Landslide Hazard Assessment Under Record-Breaking Extreme Rainfall: Integration of SBAS-InSAR and Machine Learning Models

Abstract

Global climate change has led to a marked increase in the frequency of record-breaking extreme rainfall events, which often surpass historical benchmarks and pose significant challenges to conventional geological hazard risk assessment methods. This study used a record-breaking extreme rainfall event in Zhenba County, Shaanxi Province, in July 2023 as a case study to develop a tailored risk assessment framework for geological hazards under extreme rainfall conditions. By integrating high-resolution Planet satellite imagery, millimeter-scale surface deformation data derived from SBAS-InSAR, and detailed field investigation results, a comprehensive disaster inventory containing 1012 landslides was compiled. The proposed framework integrates cumulative extreme rainfall metrics with subtle ground deformation indicators and applies four advanced machine learning algorithms—DNN, XGBoost, RF, and LightGBM—for multidimensional hazard assessment. Among these, the DNN model exhibited the highest performance, achieving an AUC of 0.82 and Kappa coefficients of 0.833 (training) and 0.812 (prediction). Further analysis using SHAP values identified distance to rivers, cumulative rainfall, and the Topographic Wetness Index (TWI) as the most influential factors governing landslide occurrence under extreme rainfall conditions. Validation using representative case studies confirmed that the framework effectively identifies high-hazard zones, particularly in areas severely impacted by debris flows and landslide deformation zones. These findings provide a robust scientific foundation and technical basis for early warning, disaster prevention, and mitigation strategies in geologically complex regions increasingly affected by extreme rainfall events.

1. Introduction

Regions with complex geological structures, such as fold-and-thrust belts, are particularly susceptible to geological hazards owing to their distinctive tectonic configurations [1,2]. The development of folds and other structural complexities often results in extensively fractured rock masses, substantially increasing the likelihood of landslide occurrence [3,4]. Superimposed upon these inherently fragile tectonic settings, the escalating impacts of global climate change have further intensified geological disaster risks [5]. With ongoing global warming, both the frequency and intensity of extreme rainfall events have shown a marked increase [6,7,8]. Pronounced changes in the hydrological cycle have led to increasingly irregular spatiotemporal precipitation patterns, with record-breaking extreme rainfall events occurring more frequently [9,10]. These alterations have significantly increased the complexity of hazard assessment and risk management [11,12].

As a vital tool in contemporary Earth observation, remote sensing technologies offer substantial advantages for monitoring geological hazards. Compared to conventional ground-based monitoring approaches, which are often constrained by limited spatial coverage, poor temporal continuity, and high costs, remote sensing facilitates the acquisition of large-scale, multi-temporal, high-resolution surface data [13,14]. Among these, active microwave remote sensing is particularly effective owing to its all-weather, day-and-night acquisition capability, unaffected by cloud cover or adverse weather, thereby ensuring reliable, continuous monitoring of dynamic geological hazards [15]. Interferometric Synthetic Aperture Radar (InSAR), capable of detecting ground deformation with millimeter-scale precision, has emerged as a highly effective remote sensing technique. By analyzing phase differences between temporally separated SAR images, InSAR enables the precise detection of subtle ground displacements [16,17]. The Small Baseline Subset InSAR (SBAS-InSAR) technique, developed to refine conventional InSAR, utilizes optimized spatiotemporal baseline configurations and multi-master image processing to effectively mitigate key limitations such as temporal decorrelation and atmospheric delays [18]. This approach markedly enhances the accuracy and reliability of surface deformation monitoring [19,20]. This technique can identify actively deforming hazard zones and detect subtle precursory deformations in areas that have not yet developed into full-scale disasters [21,22]. Such zones often constitute latent high-risk areas and serve as critical early-warning indicators in disaster monitoring frameworks [23]. In the context of intensifying climate change, there is an urgent need for comprehensive, systematic investigations of hazard characteristics associated with record-breaking rainfall events, especially in tectonically complex regions. The integrated application of advanced monitoring technologies, such as SBAS-InSAR, in hazard assessment and risk prediction constitutes a significant advancement in disaster prevention and risk management. This approach is crucial for enhancing regional disaster resilience, improving early-warning systems, and mitigating the adverse impacts of climate change-induced geological hazards.

Mounting evidence suggests that extreme rainfall events are a principal triggering factor for regional geological hazards, making their accurate assessment vital for disaster prevention and mitigation [24,25,26,27]. Huang et al. [28] assessed landslide susceptibility by integrating frequency ratio analysis with logistic regression, support vector machines, and Random Forest models. Their approach incorporated critical rainfall thresholds and susceptibility mapping. Similarly, Achu et al. [29] applied a Deep Neural Network (DNN) to evaluate landslide susceptibility under extreme rainfall conditions in Kerala, India. Their findings indicated that the DNN model outperforms traditional methods in capturing the nonlinear relationship between rainfall intensity and landslide occurrence, thereby improving spatial prediction accuracy. However, several limitations remain in current hazard assessment methodologies for extreme rainfall. First, studies focusing explicitly on record-breaking extreme rainfall events are scarce, and current assessment frameworks often inadequately address the rising frequency of historically unprecedented rainfall under climate change. Second, the integration between ground deformation monitoring technologies and rainfall data remains limited. Specifically, the capacity of SBAS-InSAR to detect subtle surface deformations induced by extreme rainfall is underutilized. From a long-term perspective, establishing an extreme rainfall hazard assessment system is crucial for early warning of disasters in geologically complex regions. Such a system would not only help identify the hazardous characteristics of record-breaking rainfall events but also provide a scientific basis for developing targeted disaster prevention and mitigation measures through integration with SBAS-InSAR deformation monitoring results. This integrated approach would effectively reduce geological disaster risks induced by extreme rainfall.

