Landslides are pervasive geological hazards threatening lives, properties, and infrastructures worldwide, especially in mountainous regions (Chae et al. 2017). As an intrinsic process in the Critical Zone, landslides are of the primary concern in disciplines such as geology, ecology, and environmental sciences (Lee et al. 2023; Silhán et al. 2021). The rise in landslide frequency and intensity worldwide is largely attributed to the compounding effects of intensifying anthropogenic climate changes and human activities in recent decades (Emberson et al. 2021; Lacroix et al. 2020b; Froude & Petley 2018).

A landslide refers to the downslope movement of rocks, soils, debris, or a combination of these materials under the influence of gravity along a slope or hillside (Cruden 1991; Varnes 1958). The occurrences of landslides result from the interplay of various geological and environmental factors (Lacroix et al. 2020b). Among which, geological conditions, such as the presence of weak rock layers or faults, play a crucial role, as these layers may gradually become unstable under persistent stress (Schulz et al. 2018; Mainsant et al. 2012). Additionally, climatic factors, such as prolonged rainfall or heavy rainstorm, increase soil pore pressure, reducing slope stability (Handwerger et al. 2019; Schulz et al. 2009). Tectonic activities, including earthquakes, also contribute to landslide susceptibility by destabilizing rock and soil masses (Keefer 1984; Bontemps et al. 2020). Human activities, particularly the land use change (e.g. road construction, agricultural activities), could reshape surface morphology and vegetation cover (Petley et al. 2007; Lacroix et al. 2020a). Seismic activities also have significant synergistic effects on landslide occurrence (Frigerio et al. 2021). Therefore, landslide susceptibility varies across different natural and societal conditions (Zhi et al., 2023; Goetz et al., 2015), which interact in complex ways, influencing regional variations in landslide triggers (Pacheco Quevedo et al., 2023).

Landslide studies primarily adopt mechanism-driven and data-driven approaches to address causative analysis, susceptibility assessment, and risk evaluation (Dai et al., 2002; Crozier & Glade, 2005; Sim et al. 2022). Mechanism-driven methods, such as limit equilibrium analysis and finite element modeling, simulate slope stability dynamics through physical laws (Duncan, 1996; Griffiths & Lane, 1999). These methods provide explicit physical interpretability but require precise parameter calibration, and are limited for large-scale applications (Van Weste et al., 2008; Zhang et al. 2018). In contrast, data-driven methodologies span statistical models like logistic regression (LR) for spatial susceptibility assessment (Guzzetti et al., 2005), machine learning algorithms including support vector machine (SVM) and random forest (RF) for nonlinear pattern recognition (Tehrani et al., 2022; Merghadi et al. 2020), and deep learning architectures such as convolutional neural networks (CNNs) have been used to decipher landslide imagery (Sameen et al., 2020). While these techniques excel at handling complex relationships, their reliance on data quality and inherent "black-box" nature limit the interpretability. Recent advancements integrate statistical frameworks with machine learning, such as CNN and RF incorporating physical constraints (Liu et al., 2024), effectively leveraging domain knowledge and data adaptability to enhance model robustness and generalization in landslide susceptibility assessment.

Landslide studies across different scales yield varying interpretations in triggering mechanisms. Slope-scale studies generally focus on the landslide process and small-scale triggers, such as topography and lithological types, variations in pore water pressure, rock mass fracturing, and the impacts of localized rainfall (Liu et al. 2021; Perrone et al. 2008; Tan et al. 2023; Feng et al. 2024). In contrast, regional or global studies would further take into account the influence of climate change, geological process (e.g. earthquakes), and human activities on the frequency and intensity distributions of landslides. The increase in extreme weather events associated with global warming, such as intense rainstorms and snowmelts, significantly elevates landslide occurrences (Iverson 2000; Huggel et al. 2012; Gariano & Guzzetti. 2016). Regional-scale studies often leverage climate models and extensive geological surveys to explore strategies for mitigating the heightened landslide hazards under future climate scenarios (Hürlimann et al. 2022; Madhu et al. 2024). On a global level, research also highlights how human interventions, such as urban expansion and agricultural irrigation, alter regional hydrological characteristics and vegetation cover, consequently affecting landslide occurrence (Lacroix et al. 2020b; Froude and Petley 2018; Fidan et al. 2024).

China, a mountainous country with 64.89% of its land area consisting of mountains and plateaus (Deng et al. 2015), is also one of the countries that are most severely affected by landslides (Gómez et al. 2023). The Southwest China is especially exposed to frequent and intensive landslides (Zhang et al. 2018), ranking as the most catastrophic natural disaster in the area (Fan et al. 2019; Yang et al. 2020b; Chen et al. 2012). Neotectonic movements in this region have led to geomorphological alterations, forming numerous faults and folds, resulting in extensive rock fractures and unstable slopes (Zhang et al. 2007). The interaction between rugged topography and the monsoon climate further contributes to the regional intense rainfall (Huang 2009). Over the past decades, rapid increases in mining, road construction, and urbanization have further destabilized slopes, leading to an increase in landslide events (Yin et al. 2010; Li et al. 2012; Ma et al. 2018; Fan et al. 2017; Cui et al. 2022). In Yunnan Province, a representative hotspot in Southwest China, the coupling of active tectonic uplift (Xu et al., 2003), rainfall seasonality (Yang et al., 2020a), and intensive anthropogenic disturbances (Wang et al., 2022) has created a favorable environment for landslide development. A recent research shows that a RF model fed with geomorphologic variables and real-time rainfall records can provide early-warning for rainfall-driven landslides in Yunnan (Kang et al., 2024).

