Machine Learning for Rainfall-Induced Slope Stability and Landslide Susceptibility: A Systematic Review

Slope failure is one of the most common and severe geohazards worldwide and is responsible for loss of life, infrastructure destruction, ecosystem degradation, and significant economic losses caused by rainfall. While the rain persists or is heavy enough for a long time, the rainwater can penetrate into soil masses, leading to an increase in soil moisture and the pore-water pressure of the soil, a decrease in matric suction and effective stress, and a corresponding decrease in the shear strength of soil slopes [1]. Over time, these hydrological and geotechnical conditions can cause slopes to become increasingly unstable, eventually leading to landslides or other slope failures. With the warming climate and increasingly harsh, erratic precipitation patterns, coupled with high rates of urbanisation, deforestation, land-use change, and infrastructure construction, rainfall-induced slope failures are of increasing concern in many areas [2]. Therefore, understanding the mechanisms that control rainfall-induced slope instability is an important aspect of Disaster Risk Reduction, Geotechnical Hazard Assessment, and Resilient Infrastructure Planning. The interaction among the various factors involved in rainfall-induced slope failure, namely rainfall characteristics, infiltration, pore-water pressure, soil properties, terrain, vegetation, and anthropogenic factors, further complicates the prediction of rainfall-induced slope failure [3].

For infrastructure projects situated on natural or engineered slopes, slope stability assessment is crucial for planning, design, construction, operation, and slope management [4]. Both deterministic and probabilistic methods have traditionally been used to evaluate slope stability, with soil strength, unit weight, slope geometry, groundwater conditions, and rainfall characteristics as parameters for calculating the factor of safety (FoS) or the probability of failure. Traditional methods of analysis and numerical modelling, such as the Limit Equilibrium Method (LEM), Finite Element Method (FEM), and Finite Difference Method (FDM), have made significant contributions to understanding and predicting the behaviour of slopes [5].

However, the spatial variability and uncertainty in soil and rock properties, rainfall infiltration, groundwater conditions, boundary conditions, and failure mechanisms can affect the reliability of these approaches [6]. Physical models of landslides occurring as a consequence of rainfall can be useful to gain insights into hydro-mechanical processes controlling landslide instability, but they are typically difficult to apply, especially at a regional scale, due to data requirements, such as detailed geotechnical, hydrological, geological, and topographical data, which may not be available on a consistent basis [7]. They may also be constrained in their use of large and complex slope systems due to computational needs, parameter calibration, model uncertainty, and difficulties in representing heterogeneous field conditions [8].

This has led to growing interest in data-driven methods, including machine learning (ML), for analysing and forecasting rainfall-induced slope instability. ML algorithms can learn from past and observed data to determine complex non-linear relationships among governing variables, a capability not possible with traditional methods, which typically rely on explicit assumptions about those relationships. This is an important capability in the context of rainfall-induced slope problems, as slope response is governed by interactions among a variety of factors: rainfall intensity and duration, antecedent rainfall, soil and rock properties, terrain morphology, groundwater conditions, land cover, and anthropogenic activity. In the past two decades, the applications of algorithms like Artificial Neural Network (ANN), Decision Trees (DT), Random Forest (RF), Support Vector Machines (SVM), Gradient Boosting approaches, and Extreme Gradient Boosting (XGBoost) have become more and more common in slope stability and landslide prediction problems, respectively [9]. With the growing availability of increasingly large datasets derived from geotechnical investigations, remote sensing, geographic information systems (GIS), digital elevation models (DEMs), rainfall observations, and slope-monitoring systems, the potential for ML to support geotechnical hazard assessments has been further strengthened.

