Rainfall-induced landslides (RIL) pose a significant threat to communities worldwide, causing loss of life, property damage, and disruption of essential infrastructure. RIL occurrence and magnitude are expected to intensify under changing climatic conditions, as alterations in precipitation regimes increase the likelihood of hydrologically driven slope failures. Despite substantial progress in landslide research, the development of reliable early warning systems that mitigate their effect is still under development due to limited data and the absence of standardised approaches for identifying the key environmental controls on slope instability. In this context, the present study advances a systematic methodology for determining the dominant features governing landslide initiation using as study case the Southern Andes. Establishing and quantifying these controlling variables is essential for improving monitoring strategies, enhancing predictive capabilities, and ultimately strengthening risk-reduction efforts in susceptible landscapes.

Recent studies have established that quantitative and data-driven methodologies markedly enhance the robustness and predictive skill of landslide models (Kumar et al., 2023; Merghadi et al., 2020). Within this framework, Steger et al. (2023) demonstrate that landslide forecasts are fundamentally constrained by the precision and representativeness of input data, showing that precipitation exerts seasonally variable controls on slope failure and therefore requires high-quality climatic observations. Complementary studies underscore the pivotal role of numerical simulations in landslide hazard assessment integrating geotechnical uncertainty to quantify the failure mechanisms and potential impacts (Jia et al., 2023; Fustos-Toribio et al., 2022, 2025). Moreover, hydrological modelling efforts further confirm the centrality of rainfall dynamics, with rainfall-driven frameworks demonstrating strong capability in the prediction of shallow landslides (Bezak et al., 2019). Collectively, these works highlight a critical need for methodological refinement that moves beyond broad variable inclusion and instead isolates the most influential features governing landslide initiation, thereby enabling the development of more accurate, transparent, and operationally meaningful predictive models.

Landslide generation is a complex, multi-variable process governed by the interplay of geological, climatic, and anthropogenic factors, with substantial spatial variability in the dominant controls (Chen et al., 2024; Fidan et al., 2024). Topography constitutes a primary conditioning factor; however, its influence can be subordinated by local subsurface dynamics, particularly rapid variations in soil moisture at specific depths, which are especially relevant in rainfall-triggered failures (Margaño-Carmona et al., 2025; Fustos-Toribio et al., 2025). Additional contributors such as lithology, drainage density, and structural discontinuities interact with triggering mechanisms – most notably intense or prolonged precipitation – to determine slope stability. Consequently, landslide susceptibility assessment becomes inherently challenging, as it must incorporate a large number of interdependent variables whose combined effects are difficult to disentangle (Margaño-Carmona et al., 2023; Zighmi et al., 2025). The central methodological difficulty lies in processing this heterogeneous information and isolating the variables that most strongly control landslide initiation.

Climate change further exacerbates this complexity. Increasing temperatures and shifts in precipitation regimes are already amplifying the frequency and intensity of hydrological extremes, elevating landslide hazard in susceptible regions (Crozier, 2010). Such dynamics are particularly evident in mountain belts like the Himalayas, where accelerated glacial melt and heightened rainfall intensity have led to pronounced increases in slope failures (Ballabh et al., 2014). At the same time, human-induced landscape alterations, including deforestation, urbanisation, and land-use change, destabilise natural soil structures and modify drainage networks, increasing the likelihood of failure under otherwise moderate hydrological forcing (Xu et al., 2024; Wang et al., 2023).

The convergence of these natural and anthropogenic drivers underscores the need for region-specific, data-driven approaches capable of distinguishing the most influential conditioning and triggering factors. Stand out the topographic influence in landslide susceptibility (Lin et al., 2017), meanwhile soil moisture emerges as a critical factor on a global scale, particularly in regions prone to rainfall-triggered landslides (Maragaño-Carmona et al., 2025). Additionally, geological characteristics such as lithology and drainage density are crucial conditioning factors that interact with triggering mechanisms like extreme precipitation events (Bisht and Rawat, 2023). Such precisions are essential for developing accurate, operationally relevant predictive models and for designing effective landslide early warning and risk-management strategies. To address this multi-variable complexity and enhance predictive accuracy, advanced methodologies must be combined with classification techniques, to refine model inputs and achieve more precise, actionable spatial predictions (Kumar et al., 2023; Ge et al., 2022).

