This study uses the national geohazard records of China for 2008–2023, provided by the Ministry of Natural Resources of the People's Republic of China, as the basic dataset for landslide susceptibility modelling and future trend analysis. The dataset was compiled annually and covers the whole of China each year, rather than only small areas affected by individual events. Therefore, the study area of each annual inventory is the whole of China. It includes the occurrence time and spatial location of landslides, debris flows, and collapses, which were collectively analysed as rainfall-induced landslide-related records in this study. The dataset includes both post-event records of occurred landslide-related geohazards and geohazard potential sites. The records of occurred hazards form a post-event inventory of landslide-related geohazards, compiled through post-event field investigation and verification by professional geological survey teams organized by local governments, together with multiple sources of information such as remote-sensing interpretation and UAV surveys, and then reported level by level. The dataset is mainly used for hazard verification, loss assessment, hazard identification, risk management, and geohazard prevention and control. Therefore, the recorded landslide-related records are mainly distributed in areas where slope failures may pose risks to people, infrastructure, or assets. The investigation and compilation procedures followed the industry technical specification, “Specification of Geological Hazard Risk Survey and Assessment (1 : 50 000)” (Ministry of Natural Resources of the People’s Republic of China, 2023). The minimum identifiable landslide area in the inventory is approximately 10 m². Because the dataset was compiled annually within a unified national reporting framework, the standards for investigation, identification, recording, and compilation were generally consistent across years. However, because these operational records are mainly compiled for hazard verification, loss assessment, risk management, and geohazard prevention, their spatial representativeness is more defensible in areas with human presence. Therefore, the modelling domain of this study was restricted to inhabited areas of China, rather than the entire national territory. As a nationwide operational geohazard dataset, it may still be affected by differences in regional accessibility, reporting practices, and investigation intensity, which may lead to some underreporting in remote areas or for small-scale events. Figure 1 presents the spatial distribution of landslide-related records in the dataset. More detailed information is provided in Fig. S1 in the Supplement.

In this study, 14 factors that may influence landslide susceptibility were selected (Table 1), with their spatial distributions provided in Fig. S2. All variables were resampled to a resolution of 1 km × 1 km. For continuous variables such as elevation and distance to river, we used bilinear interpolation, while for discrete variables including lithology and land use type, we employed mode sampling. Specifically, taking land cover as an example, the original land-cover data had a spatial resolution of 30 m. For each 1 km grid cell, all 30 m land-cover pixels within the cell were counted by category, and the category with the highest frequency was assigned as the representative land-cover type of that 1 km cell. This approach avoids generating artificial class values and ensures that the resampled categorical variables retain their physical meaning.

The assessment of rainfall-triggered landslide susceptibility within inhabited areas of China under climate change is structured into six key steps (Fig. 2): (1) extracting landslide-related records from the inventory and generating negative samples by random sampling within the inhabited-area modelling domain and outside buffer zones around recorded locations; (2) applying feature selection techniques to identify the conditioning factors from landslide conditioning variables; (3) constructing a feature matrix by extracting static and dynamic conditioning factors at each sample point, and dividing the dataset into training (70 %) and testing (30 %) sets; (4) training and optimizing an annual landslide susceptibility model using the RF algorithm; (5) LSM for both historical and future periods by applying downscaled climate projections under different climate change scenarios; (6) analysing the spatiotemporal changes in landslide susceptibility within inhabited areas of China under various climate change scenarios.

