Introduction
Landslides have long been persistent natural disasters that cause significant damage to society. Many landslides have caused substantial harm to human life and property, resulting in profound societal and environmental costs. Moreover, due to climate change, an increase in rainfall and intensified rain patterns is anticipated (IPCC, 2021), which could increase the risks associated with landslides.
Efforts to safeguard the population from damage caused by landslides are continually underway. Every year, substantial budgets are allocated for landslide mitigation projects. For example, in 2023, the Korea Forest Service(KFS) allocated 15.7 billion won to survey and manage landslide-prone areas, and an additional 163 billion won was allocated for related initiatives (KFS, 2023). However, considering the vast and varied topographical landscape of the nation, inherent challenges are present in implementing comprehensive landslide response measures, particularly budgetary and human resource limitations.
Therefore, predicting areas vulnerable to landslides and assigning priorities for landslide response is of utmost importance. To implement landslide mitigation effectively, it is vital to assign priorities cautiously, considering both financial and temporal constraints. This strategic approach aids in reducing landslide damage and bolsters the nation’s overall resilience against such disasters. Moreover, delineating priority zones can pave the way for improved resource allocation.
Several domestic and international methodologies used historical data to forecast landslide-prone zones. In South Korea, Geographic Information Systems (GISs) is used to assess the environmental variables which influence landslide incidents, providing a framework for potential risk modeling(Kim et al., 2005; Yang et al., 2007: Koo et al., 2018). Globally, researchers have merged satellite imagery with geographical data, furthermore, diverse modeling and simulation tools have been used to discern the link between climate change and landslides(Hong et al., 2007; Weirich and Blesius, 2007; Kirschbaum et al., 2018).
Although numerous studies exist, methodologies that integrate climate-change scenarios into landslide predictions are not without flaws. Considering the complexity of climate change, its integration into landslide prediction models poses significant challenges. This underlines the urgent need to deepen our understanding of the interplay between climate change and landslides, coupled with a greater emphasis on applying relevant climate scenarios. This could pave the way for more precise landslide predictions.
Therefore, the objectives of this study are as follows: First, to construct a Landslide Susceptibility Model (LSM) for the optimal analysis of landslide-prone areas in South Korea. Landslide occurrence data from 2010 to 2019 were collected, and environmental variables were adjusted accordingly. Second, to compare methods for predicting future landslide-prone areas. To achieve this objective, this study compared methods predicting climate change in the mid-21st century (2040–2049) and the late 21st century (2090–2099) annually and over a 10 year average. Third, to identify the priority for landslide response at the city and county levels by analyzing the increase and decrease in sensitive areas in the present and future. The research objectives presented in this paper are expected to offer a distinctive approach compared with existing studies, and contribute effectively to minimizing landslide damage and enabling the preparation of response measures.
Research Methods
Study area and geological setting
The study area is South Korea, located in the eastern region of the Asian continent between longitudes 124°11′ to 131°52′ and latitudes 33°06′ to 38°27′ (Fig. 1). Based on local government administrative divisions (cities/counties/districts), South Korea comprises of six metropolitan cities and nine provinces, a total of 77 autonomous cities and 82 counties, thus encompassing 165 administrative regions. This area covers an extent of 100,210 km2. The temporal scope of this research aims to predict future landslide-prone areas based on historical data from 2010 to 2019 (Period A). For the future predictions, this study describes two periods: the mid-21st century, 2040–2049 (Period B), and the late-21st century, 2090–2099 (Period C).
The general climatic characteristics of South Korea are influenced by its position within the mid-latitude temperate climate zone, which has four distinct seasons. According to the Korea Meteorological Administration, the annual precipitation averages approximately 1306.3 mm, with summer rainfall amounting to 710.9 mm, accounting for 54% of the annual total(KMA, 2023). Geologically, the region is predominantly comprises of igneous and metamorphic rocks that generally exhibit greater ground stability than sedimentary rocks. However, areas with sedimentary rocks, especially those with a high water influx or weak ground conditions, are at an increased risk of landslides. Statistics indicate that approximately 20% of the nation’s land area comprises of these sedimentary regions. Furthermore, 63.2% of the land is mountainous and inherently susceptible to landslides (NSDI, 2021).
