1. Introduction
Global climate change increases the frequency and severity of typhoon events [
1]. In recent years, catastrophic typhoons have occurred frequently in coastal areas around the world, resulting in numerous geological disasters [
2,
3,
4,
5]. The southeastern coast of China, which is bordered by the Pacific Ocean, is susceptible to landslides caused by typhoons. Every year, there are a large number of casualties and economic losses [
6,
7,
8]. To date, the Chinese government has attempted to prevent landslide disasters by establishing a geological hazard warning system that includes weather forecasting [
9,
10,
11,
12]. However, the annual economic losses are still increasing. A large part of this is due to the continuous transformation of the geological environment by human engineering activities [
13,
14]. In the past few years, China’s urbanization index has rapidly increased from 0.157 in 2000 to 0.438 in 2015 [
15]. In areas with strong engineering activities, the Typhoon-triggered rainstorm was more likely to cause landslide disasters. Therefore, we urgently need to explore the characteristics of changes in human engineering activities on the occurrence of landslides.
Land-use and land-cover are often used to describe the impact of human engineering activities, including natural elements (water bodies, forests, cropland, and grassland) and human-modified surfaces (building activity areas). Many researchers found that LULC may alter the distribution of landslide susceptibility caused by rainfall [
16,
17,
18]. For instance, Pisano et al. [
19] analyzed the hill and low mountain areas of the Italian Southern Apennines land-use change and concluded that increasing forest and cropland area can reduce the susceptibility to landslides. Rohan et al. [
20] focused on the Southern Pennsylvania area and analyzed that urbanized areas are usually more prone to landslides, which are closely related to distance from roads and terrain curvature. Hao et al. [
21] analyzed the LULC dataset in Kerala, India, from 2000 to 2018 and found that landslides mainly occur in forest areas. In China, human activities and urban expansion have become the main causes of landslides in recent decades, accounting for approximately 70% of the total number [
22]. Chen et al. [
23] compiled two types of LULC maps with a time interval of 21 years (1992–2013) and analyzed that even if cropland and grassland were converted to forestland, the susceptibility of landslides still increased. Xiong et al. [
22] analyzed the changes in landslide susceptibility in Enshi City using landslide inventories from 2000 and 2020 and changes in LULC maps. Given that human activities can rapidly change vast areas, the LULC map may require more predictive scenarios. Promper et al. [
24] used the Dyna-CLUE model to analyze LULC changes in Austria over 138 years (1962–2100) to simulate the evolution of landslide risk. Tyagi et al. [
18] proposed the ANN-CA model to predict the LULC in 2030 and analyzed the variation patterns of landslides. Guo et al. [
25] utilized the land change modeler (LCM) integrated into IDRISI Selva software to obtain LULC maps for 2040 and 2080 and analyzed the susceptibility of shallow landslides in Wanzhou, China. Shu et al. [
26] also used the LCM software to explore how LULC in the Varan region of Spain has changed with landslide susceptibility over the past 150 years (1946–2097). Past prediction models had limited capabilities in investigating the causes of land-use changes. Moreover, using dynamic simulations to evaluate patch-level alterations of various land-use types over time and space was challenging [
27,
28]. In 2021, the High performance Spatial Computational Intelligence Lab of China University of Geosciences (Wuhan) developed the patch generating land-use simulation (PLUS) model to explore the driving factors of multiple types of land expansion and predict patch-level evolution of LULC [
28]. This model has made outstanding progress in the study of land-use analysis in China [
27,
29,
30,
31]. This article is the first attempt to explore the impact of LULC change on landslides using the PLUS model. In addition, a point often overlooked in many studies is to consider road distance as a static factor without considering its dynamic characteristics [
18,
19]. Generally speaking, road construction directly affects slopes through changes in road cut or surface water runoff. Therefore, this study explores the effects of road network and LULC changes on landslide susceptibility.