All officially documented landslide-related records were spatialized in ArcGIS Pro, resulting in 93 935 point-based records, all of which were retained as positive samples. This is because the inventory represents post-event operational geohazard records collected through on-site investigation and reporting procedures. The inhabited-area mask was used to define the modelling and prediction domain and to constrain the generation of negative samples. Each recorded location was treated as one point-based positive sample for model training. A 3 km buffer was then established around each recorded location. It should be noted that the pixels within these buffers were not converted into additional positive samples. Instead, the buffers were used only as exclusion zones to prevent negative samples from being selected too close to recorded locations, where the absence of documented landslide-related records may be uncertain. Negative samples were randomly generated within the inhabited-area modelling domain and outside all buffers around the recorded locations. The number of negative samples was set to twice the number of positive samples. Therefore, the 1 : 2 ratio refers to the number of point-based training samples, rather than to the total number of pixels within the buffers. The larger number of negative samples was used to better capture the environmental heterogeneity of stable areas while maintaining a manageable and relatively balanced training dataset. Because true absence locations are generally not directly available in landslide susceptibility modelling, the selection of negative samples can substantially affect model performance and prediction reliability (Zhu et al., 2019; Guo et al., 2024). To ensure the representativeness of negative samples, they were generated only within areas with human presence. Human presence was defined using the WorldPop R2025A population dataset. The population data were converted into a binary inhabited-area mask, in which grid cells with human presence were assigned a value of 1 and grid cells without human presence were assigned a value of 0. This binary mask was used only to constrain the inhabited-area modelling domain and negative sample generation (Bondarenko et al., 2025). Population data are publicly available at https://hub.worldpop.org/ (last access: 14 August 2026). To maintain temporal consistency, each negative sample was randomly assigned a year between 2008 and 2023 according to the temporal distribution of the positive samples.

Feature selection is an essential preprocessing step for high-dimensional data analysis, visualization, and modelling. In this study, two methods were employed for feature selection to enhance computational efficiency while preserving the accuracy of the model: Pearson correlation coefficient and tolerance (TOL) and variance inflation factor (VIF).

Relevant features were extracted for each sample point. Dynamic variables were obtained by constructing a time-series three-dimensional matrix (time × location × variable) in Python. Static variables, such as elevation and slope, were extracted using the Extract Multi Values to Points tool in ArcGIS Pro. All extracted variables were then integrated into a single feature matrix for model training.

RF is an ensemble learning algorithm that improves predictive performance by building and aggregating multiple decision trees (Breiman, 2001). The algorithm employs bootstrap aggregation, whereby each tree is trained on a randomly drawn bootstrap sample from the original dataset, and at each node, a random subset of features is selected for splitting. This randomization process promotes model diversity and robustness, enhancing accuracy and reducing sensitivity to noise and outliers (Stumpf and Kerle, 2011). By averaging the results of multiple trees, RF effectively reduces variance and the risk of overfitting, making it more robust than individual decision trees. In addition, RF can efficiently handle high-dimensional datasets and categorical variables with minimal preprocessing or encoding. Importantly, RF provides feature importance measures, which facilitate the identification of key variables influencing model predictions (Bureau et al., 2003). These strengths make RF particularly suitable for tasks such as landslide susceptibility mapping, where high accuracy, robustness, and interpretability are essential.

Accurate evaluation of the RF model is essential for predicting future changes in landslide susceptibility. In this study, we employed several widely used evaluation metrics, including accuracy, precision, recall, F1-score, the receiver operating characteristic (ROC) curve, and the area under the curve (AUC). These metrics are derived from the confusion matrix (Hossin and Sulaiman, 2015), with values closer to 1 indicating better model performance. Specifically, the ROC curve visualizes the trade-off between the false positive rate (FPR) on the x-axis and the true positive rate (TPR) on the y-axis.

For model development and validation, we randomly split the dataset into training (70 %) and testing (30 %) subsets. RF model was trained on the training subset, and all performance metrics were evaluated on the independent testing subset to avoid overfitting. Model hyperparameters were optimized and set to n_estimators = 886, max_depth = 20, max_features = “sqrt”, and min_samples_split = 6, using the Gini index as the splitting criterion. 4Accuracy=TP+TNTP+TN+FP+FN,5Precision=TPTP+FP,6Recall=TPR=TPTP+FN,7F1-score=2TP2TP+FP+FN,8FPR=FPFP+TN. Here, TP (true positive) and TN (true negative) represent the number of positive and negative samples that are correctly classified, respectively. FP (false positive) refers to the number of negative samples that are incorrectly classified as positive samples, while FN (false negative) refers to the number of positive samples that are incorrectly classified as negative samples.

