Production and Analysis of a Landslide Susceptibility Map Covering Entire China

Abstract

China, with its complex geology and diverse climate, is highly prone to landslides, endangering public safety and infrastructure. To address disaster prevention needs, this study comprehensively assesses national landslide susceptibility. We divided China into 37 geomorphic districts, diverging from traditional methods. By using a 2018–2022 surface deformation dataset, we introduced a rarely—considered dynamic aspect for more accurate mapping of landslide—prone areas. Nine key environmental factors were carefully considered, including terrain, geology, meteorology, hydrology, seismic activities, and engineering activities. Based on these innovative methods and data, we created a 40 m—resolution landslide susceptibility map (LSM) for the whole country. Our assessment showed high accuracy, with an AUC of 0.927, precision of 0.859, recall of 0.815, F1—score of 0.828 and Matthews correlation coefficient of 0.773. Seven high—risk regions, like the Tianshan Mountains and the southern Tibetan valleys, were analyzed. The study revealed regional differences in landslide occurrences and key influencing factors. The LSM and findings enrich landslide susceptibility theory and offer a valuable resource for engineering, disaster management, and mitigation in China, helping reduce potential landslide losses.

1. Introduction

A landslide, also known as a landslip, is the downward slope of a mass of rock, debris, earth, or soil [1]. Landslides occur when gravitational and other types of shear stress within a slope exceed the shear strength (resistance to shearing) of the materials that form the slope [2]. Landslides are classified according to the type of movement (slides, flows, or rock falls) and the type of material (rock, debris, or earth) [3]. Occasionally, more than one type of movement occurs within a single landslide site.

China is one of the countries that is most affected by landslides [4]. Part of China is located at the junction of continental plates, and the influence of plate movement has resulted in complex and diverse topography, geological features, and climates, as well as strong tectonic and seismic activities [1]. In addition, in recent years, the increasing intensity of human engineering activities such as infrastructure construction, increase in localized heavy rainfall caused by climate change, and frequent occurrence of earthquakes have led to the frequent occurrence of landslide disasters in China, which seriously threaten the safety of people’s lives and property, causing great damage to the ecological environment, thereby hindering the sustainable development of human society. From January 2011 to June 2022, 105,000 geological disasters have occurred in China, of which 102,000 (96.6%) were landslides, resulting in 4841 deaths and economic losses of approximately 50 billion yuan.

Landslide susceptibility is the likelihood of a landslide occurring in an area based on the local terrain conditions [5]. It predicts “where” landslides are likely to occur [6]. Landslide susceptibility assessment (LSA) is widely used as an effective tool for landslide prevention and mitigation [7,8,9].

Landslide susceptibility modeling typically employs classification algorithms to establish empirical relationships between a binary dependent variable (landslide vs. no landslide) and various static environmental factors, such as slope and lithology. The established empirical relationships can then be utilized to estimate relative spatial landslide susceptibility indices and generate a final landslide susceptibility map [10,11,12,13,14,15,16,17,18]. Several researchers have conducted LSA at smaller watershed scales or at the city and county levels, achieving noteworthy results [19,20,21,22].

Many current LSA studies primarily focus on static environmental factors, often neglecting the impact of factors with dynamic characteristics [23,24,25,26]. In recent years, InSAR technology has gained widespread use in displacement monitoring and early identification of disasters [27]. It proves to be an effective method for obtaining large-scale and high-resolution surface deformation information [28,29].

InSAR deformation reflects the motion state of landslides, serving as a dynamic characterization factor in Landslide Susceptibility Analysis. The results of InSAR deformation can address the limitations of traditional LSA in analyzing the dynamic characteristics of landslides. Numerous scholars have incorporated InSAR techniques into LSA, demonstrating improved accuracy in predicting landslide susceptibility when considering InSAR surface deformation results alongside static environmental factors [30,31,32].