The Qinling-Daba Mountains constitute one of China’s most geologically diverse and tectonically active regions, forming a major orogenic belt in central–western China [30,31]. Due to its intricate tectonic framework and distinctive geomorphological characteristics, the region is highly prone to geological hazards [32,33]. The Southwest China Vortex (SWV) ranks among the most intense rainstorm-generating systems in China, second only to typhoons and their associated remnant low-pressure systems in terms of rainfall intensity, frequency, and spatial coverage [34,35]. SWV movement is typically accompanied by widespread torrential rainfall [36]. Each hydrological year, the Qinling-Daba region is significantly affected by SWV activity, leading to exceptionally high and seasonally concentrated rainfall, particularly during the summer months [37]. As a key rainfall center in the Qinling-Daba region, Zhenba County experienced a record-breaking extreme rainfall event in 2023. This event not only broke historical precipitation records but also marked the highest daily rainfall since the onset of meteorological observations, triggering widespread landslides. The unprecedented magnitude of such extreme rainfall events has severely impacted the regional geological environment and posed significant challenges to conventional rainfall hazard assessment frameworks. Although landslide and geohazard susceptibility assessments under typical rainfall conditions have been extensively studied [38,39], the distinctive hazard characteristics of record-breaking extreme rainfall events remain insufficiently explored. Moreover, existing rainfall hazard assessments seldom integrate SBAS-InSAR-derived ground deformation data with extreme rainfall observations. This limitation restricts comprehensive understanding of hazard dynamics under such conditions. Most current models inadequately capture the complex interactions between rainfall intensity and surface deformation, resulting in deficiencies in identifying high-risk zones. This shortcoming substantially undermines the applicability and accuracy of current methods in structurally complex regions such as the Qinling-Daba Mountains. To enhance hazard prediction and early warning, it is essential to develop novel assessment frameworks that integrate extreme rainfall dynamics with high-resolution surface deformation monitoring. Such frameworks would establish a more robust and adaptive basis for risk management in the context of ongoing climate change.

This study investigates the record-breaking extreme rainfall event of 1 July 2023 in Zhenba County and introduces an integrated multi-source data fusion framework for geohazard assessment. High-resolution Planet satellite imagery, combined with visual interpretation, was employed to assess pre- and post-rainfall surface changes. This approach facilitated the identification of landslides and other hazards and enabled the development of a rainfall-induced geohazard inventory. Simultaneously, SBAS-InSAR technology was utilized to monitor ground deformation throughout the study area, detecting surface displacement in both known hazard zones and areas of latent instability, thus offering critical input for hazard evaluation. Field verification and follow-up investigations were carried out to validate remote sensing interpretations, thereby ensuring the accuracy and completeness of the disaster inventory. Key evaluation indicators—such as topographic, hydrological, and geological variables—were selected and integrated with SBAS-InSAR deformation outputs to construct a multidimensional factor dataset. Cumulative rainfall was specifically extracted as a core input variable to capture the distinctive hydrometeorological characteristics of the extreme event. Kernel-based classifiers, including support vector machines (SVMs) and Kernel Extreme Learning Machines (KELMs), along with the single-layer Extreme Learning Machine (ELM), were excluded due to their limited compatibility with the SHAP framework, which restricts factor-level interpretability. Additionally, ELM’s random weight initialization may lead to unstable generalization compared to the deeper or ensemble architectures employed in this study. Consequently, four algorithms were employed that integrate SHAP-compatible outputs with strong capabilities to capture complex nonlinear relationships: Deep Neural Network (DNN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM). These algorithms were used for modeling and predicting landslide hazard under extreme rainfall conditions. Furthermore, model integration and comparative performance analysis were conducted to assess predictive accuracy, computational efficiency, and model applicability. This evaluation focused particularly on scenarios characterized by complex geological settings and the combined use of SBAS-InSAR deformation data and extreme rainfall inputs. The findings of this research enhance the precision of hazard assessments for record-breaking rainfall events in the Qinling-Daba Mountains and offer a scientific basis for early-warning systems. Moreover, the proposed methodology presents a technically robust and theoretically grounded framework applicable to other geologically complex regions facing increasing exposure to extreme rainfall events under ongoing climate change.

2. Overview of Regional Geological Background

2.1. Regional Geological Overview

The study area is situated in the southwestern part of Zhenba County, Shaanxi Province, China (Figure 1a), encompassing an area of 512.17 km². The highest point in the study area is located in the northeast, with an elevation of 1734 m, while the lowest point has an elevation of 556 m (Figure 1b). The Zhenba Fault, which separates strata of varying ages, extends along the eastern side of the study area. East of the fault, the strata primarily belong to the Sinian and Cambrian systems, with lithologies dominated by tuffaceous sandstone, dolomite, shale, and sandstone. To the west of the fault, Triassic and Jurassic strata predominate, with lithologies such as sandstone, limestone, and mudstone [40]. The study area lies at the boundary between the Yangtze Plate and the Qinling microplate, characterized by a complex geological structure (Figure 1c) [41]. The area is situated within the second-level structural unit known as the “Changling Depression”, which is marked by multiple reverse folds and back-thrusting structures, typical of a faulted fold belt [42]. Folding deformation in the region predominantly took place during the late Jurassic to early Cretaceous periods, with these structural features being closely linked to deep detachment surfaces [43]. The combined effects of compression and deep shear in the detachment zone have resulted in the deformation of the shallow strata, which primarily include the Triassic Jialingjiang Formation (T₂j), Xujiahe Formation (T₃x), and the Jurassic Baitianba Formation (J₁b), Qianfoyan Formation (J₂q), and Shaximiao Formation (J₂s) [44,45]. The shallow strata are predominantly composed of continental clastic rocks, with alternating or intercalating layers of sandstone and mudstone. The mountains have undergone significant uplift and substantial terrain changes due to multiple stages of tectonic activity. The complex fold structures and steep topographic relief within the study area create highly favorable geological conditions for the initiation and evolution of geohazards. The substantial elevation difference of up to 1178 m generates considerable gravitational potential energy, while the joint and fracture networks produced by folding deformation act as preferred sliding surfaces. In addition, the rhythmic interbedding of sandstone and mudstone creates mechanically contrasting weak zones. These combined factors render the study area highly susceptible to landslides under extreme rainfall conditions.