Landslides in the Southern Andes are predominantly influenced by a combination of geological, climatic, and tectonic factors that interact to create a complex landscape prone to instability (Fustos-Toribio et al., 2021; Ochoa-Cornejo et al., 2025; Maragano-Carmona et al., 2025). The region's geological composition, characterised by steep slopes and diverse lithologies, plays a significant role in landslide susceptibility. Studies indicate that the presence of weak lithology along the margin and the geomorphological features could predispose areas to landslides, particularly during extreme precipitation events (Maragano-Carmona et al., 2025). Additionally, climatic factors, particularly rainfall, are pivotal in triggering landslides in the Southern Andes. Research has shown that intense and prolonged rainfall events significantly increase the likelihood of landslides, especially in areas lacking adequate drainage systems (Fustos-Toribio et al., 2022; Islam et al., 2021). The development of rainfall-induced landslide early warning systems (RILEWS) has been proposed as a means to mitigate the risks associated with these events. However, such systems are not yet operational in the Southern Andes (Fustos-Toribio et al., 2022). Furthermore, the interaction between vegetation and slope stability is also noteworthy; while vegetation can stabilise slopes, its removal or degradation can lead to increased landslide occurrences (Vorpahl et al., 2012).

To understand the impact of geomorphological, hydrological and geotechnical features in landslide susceptibility delimitation, machine learning (ML) has emerged as a pivotal tool, particularly in the development of regional susceptibility analysis or early warning systems (Merghadi et al., 2020; Fustos-Toribio et al., 2022). The complexity of landslide phenomena, influenced by various geological and climatic conditions, requires tailored approaches to make accurate predictions. Recent studies indicate that ML models, such as Support Vector Machines (SVM) (Huang et al., 2019), Random Forests (RF) (Park and Kim, 2019), and Extreme Gradient Boosting (XGB) (Yang et al., 2024), among others, outperform traditional statistical methods in terms of predictive accuracy and robustness (Li et al., 2019; Bravo-López et al., 2023). Approaches that integrate various ML techniques have been shown to enhance spatial agreement in mapping landslide susceptibility, thus improving the reliability of predictions (Adnan et al., 2020; Bravo-López et al., 2023). Feature selection plays a crucial role in optimising model performance by identifying the most relevant variables that contribute to landslide occurrences, thus reducing noise and enhancing interpretability (Halder et al., 2024; Ge et al., 2022; Pham et al., 2020). This not only improves classification performance but also provides keys to understanding the phenomenon and deciding which variables to monitor. Many strategies can be used to perform the selection: filter methods select features based on statistical criteria obtained from their individual behaviour while wrapper methods evaluate the contribution of feature subsets on the classification performance, finally, embedded methods perform feature selection internally during model training through algorithm-specific mechanisms related to the specific ML model (Theng and Bhoyar, 2024; Guyon and Elisseeff, 2003). Another important challenge is the definition of negative examples. Regional data scarcity complicates the development of ML models, and this limitation is exacerbated by the fact that available databases typically contain only positive cases, requiring negative samples to be defined empirically. Consequently, it becomes essential to adopt appropriate strategies to construct balanced datasets. Buffer Control Sampling and PUBagging (Positive-Unlabeled Bagging) has been effectively used to address this issue (Gu et al., 2024; Wu et al., 2021).

This manuscript presents a methodological framework to understand the complex and dynamic processes of landslide generation in the Southern Andes (38–42° S), employing for the first time a rigorous, data-driven methodology in this region. We aim to identify and delineate the key variables that must be considered in the Southern Andes to adequately characterise landslide triggering and controlling conditions. To improve the performance and explainability of the machine learning models, two feature-selection strategies were applied in parallel: a filter-based approach that ranks the features according to their importance, using Classification and Regression Trees (CART), and a wrapper-based approach employing Genetic Algorithms (GA) for feature-subset optimisation. Both strategies were evaluated using different classification techniques, enabling a detailed assessment of the relative importance of climatic variables, soil properties, and geological features. This targeted feature selection process is vital for identifying the true local critical factors that govern slope instability in this climatically active and geologically heterogeneous environment, effectively translating complex hydro-geomorphological processes into quantifiable variables. The prioritised factors are then strategically repurposed as high-quality proxies to refine and improve variable mapping, enabling a crucial regional-to-local scale approach that captures the spatial variability of landslide susceptibility. The provision of an in-depth, data-validated analysis of the critical variables that interplay among triggering mechanisms will support the design of next-generation, ML-based monitoring networks and accurate early warning systems (EWS) for this vulnerable mountain range.