The Central Yunnan Plateau (CYP), located in the north-central part of Yunnan Province, is surrounded by large mountains and rivers. It has continuously served as the socioeconomic center of Yunnan and played a key role in the rapid development of Southwest China over the past decades (Li et al., 2024). CYP is also the most vulnerable region of geohazards in Yunnan, jointly caused by the high landslide frequency, impacts of natural and artificial triggers, and the typical fragility to geohazard in a densely-populated region. However, landslide research in southwest China has predominantly focused on Sichuan Province and northwestern Yunnan, with limited comprehensive studies on the CYP (Wu et al. 2022; Wang et al. 2024). Given its inclusion of the Central Yunnan Urban Cluster and the pivotal role in Yunnan’s socioeconomic development, it warrants closer investigation to understand landslide susceptibility and influencing factors.

This study aims to quantify landslide susceptibility by combining mechanism-based models with machine learning approaches. First, we used point biserial correlation and decision tree analysis to optimize temporal resolution of rainfall records, geological, and land-use parameters based on their physical relevance with landslide events. Subsequently, we compared RF, SVM, and LR models to evaluate the model fitting performance of different parameter combinations. Next, we used the optimized RF model to analyze the weight of each factor's impact on landslide occurrence, and implemented mechanism-based explanations. Finally, using the prediction results from the RF, SVM, and LR models, we generated a comprehensive landslide susceptibility map for the study area, and proposed corresponding mitigation strategies for landslide hotspots. Based on the mentioned approaches, this study aims to address the following key questions:

1) What is the spatial pattern of landslide susceptibility in the CYP? Which influencing factors are most significant?

2) How can we enhance the interpretability of landslide occurrence by take into account the nonlinear effects and interactions of the predictive variables?

3) What managing strategies are need to mitigate landslide risks in the study area and similar regions?

By addressing these questions, the study contributes to robust assessment of landslide susceptibility, and improves early-warning systems for geohazards in mountainous regions.

Existing landslide studies have identified precipitation, stratigraphy, and land use as key factors influencing landslide occurrence. However, these factors show different patterns and critical thresholds across regions (Caine 1980; Aleotti 2004; Guzzetti et al. 2008). We explored the relationships of these factors through various analytical methods.

We adopted a data-driven workflow of “parameter optimization – multi-model comparison – result integration” to capture the multifactor nature of landslide susceptibility in the CYP. We first used point-biserial correlation to examine rainfall data with different time resolutions for a preliminary comparison, we then applied a decision-tree algorithm to refine the categorical groups of lithology and land use, yielding variables optimized by both temporal solution and categories. We fed these predictors to three models—RF, SVM, and LR, evaluated their performance with balanced accuracy and AUC to verify the effectiveness of the optimization. The best-performing RF model was then employed to rank the relative importance of each factor. The arithmetic mean of the susceptibility outputs from the three models was used to generate a landslide susceptibility map, integrating results from statistical and machine-learning approaches. This framework preserves the interpretability and also takes advantage of the non-linear analysis of machine learning, providing a replicable route for landslide susceptibility assessment in complex mountainous terrains.

The Pearson correlation coefficients between the 24-hour cumulative rainfall prior to landslide events and the 7-day, 15-day, 21-day, and 30-day cumulative rainfall are 0.35, 0.36, 0.34, and 0.46, respectively. In contrast, the correlation coefficients among the other cumulative rainfall data of different time resolutions range from 0.61 to 0.94, significantly higher than those between the 24-hour cumulative rainfall and other time periods (see Fig. 3a).

Point-biserial correlation analysis reveals that the correlation coefficients between landslide occurrence and the cumulative rainfall for 24 hours, 7 days, 15 days, 21 days, and 30 days are 0.546, 0.292, 0.226, 0.268, and 0.303, respectively (see Fig. 3b). Notably, the 24-hour and 30-day cumulative rainfall exhibit the highest correlation coefficients with landslide occurrence.

Based on these results, the 24-hour and 30-day cumulative rainfall were selected as the optimized rainfall data parameters for further analysis.

This study optimized geological data, land use classifications, and combinations of rainfall data at different resolutions to create an average landslide susceptibility map for the CYP region using LR, SVM, and RF models. The findings indicate a generally low landslide susceptibility in the central plateau, consistent with previous studies by Wang et al. (2014) and Liu et al. (2023), which found a significantly lower landslide occurrence in this region compared to surrounding areas. This suggests that the CYP’s stable geological environment may reduce the threat of landslide hazard, benefiting regional economic and social development.