2.2. Meteorological and Extreme Rainfall Characteristics

The study area experiences a subtropical monsoon climate characterized by distinct seasons and abundant rainfall, with an annual average precipitation of 1260.4 mm. In the summer, the area is predominantly influenced by the southwest monsoon, and the complex topography significantly obstructs airflow, resulting in a pronounced windward slope effect [46]. This results in rainfall amounts and intensities in the study area significantly higher than those in surrounding regions, establishing it as a key rainfall center for the Qinling Mountains and the entire Shaanxi Province (Figure 2). The hourly precipitation data from four monitoring stations during the extreme rainfall event under study are shown in Figure 3. Specifically, 3–h precipitation reached 124.8 mm (Figure 3a,c), 24–h precipitation amounted to 228.7 mm, and 72–h precipitation totaled 408.9 mm (Figure 3b). In this study, cumulative rainfall over 3 h, 24–h, and 72–h intervals was selected for statistical analysis to comprehensively characterize the extreme nature of the rainfall event across multiple temporal scales. The 3–h cumulative rainfall reflects the peak intensity of short-duration extreme precipitation; the 24–h cumulative rainfall represents the daily maximum total; and the 72–h cumulative rainfall captures the cumulative effect of prolonged, continuous precipitation. Notably, 3–h precipitation was 170% higher than the historical maximum for the region, 24–h precipitation exceeded the historical maximum by 113%, and 72–h precipitation was 141% higher than the historical maximum (Figure 3d,e). These figures highlight that precipitation during all statistical periods (3–h, 24–h, and 72–h) surpassed historical records, underscoring the exceptional nature of this rainfall event. The high intensity and prolonged duration of rainfall triggered severe flash floods, causing widespread damage to buildings, roads, and infrastructure, and resulting in traffic disruptions. Secondary disasters, including landslides and debris flows, were also triggered. Preliminary statistics show that 223 roads were damaged, 99 of which were severely impacted, 4 bridges were destroyed, and 731 houses were damaged, resulting in direct economic losses totaling 1.52 billion yuan. The scale and impact of the geological disasters caused by this rainfall event clearly indicate its extreme nature. This event profoundly affected the local ecosystem and residents’ livelihoods, which is consistent with the characteristics of extreme rainfall events that cause severe consequences. Therefore, we classify this rainfall event as an extreme event.

2.3. Landslide Inventory and Classification

The accuracy and completeness of the landslide inventory are fundamental to ensuring the predictive capability and reliability of machine learning models. Errors or omissions in the landslide inventory may introduce noise during model training, impair generalization, and result in misclassification of landslide types. Such inaccuracies can lead to learning spurious causal relationships and ultimately distort the spatial distribution of hazard predictions. To mitigate these biases, this study implemented a refined landslide identification and verification procedure. Specifically, post-event landslides were delineated using 3 m resolution Planet imagery acquired before (Figure 4a,c) and after (Figure 4b,d) the extreme rainfall event, based on criteria such as boundary morphology, macro-textural disturbances, spectral anomalies, and topographic deformation patterns. Cross-validation was conducted by three independent interpreters, and approximately 80% of the interpreted landslides were verified through field investigations, resulting in an overall verification accuracy of 94.8%. Additional landslides identified during field surveys were also incorporated into the database, substantially enhancing its completeness and reliability. This robust inventory provides a reliable foundation for subsequent model training and performance evaluation. Through systematic interpretation of Planet satellite imagery before and after the extreme rainfall event in the study area, combined with field surveys, we established a comprehensive disaster database comprising 1012 landslides (Figure 5a–d). The database includes all primary types of geological disasters identified in the study area. Following Varnes’s landslide classification methodology, we classified these hazards into three main categories: debris flows, rockslides (including rockfall), and soil slides. This classification system considers both the type of movement mechanism and the material properties involved in each type of slope failure. The database records essential information for each disaster, such as precise geographic coordinates, disaster type, and scale, forming a robust data foundation for subsequent hazard assessments. Rockslides are the most common disaster type in the study area, accounting for 594 points, or 58.7% of the total. These rockslides are easily identifiable on satellite imagery by their characteristic slope boundaries and sliding masses. Field surveys reveal that these rockslides exhibit distinct morphological features, including well-defined rear scarps, sliding masses, and accumulation zones at the toe. The sliding surfaces typically form along joint fractures or bedding planes. Debris flows are the second most frequent disaster type, with 293 points identified, representing 29.0% of the total. These debris flows are primarily located in existing valley systems and exhibit typical channel erosion and fan-shaped deposition characteristics. Image interpretation shows that debris flows often display a three-stage structure—source area, flow path, and accumulation area—with distinctive textural features. Field surveys indicate that debris flows possess significant destructive power and widespread impacts. In the study area, the largest debris flow spanned an area of 1.2 × 105 m². Debris flows not only caused significant damage to infrastructure along their paths but also buried large areas of farmland, roads, and residential areas, leading to direct economic losses and affecting a much higher number of people compared to other geological disasters. Soil slides were identified at 124 points, accounting for 12.3% of the total. These disasters primarily occur in areas with thicker soil layers, where distinct arcuate sliding traces and fault movements can be observed in the field. Soil slides are more likely to occur in regions with relatively sparse vegetation cover and are often closely linked to local topographic and hydrological conditions.

3. Methodology

To systematically develop a landslide susceptibility assessment framework under extreme rainfall conditions, this study designed a comprehensive technical approach based on multi-source data fusion, which integrates key methodological components including high-resolution remote sensing image interpretation, SBAS-InSAR surface deformation monitoring, field investigation and validation, and machine learning modeling. The overall methodological workflow is illustrated in Figure 6.