We selected 31 GCMs from the NEX-GDDP-CMIP6 dataset. Each of these models includes the SSP1-2.6, SSP2-4.5, and SSP5-8.5 scenarios. Then, we evaluated the performance of these models against historical observations using the root mean square error (RMSE) to identify those suitable for China (Table S1). Subsequently, based on the selected models, we constructed a multimodel ensemble of annual precipitation, which was used to establish the historical baseline and to project future landslide susceptibility under different climate scenarios (SSP1-2.6, SSP2-4.5, and SSP5-8.5) and future periods, including the near-term future (2026–2050), mid-term future (2051–2075), and long-term future (2076–2100). To further account for uncertainty in rainfall-induced landslide susceptibility inhabited areas of China, we applied a 95 % confidence interval derived from the T-distribution. The landslide susceptibility classes were defined based on the mean susceptibility map of the historical baseline period. Specifically, annual landslide susceptibility maps from 1950 to 2014 were first averaged pixel by pixel to generate the baseline mean susceptibility map. The Jenks natural breaks method was then applied to this baseline map to classify landslide susceptibility into five levels: very low, low, median, high, and very high. The corresponding thresholds were 0–0.0086, 0.0086–0.2582, 0.2582–0.4617, 0.4617–0.6769, and 0.6769–1.0, respectively. To ensure comparability across different periods and climate scenarios, the classification thresholds derived from the baseline period were consistently applied to all subsequent susceptibility assessments. Finally, we assessed the impact of climate change on landslide susceptibility by analysing the changes in susceptibility relative to the baseline period across the different future periods and climate change scenarios.

First, Pearson correlation coefficients were used to select the most suitable landslide influencing factors from 12 continuous variables. The correlation coefficients among these 12 factors are shown in the correlation matrix (Fig. 3a). A strong positive correlation was observed between plan curvature and curvature (r = 0.85, p < 0.05); therefore, curvature was removed to avoid multicollinearity. To further verify whether the remaining 11 continuous variables and 2 categorical variables were suitable for landslide susceptibility modelling, multicollinearity tests were conducted (Fig. 3b–c). All factors had TOL values greater than 0.1 and VIF values less than 10, indicating no multicollinearity among the 13 influencing factors. The final selected factors included annual total precipitation, TWI, elevation, 475-year RP PGA, NDVI, plan curvature, profile curvature, slope, distance to road, distance to fault, distance to river, landcover, and lithology.

Model performance evaluation is a crucial component in the analysis of future rainfall-induced landslide susceptibility. In this study, 30 % of the sample data were utilized as a test set to assess the performance of the RF model. As illustrated in Fig. 4, model performance was visually analysed through the ROC curve and the confusion matrix. The RF model demonstrated good predictive performance, with an AUC value as high as 0.96. In the test set, the model achieved an accuracy of 0.89, a precision of 0.87, a recall of 0.80, and an F1 score of 0.83, validating its robust predictive performance. Additionally, through the analysis of Gini coefficients in the RF model, we explored the interpretability of the model. The results revealed that annual total precipitation was the key factor influencing model predictions, with a contribution rate as high as 16 %, significantly higher than other factors. Slope contributed 12 % and elevation contributed 10 % to the model (Fig. S3).

We analysed the relationship between landslide susceptibility and key influencing factors within inhabited areas of China during the historical baseline period (1950–2014). The top six continuous variables identified by RF importance ranking were categorized, and the area proportions of each category within different landslide susceptibility classes were calculated (Fig. 5). The results show that mean annual precipitation, slope, and NDVI are strongly positively correlated with landslide susceptibility. In areas with mean annual precipitation greater than 800 mm, the proportions of very high, high, and median susceptibility zones are 82.64 %, 86.56 %, and 75.33 %, respectively (Fig. 5a). Landslide susceptibility also increases markedly with slope. In areas with slopes steeper than 20°, the proportions of very high, high, and median susceptibility zones reach 51.43 %, 34.38 %, and 20.57 %, respectively (Fig. 5b). Similarly, in regions where NDVI exceeds 0.5, these proportions are 84.46 %, 79.80 %, and 66.57 %, respectively (Fig. 5f). These findings indicate that high landslide susceptibility is closely associated with humid environments, steep terrain, and relatively high NDVI values. Elevation exhibits a nonlinear relationship with landslide susceptibility, with higher susceptibility concentrated in low-elevation regions. Specifically, 65.53 %, 64.27 %, and 55.08 % of the very high, high, and median susceptibility zones are distributed below 1000 m (Fig. 5c). Landslide susceptibility is generally higher within 0–50 km of faults, decreases at intermediate distances, and shows a secondary increase at 200–500 km (Fig. 5e). The relationship between 475-year RP PGA and landslide susceptibility is nonlinear (Fig. 5d). In general, areas with very high and high susceptibility are more concentrated in the lowest PGA class, while their proportions decrease in the intermediate classes and increase slightly again in the highest PGA class. This suggests that seismic background conditions may not act as a direct trigger of rainfall-induced landslides, but may indirectly increase slope sensitivity to subsequent rainfall by damaging geologic structures and reducing slope stability.