Currently, most of the literature evaluates the spatial probability (or susceptibility) of landslides by focusing on environmental and geographical factors [
32,
33,
34,
35,
36]. The production of landslide susceptibility maps with accurate, up-to-date, and reliable information is the focal point. With the rapid development of artificial intelligence and computer science, machine learning models have been widely used for landslide susceptibility assessment. More specifically, the shift from traditional statistical models to data-driven models fundamentally directly affects the predictive performance of susceptibility modeling [
37]. The widely used machine learning models currently include decision trees (DT) [
38,
39,
40], multi-layer perceptron network (MLPNN) [
41,
42], support vector machine (SVM) [
43,
44,
45], etc. Furthermore, some scholars have attempted to use ensemble models to enhance and improve the prediction errors of single classifiers. The Random Forest model is the earliest and most widely used ensemble model. The prediction accuracy of landslide susceptibility has been greatly improved by the integration of the bagging algorithm and decision tree model [
46,
47]. Similarly, based on gradient enhancement methods (i.e., gradient boosting decision tree (GBDT), XGBoost, and LightGBM), the model with predicted errors is strengthened and improved by the principle of reusing residual patterns [
48,
49]. Dou et al. [
50] integrated the basic model SVM using bagging, boosting, and stacking algorithms to explore the susceptibility distribution of shallow landslides in Japan. However, the limitations of these models are evident, i.e., optimal performance varies across different research areas. For example, He et al. [
48] analyzed the susceptibility of landslides and wildfires in Southeast Asia and obtained that the prediction accuracy of Random Forest model is always higher than GBDT and Adaptive Boosting model. Wang et al. [
51] found that the RF model (area under the receiver operating characteristic (ROC) curve, AUC = 0.88) outperformed XGBoost (AUC = 0.86) in the study area of Wuqi County in the hinterland of the Loess Plateau. Liu et al. [
52] found that the GBDT model is better than the RF model in spatial modeling of shallow landslides near Kvam, Norway. Due to variations in model operation strategies and sensitivity to datasets, it is challenging to identify a model with strong generalization capabilities that consistently performs optimally across diverse research areas [
34]. Therefore, some scholars have attempted the hybrid strategy of strong classifiers. Li et al. [
53] adopted a stacking strategy to integrate the convolutional neural network (CNN) and recurrent neural network (RNN), and the results showed that the proposed framework could retain the best predictive capability (AUC = 0.918) compared to CNN (AUC = 0.904) and RNN (AUC = 0.900). Arabameri et al. [
54] compared the modeling AUC values of credal decision tree (CDT), Alternative Decision Tree (ADTree), and their ensemble method (CDT-ADTree) of 0.837, 0.867, and 0.981, respectively. In the analysis of landslide susceptibility in the Three Gorges Reservoir area, Zeng et al. [
34] compared and analyzed various ensemble strategies (bagging, boosting, and stacking) and basic models (DT, SVM, MLPNN, XGBoost, and RF). They found that the stacking strategy, which ensembles strong models like XGBoost and RF, yielded optimal prediction accuracy and generalization ability. Therefore, appropriate ensemble strategies can help researchers obtain models with stronger robustness. In particular, more and more attention has been paid to the application of the hybrid model to landslide susceptibility analysis. We should explore the impact of the hybrid model and dynamic human engineering activities on landslide occurrence.
The southeast coastal mountainous areas of China are affected by typhoon-triggered rainstorms all year round. Land-use change (such as deforestation, agricultural, and urban regional expansion) and road network expansion have aggravated the occurrence of landslide disasters. In this study, we used high-precision optical remote sensing images to obtain the landslide inventory triggered by Typhoon “Megi” in 2016. The main purpose of this study is to explore the relationship between LULC and road network on the susceptibility of landslides triggered by typhoons. The innovations include the following: (i) We adopted the newly developed PLUS model to simulate various factors, and utilized historical LULC imagery to forecast the LULC long-term dynamic factors for both 2030 and 2060; concurrently, the road network distribution from 2018 to 2020 was harnessed as a short-term dynamic factor. (ii) In a novel approach, we combined the RF, XGBoost, and LightGBM models through a stacking strategy to delineate the susceptibility relationship between influential factors and landslides. (iii) Furthermore, we assessed the susceptibility distribution of typhoon-induced landslides under diverse human engineering activity scenarios. To our knowledge, this marks the pioneering application of both the PLUS and the stacked RF-XGB-LGBM models in landslide susceptibility research.
