Statistical Models Map Landslide Risks in Taiwan’s Yilan County

The study, published in Geoenvironmental Disasters, applies three multivariate statistical methods to a large-scale inventory of past landslide events to identify areas most prone to future slope failures. By integrating geographic information systems with statistical algorithms, the team aimed to provide a reproducible approach for hazard assessment that can be adapted to similar geological settings worldwide.

The investigation focused on Yilan County, which covers an area of 2,143 square kilometers and is characterized by a mix of plains and mountainous terrain. The region sits within the collision zone between the Philippine Sea Plate and the Eurasian Plate, leading to significant geological deformation and frequent earthquakes. To build a robust dataset, the researchers compiled a catalog of 12,047 landslide events recorded between 2004 and 2017. These events were derived from the semi-automatic interpretation of high-resolution satellite imagery, including SPOT and Formosat-2 data, with a minimum detectable area threshold of 0.1 hectares. The inventory was validated through field surveys and aerial photograph comparisons to ensure data quality.

The study incorporated twelve causative factors spanning topography, hydrology, geology, and seismicity. Topographical variables, such as slope degree, aspect, and curvature, were extracted from a 30-meter resolution Digital Elevation Model derived from ASTER satellite data. Hydrological factors included annual rainfall, sourced from radar-gauge integrated data, and distance from river networks. Geological considerations involved lithology and distance from fault lines, while seismic parameters included peak ground acceleration and time-averaged shear-wave velocity to 30 meters depth. This multi-factorial approach allows for a detailed analysis of the complex interactions that drive landslide occurrence in the region.

To ensure the reliability of the statistical models, the researchers performed multicollinearity and Pearson correlation analyses. These tests confirmed that all twelve factors were suitable for inclusion, with no variance inflation factors exceeding 10 and correlation coefficients remaining below 0.7. A chi-square ranking was then used to determine the relative importance of each factor. The analysis identified slope gradient and annual rainfall as the two most influential determinants of landslide susceptibility in Yilan County. This finding aligns with the region’s climatic and topographic characteristics, where steep slopes and heavy precipitation create ideal conditions for slope failure.

The core of the methodology involved applying three distinct multivariate statistical methods: the Analytical Hierarchy Process (AHP), Frequency Ratio (FR), and Simple Additive Weighting (SAW). AHP is a multi-criteria decision-making technique that uses pairwise comparisons to assign weights to different factors based on their relative importance. Frequency Ratio calculates the likelihood of landslide occurrence by comparing the proportion of landslide pixels in each class to the proportion of total pixels. SAW, another decision-making framework, calculates a weighted sum of performance ratings to prioritize alternatives based on indicator scores. Each method was implemented using ArcGIS and R software to generate pixel-based susceptibility maps.

The dataset was partitioned into 70% training samples and 30% testing samples to evaluate the predictive capability of the models. The susceptibility maps were classified into four levels—low, moderate, high, and very high—using the Jenks natural breaks method. The results indicated that high and very high susceptibility zones comprised approximately 37.5% to 49.5% of the study area, depending on the method used. These high-risk areas were concentrated in southern Yilan, particularly in Tiansongpi Township, where steep slopes exceeding 30 degrees coincide with annual precipitation ranging from 3,100 to 4,700 millimeters.

Model validation was conducted using the Area Under the Curve (AUC) metric and a suite of error metrics, including mean squared error, root mean square error, and correlation coefficients. The AHP method achieved the highest AUC values, with 0.830 for the training set and 0.839 for the testing set. The FR method followed closely with AUC values of 0.828 for both training and testing, while the SAW method recorded 0.824 for training and 0.814 for testing. Although the differences in AUC values were relatively small, with a maximum difference of 0.025, the AHP method consistently outperformed the others across all evaluation criteria.

Further analysis of error metrics reinforced the superior performance of the AHP model. It yielded the lowest mean squared error of 1.7 × 10⁻³ and the lowest root mean square error of 2.7 × 10⁻². Additionally, the AHP method produced the highest correlation coefficient between predicted and observed values, reaching r = 0.9974. The FR and SAW methods exhibited slightly higher error values and lower correlation coefficients, suggesting that AHP provides a more stable and accurate representation of landslide susceptibility in this specific geological context.

The spatial distribution of susceptibility classes revealed that the southwestern part of Yilan is particularly vulnerable due to high average precipitation and steep topography. The study noted that the mountainous regions in the south and west are subject to a humid monsoon climate, with intense rainfall during typhoon season from June to October. Historical meteorological records show peaks in landslide occurrences during September, when saturated soils and elevated pore water pressure lead to slope failure. The 2010 Typhoon Megi served as a notable example, where landslide debris was transported downstream to form coastal alluvial fans, underscoring the empirical relationship between rainfall intensity and upstream slope failure.