An important difference, however, should be noted: rainfall-induced landslide research applications are not a single prediction problem but a number of them. Based on the different goals of prediction, spatial and temporal scales, and engineering goals, these can be categorised into four interconnected yet distinct application domains. The first domain is the prediction of slope-scale stability and factors of safety, where ML models are used to predict FoS, assess slope stability or instability, or approximate the results of traditional stability analysis. This application primarily focuses on the engineering evaluation of a specific slope or a defined slope system. The second domain is regional landslide susceptibility mapping, in which ML models integrate rainfall, topographic, geological, environmental, and anthropogenic data, as well as GIS and remote sensing data, to assess the spatial probability of landslide occurrence [10]. Susceptibility mapping focuses on the spatial distribution of landslide susceptibility across a landscape, rather than the mechanical stability of a particular slope, and is therefore not the same as FoS prediction.

The third domain is the prediction of slope displacement and deformation, which addresses the temporal behaviour of slopes using monitoring data, including displacement, deformation, rainfall, pore-water pressure, and other time-dependent measurements. The use of ML techniques can help detect temporal patterns and predict future movement on slopes, thereby aiding in the monitoring and assessment of slope deformation. The fourth domain concerns rainfall-induced landslide forecasting and warning, in which ML models are built to detect temporal conditions that increase landslide risk and to provide information to support timely decision-making for warning and risk management [11]. These four domains are very similar, but are not synonymous. The mechanical stability at the slope scale is predicted by FoS prediction, spatial patterns of hazard by susceptibility mapping, temporal deformation behaviour by displacement prediction, and early-warning models for forecasting impending or near-term hazard conditions. Thus, they have various input variables, output variables, datasets, spatial and temporal resolutions, validation strategies, and applications.

This difference is crucial when assessing the reported performance of ML models. A laboratory-based model used to predict FoS cannot be compared with a susceptibility model based on remotely sensed and GIS variables at the regional scale, or with the model predicting slope displacement over time, or with the early-warning model based on rainfall. Likewise, a high predictive accuracy in one application domain doesn't necessarily mean that the model can be transferred to another spatial or temporal scale. The size and quality of a dataset, class imbalance, representation of rainfall, feature selection, spatial and temporal dependence, model calibration, uncertainty, and validation strategy can significantly impact performance reported [12]. Therefore, a systematic evaluation of ML applications should take into account, in addition to the best-performing algorithm, the type of prediction, the scale of the prediction, the data type, rainfall, the geotechnical conditions, and the purpose of the engineering or disaster-management application.

Previous reviews have successfully analysed the current state of the art in ML for slope stability, landslide susceptibility prediction, landslide prediction, and early-warning applications. Previous review studies have focused on the development of ML approaches to slope stability and landslide-of-safety (LOS) prediction, and other reviews have focused on ML- and GIS-based landslide susceptibility mapping. Other reviews have also examined the use of ML in landslide monitoring, prediction, prevention, and early-warning systems. The contributions illustrate the increasing role of a data-driven approach in managing geotechnical hazards. However, the current literature is rather disjointed with respect to the application domain and to the analysis of slope stability, susceptibility, displacement prediction, and early warning, whether treated separately or within a broader landslide-hazard context [13]. Therefore, a systematic synthesis is still needed that specifically addresses the application of ML to the prediction of rainfall-induced slope instability, covering a range of geospatial prediction domains, and taking into account differences in data sources, prediction objectives, spatial/temporal scales, model architectures, model validation procedures, and integration with physically based approaches.

Moreover, the emergence of hybrid and intelligent modelling solutions has introduced new opportunities and methodological challenges. Hybrid approaches in which physically based geotechnical or hydrological models are coupled with ML algorithms have been shown to effectively enhance prediction reliability by integrating both physical knowledge and pattern-recognition capabilities [14]. Likewise, the incorporation of ML into remote sensing, GIS, Internet of Things (IoT)-based monitoring, Digital Twin, and real-time rainfall observation is developing ever more complex solutions for monitoring and forecasting of hazards on slopes. There are, however, questions to be addressed regarding the interpretability of the models, uncertainty quantification, data needs, generalisation across geological and climatic environments, transferability between sites, and the potential to use data-driven models in tandem with, but not in place of, traditional geotechnical models. For practical engineering applications such as these, it is critical that model predictions are reliable, interpretable, and robust enough for use in safety-critical decision-making.