5. Discussion
The current research focus involves considering human engineering activity scenarios and evaluating how they alter landslide susceptibility [
63]. Now, ample evidence suggests that landslide susceptibility is dynamic, particularly in the context of rapid urban and road network expansion—a prevalent issue in many developing countries, including China. Especially in the eastern coastal areas of China, policy adjustments have driven a large number of engineering activities [
74,
75]. Firstly, it is necessary to consider the influence of LULC changes on landslides of mountain area. Accurately predicting future LULC scenarios is crucial for improving susceptibility reliability. Therefore, this study used the PLUS model for the first time in landslide research. This model has high simulation and prediction accuracy for larger regions and can better predict the evolution of future land-use patterns [
28,
67]. Many studies [
27,
29,
30,
76] have confirmed that the PLUS model can simulate LULC changes at different scales and scenarios, providing guidance for sustainable urban development. Eleven conditioning parameters were used to predict the 2060 LULC map: the forestland and cropland areas decreased by 28.53 km
2 and 18.07 km
2, respectively, and the built-up area increased by 45.88 km
2. The land-use in this area is mainly transformed from forests to built-up areas. The increase in built-up land is mainly distributed around government residences with relatively flat terrain and frequent socio-economic activities. Numerous studies have assessed the impact of LULC changes on landslide susceptibility [
18,
20,
22,
25]. However, human activities are carried out by road network [
77]. The extended road network may lead to rapid changes in urban land-use. Generally, roads enter the susceptibility model based on their distance from the plotted landslide [
18,
21]. But few researchers have explored the influence of road network changes on landslides. Roads may introduce deviations in the distribution of landslide susceptibility [
20]. Therefore, this article has updated the list of road networks and LULC. It should be noted that many researchers have explored the application of different landslide inventory in land-use change, such as Xiong et al. [
22] used LULC and landslide inventories from 2000 and 2020 and Tyagi et al. [
18] used 218, 243, and 387 landslide events from 2005 to 2010, 2010 to 2015, and 2015 to 2020, respectively. On the contrary, Hao et al. [
21] only used the 2018 landslide inventory for dynamic research. The update landslide inventory inevitably has a higher density and quantity of landslides, making it difficult to explore the impact of LULC on susceptibility alone. Therefore, this study focused solely on the landslides triggered by the 2016 Typhoon event to explore the impact of LULC and road network changes on landslide susceptibility, employing nonlinear relationships constructed by hybrid models.
People are more and more interested in the study of the application of hybrid model in landslide susceptibility analysis, especially the effect of human engineering activities on landslide occurrence [
63]. In the realm of landslide susceptibility analysis, the integration of models offers a potent approach to enhance prediction capabilities. Our choice of employing a hybrid model, specifically combining RF, XGBoost, and LightGBM, was motivated by several factors. Firstly, each of these models has shown superior predictive prowess individually in previous landslide susceptibility studies [
49,
78]. While they are adept on their own, their combination allows us to harness their unique strengths and offset individual weaknesses, offering a more robust and comprehensive modeling approach. Secondly, the stacking strategy offers a layer of meta-learning where predictions from individual models are used as input to train a higher-level model, thereby refining predictions. However, no papers attempted to integration and application. The Stacking strategy has started to be applied to integrate basic models. For instance, Li et al. [
53] adopted a stacking strategy combined with the convolutional neural network and recurrent neural network to explore the distribution of landslide susceptibility in the Three Gorges Reservoir area. Zeng et al. [
34] compared and analyzed the application of different ensemble models and strategies in landslide susceptibility, and found that stacking strategy combined with strong classifiers can maximize the efficiency of improving prediction accuracy. We can find the differences from the factor contribution of ensemble models: the RF model tended to use aspect, distance from road and DEM. The XGBoost model focused more on soil, LULC, hydrology, and distance from road. The LightGBM model generated higher scores for distance from road, DEM, and aspect. A major difference between this study and previous studies is that both RF and LightGBM have low rankings for LULC. This may reflect differences in environmental and/or climatic conditions and LULC categories between this study and previous studies. The results show that distance from road is highly correlated with landslides occurrence. Finally, the stacking RF-XGB-LGBM model achieved the optimal AUC value of 0.915 on the test set. The hybrid model in this study area can be proved to improve the prediction accuracy and generalization ability of landslide susceptibility. We also encourage researchers to attempt the use of more hybrid model. These studies are good attempts to map the susceptibility of landslides in dynamic environments.
The hybrid model explores the road network and LUCL expansion on landslide susceptibility. The density of landslides gradually decreases with the increase in road distance (
Figure 7e). This correlation reflects the richness of roads along the valleys in the study area, and nearby slopes are prone to slide. For the whole research area, the regional susceptibility of road network expansion has changed from 2016 to 2020. In 2018, the moderate and high susceptibility areas increased by 15.94% and 7.04%, respectively. Early large-scale engineering construction increased the probability of landslides. Urban areas have been more affected: from 2016 to 2020, the very low susceptibility decreased by 11.96%, while the low, moderate, and high susceptibility increased by 11.15%, 26.88%, and 17.72%, respectively. Most of the landslides that happened in 2016 occurred in forest areas. This is contrary to many other findings in which forest areas were considered to have a stabilizing ability [
24,
79,
80]. However, Lan et al. [
81] found that the weight of trees may increase the sliding force in parallel directions. The vegetation structure with different root systems can increase or reduce the susceptibility of shallow landslides [
82]. Especially bamboo forests and shrubs are mostly distributed in the research area. The extreme rainfall event in 2016 is more prone to causing a large number of planer slides in forest areas.