Wide-area or national-scale LSA holds the potential to comprehensively and sustainably identify landslide-prone areas on a national scale, fostering enhanced cooperation between local and national decision-makers in responding to major landslide events [33]. Despite the significance of this approach, there are relatively few studies focusing on LSA at the national scale.

Noteworthy examples include Italy [34], India [35], Romania [36], Iran [37], and China [38,39], which have successively completed LSA at the national scale, providing valuable information for land management and territorial planning in their respective countries. Additionally, some researchers have undertaken LSAs in continental Europe [40,41], Central and North America, and the Caribbean [42], contributing to a broader understanding of landslide susceptibility globally.

National—scale LSA encounters more pronounced challenges in contrast to regional—scale LSA. Firstly, spatial heterogeneity is a conspicuous issue in large—scale areas [43]. For instance, the Qinghai—Tibet Plateau and the southwestern coastal regions of China exhibit vast disparities in natural environments, giving rise to divergent landslide distributions and distinct primary controlling factors. Employing a uniform model for landslide susceptibility evaluation across regions with highly variable geomorphologies inevitably compromises the accuracy of the assessment outcomes. Secondly, previous national—scale LSA investigations have predominantly concentrated on static features associated with landslides, while overlooking their dynamic aspects. Although InSAR technology has the capability to capture the dynamic characteristics of landslides, its intricate computational process has deterred its extensive application in national—scale research.

Innovatively, this study designates the entire territory of China as the research target. To tackle the problem of spatial heterogeneity, the Geomorphological Zoning Map of China is introduced, and nine categories of high—precision environmental factors are meticulously extracted. Furthermore, a breakthrough is made by integrating the dynamic characteristics of landslides as manifested in the surface deformation rate maps of China from 2018 to 2022 to construct a national—scale LSM. Through lattice—zoning comparative analyses, the enhancement effect of incorporating dynamic characteristics on the assessment accuracy is elucidated, thereby validating the efficacy of the proposed approach. Ultimately, the characteristics and primary controlling factors of seven typical landslide—prone areas are analyzed in depth, offering robust support for landslide prevention and mitigation endeavors.

2. Materials and Methods

The schematic flow chart of our research is shown in Figure 1.

2.1. Geomorphological Regionalization

Geomorphological regionalization involves comprehensive research on regional geomorphology, including aspects such as surface constituents, internal force processes dominated by neotectonic movements, external force processes reflected in denudation and accretion, and the evolutionary history of geomorphological formation [44].

Landslides are affected by both hazard-containing and -causing factors. China has a vast land area, wide latitudinal span, complex and diverse climate, and different terrain heights; therefore, the spatial distribution patterns and triggering factors of landslides in different regions are different. The geomorphological regionalization scheme considers topographic and geological features, and in the same geomorphological district, it can ensure that landslides have similar hazard-containing environments to a certain extent, which solves the problem of spatial heterogeneity that exists in wide-area LSA. Therefore, it is necessary to achieve geomorphological regionalization prior to the LSA.

Based mainly on the differences in the combinations of secondary basic geomorphological types, such as large-scale mountains, plateaus, tablelands, basins, and plains, which are caused by endogenic forces, Li et al. divided China into 6 geomorphological regions and 37 geomorphological sub-regions [45]. This geomorphic zoning scheme was adopted in this study. The regionalization is shown in Figure 2, and the names of the geomorphological regions and districts are shown in

Table 1. Geomorphological regionalization of China.