3.1. SBAS-InSAR Processing Workflow

In this study, Sentinel-1 C-band (wavelength: 5.6 cm) VV-polarized synthetic aperture radar data were employed, including 52 high-coherence SAR scenes acquired over a 2.5-year period (January 2021 to August 2023). The Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technique was adopted for ground deformation monitoring. Data preprocessing involved precise orbit correction, image co-registration, and interferogram generation. A temporal baseline threshold of 48 days was established to ensure temporal coherence, while a spatial baseline threshold of 300 m was applied to mitigate spatial decorrelation and preserve geometric accuracy. The 12.5 m resolution Advanced Land Observing Satellite Digital Elevation Model (ALOS DEM) was employed for topographic phase removal. Phase unwrapping was conducted using the minimum cost flow algorithm. For quality control, a coherence threshold of 0.4 was applied to select high-quality pixels, while water bodies, shadow regions, and layover areas were masked. Multilooking was applied using a range window of 4 pixels and an azimuth window of 1 pixel. Singular value decomposition (SVD) was employed for time-series analysis to derive deformation velocity fields, and atmospheric phase delays were corrected using height-correlated correction and spatiotemporal filtering. The results indicate that the annual line-of-sight (LOS) deformation rates in the study area ranged from −56 mm/year to 31 mm/year (Figure 7). Cross-validation was conducted using GNSS benchmark data from geohazard monitoring stations established by the Zhenba County Natural Resources Bureau, as well as field geological investigations, to ensure the accuracy and reliability of the InSAR results.

3.2. Selection of Evaluation Factors

The occurrence of geohazards is controlled by the complex interplay of multiple contributing factors. Given the spatial extent of the study area (512.17 km²), variable selection was used to address spatial heterogeneity and ensure predictive accuracy. Based on field investigations, spatial distribution, and the environmental characteristics of landslides, 12 factors were selected for comprehensive analysis. The selected factors included topographic factors (plan curvature, slope, Topographic Position Index, profile curvature) (Figure 8e,f,i,j); geological structural factors (folding zones) (Figure 8d); hydrological factors (Stream Power Index, Topographic Wetness Index, distance to rivers) (Figure 8b,h,k); vegetation factors (NDVI) (Figure 8c); deformation rate (SBAS-InSAR) (Figure 8a); and precipitation factors (total cumulative rainfall from extreme rainfall events) (Figure 8l). These factors were derived from 12.5 m resolution DEM data, 1:50,000 geological maps, and Sentinel-1 satellite imagery. The selection of these factors was primarily guided by their explicit physical relevance to the mechanisms governing geohazard formation. The velocity map was selected over cumulative deformation as it provides normalized annual rates that facilitate consistent multifactor analysis and long-term trend assessment. Slope directly regulates the gravitational driving forces that control mass movements along hillslopes. Plan curvature and profile curvature affect surface water accumulation and redistribution processes. Lower plan curvature values indicate concave convergent surface morphology that enhances runoff concentration capacity, while lower profile curvature values represent concave slope profiles that decelerate flow and increase the potential for water accumulation. The Topographic Position Index (TPI) characterizes local terrain variations. The Topographic Wetness Index (TWI), integrating slope and upslope contributing area, effectively represents the potential surface water storage capacity per unit area. This metric closely aligns with the pore water pressure response mechanism triggered by extreme rainfall events and has demonstrated superior explanatory power relative to other hydrological factors in numerous landslide susceptibility studies [47,48]. The Stream Power Index (SPI) quantifies the erosive capacity of slope runoff; elevated SPI values signify intensified fluvial erosion, which may compromise slope stability and increase landslide susceptibility. Distance to rivers captures the scouring and toe-erosion effects of surface water on slope bases during extreme rainfall events. The fold zone variable reflects the influence of pre-existing tectonic deformation on the orientation and structural configuration of strata, directly influencing slope failure mechanisms. The normalized difference vegetation index (NDVI) quantifies vegetation cover, which stabilizes slopes through root reinforcement and hydrological interception processes. Additionally, variable selection considered data availability and reliability constraints. Soil type and land use were intentionally excluded from the evaluation framework. For soil type, the study area is uniformly dominated by silty clay, exhibiting minimal spatial heterogeneity in geotechnical properties, thereby limiting its discriminatory capacity for hazard prediction. For land use, the region is predominantly covered by continuous forest, with forest coverage exceeding 85% and minimal anthropogenic disturbance. Moreover, land-use type is highly redundant with the NDVI, which already serves as an effective proxy for vegetation coverage and land-use characteristics. Consequently, these two candidate factors were excluded from the final predictor set. The remaining twelve factors were subjected to Pearson correlation analysis, with the highest absolute correlation coefficient observed at 0.48894 (Figure 9). According to multivariate statistical theory, absolute Pearson correlation coefficients below 0.5 generally suggest the absence of significant multicollinearity [49,50]. Therefore, the selected evaluation factor set exhibits strong independence, with each factor capturing distinct dimensions of geohazard susceptibility, thereby minimizing prediction bias arising from variable redundancy.