We further explored the changes in susceptibility using the predicted LULC maps for 2030 and 2060. For the whole study area, the susceptibility of road networks and LULC to landslides showed significant percentage changes: a decrease of 9.05% for very-low levels, an increase of 5.29% for low levels, a moderate increase of 2.96%, and an increase of 0.79% for high susceptibility. For the urbanized areas in 2015, the expansion of road network and LULC increased the number of landslides with high susceptibility. The urbanization areas in 2030 and 2060 show that LULC has a positive impact on landslides. This may be due to our data lamination of 2020 road network. It is difficult for us to obtain data on future road planning. Therefore, we can conclude that for our research area, road expansion is the main dynamic factor, followed by LULC. The influence of LULC changes on the stability conditions of our study area is positive, but relatively limited. About this point, controversial findings have been reported in the literature. In high mountain areas, the improvement in stability conditions may be significant [
83], while other researchers have found that specific changes in LULC (such as deforestation) may also lead to an increase in regional landslide susceptibility [
80]. However, there is no doubt that land-use changes in urban areas have brought about unstable factors. The current policy of China will continue to lead to a large number of road network and urban regional expansion in many years. We suggest that the government can provide more road network planning to more precisely describe the landslide area. Through policy interpretation, landslide managers can obtain future land information, such as potential landslide risks in future roads and land built-up areas.
Although this study systematically quantifies the long-term impact of urbanization on landslide occurrence, it is still limited by the dataset and methods used. Our finding used a limited inventory induced by extreme rainfall events. A larger inventory of multi-temporal landslides and more detailed information on the occurrence of landslides at different stages of urbanization may help to better quantify the impact of urbanization on susceptibility. This article does not consider the impact of climate and other environmental factors on the stability conditions of future slopes, especially the prediction of Typhoon induced rainfall events. As future work, we plan to include more aspects in the prediction of human engineering activities. In addition, applying current programs in other regions to test the impact of environmental changes will reveal better insights into this topic.
6. Conclusions
This study explores the individual and joint effects of road network and LULC changes on the stability of Typhoon-triggered landslides in the mountainous areas along the southeastern coast of China. The hybrid model simulates the regional landslide susceptibility of seven engineering scenarios. Finally, we compared the results obtained in each scenario with the reference scenario.
From 2010 to 2020, the rapid urbanization of the research area led to a maximum built-up area of 13.433 km2, mainly expanding forestland and cropland, with 8.28 km2 and 5.99 km2, respectively. The novelty PLUS model was used to predict changes in LULC, and the results indicate an increase in built-up land in the future. The increase in built-up land is mainly distributed around government residences in areas with relatively flat terrain and frequent socio-economic activities. The predicted 2060 LULC map shows that forestland and cropland will decrease by 28.53 km2 and 18.07 km2, respectively, while the built-up area will increase by 45.88 km2.
The hybrid model has improved the prediction accuracy and generalization ability of landslide susceptibility. In the training set, the optimal performance model shown in the 10-fold cross-validation was Stacking RF-XGB-LGBM model (mean AUC = 0.908), followed by LightGBM model (mean AUC = 0.9) and XGBoost model (mean AUC = 0.898), and finally RF model (mean AUC = 0.889). The final testing set showed that the Stacking model achieved an excellent AUC value of 0.915.
The prediction model shows that the factor contribution of distance from road was greater than that of LULC. From 2016 to 2020, the prediction results show that the distance from road has decreased by 11.96% from 2016 to 2020, and the low, medium, and high susceptibility have increased by 11.15%, 26.88%, and 17.72%, respectively.
Ultimately, changes in LULC and road network may disrupt the stability of mountainous areas, endanger natural resources, and damage the environment. These results require more rational land-use and road planning in future urbanization processes, and suggest incorporating LULC changes more systematically in disaster assessment to implement preventive measures from the beginning. This adopted method is novel for landslide susceptibility research in this region.