Despite the strong performance of the statistical models, the authors acknowledged several limitations. The random partitioning of training and testing datasets did not account for spatial autocorrelation, which may inflate validation accuracy when samples are drawn from adjacent areas. Furthermore, the small differences in AUC values among the three methods were not formally tested for statistical significance. The study also relied on a single-epoch Digital Elevation Model, assuming static topographic conditions, whereas multi-temporal data could better capture the evolving landscape in areas with frequent large landslides.

The findings suggest that multivariate statistical approaches remain a practical and effective tool for landslide hazard assessment, particularly in regions where comprehensive geotechnical data are limited. The integration of GIS analysis with R software established a reproducible workflow that can be applied to other tectonically active and rainfall-prone regions, such as Southeast Asia, the Himalayas, and Central America. The identification of slope gradient and annual rainfall as dominant factors supports the integration of precipitation thresholds into early warning systems, enhancing the ability to predict future landslide hazards and mitigate their destructive effects.

Future research should address the limitations identified in this study by employing spatially stratified sampling or spatial cross-validation techniques to provide more conservative accuracy estimates. A direct comparison between these statistical methods and advanced machine learning techniques, such as random forest or deep learning, could help quantify the performance gains achievable with more complex models. Additionally, incorporating multi-temporal Digital Elevation Model data and dynamic landslide inventories would better capture the changing topographic conditions. Extending this framework to incorporate future climate projections would support long-term landslide risk assessment under changing precipitation patterns, contributing to the global effort in landslide risk reduction.

The study represents the first comprehensive, county-scale landslide susceptibility assessment for Yilan County that integrates twelve causative factors, including seismic parameters such as peak ground acceleration and shear-wave velocity that are frequently omitted in regional statistical studies. Unlike previous investigations that relied solely on the Area Under the Curve metric for model comparison, this research employs a multi-dimensional evaluation framework combining AUC with root mean square error, mean squared error, mean absolute error, standard deviation, and correlation coefficient. This approach provides a more comprehensive assessment of model performance, addressing the limitation that AUC alone may not fully capture spatial prediction accuracy in complex geological settings.

The integration of geographic information system analysis with R software for pixel-based data construction establishes a reproducible workflow applicable to other regions with similar tectonic and climatic conditions. The methodology involved converting the landslide inventory and thematic maps of causative factors into images, applying a 30 by 30 pixel aperture to generate over 30,000 samples. This pixel-based approach allows for a detailed analysis of landslide probability, calculated by dividing the count of landslide pixels by the total pixel count within each sub-region. The resulting matrix facilitates a robust statistical analysis of susceptibility across the entire study area.

Chi-square ranking identified five key factors as the most significant influences on landslide susceptibility: slope, rainfall, lithology, elevation, and slope aspect. In mountainous regions, steep gradients intensify gravitational forces on slope materials, while rainfall saturates soil layers, increasing pore water pressure and reducing shear strength. Lithology governs the mechanical and hydrological behavior of slope materials, with weathered Miocene shale, argillite, and volcanic breccia being particularly prone to failure when saturated. Elevation affects weathering and vegetation density, with higher elevations undergoing more intense weathering and having sparse vegetation, which further reduces slope stability. Slope aspect influences solar exposure and moisture retention, affecting the duration of soil saturation.

The study compared the performance of the three statistical models with similar studies conducted in other regions, such as the Darjeeling Himalayas and the Maze catchment in southern Ethiopia. The AUC values obtained in this study are comparable to those reported in other comparative studies, confirming that traditional statistical methods generally produce reliable but broadly similar results for regional-scale landslide susceptibility mapping. The relatively narrow range of AUC values observed across the three methods is consistent with findings from other regions, suggesting that larger performance gains typically require the introduction of machine learning-based approaches. This contextual comparison highlights the consistency of the multivariate statistical framework across different geological settings.

The framework presented in this study provides a practical and reproducible approach for landslide hazard assessment and risk management in similar geological settings worldwide. The identification of slope gradient and annual rainfall as the dominant causative factors reinforces findings from global landslide studies and supports the integration of precipitation thresholds into early warning systems. The reproducible workflow developed in this study can serve as a practical reference for researchers and practitioners in developing countries where advanced computational resources may be limited. By providing a detailed and validated susceptibility map, the study offers valuable insights for sustainable development planning and disaster preparation in Yilan County and similar regions.

Subject of Research: Landslide susceptibility mapping in Yilan County, Taiwan, using multivariate statistical methods

Article Title: Multivariate statistical assessment for mapping landslide susceptibility: a case study of the Yilan area, Taiwan

Article References: Putriani, E., Wu, Y.-M., & Chen, C.-W. (2026). Multivariate statistical assessment for mapping landslide susceptibility: a case study of the Yilan area, Taiwan. Geoenvironmental Disasters, 13(1), Article 31. https://doi.org/10.1186/s40677-026-00384-6

Image Credits: AI Generated

DOI: 10.1186/s40677-026-00384-6

Keywords: Landslide susceptibility, Yilan County, Taiwan, Multivariate statistics, AHP, Frequency Ratio, GIS, Natural hazards, Seismic activity, Rainfall, Multivariate, statistical