Materials and methodology
Flow of study
The methodology for evaluating landslide susceptibility in this study is as follows:
(1) Collection and development of a landslide inventory map and controlling factors: a landslide inventory map was systematically gathered and constructed. The inventory contained landslides which occurred within a predefined time period and area in the past in a given area. within addition, factors related to landslide occurrence, such as topography, soil, vegetation, and rainfall were collected. (2) Construction of the LSM using the landslide inventory map and controlling factors: Using the consolidated landslide inventory map and landslide controlling factor data, 10 diverse statistical and machine learning models were applied. The ensemble method combined the prediction results of individual models to develop the LSM. (3) Projecting rainfall factors of periods B and C onto the LSM of past period A: After applying past landslide-prone areas to the LSM, rainfall factors were subsequently extracted from future scenarios of periods B and C were projected onto the LSM. This allows for the prediction of changes in landslide susceptibility owing to future rainfall variations. (4) Comparison and analysis of landslide-pone areas for each scenario and year: Through the projected LSM results, the landslide-prone areas derived for each scenario and year during Periods B and C were compared and analyzed. This provides assistance to understand the potential occurrence of landslides in the future based on different climate scenarios and changes in landslide-prone areas. (5) Identification of highly landslide- prone areas: Based on the scenarios that resulted in the most landslide-prone areas during periods B and C, regions vulnerable to landslides were identified (Fig. 2).
Data preparation
In the context of the landslide susceptibility assessment, the fundamental and crucial steps of data collection are explained. A landslide inventory was constructed by processing landslide occurrence data provided by the KFS. The landslide conditioning factors were categorized based on terrain, soil, vegetation, and rainfall criteria, resulting in a total of 16 environmental variables being selected. Each factor was extracted as follows:
Terrain factors were extracted from the 30m digital elevation model (DEM) data obtained from the National Geographic Information Institute. The altitude, aspect, slope, distance from main ridge (DMR), planar curvature (plancurv), profile curvature (profilecurv), stream power index (SPI), and topographic wetness index (TWI) were determined. Geological data were extracted from a 1:250,000 geological map provided by the Korea Institute of Geoscience and Mineral Resources. Soil factors, including soil depth and soil type, were extracted from the 1:5,000 Forest Site Soil Map provided by the KFS. Vegetation factors such as forest type, diameter class, and age class were extracted from the 1:25,000 Forest Map provided by the Korea Forest Service. Climatic factors were extracted from the MK-PRISM v2.1 model developed by the Korea Meteorological Administration, included summer rainfall, daily maximum rainfall, and five-day maximum rainfall data from June to August.
For the future landslide susceptibility assessments, future rainfall factors were extracted from the Shared Socioeconomic Pathways (SSP) scenarios proposed in the Intergovernmental Panel on Climate Change(IPCC) Sixth Assessment Report. All landslide controlling factors were sampled using ESRI’s ArcGIS 10.8 software, creating a 1 × 1 km pixel grid.
Landslide inventory
Data of landslides caused by heavy summer rainfall and typhoons was collected to construct a landslide inventory for the research area. The landslide inventory is an essential resource that accurately represents past landslide occurrence locations and the current distribution of landslides. Thus, the understanding of the relationship between landslide conditioning factors and sensitive areas can be enhanced.
The data were obtained using field surveys and satellite imagery. Field surveys were conducted to confirm the actual locations of the landslides and satellite imagery was used to estimate the landslide locations.
Using this method, landslide data from 2010 to 2019 was collected and a landslide inventory consisting of 2,995 cases was constructed (Fig. 1). The compiled landslide inventory allowed for the comprehensive of the patterns of past landslide occurrences and the current distribution of landslides.