3.3. Analysis and Validation of Regional Geological Hazard Risk Under Extreme Rainfall Conditions

The formation of landslides involves the dynamic interaction of multiple factors, including terrain, geological structures, and hydrological conditions, with extreme rainfall events exacerbating this complexity. The nonlinear interactions among these factors pose significant challenges for geological hazard risk assessment [51]. This requires the evaluation model to effectively address the synergistic effects of multidimensional data while accurately capturing their inherent nonlinear relationships. In this study, four advanced machine learning algorithms—Deep Neural Network (DNN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM)—are employed to construct a risk evaluation model for extreme rainfall events. These methods possess strong nonlinear mapping capabilities, making them well-suited for handling the multifactorial interactions in complex geological environments [52,53,54,55]. The DNN model was developed with a multilayer architecture comprising an input layer, three hidden layers (with 50, 30, and 20 neurons, respectively), and an output layer. Through multiple layers of nonlinear transformations, the model effectively captures the complex coupled interactions among rainfall, ground deformation, and geohazard occurrences. The ReLU activation function was utilized in the hidden layers to mitigate vanishing gradient problems, whereas the Sigmoid function was adopted in the output layer to map predictions to probability values ranging from 0 to 1. Systematic hyperparameter optimization was performed using grid search in conjunction with five-fold spatial block cross-validation. The search space included the learning rate (0.001–0.1, on a logarithmic scale), the number of hidden layers (2–6), the number of neurons per layer (20 to 128, decreasing progressively across layers), and dropout rates (0.1–0.5). The optimal hyperparameter configuration was identified as a learning rate of 0.005 (with the Adam optimizer) and a dropout rate of 0.3. To mitigate spatial autocorrelation bias often inherent in conventional random sampling, the study area was subdivided into 2 × 2 km geographical blocks. A spatial block cross-validation strategy was applied to ensure full spatial independence between training and validation datasets. This approach was validated using block distance analysis and spatial uniformity tests, substantially reducing spatial dependence between datasets and producing more robust and conservative evaluations of model performance (Figure 10a). The XGBoost model is based on the decision tree ensemble method and optimizes prediction results iteratively using gradient boosting. In this study, the model is configured with a learning rate of 0.1, a maximum depth of 6, a subsampling rate of 0.7, and regularization parameters λ = 1 and α = 0.1, using 300 trees (Figure 10b). The RF model constructs multiple decision trees and makes predictions by averaging their results. It fully utilizes bootstrap sampling and random feature selection to enhance model stability. In this study, the model consists of 1000 trees, with a maximum depth of 15 for each tree (Figure 11b). The LightGBM model utilizes a histogram-based decision tree algorithm and a leaf-wise growth strategy, significantly enhancing computational efficiency while maintaining high accuracy. The model is configured with a maximum number of leaves of 31, 300 trees, a learning rate of 0.05, and a minimum gain of 0.1 for each iteration (Figure 11a). To ensure the model’s effectiveness and predictive capability, this study employed a stratified random sampling method to split the disaster data into a training set and a validation set with a 7:3 ratio. Specifically, 70% of the disaster points (709 total) were randomly selected for model training, while the remaining 30% (303 total) were used for model prediction. Simultaneously, a random sampling approach was applied to generate 1012 non-landslide sample points, matching the number of landslide samples, within stable zones. To minimize potential spatial interference, each non-landslide point was positioned at least 100 m away from any identified landslide. The non-landslide samples were subsequently partitioned into training and validation subsets using the same 7:3 split ratio. To comprehensively evaluate the model’s performance, this study established a systematic framework for accuracy validation. First, the Receiver Operating Characteristic (ROC) curve and its Area Under the Curve (AUC) were used as the primary metrics for evaluating model performance. The ROC curve evaluates the model’s performance by plotting the relationship between the true positive rate (sensitivity) and the false positive rate (1-specificity). The AUC ranges from 0 to 1, where larger values indicate better model performance [56]. Additionally, sensitivity and specificity metrics were introduced to assess the model’s classification ability quantitatively. Sensitivity reflects the model’s ability to correctly identify landslide pixels, calculated as the ratio of true positive samples (TP) to the sum of true positives and false negatives (TP + FN). Specificity indicates the model’s ability to correctly identify non-landslide pixels, calculated as the ratio of true negative samples (TN) to the sum of true negatives and false positives (TN + FP).

$$\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(1)

$$\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}}}$$

(2)

Here, TP (true positive) and TN (true negative) represent the number of pixels correctly classified as landslides and non-landslides, respectively, while FP (false positive) and FN (false negative) represent the number of pixels incorrectly classified. Furthermore, the Kappa coefficient was used to evaluate the model’s performance in predicting landslides and non-landslides. This coefficient measures classification effectiveness by comparing the observed agreement (P_obs) with the expected agreement (P_exp) [57]. P_obs represents the proportion of correctly classified pixels in the total sample, while P_exp accounts for the distribution of pixels in each category. The formula for calculating the Kappa coefficient is presented as follows:

$$\mathrm{K}={\displaystyle \frac{{\mathrm{P}}_{\mathrm{o}\mathrm{b}\mathrm{s}}-{\mathrm{P}}_{\mathrm{e}\mathrm{x}\mathrm{p}}}{1-{\mathrm{P}}_{\mathrm{e}\mathrm{x}\mathrm{p}}}}$$

(3)

$${\mathrm{P}}_{\mathrm{o}\mathrm{b}\mathrm{s}}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{N}}}$$

(4)

$${\mathrm{P}}_{\mathrm{e}\mathrm{x}\mathrm{p}}={\displaystyle \frac{\left(\right(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})\times (\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P})+(\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N})\times (\mathrm{F}\mathrm{N}+\mathrm{T}\mathrm{N}\left)\right)}{\sqrt{\mathrm{N}}}}$$

(5)

Based on previous studies, Kappa coefficient values can be categorized into the following levels: near-perfect (0.81–1.00), substantial (0.61–0.80), moderate (0.41–0.60), fair (0.21–0.40), poor (0.01–0.20), and very poor (0.00) [58,59]. Through this comprehensive evaluation framework, the model’s performance can be objectively assessed from multiple dimensions, providing scientific evidence for verifying the model’s reliability.

Figure 10. (a) DNN model setup. (b) XGBoost model setup.

Figure 10. (a) DNN model setup. (b) XGBoost model setup.

Figure 11. (a) LightGBM model setup. (b) RF model setup.

Figure 11. (a) LightGBM model setup. (b) RF model setup.

