Using Machine Learning and LAHARZ to Develop a Landslide Risk Analysis Model for Buildings

Despite the plethora of studies on landslide analysis and prediction, buildings are often the structures that endure the most tangible harm and must address the aftermath. In Korea, landslide damage attributable to climate change is escalating, particularly impacting buildings and residences. To mitigate this issue, it is imperative to forecast the areas where landslides are likely to occur and identify structures within their potential damage range. Consequently, this study aims to develop a landslide risk analysis model for buildings.

This landslide risk analysis model consists of three steps: (1) deriving landslide-susceptible areas, (2) deriving landslide damage areas, and (3) identifying buildings expected to be damaged by landslides.

To derive landslide-susceptible areas, data on past landslide occurrences and environmental variables related to topography, soil, vegetation, and climate were utilized. To enhance the reliability of the dependent variable, Pearson's correlation coefficient was employed to exclude variables with high intercorrelation. Machine-learning-based ensemble models—namely artificial neural networks (ANN), extreme gradient boosting (XGBoost), and generalized linear models (GLM)—were then applied to analyze these landslide-susceptible areas. The area under the curve (AUC) for the final model’s accuracy analysis was 0.934, indicating a high degree of predictive accuracy.

To derive the landslide damage area, various runout models were considered, and LAHARZ was ultimately selected as the analysis tool. LAHARZ, developed by the United States Geological Survey (USGS), can simulate debris flow behavior and is frequently used for landslide damage analysis. In this study, potential landslide initiation points—identified from the landslide-susceptible area results—were combined with weather, topography, geology, soil, and vegetation data to determine the extent of debris flow damage in the event of a landslide.

In the final stage of the analysis, buildings located within the debris-flow damage area were extracted. To achieve this, building register information was geocoded and converted into spatial data, using the geocoding tool on a selected sample area. The analysis revealed that in 10 of the 19 potential landslide sites, buildings are situated within the damage range in the event of a landslide. However, in the remaining 9 sites, no buildings are damaged even if a landslide occurs. Consequently, a total of 67 buildings in the sample area are likely to be damaged. These include 14 apartments, 6 multi-family/multi-unit houses, 2 single-family houses, and 1 apartment complex. The model developed in this study can serve as a foundation for residents and building users to respond more effectively to potential landslide damage.