We consider a sparse data area, focused in the Southern Andes, spanning latitudes from 38 to 43° S (Fig. 1). The area presents a unique and complex geological and climatological landscape that significantly influences soil moisture dynamics and mass wasting processes. The diverse geological composition of the study area yields a complex landscape, formed by numerous relief changes related to volcanic and glacial processes during the Holocene and more recently, influenced by different-scale mass wasting processes (Maragaño-Carmona et al., 2025; Fustos-Toribio et al., 2022, 2025). From a geological perspective, the area has a wide range of lithological units, encompassing volcanic rocks (andesites, basalts), metamorphic rocks (Horton, 2018), sedimentary formations, and glacial deposits (Cuzzone et al., 2024). This variability introduces potential soil sources from weathered rocks and sediment sources that will send their material to sedimentary environments to the subduction area along the central valley. However, the amount of the source is still poorly quantified.

The geological framework of the Southern Andes is predominantly shaped by tectonic activity, resulting in a range of rock types, including andesites, basalts, and sedimentary rocks. Stand out the volcanic activity, where this area contributed to volcanic eruptions, ranging from plinian to strombolian, which formed tephra deposits with different mineralogies (Moreno-Yaeger et al., 2024). The tephra degradation during the Holocene triggered high soil variability, characterised by different textures and water retention capacities, which could control the landslide triggering processes. Moreover, the different soil textures generated from lithological and tephra deposits control the water storage and the following organic matter content.

The area was covered by two pulses of glacial periods (Cuzzone et al., 2024), which modified the relief and eroded previous deposits leaving young soils prone to be affected by landslides on the eastern side. Meanwhile, ancient soils are present in the western area, creating a good contrast between soil ages. The presence of glacial deposits from the last glacial maximum has resulted in the formation of heterogeneous soil profiles, which vary in texture and composition throughout the region (Vásquez-Antipán et al., 2025). These features influence the water retention and drainage properties, thereby affecting soil moisture levels and slope susceptibility to mass wasting events. Understanding the main control of the soil features becomes essential for assessing the stability of slopes and predicting potential landslide occurrences. In addition to the wide range of lithological and soil properties and features existing in the area, the climate of the Southern Andes exhibits an extreme range of climates, from arid to humid, influenced by altitude, latitude, and patterns of western winds and atmospheric rivers (Tetzner et al., 2025). The main precipitation amount occurs primarily in the austral winter months (May–September), reaching over 4000 mm yr⁻¹. The summer dry seasons modify soil moisture dynamics at different depths, with an unknown impact on landslides. Recently, climate change has posed additional challenges, as shifts in precipitation patterns and increased frequency of extreme weather events may exacerbate soil erosion and landslide risks.

Nowadays, accurately assessing landslide risk in this region presents significant challenges. The high variability in multiple variables introduces a complex dynamic that acts as a primary control on slope stability, which is currently poorly understood. The interplay of these varying variables creates a highly dynamic environment where the key relationships governing stability can change dramatically over short distances. The study area has been affected by large-scale mass wasting events (Fustos-Toribio et al., 2021; Ochoa-Cornejo et al., 2025; Maragano-Carmona et al., 2025; Vásquez-Antipán et al., 2025), underscoring the need to develop a robust early warning system and enhance assessment accuracy. This requires comprehensive, integrated studies to create models that fully incorporate geological, climatic, and soil data, including the often-overlooked parametric controls, enabling a reliable landslide early warning system. A network of advanced sensors is needed to monitor real-time changes in all the variables across the varying lithological units. However, it needs to focus on the main variables, which have not yet been established.

To assess the interplay between rainfall intensity, soil moisture, and geological factors that underlie landslide occurrences, it is essential to construct a comprehensive and reliable database. We designed and developed databases that synthesise data sets, including meteorological records, modelled soil moisture, and detailed hydraulic soil properties. We used an approach to identify critical conditioning factors that could be systematically captured and quantified, providing a solid empirical foundation for subsequent machine-learning-based risk assessments using soil features extracted from high-resolution Chilean databases via pixel-to-point processing (Dinamarca et al., 2023). The generated database enabled constraining the soil features to the exact emplacement of the landslide and its date, considering the lithological and climatic features within the study area. The selection allowed capture the range of soil and lithologic features under different hydrometeorological and climatic conditions in the Southern Andes. The selected data provided a robust framework for analysing the dynamic relationships between hydrological, geological, and climatic controls in landslide generation.