Landslide conditioning factors
Landslide conditioning factors were selected to perform the landslide susceptibility assessments based on the characteristics of the research area (Table 1). These controlling factors, which can directly or indirectly influence the occurrence of landslides, include topography, soil, vegetation, and rainfall. These factors represent the complex interactions that influence landslide occurrence. Careful selection and analysis was conducted to construct a reliable landslide model. The selected landslide conditioning factors are listed in Table 2.
The altitude topographical factor is directly influenced by various environmental conditions (Moharrami et al., 2020). Aspect determines the degree of sunlight exposure, which affects soil moisture and slope stability(Camilo et al., 2017). Slope indirectly influences landslides under hydrological conditions(Tang et al., 2020). The DMR suggests that factors influencing landslides can vary based on distance. Planform curvature relates to the convergence and divergence of cross flows(Pourghasemi and Kerle, 2016). Profile curvature affects the driving and resisting stresses in the direction of movement within landslides (Pourghasemi and Kerle, 2016). Geology describes the geological engineering characteristics(Ciurleo et al., 2017). The SPI is an index that measures the surface runoff erosion capacity, while ‘TWI’ predicts areas which are sensitive to hydrological conditions on the soil surface(Tang et al., 2020; Sörensen et al., 2006). In areas with a thin effective soil depth, the soil becomes weathered, which limits the surface soil and increases the risk of landslides. Soil type indicates the characteristics of the surface soil affecting landslide occurrences. Vegetation factors significantly influenced the slope failure(Aditian et al., 2018). Factors such as ‘foresttype’, ‘diamclass’, and ‘ageclass’ are closely related to soil-root binding strength. Rainfall factors(Lee et al., 2020; Nishioka et al., 2023), including the amount of rainfall affecting the shear strength of a slope, are among the primary factors that cause landslides(Chen et al., 2017; Chen et al., 2018; Ciurleo et al., 2017). A systematic analysis of the relationships between the selected landslide conditioning factors and landslides is expected to enhance the accuracy and reliability of landslide models.
Multicollinearity arises when there is a high correlation between variables, leading to instability in predictive out-comes. Thus, the assessment and resolution of multicollinearity are crucial for landslide modeling. A Pearson correlation analysis is used to understand linear correlations between variables, and a correlation coefficient absolute value of 0.7 or above, or −0.7 or below, is considered to indicate multicollinearity(Booth et al., 1994 ; Bui et al., 2016). The Variance Inflation Factor (VIF) analysis is an index for assessing multicollinearity. Variables with a VIF value greater than 10 were considered to exhibit ‘high’ multicollinearity (O’Brien, 2007). Multicollinearity was analyzed using Pearson’s correlation analysis and VIF.
Individual model
Various statistical and machine learning models were applied to construct individual models for the landslide susceptibility assessment. The characteristics and types of the 10 models used are summarized in Table 3.
A generalized linear model (GLM) is a statistical model that describes the linear relationships between dependent and independent variables(McCullagh, 1984). The generalized additive model (GAM) is an extension of the GLM that is capable of modeling nonlinear relationships(Hastie and Tibshirani, 2004). The Gradient Boosting Machine (GBM) is an ensemble technique that combines multiple decision trees to enhance the predictive performance(Elith et al., 2008). Artificial neural networks (ANNs) are machine learning models that mimic biological neural networks and can model complex nonlinear relationships. Stepwise regression (SRE) is a variable selection technique that constructs a model by selecting significant variables(Busby et al., 1991). The CTA (Classification tree analysis) is a model that learns decision-making rules based on data for prediction(Breiman, 1994). A Flexible discriminant analysis (FDA) is an extension of discriminant analysis techniques, performing classifications considering non-linear relationships(Breiman, 1994). Random Forest (RF) is a model that enhances predictive accuracy by creating multiple decision trees and assembling them(Ripley, 2007). The Multivariate Adaptive Regression Splines (MARS) is a model that identifies nonlinear relationships in multivariate data(Friedman, 1991). Maximum Entropy (MAXENT) is a probabilistic model based on the maximum entropy principle that is used for ecological distribution modeling(Phillips et al., 2006).