We further analysed the effects of multi-factor combinations on the spatial distribution of landslide susceptibility within inhabited areas of China during the historical baseline period (1950–2014). Using the top three variables ranked by feature importance derived from the RF model (mean annual precipitation, slope, and elevation), we calculated the area proportion of regions with median to very high landslide susceptibility under different combinations, and visualized the results as heatmaps (Fig. 6). In the precipitation–slope combinations (Fig. 6a), the proportion of median to very high susceptibility areas generally increases with increasing slope, and this tendency becomes more pronounced under relatively high precipitation conditions. The highest value (96.33 %) is observed in areas with slopes greater than 30° and mean annual precipitation of 800–1200 mm. In the precipitation–elevation combinations (Fig. 6b), high susceptibility is mainly concentrated in areas with relatively high precipitation, the effect of elevation is nonlinear; the maximum proportion (93.71 %) is observed in areas above 3000 m with precipitation of 800–1200 mm. The elevation–slope combinations (Fig. 6c) further indicates that highly susceptible areas tend to occur where steep slopes coincide with relatively low and high elevations. In particular, the 0–500 and 2000–3000 m elevation ranges exhibit higher landslide susceptibility under steep-slope conditions.

Figure 7 shows projected changes in mean annual precipitation across China and its geological environmental zones under three climate scenarios and future periods, relative to the baseline. Overall, multimodel ensemble simulations show that future mean annual precipitation across China exhibits an increasing trend. The relative increases and associated uncertainties are greater under the SSP5-8.5 than under the SSP1-2.6 and SSP2-4.5. Compared to the near-term (2026–2050) and mid-term future (2051–2075), the long-term future (2076–2100) generally shows larger increases. According to projections from the NEX-GDDP-CMIP6 multimodel ensemble, mean annual precipitation in China is expected to increase by 59, 63, and 111 mm under the SSP1-2.6, SSP2-4.5, and SSP5-8.5, respectively, during the long-term future from 2076 to 2100. These correspond to increases of 10.46 %, 11.18 %, and 19.14 % relative to the baseline.

Projected changes in mean annual precipitation exhibit substantial spatial heterogeneity across different geological environmental zones. In the NW, where the baseline mean annual precipitation is only 136 mm, the relative increase in future mean annual precipitation remains low. During the long-term future under the SSP5-8.5, the increase is 59 mm. In geological environmental zones where baseline mean annual precipitation ranges from 400 to 900 mm, the relative increases in mean annual precipitation are generally higher, particularly in the Tibet and Yangtze. For instance, during the long-term future, under the SSP5-8.5, the largest increases are projected in these regions, reaching 170 and 147 mm, respectively. In the Southern region including SC and SW, where baseline mean annual precipitation exceeds 1000 mm, the absolute increase is relatively smaller compared to areas with baseline mean annual precipitation between 400 and 1000 mm. In the geological environment division, the associated uncertainties are the highest.

Figure 8a illustrates the spatial distribution of landslide susceptibility within inhabited areas of China during the baseline period (1950–2014). The results show that approximately 61.34 % of the inhabited-area modelling domain exhibited low to very low susceptibility, primarily located in northern Tibet, southern NW, the northwestern Loess, and central NE. Regions with median to very high susceptibility accounted for 38.66 % of the total simulated area, mainly distributed across SC, SW, eastern Tibet, and the Loess–Taihang Mountain area. Notably, in the SC, approximately 70.42 % of the area was exposed to median to very high landslide susceptibility. Figure 9 illustrates the increase in the area of median to very high landslide susceptibility zones within inhabited areas of China compared to the baseline period across future scenarios and periods. Overall, the spatial extent of median to very high susceptibility zones is projected to expand under all future scenarios and periods. The expansion is more pronounced under SSP5-8.5 compared to SSP1-2.6 and SSP2-4.5. In addition, long-term future shows a more significant increase in susceptibility area than near-term and mid-term future. In the baseline period, the proportion of median to very high susceptibility areas within the inhabited-area modelling domain is 38.66 %. During the long-term future under the SSP5-8.5, it increases to 40.98 %, corresponding to a spatial expansion of approximately 1.2 × 10⁵ km².

