
Abstract
Background and objective
Climate change driven increases in the frequency and intensity of localized heavy rainfall have undermined slope stability in the Seoul metropolitan area, where urban development meets steep terrain. This study develops a robust, ensemble-based landslide susceptibility model and applies it at the building scale to quantify damage potential, providing municipalities with precise, actionable guidance for risk mitigation.
Methods
We compiled 883 georeferenced landslide events and ten spatial predictors covering topography, soil properties, climate variables, and vegetation cover. To correct presence-only sampling bias, pseudo absence points were generated at a 3:1 ratio using the Small Range Envelope method. The data were divided into 80% for training and 20% for testing, with five-fold cross-validation. Ten machine learning algorithms were trained; models achieving a validation AUC ≥ 0.70 were combined into an ensemble via weighted averaging and committee voting. Ensemble outputs were reclassified into five quantile-based susceptibility classes (1: Very High; 2: High; 3: Moderate; 4: Low; 5: Very Low) and overlaid on building point data to assign each structure a class by use (residential, commercial, public).
Results
The ensemble achieved a validation AUC of 0.968. Areas of high landslide susceptibility are concentrated in the northern and eastern parts of Gyeonggi Province. Among municipalities, Yongin-si, Yangpyeong-gun, Gwangju-si, Namyangju-si, and Seoul Special City exhibit the greatest numbers of high-susceptibility buildings, with residential structures comprising 53% of these assets. This building-scale framework supports targeted, data-driven disaster mitigation planning.
Conclusion
Integrating an ensemble-based landslide susceptibility surface with building-use attributes provides a practical, building-scale measure of susceptibility-based damage potential across the Seoul metropolitan area. High-susceptibility assets are disproportionately single-family houses and Neighborhood Living Facilities, indicating clear priorities for mitigation. We also outline a policy-ready extension Potential Risk = (Susceptibility + Runout) × Exposure × Vulnerability to support pre-disaster screening and hazard-informed planning.
Introduction
Recent shifts in rainfall patterns and the increasing frequency of intense, localized downpours driven by climate change have markedly undermined the stability of steep slopes in Korea’s mountainous regions (Kim et al., 2023; Ding et al., 2023). In particular, extreme rainfall concentrated over short periods rapidly reduces soil shear strength, thereby elevating both the occurrence rate and potential severity of landslides (Oramas Dorta et al., 2007; Magirl et al., 2010). Climate scenario analyses project that the intensity and frequency of such localized heavy rainfall events will continue to rise on the Korean Peninsula, further heightening landslide-related hazards.
The Seoul metropolitan area characterized by high population density and extensive infrastructure is especially vulnerable, as urban developments often abut steep terrain. In the event of a landslide, not only are lives and property at stake, but critical transportation and utility networks (roads, railways, power grids) may also suffer widespread disruption. Consequently, there is a pressing need to shift from post-disaster recovery to proactive landslide risk management, underpinned by pre-event hazard assessment and mitigation planning.
Internationally, sophisticated studies have combined debris-flow runout physical models with building vulnerability functions to achieve detailed risk analyses. For example, Uzielli et al. (2015) developed vulnerability functions to estimate damage probabilities for individual structures, while Pereira et al. (2017) integrated rainfall scenarios into urban-scale risk assessments. More recently, Sun et al. (2023) employed landslide flow modeling to calculate mobility and energy flux, linking these to building damage likelihoods, and Del Soldato et al. (2017) paired field surveys with physical modeling to derive systematic damage grading for buildings.
Domestically, however, the scarcity of high-resolution topographic and soil data, alongside limitations in large-scale computational infrastructure, has constrained research to GIS-based susceptibility analyses. In complex environments like the Seoul metropolitan area where terrain, soil properties, and urbanization patterns interact there remains a clear need for a practical methodology that couples landslide susceptibility mapping with detailed building attribute data to quantitatively assess damage potential.
Accordingly, this study aims to evaluate building damage potential based on landslide susceptibility analysis across the Seoul metropolitan area. Focusing on building use attributes, we estimate susceptibility-based building damage potential at the structure level and generate foundational data to inform municipal disaster mitigation policies.