Areas with a higher landslide susceptibility are mainly located along the eastern and southern margins of the plateau, where rainfall is significantly higher than in the central and northwestern regions. Although the terrain in the eastern region is relatively gentle, a high density of roads and substantial agricultural land use indicates considerable human disturbance. Additionally, landslide-prone eastern areas primarily consist of carbonate rock with 30–70% composition. This region is also characterized by generally low vegetation cover (Rao et al., 2023). Due to erosion, mixed composition-induced instability, and heterogeneous physical properties, this type of rock is susceptible to landslides under certain conditions (Ehrenberg & Nadeau, 2005). This indicates that landslides in the eastern region result from combined effects of climate, human activity, and geological factors.

In the southern region, characterized by river valleys with steep slopes and high soil organic matter content, rainfall is the highest in the plateau. Although organic matter can enhance soil aggregation and water retention (Esmaeilzadeh & Ahangar, 2014), excessive organic matter can significantly increase soil water absorption capacity. During extreme rainfall, this can lead to soil saturation and reduced shear strength, thus raising landslide risk (Xuan et al., 2023). Metamorphic rocks typically exhibit good stability (Eberhardt et al., 2004), but complex fractures and faults may destabilize under external stress or increased water pressure from rainfall, triggering landslides (Ehteshami-Moinabadi, 2022). This suggests that the southern region is particularly susceptible to landslides following extreme rainfall.

Different studies utilizing machine learning for landslide modeling vary in the input factors (Dehnavi et al., 2015; Zhao et al., 2022; Jia et al., 2019; Song et al., 2025). Rainfall remains a common predictor, with researchers proposing rainfall thresholds based on duration and intensity (Caine, 1980). Although some argue that landslides are caused by water infiltration rather than rainfall volume (Terlien, 1998), limited data availability often restricts models to factors like annual precipitation, monthly averages, or cumulative event rainfall (Janizadeh et al., 2023; Smith et al., 2023). This study proposes a combined 24-hour and 30-day cumulative rainfall approach for landslide modeling, which yielded the best performance in predicting landslide susceptibility. The 24-hour rainfall reflects the impact of sudden events, while the 30-day rainfall represents cumulative effects, both contributing independently to landslide prediction (Smith et al., 2023).

On the lithology, the classification standards vary due to different academic viewpoints and data accessibility, often lacking geological relevance for landslide susceptibility assessment (Dehnavi et al., 2015; Chen & Zhang, 2021). To address this, a decision tree was used to reclassify the lithological types, yielding practical and reliable results. In the CYP, metamorphic, volcanic, intrusive, clastic, and carbonate rocks with more than 70% carbonate content are less prone to landslides, while unconsolidated rocks and carbonate rocks with 30–70% carbonate content are more prone. Metamorphic rocks typically exhibit high strength and low porosity, providing stability and reducing landslide risk (Eberhardt et al. 2004). Volcanic and intrusive rocks are generally dense and hard, with strong resistance to weathering and erosion, making landslides less likely (Ladygin et al. 2023). Clastic rocks have tightly arranged particles, low porosity, and permeability, which contribute to stability and reduce the risk of large-scale landslides (Smith et al. 2023). In contrast, unconsolidated rocks, with high porosity and permeability, have loose internal structures and are susceptible to landslides under water flow. While carbonate rocks with more than 70% carbonate content typically have high permeability, good drainage, and high strength, rocks with 30–70% carbonate content are prone to landslides due to erosion, mixed composition-induced instability, and heterogeneous physical properties (Ehrenberg & Nadeau 2005).

Based on these findings, the study proposes the following recommendations for landslide prevention and control in the CYP region:

While this study contributes to data optimization for landslide susceptibility quantification, some limitations remain. Due to data availability constraints, the landslide events include various types (e.g., slides, flows, and failures) with differing trigger mechanisms, which warrant further investigation (Hungr et al., 2014). Additionally, landslide records may be influenced by subjective factors and local economic conditions, introducing potential biases. Future research could benefit from integrating remote sensing, geological survey, and socioeconomic data to enhance data accuracy and reliability through data fusion and correction methods.

This study analyzed the spatiotemporal patterns and key factors influencing landslides in the CYP. By optimizing geological and land use classifications and integrating rainfall data at different resolutions, we improved the performance of the landslide susceptibility quantification. The results highlight that 24-hour cumulative rainfall, 30-day cumulative rainfall, human activities (such as road construction), and soil organic carbon content are critical factors affecting landslide occurrence. Landslide-prone areas are primarily concentrated in the eastern and southern regions of the plateau, where climatic conditions, human activities, and geological features are more favorable.

Based on these findings, this study recommends strengthening rainfall monitoring, with a focus on both daily and monthly rainfall dynamics. Additionally, enhancing the maintenance of roads and other infrastructure can help mitigate the impact of human activities in high-risk areas. Finally, it is also essential to prevent overdevelopment in vulnerable regions.