Considering the above, this study is a systematic literature review based on the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, which will comprehensively cover the development and application of ML for rainfall-induced slope instability. The literature is reviewed in a systematic fashion and organised into four major application areas: (i) factor-of-safety and slope-stability prediction, (ii) regional landslide susceptibility mapping, (iii) slope displacement and deformation prediction, and (iv) rainfall-triggered landslide forecasting and early-warning systems. In these aspects, the review covers the number of publications, geographical spread, predictors of precipitation (rainfall and geotechnical), precipitation data sources, ML algorithms, model architectures, evaluation metrics, and reported predictive performance. The differences between application domains are highlighted, and in particular, the impact of data characteristics, spatial and temporal scales, and data-validation dimensions on the model performance is discussed [15].

The review also examines new hybrid physics–ML methods and methodological challenges related to data availability and uncertainty, interpretability, transferability, and real-world implementation. Instead of expecting a single ML algorithm to be best for all rainfall-induced slope instability problems, this study aims to develop a structured understanding of when each ML algorithm is most suitable for each such problem [16]. The results will provide researchers and geotechnical practitioners with an evidence-based toolkit to understand the state of the art in ML applications, recognise persistent research gaps, and inform the design of more reliable, interpretable, and physically sound data-driven tools for rainfall-induced slope stability assessment and landslide risk management. The aim of the study is to systematically review literature on ML for rainfall-induced slope stability and landslide susceptibility. In order to achieve the aim, the study has the following objectives, which are to:

1. Carry out a systematic review of the application of ML in rainfall-induced slope stability analysis.

2. Find out the most used ML algorithms and assess their performance for the prediction of slope stability and landslide susceptibility.

3. Identify the research gaps for Rainfall-Induced Slope Stability and Landslide Susceptibility.

1.1 Statement of the problem

Many communities around the world remain at risk from rain-induced landslides, which can threaten life, property, and the environment. The number of extreme precipitation events has also increased due to climate change, intensifying the frequency of slope failures and further increasing the importance of prediction and hazard assessment. Although major progress has been made in geotechnical engineering, predicting rainfall-induced slope instability is a complicated problem [17]. This is because many factors that affect slope failure are interdependent, including rainfall intensity and duration, soil properties, geologic factors, groundwater dynamics, vegetation, and anthropogenic activities.

The LEM, FEM, and FDM are widely used in engineering practice for assessing slope stability. These approaches have been shown to be useful tools for understanding slope behaviour, but they often rely on simplifying assumptions and require a high level of site-specific information that is not always available [18]. Additionally, the complex and stochastic behaviour of rainfall-induced slope failures may make traditional analytical and numerical failure models less predictive, especially in large-scale or data-rich settings.

However, in the past few years, ML techniques have proven valuable tools and potential alternatives to traditional approaches for predicting slope stability, owing to their ability to learn complex patterns from vast amounts of data and model nonlinearities without the need for physical models. Therefore, many studies have used ML techniques such as ANNs, SVMs, RF, DT, and XGBoost to predict rainfall-induced landslides [19]. The current literature, however, is quite heterogeneous across datasets, input parameters used, model architectures, validation approaches, and performance evaluation metrics.

This fragmentation has made it difficult to determine the most successful machine-learning techniques, the factors that affect model performance, and the best practices for rainfall-induced slope stability analysis. In addition, thorough and updated systematic reviews that incorporate trends in the literature, methodological innovations, model performance evaluations, and promising directions for research within a single study have not yet been published, and research in this area continues to grow. It is therefore essential to conduct a systematic literature review to synthesise existing knowledge, identify gaps, and suggest directions for future studies on slope stability assessment using ML approaches.

To address the research gaps above and achieve the objectives of this study, the following research questions have been formulated:

RQ1: What ML techniques are mostly used for rainfall-induced slope stability analysis?

RQ2: What data sources, input parameters, and performance evaluation measures are commonly used in ML-based slope stability studies?

RQ3: How effective is ML when compared with conventional geotechnical methods for rainfall-induced slope stability prediction?