Research Methods
Study area
The study area encompasses the Seoul Special City metropolitan region of South Korea, including Gyeonggi Province. To ensure analytical consistency and a clearly defined scope, Incheon Metropolitan City, Gimpo-si, Paju-si, Yeoncheon-gun, and Pocheon-si areas characterized by military restrictions and limited access to topographic data were excluded from the analysis (Fig. 1).
Study area with landslide inventory and administrative boundaries.
The Seoul Special City metropolitan region covers approximately 11,875.9 km2 and hosts a population of 26,076,470, making it the nation’s largest urban agglomeration with extensive socio-economic infrastructure (KOSIS, 2024). Although Seoul Special City has relatively little steep terrain, its high population density and concentrated infrastructure amplify the potential impacts of landslides. Accordingly, this study conducts a building-level landslide damage potential assessment that accounts for these geographic and social characteristics.
Study Data
This study integrates historical landslide records with environmental data to quantitatively assess landslide susceptibility in Gyeonggi Province. The inventory comprises 883 georeferenced events 145 in 2011, 110 in 2013, and 628 in 2020 merged into a single dataset. By reflecting long-term landslide history under diverse climatic and topographic conditions, this integrated approach enhances model generalizability (Sun et al., 2023).
Environmental predictors were categorized into topography, soil, vegetation, and climate variables, each known to influence landslide initiation (Table 1). Topographic variables represent slope morphology and stability; soil variables affect infiltration and pore-water pressure dynamics during rainfall; vegetation variables account for root reinforcement and surface water retention; and climate variables capture key rainfall characteristics that trigger slope failures (Ding et al., 2023; Jiménez-Perálvarez et al., 2009; Kim et al., 2023; Magirl et al., 2010; Oramas Dorta et al., 2007; Youberg et al., 2014). All layers were resampled to a consistent 30 m × 30 m grid to ensure analytical uniformity.
Environmental variables for landslide susceptibility modeling
Finally, to mitigate multicollinearity, we applied Pearson correlation analysis and removed predictors with correlation coefficients of 0.7 or higher, resulting in a final set of ten variables. This preprocessing widely adopted in landslide susceptibility research improves model stability and removes redundant information (Del Soldato et al., 2017).
Methodological Workflow
This study is structured into four sequential steps (Fig. 2). First, we collected and preprocessed key environmental variables topography, soil, climate, and vegetation alongside historical landslide occurrence records to build the analysis dataset. Second, we applied Pearson correlation analysis to identify and remove multicollinear predictors, selecting the core variables for modeling. Third, we trained and validated ten individual machine learning models, then combined those meeting performance criteria into an ensemble to generate a landslide susceptibility map. Fourth, we overlaid this susceptibility map with building location data and integrated building-use attributes to assess the susceptibility-based damage potential for each structure.
Workflow for ensemble-based landslide susceptibility mapping and building-level damage potential assessment.
Spatial Distribution Modeling
Ten individual statistical and machine learning models were constructed to predict landslide susceptibility: Artificial Neural Network(ANN), Classification Tree Analysis(CTA), Flexible Discriminant Analysis(FDA), Generalized Additive Model(GAM), Generalized Boosting Model(GBM), Generalized Linear Model(GLM), Multiple Adaptive Regression Splines(MARS), Maximum Entropy(MAXENT), Surface Range Envelope(SRE), and eXtreme Gradient Boosting (XGBOOST) (Stanley et al., 2020).
To mitigate data imbalance, pseudo absence samples were generated at a 3:1 ratio relative to presence points using the SRE method before model training (Jiménez-Perálvarez et al., 2009). The full landslide inventory was partitioned into 80% for training and 20% for validation, and five-fold cross-validation was performed to assess generalization performance (Stanley et al., 2020).