We considered the soil moisture, precipitation (PP), and slope from the ERA5 database (ERA5, 2023), one of the most widely used climate datasets, with 10 km resolution. The slope was obtained by combining the ERA5 database with high-resolution digital elevation models from the shuttle Radar Topography Mission (SRTM), at 30 m spatial resolution (NASA JPL, 2013). The Chilean soil properties were extracted from the CLSoilMaps database (Dinamarca et al., 2023). The ERA5-Land dataset corresponds to the land component of the European Centre for Medium-Range Weather Forecasts' (ECMWF) fifth-generation reanalysis, executed under the Copernicus Climate Change Service (C3S) mandate, being generated via offline, high-resolution integrations of the ECMWF land surface model, CHTESSEL. The offline generation approach offers improved temporal consistency in land surface fields, thereby representing global water and energy cycles more accurately (Muñoz-Sabater et al., 2021). The dataset delivers approximately 50 variables essential for describing hydrological and surface energy processes, including soil moisture, runoff, snow cover, skin temperature, evapotranspiration, and land carbon fluxes at high resolution, providing a 9 km global horizontal resolution in comparison to 31 and 80 km resolutions of ERA5 and ERA-Interim, respectively. The currently accessible public record spans from January 1981 to the present, subject to a typical latency of 2–3 months. To mitigate boundary discontinuities, each stream is preceded by multi-year spin-up periods. However, residual inconsistencies are acknowledged, particularly within variables with long memory, such as deep soil moisture and permanent snow regions. The internal consistency of ERA5-Land is intrinsically tied to its reliance on ERA5 for meteorological forcing, through a 4D-Var data assimilation system. Near-surface air temperature, pressure, and humidity are, however, corrected for elevation discrepancies between the ERA5 and ERA5-Land grids utilising daily lapse rates. Validation against diverse observational datasets, including in situ networks, satellite products, and hydrological records, unequivocally demonstrates that ERA5-Land surpasses ERA5 performance in the representation of soil moisture, temperature, runoff, and lake dynamics (Muñoz-Sabater et al., 2021; Bonshoms et al., 2022; Yilmaz, 2023). Moreover, we utilised the new CLSoilMaps database, an improvement over global soil databases in Chile. CLSoilMaps provides spatially explicit and high-resolution predictions of critical soil physical and hydraulic properties specifically for continental Chile and its shared transboundary basins with Argentina. The database was rigorously developed using Digital Soil Mapping (DSM) techniques within the SCORPAN framework (McBratney et al., 2003), trained on over 4000 soil profile observations that encompass diverse and historically underrepresented ecosystems, including those in the Andes and Patagonia. Modelled soil attributes include clay, sand, and bulk density, with silt content derived indirectly from these parameters. Hydraulic properties, such as field capacity, permanent wilting point, and available water capacity, were estimated using the well-established Rosetta V3 pedotransfer function (Zhang and Schaap, 2017). The predictions were executed using Random Forest algorithms that incorporated over 200 environmental covariates, spanning climate, topography, and satellite-based reflectance indices, and adhered to the six standardised depth intervals (0–200 cm) stipulated by the GlobalSoilMap project.

We used a rainfall-induced landslide database developed using the Xterrae database (https://www.xterrae.cl/, last access: 1 June 2024) maintained by the National Geology and Mining Service (SERNAGEOMIN) of Chile. Positive landslide events were obtained from this source, while negative cases were generated using a Buffer Control Sampling strategy (Gu et al., 2024). In this approach, a buffer zone of 100 m was established around each landslide event, and sampling points were extended outward up to a 20 km radius. This distance was intentionally chosen given the 10 km resolution of the climatological variables, enabling the capture of spatial variability of the precipitation and soil moisture content, differentiating between landslide and non-landslide conditions. Subsequent refinement of the negative examples was performed via a modified PUBagging (Gu et al., 2024), built on the idea of bagging (bootstrap aggregating) by repeatedly sampling subsets of the unlabeled data and combining them with the positive set to train multiple base classifiers. In our setting, many candidate negative samples can exhibit characteristics that are very similar to the positives, and including them could make the decision boundary more difficult to find and the classification unnecessarily more complex. PUBagging provides a way to mitigate this risk by leveraging the positive set and treating each subset as if the unlabeled instances were negative (which may introduce noise), and by aggregating the predictions across many such classifiers, PUBagging reduces the bias introduced by this assumption and improves robustness. The final decision is typically made through majority voting or averaging across these classifiers. Here a Random Forest model – trained exclusively on positive examples – was used to evaluate the negative cases. Negative instances yielding a prediction score above 0.5 were removed, ensuring a more robust set of non-landslide cases.

After excluding records with missing values from the initial 3388 entries, the final database comprises 3148 examples, including 1618 positive and 1530 negative instances. Each record contains 136 normalised feature columns and an additional column indicating binary landslide occurrence. Although explicit vegetation indices, such as NDVI, were not included in the dataset, vegetation effects are implicitly incorporated through predicted soil properties influenced by NDVI measurements, reflecting parameters such as organic matter content and soil moisture. Notably, approximately 95 % of the data is concentrated in the Los Lagos and Los Ríos regions, resulting in a regionally focused dataset for subsequent analysis.

