Yunnan Plateau Landslide Risk Mapped Using Optimized Machine Learning

Researchers have mapped the landslide susceptibility of the Central Yunnan Plateau in Southwest China by integrating optimized statistical and machine learning models. Published in Geoenvironmental Disasters in April 2026, the study identifies specific climatic and geological conditions that significantly increase the risk of slope failure in this densely populated region. The team developed a data-driven framework that combines point-biserial correlation and decision tree analysis to refine input variables before feeding them into three distinct predictive models: random forest, support vector machine, and logistic regression. This approach allows for a more robust assessment of where landslides are most likely to occur, moving beyond simple historical records to identify underlying environmental drivers.

The Central Yunnan Plateau, spanning approximately 113,000 square kilometers, is particularly prone to geohazards due to its complex geology and monsoon climate. The region experiences intense, short-duration rainfall events that can destabilize slopes, especially where human activities such as road construction and agricultural expansion have altered the natural landscape. Previous studies in Southwest China have often focused on other provinces, leaving a gap in comprehensive susceptibility mapping for the Central Yunnan Plateau. This study addresses that gap by utilizing a geotechnically confirmed landslide inventory from 2021 to 2023, ensuring that the dataset reflects recent and verified events rather than potentially biased historical reports.

To build the susceptibility model, the researchers first optimized the rainfall data used as a predictor. Landslides are often triggered by specific rainfall patterns, but the optimal time window for measuring this trigger can vary. The team compared cumulative rainfall over 24 hours, 7 days, 15 days, 21 days, and 30 days prior to landslide events. Using point-biserial correlation, they found that 24-hour and 30-day cumulative rainfall had the strongest association with landslide occurrence. The 24-hour rainfall captures the immediate triggering effect of intense storms, while the 30-day rainfall represents the antecedent wetness that can saturate the soil over time. Combining these two metrics provided a more comprehensive picture of the hydrological conditions leading to slope failure.

In addition to rainfall, the study optimized the classification of geological and land-use factors. Traditional lithological maps often group rocks in ways that do not align with their actual susceptibility to landslides. The researchers applied a decision tree algorithm to reclassify rock types based on their observed relationship with landslide occurrence. The analysis revealed that unconsolidated rocks and carbonate rocks with 30–70% carbonate content were significantly more prone to landslides than metamorphic, volcanic, or intrusive rocks. Similarly, for land use, the decision tree identified cropland and artificial surfaces as high-risk categories, while forests, grasslands, and shrublands were associated with lower susceptibility. This reclassification allowed the models to better capture the physical reality of slope stability in the region.

The optimized variables were then used to train and evaluate three machine learning models: random forest, support vector machine, and logistic regression. The models were trained on a dataset of 108 confirmed landslide points and 108 randomly generated control points, with the data split into training and validation sets to prevent information leakage. The random forest model, which uses an ensemble of decision trees to capture non-linear relationships, showed the most significant improvement in performance after the optimization steps. Specifically, the area under the curve (AUC) for the random forest model increased from 0.78 to 0.83 when lithological and land-use categories were reclassified and the combined rainfall metrics were used. This improvement indicates that the optimized features provided a stronger predictive signal than the raw data.

The final landslide susceptibility map was generated by averaging the predictions from the three models, weighted by their respective performance. This ensemble approach helps to mitigate the biases of any single algorithm and provides a more stable probability surface. The resulting map shows that landslide susceptibility is not uniform across the plateau. The central and northwestern parts of the region exhibit generally low susceptibility, while the eastern and southern margins are identified as high-risk zones. This spatial pattern aligns with the distribution of high rainfall and significant human activity in those areas. The eastern region, in particular, is characterized by a high density of roads and agricultural land use on susceptible carbonate rock, creating a perfect storm of environmental and anthropogenic risk factors.