RQ4: What are the main research gaps, challenges, and future benefits in the application of ML to rainfall-induced slope stability analysis?

Therefore, the current study developed a conceptual framework diagram (Figure 1) showing how the technologies mentioned (ML-GIS, deep learning, Explainable Artificial Intelligence (XAI)) could be used for rainfall-induced slope stability prediction. Also, the conceptual framework illustrates the pattern of the current study.

1.png

Figure 1. Intelligent hybrid framework for rainfall-induced slope stability prediction

The use of ML in rainfall-induced slope stability analysis has evolved from a relatively unknown research area into an expanding interdisciplinary field that integrates geotechnical engineering, hydrology, remote sensing, and artificial intelligence. To study ML techniques and their performance, it is essential to first understand the fundamental processes governing rainfall-induced slope failure. These processes involve infiltration of rainwater, development of pore-water pressure, reduction in the shear strength of a slope, and complex interactions among geological, hydrological, and environmental processes that affect slope behaviour. These concepts form the physical underpinning for the development of modern data-driven prediction models. Thus, this section reviews the basic mechanism of rainfall-induced slope failure, the key factors, and the problems in predicting and monitoring slope failure.

3.1 Rainfall-induced slope stability analysis: Concepts and challenges

In unsaturated soils, where matric suction can significantly contribute to apparent cohesion, precipitation infiltration has been shown to be the most important factor in triggering both shallow- and deep-seated landslides. When rainfall enters the soil profile, suction slowly decreases, causing effective stress and shear strength to decline as well, potentially leading to failure [23].

In addition, data from monitoring systems may be large-scale and of various types, such as rainfall, piezometric, remote sensing, and deformation monitoring data. The lack of integration of these various data sources to build powerful predictive systems is a major challenge. Advanced data-driven methods that address uncertainty and high-dimensional data have therefore been proposed recently, leading to the growing use of ML methods in rainfall-induced slope stability analysis [24]. The main factors affecting rainfall-induced slope failure are summarised in Table 3, along with typical qualitative effects reported in recent studies.

Table 3. Selected summary of literature on some factors influencing rainfall-induced slope failure

3.2 Machine learning algorithms applied in rainfall-induced slope stability analysis

Due to uncertainty in hydraulic parameters, spatial variability in soil properties, and the complexity of computing physically based rainfall-infiltration models, researchers have turned to ML techniques as alternatives for predicting slope stability. Traditional approaches rely on explicit physical formulations, which can be circumvented by ML algorithms that learn nonlinear relationships directly from historical observations and environmental datasets. Recent research shows that ML has emerged as one of the fastest-growing fields in landslide susceptibility assessment and rain-induced slope failure prediction, with significant increases in publications since 2018 [36, 37].

ANNs were among the first ML approaches applied to slope stability problems due to their capacity to model highly nonlinear relationships between input variables and failure outcomes. The extensive use of ANN models for landslide susceptibility mapping and rainfall-induced failure prediction has been suggested to approximate complex hydro-mechanical interactions that are hard to model using conventional analytical methods. The prediction accuracy of recent studies has exceeded 85%, and interpretable neural networks have been shown to achieve performance on par with state-of-the-art deep learning models while offering enhanced model explainability. A scalable, interpretable, and high-performing neural network framework was developed by Soltaninejad et al. [38] that was able to recognise the contribution of slope gradient-precipitation interaction to landslide occurrence and predict it with high accuracy in various Himalayan datasets.

SVMs have also received significant attention in rainfall-induced slope stability analysis due to their good generalisation properties and efficiency with limited training samples. SVM-based models work well for high-dimensional classification problems, where non-linear kernel functions can separate stable and unstable slope conditions. In landslide susceptibility applications, the Area Under the Curve (AUC) value of the SVM model is often reported to exceed 0.85 [39, 40]. Lately, SVM variants based on optimisation techniques have been developed to further enhance the predictive power of SVM. As an example, Ajin et al. [41] combined Support Vector Regression with Particle Swarm Optimisation and Grey Wolf Optimisation algorithms, achieving significant improvements in susceptibility prediction compared to traditional implementations. The results suggest that optimisation-assisted SVM frameworks are still relevant despite the emergence of more advanced ensemble and deep learning models.