We executed each individual model with five repetitions to effectively mitigate the inherent uncertainties associated with a single run. A comprehensive review of the advantages and disadvantages of each model was performed by utilizing various models and repeatedly driving them. This greatly aids the selection of a more reliable LSM.
Ensemble models
This study employed ensemble modeling techniques in the spatial distribution modeling to enhance the prediction accuracy. Ensemble models integrate various models to provide stronger predictive and generalization capabilities (Rossi et al., 2010). Applying these benefits to landslide modeling allows for more accurate and efficient predictions. Ensemble models were used to evaluate the performance of the individual models deemed to have superior performance, with Area Under the ROC Curve (AUC) values of 0.7 or higher(Fielding and Bell, 1997).
The mean of the probabilities (PM) computes and combines the average landslide occurrence probabilities of the selected models. The confidence interval (CI) enhances the prediction by reflecting the confidence interval for the average landslide occurrence probability. The median of probability (PME) adjusts the predicted results using the median of the landslide occurrence probabilities of the selected models considering the influence of outliers. Committee Averaging (CA) converts the probability of selected models into binary values (0 and 1) using each model’s threshold and subsequently calculates their average to combine the predicted outcomes. The (weighted mean of the probability (PWM) applies weights based on the trustworthiness scores of the individual models when computing the average and determines the relative importance of each model(Kim et al., 2018).
The experimental results showed that among the various ensemble models, the PMW method demonstrated the best performance and was thus adopted. By applying such ensemble techniques to landslide modeling, it is possible to predict critical information, such as the possibility of disaster occurrence, with greater precision.
Results and Discussion
Correlation analysis
A model was constructed considering the multicollinearity among 16 landslide conditioning factors. No variables with high multicollinearity were identified in the study. This suggests that each variable exists independently without strong correlations with one another. Therefore, the 16 conditioning factors selected for landslide modeling independently contributed to the predictions. Incorporating these variables into the model was appropriate, enhancing the stability of the prediction outcomes, and allowing for a more precise understanding of the influences between the variables.
Validation of landslide models
A cross-validation using a random approach was used for landslide modeling. Cross-validation is a critical process for evaluating the model performance and understanding its generalization capabilities. Random cross-validation arbitrarily divides the data to generate training and validation sets by repeatedly evaluating the model’s performance. The landslide data were divided according to 80% for training and 20% for validation. Furthermore, by generating non-occurrence points three times and using them with actual landslide occurrence points for model training and validation, the model bias was reduced and its generalization capacity was enhanced.
The AUC was utilized as the evaluation metric. The AUC represents the area under the Receiver Operating Characteristic (ROC) curve, with values ranging from 0 to 1. An AUC value of 0.5 signifies the model’s predictions are at a random discernment level. The closer the AUC value is to 1, the better the predictive power, with a value near 1 representing a perfect prediction model(Fielding and Bell, 1997).
The individual model AUC values from the 150 runs are displayed in Fig. 3. The ANN, GLM, and SRE models displayed AUC values of 0.7 or lower, suggesting a relatively poor performance or insufficient predictive power compared to other models, and thus were excluded from the ensemble model. Based on these results, the constructed LSM achieved an AUC value of 0.748, effectively distinguishing between landslides and non-occurrences. Thus, an LSM model with an appropriate AUC value could deliver reliable predictions, and is considered to exhibit excellent performance as a landslide prediction model.
Landslide susceptibility map
Landslide Susceptibility model
In the constructed LSM, the Landslide Susceptible Area was derived based on the CTA, FDA, GAM, MARS, MAXEIT, and RF models, and the 16 landslide conditioning factors. To evaluate landslide susceptibility, a binary map was used. A binary map represents one of the two categories or classes, indicating whether each location or pixel belongs to a given class. The landslide susceptibility was represented for each point or pixel simply as ‘occurrence’ or ‘non-occurrence’ (Fig. 4). Using this binary representation, the potential for regional landslide occurrence was analysed and predicted.