In this study, we evaluated the performance of the individual models using Area Under the ROC Curve (AUC) and Cohen’s Kappa (Nahm, 2022). AUC is a threshold-independent summary of discrimination, and values of ≥ 0.70 are generally regarded as acceptable. Kappa is threshold-dependent but corrects for chance agreement between observed and predicted classes, and values of ≥ 0.61 are typically interpreted as indicating substantial agreement (Li and Yu, 2022). We did not collapse these two indicators into a single criterion; rather, we interpreted their performance independently. At the ensemble stage, only algorithms satisfying AUC ≥ 0.70 and Kappa ≥ 0.61 were retained, and we compared four combination strategies median probability, simple mean, committee averaging, and weighted mean of probabilities. We also considered a weighted-mean scheme in which model contributions are proportional to their respective metric values (Pyakurel et al., 2024).
Building Damage Potential Assessment Based on Landslide Susceptibility
In this study, we assessed damage potential at the building level using an ensemble-based landslide susceptibility map. Susceptibility probabilities were reclassified into five quantile-based categories Very Low, Low, Moderate, High, and Very High and, because building locations are represented as points, we applied a GIS “extract values to points” overlay to assign each building the susceptibility class of its corresponding raster cell (Jiménez-Perálvarez et al., 2009). This method enables quantitative assessment of damage potential for point-based building locations by combining GIS analysis with building attribute data, without requiring complex runout simulations.
We also incorporated building-use attributes according to the Enforcement Decree of the Building Act, classifying structures as residential (e.g., detached houses, apartments), commercial (e.g., neighborhood living facilities), or public (e.g., educational and research facilities, religious buildings). This allowed us to distinguish relatively higher-risk buildings within the same susceptibility category (Cho et al., 2024).
Results and Discussion
Results
The ensemble-based landslide susceptibility map developed in this study was classified into five categories (1: Very High to 5: Very Low) based on predicted probabilities (Fig. 3). This classification enables an intuitive assessment of spatial landslide potential. The results reveal that high-susceptibility zones are concentrated in the northern Gyeonggi area and along mountain–urban interfaces. The final model incorporated ten core environmental predictors Altitude, Aspect, Profile Curvature, Stream Power Index (SPI), Topographic Wetness Index (TWI), Forest Age Class, Soil Depth, Soil Drainage, Daily Maximum Rainfall, and Summer Rainfall (June-August) selected via Pearson correlation analysis to remove redundancy and optimize performance. This spatial distribution provides a foundation for comparing susceptibility across different regions and is particularly useful for proactive monitoring and tailored mitigation planning in high-susceptibility areas.
Landslide susceptibility mapping and building-level classification.
Fig. 4 presents the distributions of ROC–AUC and Cohen’s Kappa for each algorithm, visualizing the variability in model performance. All algorithms achieved AUC ≥ 0.70, indicating acceptable discrimination. In contrast, the Kappa panel shows that GAM fell below 0.61 and was therefore excluded from the ensemble; all other algorithms met the Kappa ≥ 0.61 threshold and were retained under the inclusion criterion.
Model performance distributions (ROC-AUC and Cohen’s Kappa) by algorithm.
Among the ensemble strategies tested, the AUC-weighted mean achieved an AUC of 0.968, while the Kappa-weighted mean achieved a Kappa of 0.835, both indicating high performance. In sum, although Fig. 4 displays the performance of individual algorithms, the comparison of ensemble averages confirms that the ensemble improves predictive consistency and reliability over any single model. We therefore adopted the AUC-weighted mean approach to produce the final landslide susceptibility map.
Overlaying the susceptibility surface with building points identified municipalities with the largest numbers of Very High (Class 1) buildings (Table 2). Yongin-si has the highest count (Class 1 = 2,222; Total = 17,653), followed by Gwangju-si (Class 1 = 981; Total = 10,892), Yangpyeong-gun (Class 1 = 626; Total = 11,181), Icheon-si (Class 1 = 408; Total = 4,741), and Yeoju-si (Class 1 = 400; Total = 5,037). These figures indicate notable clusters of high-susceptibility structures but not local majorities (for reference, Yongin-si ≈ 12.6% and Yangpyeong-gun ≈ 5.6% of all buildings are Class 1). For readers less familiar with the region, these jurisdictions are primarily located in the northern and eastern parts of the Seoul metropolitan area where mountainous terrain interfaces with urban development.