4. Results

4.1. Model Evaluation Results

The results of the geological hazard risk assessment under extreme rainfall conditions are presented in Figure 9. The study area is divided into five hazard risk levels based on the established classification framework [23,39]: very low, low, medium, high, and very high. All four hazard maps exhibit a consistent pattern: very high-risk areas are concentrated in steep mountainous regions, particularly along both sides of rivers, while low-lying plains and gentle slopes are primarily classified as low-risk areas. This suggests that under heavy rainfall conditions, areas with steep terrain and water flow convergence are consistently identified as potential landslide hotspots across all models. The consistency among the models improves the overall reliability of the extreme rainfall hazard prediction approach. Although there is general agreement in identifying the key high-risk areas, notable differences exist in the hazard range and intensity predicted by each model. The DNN model identifies the most pronounced and widespread high-risk areas, not only marking known landslide-prone slopes but also identifying newly emerging local hotspots that could be at risk during heavy rainfall events. Due to its neural network architecture, the DNN model captures the complex nonlinear relationships among multiple conditional factors, producing highly detailed hazard maps (Figure 12a). In contrast, the maps produced by LightGBM and XGBoost, while highly consistent with the DNN model in identifying major high-risk areas, show slightly lower continuity in high-risk coverage at the boundary areas (Figure 12c,d). These gradient boosting models tend to concentrate high-risk areas around historically unstable slopes, with fewer discrete high-risk pixels in between. The similarity in the results of LightGBM and XGBoost is noteworthy, as the underlying algorithmic principles of both models are similar, which explains this consistency. Meanwhile, the RF model provides a more conservative hazard estimate: it clearly marks the major landslide-prone areas as high-risk but, compared to other models, classifies a larger proportion of terrain as medium-risk (Figure 12b). Many areas identified as high-risk by the DNN model or the boosting models under extreme rainfall are categorized as medium-risk in the RF map, suggesting that the RF model may underestimate some hazard areas. This conservative prediction by the RF model may result from its ensemble averaging feature, which smooths extreme responses and could potentially overlook fine-scale hazard variations. In terms of disaster proportion (Figure 13a), the DNN model shows a significantly higher proportion of disasters in high-risk areas compared to the other models, indicating that the DNN model performs well in identifying and predicting regions more susceptible to disasters. Additionally, in the high-risk area proportion (Figure 13b), the DNN model shows a larger area proportion, with the area in high-risk zones being much higher than that in the RF model. This suggests that the DNN model provides a more precise prediction of disaster distribution in high-susceptibility regions, effectively capturing disaster occurrence and distribution characteristics.

The performance metrics of the four models correspond with the qualitative observations from the maps. Using the ROC curve and the corresponding AUC as evaluation metrics, the DNN model achieved the highest prediction accuracy, with an AUC of 0.82, indicating an excellent ability to distinguish between landslide and non-landslide locations under extreme rainfall conditions. The XGBoost and LightGBM models followed closely, with AUCs of 0.80 and 0.79, respectively. These high AUC scores demonstrate that both boosting models have nearly equivalent and strong predictive abilities, slightly trailing the DNN model. In practical applications, the differences between 0.82 and 0.80 or 0.79 are relatively small, suggesting that XGBoost and LightGBM perform similarly to the DNN model in ranking hazard zones. In contrast, the RF model had the lowest AUC of 0.76 among the four models. Although an AUC of 0.76 still indicates good predictive capability, it is notably lower than the those of other models (Figure 14). Further evaluation using the Kappa coefficient confirmed the model performance ranking (

Table 1. Kappa value for the four models.

	DNN		RF		XGBoost		LightGBM
Kappa value	Training	Prediction	Training	Prediction	Training	Prediction	Training	Prediction
	0.833	0.812	0.763	0.741	0.815	0.801	0.807	0.791

). The DNN model had a Kappa value of 0.833 during the training phase and 0.812 during the prediction phase, indicating a very high degree of consistency. Similarly, XGBoost had a Kappa value of 0.815 (training) and 0.801 (prediction), and LightGBM had a Kappa value of 0.807 (training) and 0.791 (prediction). The Kappa values around 0.8 during the prediction phase confirm that both the XGBoost and LightGBM models exhibit extremely high accuracy, consistent with their AUC scores. The minimal differences between training and prediction Kappa values suggest that these models have strong generalization abilities and do not show signs of overfitting. In contrast, the RF model had a lower Kappa value, with 0.763 during training and 0.741 during prediction. While still in the “substantial” agreement range, the RF model’s performance lagged behind that of the other models. The lower Kappa value of the RF model (consistent with the AUC result) suggests slightly weaker consistency between the model’s predictions and actual landslide occurrences.

4.2. Factor Contribution Under Extreme Rainfall Conditions

SHapley Additive exPlanations (SHAP) values were calculated for the four trained models (DNN, XGBoost, LightGBM, and RF) to quantify the contribution of each factor to landslide occurrence under extreme rainfall conditions (Figure 15). For each model, SHAP values for all features were computed across 1012 landslides, and the average absolute SHAP value for each feature was calculated as a global importance indicator. The average absolute values for the four models were then averaged and normalized (with the total sum equal to 100%) to obtain the contribution percentage of each factor. A larger average absolute SHAP value indicates a greater influence of that factor [60]. In the SHAP summary plot, the SHAP values for each sample are color-coded, with red indicating a positive contribution to landslide prediction and blue representing a negative contribution. The combined results show that distance to rivers has the highest average contribution (20% of total importance), followed by cumulative extreme rainfall (16%), Topographic Wetness Index (TWI, 12%), slope (10%), InSAR-derived deformation rate (8%), fold zones (7%), and Topographic Position Index (TPI, 3%). The high contribution of distance to rivers reflects its multifaceted hazard-inducing mechanisms under extreme rainfall conditions. Under normal conditions, lateral river erosion continuously undermines riverbank slope stability; however, this process becomes significantly intensified during extreme rainfall events. Flash floods triggered by extreme precipitation rapidly increase river discharge, dramatically altering the stress state of riverbank slopes within a short timespan. Simultaneously, river-adjacent slopes experience pronounced increases in groundwater levels during intense rainfall, substantially elevating failure risk. The TWI, representing potential water accumulation capacity, also plays a significant role, indicating that areas prone to water accumulation exhibit heightened landslide susceptibility under extreme rainfall. The interaction among distance to rivers, slope, and TWI constitutes a hazardous tripartite coupling mechanism under extreme rainfall scenarios. Specifically, elevated TWI values reflect convergent topographic features that promote surface runoff and subsurface flow accumulation along riverbanks, rapidly increasing soil saturation and pore water pressure under intense rainfall. Concurrently, high-magnitude river flows intensify slope toe erosion, further exacerbating slope instability. This synergistic interaction explains the highest SHAP contribution of 20% attributed to distance to rivers. This importance weight reflects not only the direct erosional effects of rivers but also the systemic role of river systems as convergence zones for geomorphological, hydrological, and geological processes. Additionally, SBAS-InSAR deformation rates identify pre-existing ground movements, indicating that zones undergoing creep or deformation are more prone to reaching failure thresholds under rainfall triggers. Overall, the analysis demonstrates that hydrological factors dominate landslide initiation under extreme rainfall conditions, while topographic and geological factors serve as secondary but essential contributors.