The area of the landslide-susceptible region during period A was identified to be 31,902 km2. Landslide-prone areas were predominantly found in the southern region, along the eastern coastline, and surrounding the northwestern areas. This distribution mirrors the distribution of landslide occurrence points from 2010 to 2019. This implies that the LSM provides realistic and accurate outcomes based on historical data. Ultimately, the similarity between the actual landslide distribution and the LSM results enhanced the predictive capabilities and credibility of the model, validating the LSM value in landslide response strategies. Emphasizing that the LSM accurately reflects the actual landslide occurrence points, using the LSM to predict future landslide-sensitive areas was reiterated to be a reliable tool.
The contribution of the influential variables in the constructed LSM was analyzed (Table 4). Variables such as slope, SPI, soil type, diameter class, forest type, plan curvature, and profile curvature exerted the least influence, with each contributing less than 1%. Elevation, a topographical factor, had the greatest influence (39%). Subsequently, the distance from the ridgeline and rainfall factors such as summer rainfall and daily maximum rainfall were found to influence the LSM in the range of 10 to 15%. These findings help to elucidate the impact of landslide conditioning factors on landslide occurrence.
Future landslide susceptibility map
SSP projection
To anticipate future landslide susceptibility, the LSM from Period A was projected onto scenarios SSP1-2.7, SSP2-4.5, SSP3-7.0, and SSP5-8.5. This helped to predict landslide-prone areas for periods B and C. Our analysis utilized two methods.
Mean analysis - by averaging the rainfall factors every decade for periods B and C, the LSA was analyzed.
Overlay analysis - The LSA was deduced annually for periods B and C and was then overlaid to understand the LSA for both periods.
Fig. 5 presents the results of the mean analyses. This approach was instrumental in discerning the holistic distribution of the LSA, and observing its temporal shifts. When juxtaposed with the landslide zones of period A, a discernible escalation of 50–90% was observed across all periods and scenarios. This indicates a probable intensification of landslide risk in the forthcoming periods. Intriguingly, the decline in the LSA growth rate from period B to C was exclusive to scenario SSP1-2.7 This decline reflects the scenario’s emphasis on sustainable development, coupled with proactive climate change interventions. Such insights underline the pivotal role of sustainability and proactive measures, in addition it indicates the potential efficacy of scenario SSP1-2.7 in curbing landslide hazards.
Conversely, the SSP5-8.5 scenario registered the most modest surge in landslide proneness at 35% for period B, however had a sharp escalation at 91% for period C. This trend could be linked to the somewhat gentle undertones of SSP5-8.5, which presumes unrestrained greenhouse gas emissions and a lackluster approach towards climate change mitigation. This emphasizes the heightened risk associated with a passive stance on environmental concerns, as exemplified by the SSP5-8.5 scenario.
Fig. 6 displays the results of the overlay analysis. Although the LSA trend mirrored the mean analysis, this method delved deeper, offering insights into the frequency of the LSA occurrences. The annual derivations of landslide-prone areas afforded a granular view of shifts across all periods and years. Furthermore, it enabled stratification of these areas based on recurrence. This nuanced understanding augments the precision of gauging landslide susceptibility. Thus, when strategizing landslide mitigation priorities, the overlay analysis seemingly held an edge over the mean analysis.
Landslide response priority
An overlay analysis method was adopted to determine the priority of landslide responses in the research area. The overlay analysis provided significant insights into the LSA characteristics in each region. Moreover, to enhance the differentiation of the LSA, the areas where the LSA occurred were extracted every year for each period. This approach added precision to the determination of landslide response priorities by considering the frequency of the LSA occurrence.
To address future landslide risks, the response priorities targeting future LSA were analyzed using the SSP5-8.5 scenario. The SSP5-8.5 scenario assumed a society undergoing rapid growth based on fossil fuels, and compared to other SSP scenarios, it predicted the most landslide occurrences. As it offers various paths for the future depending on socioeconomic conditions and policy decisions, analyzing the SSP5-8.5 scenario, which was expected to have the most significant increase in landslide occurrence, was prioritizes.