To understand which factors drive these susceptibility patterns, the researchers performed a feature importance analysis using the optimized random forest model. The analysis ranked 24-hour cumulative rainfall and distance to roads as the two most influential factors. This finding highlights the critical role of both natural triggers and human infrastructure in landslide occurrence. The partial dependence plots revealed that landslide susceptibility increases sharply as 24-hour rainfall reaches 40 millimeters. Additionally, the susceptibility is highest within 1,000 meters of road networks, indicating that road construction and associated disturbances significantly destabilize slopes in the vicinity.

The study also examined the role of other factors, such as soil properties and topography. Soil organic matter content and clay content showed a positive correlation with landslide susceptibility, likely because these components affect soil water retention and shear strength. In the southern part of the plateau, where rainfall is highest, the combination of steep slopes and high soil organic matter content creates conditions that are particularly vulnerable to failure during extreme rainfall events. In contrast, factors such as altitude, slope angle, and the Normalized Difference Vegetation Index (NDVI) had a comparatively smaller influence on the model’s predictions. This suggests that while topography and vegetation cover are important, the specific combination of rainfall intensity, lithology, and proximity to human infrastructure are the dominant predictors of landslide susceptibility in the Central Yunnan Plateau.

Uncertainty analysis was conducted to assess the reliability of the susceptibility map. The researchers calculated the standard deviation of the predicted susceptibility values across the three models for each grid cell. Higher uncertainty was observed in the eastern and southern margins, where the models diverged most significantly. Specifically, the logistic regression model tended to predict more extreme susceptibility values in these regions, whereas the support vector machine model yielded lower and less heterogeneous predictions. Despite these numerical differences, the spatial patterns of susceptibility remained largely consistent across models, with pairwise correlation coefficients ranging from 0.62 to 0.67. This moderate to strong agreement lends confidence to the identified hotspots and suggests that the ensemble approach effectively captures the underlying susceptibility structure of the region.

The findings have practical implications for landslide risk management in the Central Yunnan Plateau. The identification of 24-hour and 30-day rainfall as key predictors suggests that early warning systems should monitor both short-term storm intensity and cumulative wetness. Furthermore, the high importance of distance to roads highlights the need for prioritized geotechnical inspections and maintenance of drainage systems along road networks, especially in areas underlain by susceptible rock types. Integrating landslide susceptibility into land-use planning is also recommended, with a focus on avoiding large-scale development in high-risk zones and promoting vegetation restoration in areas currently used for agriculture or construction.

The study acknowledges several limitations that should be considered when interpreting the results. The landslide inventory is limited to events from 2021 to 2023, which may not capture the full range of landslide types and triggers over a longer period. Additionally, the susceptibility map was generated using long-term averaged rainfall data, which assumes a stationary relationship between mean climatic conditions and extreme rainfall events. This assumption may not hold in a changing climate, where the spatial distribution of extreme rainfall could shift. Future research should aim to integrate high-resolution, spatio-temporal rainfall estimates into susceptibility models to better account for the probability of extreme events. Despite these limitations, the study provides a valuable framework for assessing landslide susceptibility in complex mountainous regions and offers actionable insights for risk mitigation in the Central Yunnan Plateau.

Subject of Research: Landslide susceptibility assessment in the Central Yunnan Plateau by assembling optimized statistical and machine learning models

Article Title: Landslide susceptibility assessment in the Central Yunnan Plateau by assembling optimized statistical and machine learning models

Article References: Chen, J., Li, Q., Huang, C., Hu, X., & Shen, Z. (2026). Landslide susceptibility assessment in the Central Yunnan Plateau by assembling optimized statistical and machine learning models. Geoenvironmental Disasters, 13(1), Article 28. https://doi.org/10.1186/s40677-026-00380-w

Image Credits: AI Generated

DOI: 10.1186/s40677-026-00380-w

Keywords: Landslide, susceptibility, assessment, Central, Yunnan, Plateau, assembling, optimized, statistical, machine, learning, models