RF combines predictions from multiple DTs (created via bootstrap sampling and random feature selection) to reduce variance and improve classification. AUC values ranging from 0.85 to 0.95 have been reported in many studies on landslide susceptibility modelling using the RF method [42]. In contrast, in rainfall-induced landslide early-warning systems, RF models consistently outperform many traditional statistical methods, as they can handle heterogeneous geospatial data. More recently, comparisons have shown that RF often outperforms individual DT, thereby providing the advantages of ensemble learning in geotechnical prediction problems [43].

3.3 Data sources, feature engineering and model performance

ML algorithms are increasingly relevant to rainfall-induced slope stability analysis, and it is evident that dataset quality and the effectiveness of feature engineering strategies are crucial [44]. The predictive power of advanced ML models like RF, XGBoost, and deep learning models depends heavily on the quality and variety of input variables used during training. Therefore, the past few years have seen significant focus on finding appropriate data sources, fine-tuning feature selection methods, and developing accurate performance assessment methods [45]. Effective ML slope stability models require not only the selection of an appropriate algorithm but also representative rainfall, geotechnical, topographic, geological, and environmental data that capture the complex mechanisms governing slope failure.

Rainfall datasets are one of the most vital sources of information in rainfall-induced landslide prediction studies. Meteorological stations, satellite-based rainfall products, and reanalysis datasets are commonly used by researchers to access historical rainfall data [46]. Parameters of rainfall include cumulative rainfall, precipitation rate, AR, rainfall duration, and maximum hourly precipitation. Huang et al. [47] found that cumulative and antecedent rainfall are significant factors and among the most influential predictors of slope failure. Satellite-derived precipitation data such as TRMM and GPM are now more useful for regional-scale susceptibility mapping, as they can fill gaps in the spatial coverage of observed rainfall in regions with sparse ground-based monitoring networks [48]. A combination of rainfall, terrain, and geological data has greatly enhanced the predictive accuracy of ML models in various recent studies [49].

Besides rainfall datasets, geotechnical parameters are still important predictors for slope stability assessment. The mechanical resistance of slope materials is directly affected by variables such as soil cohesion, friction angle, bulk density, hydraulic conductivity, porosity, groundwater level, and shear strength [50]. As a result, field- and laboratory-derived variables are often supplemented with remote sensing and GIS variables. Merghadi et al. [51] used geotechnical variables and remotely sensed terrain attributes to significantly improve landslide susceptibility prediction accuracy compared with topographic variables alone. These conclusions were also drawn by Arabameri et al. [52], who demonstrated the utility of multi-source GIS datasets in improving machine-learning-based hazard mapping.

With the increasing size and complexity of ML datasets, feature engineering has become increasingly significant for enhancing model efficiency and interpretability [53]. Feature selection techniques are widely used to remove redundant features, reduce dataset dimensionality, and improve prediction accuracy. Common methods include those like Information Gain, Gain Ratio, Relief-F, Principal Component Analysis (PCA), and Shapley Additive Explanations (SHAP)-based importance ranking. It has been demonstrated recently, however, that feature selection can reduce computational costs without compromising, and even enhancing, predictive performance [54]. Common metrics to assess ML models' performance in rainfall-induced slope stability analysis include Qualitative metrics like Accuracy, AUC, Root Mean Square Error (RMSE), Precision, Recall, and F1-score [55]. Among these metrics, AUC remains one of the most widely reported because it provides a threshold-independent measure of classification performance [56]. Most experiments have shown that ensemble methods and powerful boosting algorithms are generally superior to standard classifiers. RF, XGBoost, LightGBM, and deep learning models often achieve AUC values greater than 0.90 and classification accuracy greater than 90% [57]. Table 4 summarises representative studies for rainfall-induced landslide susceptibility and slope stability prediction using ML. The AUC varies across studies.