4.3. Field Validation of Typical Cases

To further validate the predictive performance of the models under extreme rainfall conditions, we selected two representative field cases for testing: a giant debris-flow event and a typical landslide event (Figure 16 and Figure 17). The selection of typical validation cases was based on disaster scale and the representativeness of disaster types: we chose the largest debris-flow event in the study area to validate the models’ capability for predicting high-magnitude disasters, and a representative large-scale landslide to assess model performance on the most common disaster type. In the debris-flow case (Figure 16a,b), subplots Figure 16c–f show the hazard maps generated by the DNN, LightGBM, XGBoost, and RF models. The DNN model (Figure 16c) identifies the largest continuous high- and very-high-risk areas. Notably, these very-high-risk areas closely correspond with the actual debris-flow channels. In contrast, the RF model (Figure 16f) predicts the very-high-risk areas primarily within the identified flow channels. The LightGBM and XGBoost models (Figure 16d,e) also capture extensive high-risk areas, although their coverage is not as broad as that of the DNN model. This indicates that the DNN model provides the most extensive delineation of the disaster range for the debris-flow scenario, while the RF model presents the most conservative estimate of high-risk areas. Importantly, all four models identified the debris-flow accumulation areas as very-high-risk zones (Figure 16c–f), with these areas representing the highest potential hazard zones. This confirms the models’ strong ability to capture critical debris-flow hazard areas. Overall, these results demonstrate that, while all models accurately locate the primary hazard zones, the DNN model provides the most comprehensive delineation of the disaster extent.

In the typical landslide case (Figure 17a), the SBAS-InSAR deformation rate from January 2021 to August 2023 (Figure 17b) serves as an independent measure of ground motion. The hazard maps generated by the DNN, LightGBM, XGBoost, and RF models are shown in Figure 17c–f. In this case, the high-risk areas predicted by the DNN model (Figure 17c) align well with the subsidence areas observed by InSAR, indicating that the neural network effectively delineated the deformed slope. In contrast, the other three models (Figure 17d–f) restricted high-risk predictions to the toe of the landslide and the adjacent areas affected by river erosion, rather than covering the entire observed deformation zone. In fact, the DNN map (Figure 17c) covers the entire deformation body observed by SBAS-InSAR, while the LightGBM, XGBoost, and RF maps (Figure 17d–f) mainly emphasize the toe and river erosion zones. Therefore, the DNN model captures the complete spatial extent of slope deformation more comprehensively, while the tree-based models are more influenced by local topographic features.

5. Discussion

Record-breaking extreme rainfall events present critical challenges for disaster mitigation and prevention, especially when they occur in regions that have not historically experienced such events. Global warming has been linked to more frequent and intense precipitation, and even slight increases in rainfall can significantly heighten landslide risks [61,62]. In regions like Zhenba County in the Qinba Mountains, which are characterized by steep terrain and complex geology, the July 2023 rainfall event shattered historical records and triggered widespread slope instability. Against this backdrop, employing advanced machine learning models such as DNN, XGBoost, LightGBM, and RF for disaster hazard assessment under extreme rainfall conditions offers significant scientific and practical value. These models, with their powerful nonlinear fitting capabilities and adaptive feature learning abilities, can effectively capture the complex relationship between extreme rainfall and geological hazards, overcoming the limitations of traditional statistical models in handling high-dimensional features and nonlinear mappings. In this study, the superior performance of the DNN model compared to the tree-based models is primarily attributed to its ability to capture complex nonlinear interactions. Unlike the tree-based models, which construct piecewise linear decision boundaries, the DNN architecture employs multilayer nonlinear transformations to form continuous decision surfaces, offering clear advantages in modeling higher-order nonlinear interactions between rainfall inputs and InSAR-derived deformation data. Furthermore, the multilayer perceptron architecture of the DNN incrementally extracts and integrates hierarchical features through successive layers, employing the ReLU activation function for nonlinear transformations, thereby enabling effective modeling of intricate spatiotemporal patterns.

To capture complex geological processes under extreme rainfall conditions, this study integrates multi-temporal SBAS-InSAR observations with machine learning models, offering an innovative approach to regional disaster hazard assessment. Remote sensing and GIS technologies have long been advocated for landslide cataloging and hazard mapping; however, most previous studies have primarily relied on static factors (e.g., topography, geology, and historical rainfall data) without incorporating directly measured ground deformation data [63,64,65]. In contrast, the SBAS-InSAR analysis in this study directly observed slow surface movements under heavy rainfall and integrated this dynamic information into machine learning models, significantly enhancing predictive capabilities. The results align with recent studies, demonstrating that combining InSAR data with machine learning significantly improves disaster prediction capabilities. For instance, Kulsoom et al. [66] found that integrating SBAS deformation rates with the XGBoost model produced superior hazard maps, as deformation rates highlight actual landslide hotspots. Liu et al. [67] demonstrated that SBAS could accurately identify individual creeping landslides in high-risk areas. In comparison with the studies by Zhong et al. [68] and Gao et al. [69], this research not only focuses on single-phase InSAR deformation data but also captures the dynamic evolution of surface deformation through time-series analysis, providing a more accurate reflection of geological body stability changes under extreme rainfall conditions. In summary, SBAS-InSAR enhances the accuracy and interpretability of hazard predictions under extreme rainfall conditions by providing physical precursors.