Based on the findings presented in Table 5, notable disparities in the results of the periodic LSA for the top 10 regions in each respective period have become evident. Leveraging this information, we have categorized landslide response priorities in a manner that aligns with the format commonly employed in landslide hazard maps produced by the KFS. These categories are defined based on LSA area proportions as follows: fifth priority with an LSA area of 1–20%, fourth priority with 21–40%, third priority with 41–60%, second priority with 61–80%, and first priority with 81% and above. Applying these criteria, 238 regions were classified within the fifth priority, 85 regions within the fourth priority, 24 regions within the third priority, and 6 regions within the second priority. Notably, Goseong County in Gyeongsangnam-do stood out with a high LSA rate of 88.4%, leading to its classification as the first priority. This suggests that this region should be prioritized for future landslide risks.
Discussion
A LSM was constructed in this study using 10 individual models in conjunction with an ensemble approach. To estimate the future LSA, rainfall factors from the SSP scenarios were projected. This enabled the analysis of potential changes in landslide susceptibility during the forthcoming periods. Additionally, the two analytical methods were compared to provide a detailed assessment of future LSAs.
The mean analysis method has a clear advantage in illustrating the overall distribution of LSAs and their temporal variations. The distribution and time-based changes in the future LSAs using this method are summarized in Table 6. This facilitates a comprehensive understanding of the overall trends and distribution patterns of LSAs across different periods. This mean-based approach provided significant insights into comprehending future landslide risks, and was instrumental in contrasting and deciphering the implications of different scenarios.
The overlay analysis was instrumental in distinctly distinguishing the characteristics of the LSAs for each period, proving invaluable in determining priority areas for landslide responses. This method evaluates overlaps in LSA regions derived from every year within a selected period and identifies areas persistently exposed to LSAs in the future. Moreover, by examining the yearly growth rate of LSAs compared with past records, detailed trends in landslide occurrence could be discerned, as illustrated in Fig. 7.
The prioritization approach proposed in this study was effective in verifying the necessity of addressing areas that are most susceptible to landslides. Through the adopted methodology, we meticulously analyzed the regional characteristics of LSAs, enabling the efficient designation of landslide response priorities in the study area. Leveraging these insights, we anticipate identifying areas at imminent risk of future landslides, devising appropriate response strategies, and implementing effective disaster management and protection measures.
The findings and analyses of this study offer critical information to governmental bodies and relevant agencies for formulating landslide response policies and resource allocation. This study is expected to contribute to the safety and disaster management in local communities. Furthermore, the research methodology and approach hold potential for broader applications, making them viable for predicting and addressing natural disasters in other regions, and thus emphasizing their relevance in future predictive model development.
Conclusion
In this study, to construct the LSM, 10 individual models and an ensemble method were employed using the PMW approach to enhance the prediction generalization and accuracy. Considering future environmental changes, rainfall factors from the SSP scenarios were projected onto the LSM to evaluate landslide susceptibility. By utilizing both the average and overlap analysis methods, insights into future LSAs were provided from multiple perspectives, laying the groundwork for future landslide response strategies.
A limitation of this study is that while various controlling factors influencing landslide occurrence were considered in the susceptibility analysis, future susceptibility evaluation solely relied on changes in rainfall factors using SSP scenarios. Therefore, an integrated approach that considers a multitude of controlling factors is essential for responding to future landslides. Addressing this issue within a short timeframe may prove challenging and necessitate continued research.
In future studies, we plan to conduct precise analyses at resolutions higher than 1 km × 1 km, focusing on areas with high landslide response priorities. This would provide more detailed LSA information for these high-priority regions, greatly aiding in a more precise evaluation of landslide susceptibility and the formulation of response measures.
Building on such analyses, the authors aim to predict damage scales by focusing on urban areas, infrastructure, and agricultural lands, which are the primary targets of landslide damage. Thus, the aim is to provide reliable information on landslide susceptibility predictions, pinpoint response priorities, and formulate more effective response strategies.
PDF Links
PubReader
ePub Link
Full text via DOI
Download Citation