Ranking of municipalities by building-level landslide susceptibility
By building use (Table 3), Single-family House dominates the high-susceptibility tally across the top ten municipalities (42,482 buildings; ≈53.5% of the listed total 79,419), followed by Neighborhood Living Facility Type 2 (8,175) and Neighborhood Living Facility Type 1 (4,539), which together account for ≈16.0%. Multi-family Housing contributes 7,634 (≈9.6%), with additional counts from Factory (4,781), Animal and Plant-related Facility (3,903), Warehouse Facility (3,830), Correctional and Military Facility (1,549), Educational and Research Facility (1,372), and Accommodation Facility (1,154). This composition shows that high-susceptibility concentrations are centered on Single-family House and Neighborhood Living Facility categories, underscoring the need to prioritize mitigation for these uses.
Landslide susceptibility by building usage in high-susceptibility municipalities
Discussion
This study quantitatively evaluates landslide susceptibility in the Seoul metropolitan area and links it to individual buildings, thereby moving beyond earlier work that merely visualized areas with potential landslide occurrence. In particular, the finding that single-family houses and Neighborhood Living Facilities are concentrated in high-susceptibility zones indicates that landslide impacts are closely associated not only with topographic factors but also with the urban spatial distribution of building types.
However, a fundamental limitation of this study is that it does not explicitly account for the movement paths and spatial extent of landslide runout. Because it does not estimate which routes a landslide would follow or how far it would travel, there are constraints on quantitatively evaluating the actual probability of damage. Future research can address this by coupling the susceptibility surface with runout simulations such as LAHARZ and FLO-2D, enabling more refined predictions of damage probability and magnitude (Ding et al., 2023; Sun et al., 2023; Kim and Kim, 2025).
While a susceptibility-based analysis is effective for identifying potential initiation areas, it does not directly incorporate building exposure and vulnerability, which limits its usefulness for policy decision-making. Traditionally, risk is defined as Risk = Hazard × Exposure × Vulnerability (Uzielli et al., 2015; Pereira et al., 2017), and integrating these three components is necessary to quantitatively predict actual losses. However, estimating hazard faces practical limitations due to uncertainty in rainfall. Accordingly, we propose a concept of potential risk that couples susceptibility with runout characteristics to approximate the spatial extent of impact, and considers this together with exposure and vulnerability (Potential Risk = (Susceptibility + Runout) × Exposure × Vulnerability).
Exposure can be represented by combining building counts, floor area, and use from the Ministry of Land, Infrastructure and Transport (MOLIT) Building Register with land-use status data from the National Geographic Information Institute (NGII) to capture building density and functional importance. Vulnerability can be quantified by weighting and aggregating attributes in the Building Register number of stories, primary structural material, foundation type, year of construction, and seismic-design status to produce a quantitative index. By integrating these indicators with the susceptibility and runout results, it becomes possible to produce policy-ready potential risk maps at the individual-building scale.
Conclusion
This study employed an ensemble of ten machine-learning algorithms to quantify landslide susceptibility across the Seoul metropolitan area and link it to building-level, susceptibility-based damage potential. We find that single-family houses and Neighborhood Living Facilities are concentrated in high-susceptibility zones, providing practical evidence for prioritizing mitigation beyond simple visualization of potential occurrence areas.
We further propose a practical framework for potential risk mapping by coupling runout simulations with the susceptibility surface and integrating exposure and vulnerability: Potential Risk = (Susceptibility + Runout) × Exposure × Vulnerability. This reframes terrain-based prediction as a building-scale assessment of damage potential suitable for practice.
The framework has clear regulatory and policy applications. First, in pre-disaster impact assessments, high-susceptibility classes (for example, Classes 1–2) can serve as screening thresholds to compare avoid-mitigate-redesign options at the planning stage. Second, in urban planning, zoning, and development control, agencies can establish hazard (susceptibility) overlay zones and condition building/ earthwork permits on mitigation measures such as drainage and slope reinforcement. Third, to support national hazard and land-use databases, the outputs can be delivered as standardized GIS layers interoperable with the government Building Register (maintained by local governments), NGII base/cadastral layers, and MOLIT land-cover maps, enabling periodic updates and external validation toward operational deployment. These layers can be versioned and updated regularly, with pilot-jurisdiction validation prior to full adoption.