Table 4. Descriptive analysis of the performance of machine learning (ML) models applied in rainfall-induced slope stability

3.4 Hybrid and intelligent prediction study

The introduction of multi-source datasets and the development of computational intelligence have led to rapid progress in hybrid and intelligent prediction frameworks for assessing rainfall-induced slope stability [68]. One of the most popular hybrid approaches is the use of ML algorithms combined with GIS. GIS is a powerful tool for managing, visualising, and analysing distributed landslide conditioning factors, including slope gradient, elevation, lithology, land use, rainfall distribution, and drainage spatially. GIS-integrated ensemble models have been reported to have AUC > 0.93 in regional landslide susceptibility assessment [69]. Likewise, Nurwatik et al. [70] demonstrated that ML algorithms, combined with environmental variables derived from GIS, yielded predictive performance more than 10% better than conventional statistical methods. One of the main reasons for the broad use of ML-GIS frameworks is that they can process large-scale spatial data while still retaining high predictive accuracy and visualisation capabilities. FEM-generated datasets can be used to train machine-learning surrogate models that predict slope-stability conditions with an acceptable prediction error of less than 5% within a reasonable computational time. In a similar vein, it has been reported that RF models built with outputs from classical limit equilibrium analysis perform well in reproducing factor-of-safety (FOS) predictions and require only a small fraction of the computational effort [71].

Changes in Earth's surface can now be monitored over large geographical areas using satellite, LiDAR, synthetic aperture radar (SAR), and unmanned aerial vehicle (UAV) observations. Similarly, Li et al. [72] concluded that combining deep learning with Sentinel-1 SAR data could help detect early signs of slope deformation associated with impending rainfall-induced failures. These developments indicate the growing importance of using remote sensing and deep learning in hazard assessment and monitoring for large-scale applications [73-76]. Rainfall-induced landslide prediction has also been revolutionised by recent advances in deep learning and remote sensing [77-80].

As a consequence, various approaches to XAI have been developed, including feature attribution techniques such as SHAP and Local Interpretable Model-Agnostic Explanations (LIME) [81, 82]. Zhou et al. [83] demonstrated, using SHAP analysis, that rainfall intensity, slope angle, lithology, and proximity to drainage are the dominant factors in susceptibility prediction models. The methods are especially relevant in geotechnical engineering, where understanding causal relationships may be more important than predictive accuracy.

The purpose of this systematic literature review was to explore the application of ML techniques to the study of rain-induced slope stability, with a focus on landslide susceptibility assessment, predictive modelling, data integration, and intelligent monitoring systems. The PRISMA framework was followed in a structured review process, and 120 relevant studies were analysed to draw conclusions on general research trends, the most frequently used methodology, the main data sources, metrics used to monitor and assess technology performance, and technological developments arising from the research. The results showed that rainfall-induced slope instability is a major geotechnical and environmental issue worldwide, and that more efficient and reliable predictive models should be developed. As the amount of geospatial data, computing power, and remote sensing tools continues to increase, ML approaches offer a potential way to assess hazards and mitigate risk.

The study revealed that the ML techniques most widely applied in rainfall-induced slope stability studies are Artificial Neural Networks, SVMs, RF, and XGBoost. Among these, ensemble learning methods such as RF and XGBoost performed well at predicting experimental outcomes, as they capture nonlinear relationships and are resistant to overfitting. In addition, developments in deep learning have enabled the creation of more complex predictive models that can detect intricate spatial and temporal patterns in massive geospatial data. Prediction accuracy and operational use have been further enhanced by recent developments in GIS, remote sensing, and numerical modelling.

The descriptive and thematic analysis revealed that the research field is growing rapidly, with steadily increasing publications, international activities, and ongoing methodological development. New research areas are: explainable AI, hybrid ML–numerical simulation techniques, IoT-based monitoring techniques, digital twin techniques and physics-informed ML techniques. However, there are still a number of challenges to address, including data limitations, limited transferability of geographical models, limited interpretability of complex models, and the lack of complete real-time prediction systems. Solving these issues will be fundamental in making research advances a reality in engineering solutions.