At the regional scale, all geological environmental regions within inhabited areas show an increasing trend in the extent of median to very high landslide susceptibility areas. The Loess and the SW exhibit the most significant expansion (Figs. 8, 9, and S4). In the baseline period, the proportion of areas with median to very high susceptibility in these regions is 10.07 % and 79.81 %, respectively. During the long-term future under the SSP5-8.5, these proportions increase to 17.38 % and 81.85 %, corresponding to area increases of 5.9 × 10⁴ and 2.2 × 10⁴ km². These expansions mainly occur in northern SW, and around the Taihang Mountains in the Loess (Figs. 8 and S4). During the long-term future under the SSP5-8.5, other regions such as SC, Tibet and NE also show increases. According to the multimodel ensemble averages, the areas of median to very high susceptibility in these regions increase by approximately 1.6 × 10⁴, 1.4 × 10⁴, and 8.3 × 10³ km², respectively, compared to the baseline. The expansions are mainly located in northern SC, eastern Tibet and southern NE. The NW and Yangtze show relatively small increases in median to very high susceptibility areas based on the ensemble mean (Fig. 8).

To further investigate the relationship between landslide-related records and conditioning factors, this study selected the top nine conditioning factors based on feature importance rankings from a random forest model for a graded statistical analysis (Fig. 10). For continuous variables, the natural break method was applied for classification, and landslide frequencies were calculated for each interval. Here, the total samples consisted of landslide-related records and negative samples generated at twice the number of positive samples. The results showed a clear positive relationship between annual total precipitation and landslide frequency (Fig. 10a). Landslide frequency generally increased with annual precipitation and exceeded 50 % when annual precipitation was greater than 1147 mm. Landslide frequency increased markedly with slope and exceeded 60 % when the slope was greater than 25° (Fig. 10b). In terms of elevation, the landslide frequency exhibited an M-shaped pattern with increasing altitude. The frequency was lowest in the -217–112 m elevation range, at only 16 %, whereas it reached a peak in the 2338–3514 m elevation band, with the highest value of 59 % (Fig. 10c). For the 475-year return period PGA, rainfall-induced landslides were mainly concentrated in the low-value intervals, while the frequency showed a secondary increase in some high-value intervals. This may suggest that seismic background conditions indirectly influence rainfall-induced landslides by weakening slope stability (Fig. 10d). Landslide-related records were more numerous near fault lines, especially within 0–40 km. However, landslide frequency did not decrease monotonically with fault distance and reached relatively high values in some intermediate-distance intervals (Fig. 10e). NDVI, which reflects vegetation coverage, showed an increasing trend in landslide frequency as NDVI values increased. When NDVI exceeded 0.56, landslide frequency was generally greater than 45 % (Fig. 10f). Profile curvature showed a pattern of higher landslide frequency at both ends and lower frequency in the middle: the frequency was lowest in the near-zero interval (-0.02–0.01), at only 15 %, but increased markedly toward both negative and positive extreme ranges, reaching as high as 70 %. This suggests that nearly planar slopes are relatively stable, whereas strongly concave or convex slopes are more susceptible to rainfall-induced instability (Fig. 10g). TWI showed a negative correlation with landslide frequency: as TWI increased, the frequency decreased continuously from 60 % in the low-value interval of 14.2–15.9 to 1.5 % in the high-value interval of 25.8 and above. This indicates that landslide-related records are more likely to occur in areas with lower TWI (Fig. 10h). Among all lithological types, siliciclastic sedimentary rocks (ss) contained the largest number of landslide-related records, with 26 298 cases (Fig. 10i).