The hazard maps from this study reveal a clear spatial concentration of high-risk zones, with extremely high-risk areas clustered along steep ravines and valley walls, while flat lowlands are classified as low-risk. This pattern reflects the hydrological effects of heavy rainfall. Intense rainfall causes rapid runoff accumulation on steep slopes, leading to the saturation of rock and soil masses. Groundwater flow and river erosion connect adjacent slopes, triggering chain failures [70,71]. This finding is consistent with the studies of Bordoni et al. [72] and Mondini et al. [73], who noted that under extreme rainfall conditions, moisture exchange between slope units increases significantly, promoting the synchronous instability of multiple slopes. Furthermore, disaster patterns resulting from extreme rainfall events differ fundamentally from those under regular conditions, as large-scale rainfall events activate regions that remain stable under normal rainfall, reflecting highly nonlinear interactions among various factors. This is consistent with the work of Mtibaa and Tsunetaka [74], who pointed out that landslide density sharply increases only when multi-period precipitation reaches extreme recurrence levels (such as a centennial threshold), whereas isolated short-duration heavy rainfall events trigger fewer landslides. This explains why record-breaking large-scale rainfall triggers more and larger landslides compared to typical heavy rainfall events, which is consistent with the multifactorial nature of slope instability: intense rainfall saturates rock and soil masses under critical conditions, weakening their strength. Traditional hazard methods, which often assume slope units are independent or use fixed rainfall thresholds, fail to capture these nonlinear cascading effects [75]. As rainfall intensity and temporal distribution change, the spatial patterns of disasters can shift significantly, and extreme rainfall events may trigger slope instability in areas predicted to be low-risk by standard hazard maps. This highlights the urgency and necessity of comprehensive disaster assessments in the context of increasingly frequent extreme climate events.

As climate change drives dramatic shifts in precipitation patterns, regional extreme rainfall events are becoming more frequent and intense, with particularly significant impacts on geological hazards [76,77]. Research in the Qinling-Daba Mountains has shown that the rising frequency of extreme rainfall events significantly increases the risk of slope instability and debris flows, highlighting the urgency of conducting regional hazard assessments under such extreme scenarios [78]. This study provides a framework for such work by explicitly incorporating record-breaking rainfall amounts and surface deformation characteristics. The findings of this study offer substantial practical value for disaster risk management and mitigation within the study area. The high-resolution hazard maps and validated predictive models can be directly incorporated into local emergency planning frameworks, providing scientific support for land-use zoning, infrastructure development, and the implementation of early-warning systems. The advantages of SBAS-InSAR technology in capturing millimeter-scale surface displacements offer critical insights for identifying potentially unstable areas, providing higher spatiotemporal resolution and monitoring accuracy compared to traditional methods. This not only provides scientific support for regional disaster prevention and mitigation but also serves as a reference for geological hazard management in other regions with similar geological characteristics. Furthermore, it proposes new approaches to addressing the increasing frequency of geological hazards in the context of global climate change. However, this approach presents two main limitations. First, the model’s dependence on historical rainfall data introduces inherent uncertainties in projecting future hazard scenarios, particularly under climate change conditions where extreme precipitation events may surpass historical observational ranges. Second, the scalability of the proposed framework is significantly limited in data-scarce regions, where the lack of long-term InSAR observations or insufficient rainfall monitoring impedes the acquisition of high-quality datasets essential for robust model development. Future research should further integrate socioeconomic vulnerability and adaptive capacity assessments to establish a more comprehensive hazard evaluation framework, providing systematic scientific support for regional sustainable development and climate change adaptation.

6. Conclusions

This study takes the extreme rainfall event in Zhenba County as a case study to develop a multidimensional landslide hazard assessment framework integrating SBAS-InSAR surface deformation monitoring with machine learning algorithms. Comparative analysis of four machine learning models demonstrates that the DNN model achieves superior predictive performance, with an AUC of 0.82 and a Kappa coefficient of 0.812 during the prediction phase. SHAP-based feature importance analysis indicates that distance to rivers is the most influential factor, contributing 20% to the model output, followed by cumulative rainfall and the TWI, which also contribute substantially. Within this framework, SBAS-InSAR technology performs dual functions; it provides high-precision surface deformation data for hazard identification and serves as a key input for the machine learning models, thereby enhancing model adaptability and predictive accuracy in complex mountainous environments. The establishment of this methodological framework represents a significant advancement for geological hazard risk assessment in the context of climate change.

The deep integration of SBAS-InSAR technology with machine learning models effectively overcomes the limitations of conventional assessment approaches related to data precision and spatial coverage, establishing a predictive framework suitable for geohazard risk assessment under climate change scenarios. The SHAP-based interpretability framework not only enables high-precision hazard prediction but also offers novel scientific insights into the relative importance of hazard-inducing factors under extreme rainfall conditions. These technical advancements provide important practical guidance for the development of more targeted disaster prevention and mitigation strategies. Building upon the methodological foundation established in this study, future research will focus on two key directions: (1) integrating socioeconomic vulnerability data into the current framework to develop a comprehensive risk assessment model that incorporates both environmental and socioeconomic dimensions, thereby achieving a more holistic representation of disaster risk; (2) extending the SBAS-InSAR and machine learning integration framework to additional climate-induced extreme events, such as droughts and extreme heatwaves, with the aim of establishing a multi-hazard assessment system applicable to climate change scenarios.