Here, we present the methodology used to examine the relationships between rainfall intensity, soil moisture, and soil characteristics in rainfall-induced landslide generation, to support the development of robust monitoring networks and reliable early warning systems for landslides, as well as to identify the most significant variables. First, highly correlated variables were removed from the initial database. Then, both filter and a wrapper parallel feature selection strategies were applied. Finally, optimised data sets were used to design the landslide classifiers. All this process is shown in Fig. 2.

The CART-based filter method ranks variables based on their individual discriminatory power between positive and negative cases. In contrast, the GA-based wrapper approach optimises a subset of features by evaluating their combined contribution to classification performance. In this case, the genetic algorithm evolves to find the best feature combination that minimises an objective function (Katoch et al., 2020). This objective function was designed to reduce the number of selected variables and the classifier error, evaluated using three models: Support Vector Machines, Random Forest, and XGB. The filter and the wrapper strategies were independently evaluated by assessing the quality of the optimised datasets through the performance of the classifiers, ensuring a consistent estimation of the relevance of each contributing variable.

To understand the relationship between rainfall intensity, soil moisture, and soil hydraulic features in landslide generation, we present an in-depth analysis of the conditioning factors that influence landslide occurrence in a region of the southern Andes (38–42° S). After exploring various machine learning models with different feature configurations, we observed notable differences in performance depending on the features subsets. This process helped identify which combinations of features and modeling approaches best captured the underlying structure of the data. The results reveal that the dynamic interactions among rainfall intensity, soil moisture patterns, and geological characteristics are significant at the determined exploration depth of the soil, based on the database used during feature selection. The following section presents the results, highlighting the most promising models and the impact of feature selection on the performance of susceptibility estimation.

Accurate landslide forecasting enables early warnings, allowing authorities and communities to take preventive measures, such as evacuations or slope stabilisation, to reduce risk. It also supports better land-use planning and disaster preparedness, contributing to long-term resilience against geological hazards. Ultimately, landslide forecasting is a critical tool for safeguarding lives and promoting sustainable development in high-risk areas. For the first phase of feature selection, expert analysis was considered, along with setting a correlation elimination threshold of |0.9|. Subsequently, two strategies were employed to finalise the feature selection: CART and a Genetic Algorithm.

We evaluated machine learning models (SVM, RF and XGB) to predict landslide susceptibility in the southern Andes (38–42° S), selecting domain-relevant geotechnical, hydrological, and geomorphological variables through feature selection techniques (CART and GA). The GA–optimised models, particularly GA-RF and GA-XGB, significantly outperformed baseline methods, achieving the lowest classification errors (10.95 %) with compact feature sets that improved both accuracy and efficiency. Across experiments, slope, precipitation, and near-surface soil hydraulic properties – especially bulk density and saturated water content – emerged as the most critical factors influencing rainfall-induced landslides, underscoring their importance for future monitoring and assessment.

Our findings also underscore the importance of incorporating both shallow and deep soil moisture characteristics, as well as soil retention curve parameters, to better capture the complex subsurface dynamics that precede slope failure. The differences in feature prioritisation between GA-RF, GA-XGB, and GA-SVM reflect distinct modelling philosophies: while RF and XGB emphasised shallow hydraulic traits and retention thresholds, SVM gave greater weight to deeper soil moisture indicators and retention curve shape parameters.

From an operational early warning perspective, the compact GA-selected feature sets reduce monitoring complexity by prioritising a small number of high-impact variables (precipitation, slope, and depth-specific soil hydraulic/moisture indicators). This supports more cost-effective station design and variable selection, while maintaining low error rates. The identified shallow-versus-deep moisture controls also provide guidance for sensor placement (near-surface and deeper soil moisture) in mountainous catchments prone to rainfall-triggered failures.

We conclude that integrating data-driven models with physically meaningful features provides a robust framework for enhancing early warning systems and regional risk assessments. The superior performance of GA-optimised ensemble models suggests that future efforts should prioritise hybrid strategies that combine expert knowledge with automated feature selection. These approaches are particularly valuable in data-scarce environments, offering scalable solutions to inform risk management and decision-making in mountainous regions vulnerable to rainfall-triggered landslides.

Despite the limitations related to the spatial reproducibility of the results, the main contribution of this work lies in proposing a method that can be scaled according to data availability and applied to region-specific contexts, enabling the identification of meaningful variable sets for areas with similar geological or climatic conditions.