Geomorphologic Region	Geomorphologic District	Serial Number	Geomorphologic Region	Geomorphologic District	Serial Number
I. Eastern hilly plains	A. Wandashan Sanjiang Plain	I-A	IV. Northwestern middle and high mountains and Basins	A. Xinjiang Gansu Inner Mongolia Hilly Plains	IV-A
	B. Changbaishan low-middle Mountains	I-B		B. Altai Alpine	IV-B
	C. Ludong low mountains and hills	I-C		C. Junggar Basin	IV-C
	D. Xiaoxing’anling Mountains low-middle Mountains	I-D		D. Tianshan Alpine Basin	IV-D
	E. Songliao plain	I-E		E. Tarim Basin	IV-E
	F. Western Liaoning low-middle Mountains	I-F	V. Southwestern subalpine mountains	A. Qinling–Dabashan subalpine	V-A
	G. North and East China Plains	I-G		B. Hubei Guizhou Yunnan middle Mountain	V-B
	H. Ningzhen Plain Hills	I-H		C. Sichuan basin	V-C
II. Southwestern low-middle mountains	A. Zhejiang and Fujian Low-middle Mountains	II-A		D. Southwest Sichuan and Central Yunnan Alpine basin	V-D
	B. Huaiyang Low Mountain	II-B		E. Southwestern Yunnan Subalpine	V-E
	C. Middle Yangtze River Low mountain Plain	II-C	VI. Tibetan Plateau	A. The Aljinshan and Qilian Mountains	VI-A
	D. South China low mountain Plain	II-D		B. Qaidam Basin	VI-B
	E. Taiwan Plain and Mountain	II-E		C. Kunlun Extremely Alpine	VI-C
III. Central and northern middle mountains and plateau	A. Daxing’anling Mountains low-middle Mountain	III-A		D. Hengduan Mountains high mountain valleys	VI-D
	B. Shanxi middle mountain Basin	III-B		E. Sanjiangyuan high mountain valley	VI-E
	C. Inner Mongolian Plateau	III-C		F. Sanjiangyuan hilly plateau	VI-F
	D. Ordos Plateau and Hetao Plain	III-D		G. Qiangtang Plateau lake basin	VI-G
	E. Loess Plateau	III-E		H. Himalayan Extreme Alpine	VI-H
				I. Lama Kunlun Extremely Alpine	VI-I

.

2.2. Landslide Information

The main prerequisite for completing the LSA is the acquisition of spatial distribution information from historical landslide data [46], which characterizes the distribution of landslides through the acquisition of landslide points [47]. Liu et al. (2013) performed an LSA of China in 2012 utilizing only spatial information on 1200 landslide sites from 1949 to 2011 [38]. Wang et al. (2021) used spatial information from only 452 landslide sites during 2007–2019 [39]. Historical landslide data used by these two researchers were mainly distributed in the southwest and southeast regions of China, excluding regions IV-B and I-D as delineated in this study. Owing to the spatial heterogeneity, in wide-area LSA, there are no sample data in one area, and the results obtained by reasoning from data in other areas are not sufficiently precise.

In this study, 246,593 records of historical geohazard cataloging data, including landslides, ground cracks, ground collapses, and other geohazards in China (except Taiwan) from 2000 to 2020, were collected through the Geo-Remote Sensing Ecological Network scientific data registration and publication system (www.gisrs.cn, accessed on 5 July 2023). Global historical landslide data during 2000 to2021 were collected from the official website of the NASA Landslide Viewer (https://maps.nccs.nasa.gov/arcgis/apps/MapAndAppGallery/index.html?appid=574f26408683485799d02e857e5d9521, accessed on 5 July 2023), with a total of 644 data points pertaining to China, which also provided historical landslide data for the Taiwan Province as supplementary data. Only landslides were considered in the assessment; ultimately, 219,351 records were retained to provide a database for performing the LSA of China. The distribution of landslides in China is shown in Figure 3.

2.3. Dynamic Features of Landslides

The dynamic features of landslides are reflected by ground deformation [48]. Ground deformation is the most intuitive manifestation of landslides and the preferred indicator for landslide identification and assessment [49,50,51]. InSAR technology can monitor dynamic changes in the ground surface during the period covered by the data used. In recent years, InSAR technology has developed into a powerful space geodesy technology for landslide monitoring and has been widely used in the fields of LSA, early landslide identification, deformation monitoring, and causation analysis [52,53].

The State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing of Wuhan University used Sentinel data to produce a map of the surface deformation rate in China for a 5-year period from 2018 to 2022 (Figure 4) [54], which was used in this study as a dynamic feature in the LSA.

The effective coverage of the ground deformation rate map of China was 94.6%, and the percentage of incoherent areas was 5.4%. For the regions with incoherence, we set their values to NoData. In this way, during the LSA process, only the influence of environmental factors is taken into account for the incoherent regions.

2.4. Landslide-Related Environmental Factors

Environmental factors can reflect the causative factors of landslides and characterize their static features. In this study, nine environmental factors, namely elevation, slope, aspect, lithology, distance to fault, distance to stream, average annual precipitation, distance to epicenter, and distance to road, were selected as data inputs for the LSA based on the quality and accessibility of relevant data. To ensure the evaluation results at 40 m spatial resolution, we resampled all environmental factor layers to that resolution.

The basic information of the data used in this study is shown in

Table 2. Data information.

Primary Data	Spatial Resolution	Source	Derived Map
DEM	30 m	Geospatial Cloud website	Figure 5, Figure 6 and Figure 7
Geological map	1:250,000	Geoscientific Data and Discovery Publishing System	Figure 8 and Figure 9
Stream data	/	OpenStreetMap	Figure 10
Precipitation data	/	China Meteorological Data Network	Figure 11
earthquake data	/	China Seismic Network	Figure 12
Road data	/	OpenStreetMap	Figure 13

.

2.4.1. Elevation

Elevation controls the stress, temperature, and vegetation cover of a slope and influences landslide development [55]. In this study, China’s 30-m resolution DEM data were collected through the Geospatial Cloud website (https://www.gscloud.cn/, accessed on 8 July 2023), which was used to extract elevation factors. Figure 5 shows the elevation map of China.

2.4.2. Slope

The slope is a major factor in the occurrence of landslides [56,57]. The slope provides potential energy; when the slope increases, the potential energy obtained by the gravity of the slope body also increases. When the slope body is in an unstable state, the potential energy gradually becomes kinetic energy, which then induces landslides. The slope factor was extracted from DEM data. Figure 6 shows the slope map of China.

2.4.3. Aspect

The aspect affects the climatic conditions of slopes, the intensity of solar radiation, vegetation cover, evapotranspiration, and the degree of slope erosion. These factors affect the degree of joints and fissures and determine the stability of slopes. The aspect factor was extracted from DEM data. Figure 7 shows the aspect map of China.

2.4.4. Lithology

The lithology provides an environment conducive to landslides, and all types of rocks and soil may constitute landslides [58,59]. For the same slope shape (height and angle), the looser the structure, the lower the shear strength and weathering resistance and the more prone it is to changes in its properties under the action of water. Landslides are less likely to occur in hard rock formations.

In previous wide-area LSA studies, the geological maps that were used to extract lithological and tectonic features were of small scale and low precision, and did not adequately demonstrate local conditions. In this study, 729 geological maps of China at the scales of 1:250,000 were collected through the Geoscientific Data and Discovery Publishing System of the China Geological Survey (http://dcc.ngac.org.cn/, accessed on 8 July 2023). Geological maps contain many elements, and only lithological and tectonic information is required for the LSA process. Each map was produced by a different geological team and the integration process revealed the existence of the same name in the splices. We analyzed the lithology in the geological map legend, integrated and classified all the geological maps according to the names of the rocks, and divided the rocks in China into 54 categories (as shown in Figure 8). Finally, the vector files from the geological map were converted into raster files and resampled to a spatial resolution of 40 m (Figure 8), intended for subsequent use in LSA.

2.4.5. Distance to Faults

Fault activity can create a destructive surface for slope instability and directly control the form and scale of slope deformation and destruction. Landslides are significantly more frequent in areas with a high distribution of faults, where slope stability is poor, and hazardous rock bodies are densely developed.

Fault distribution features were extracted from geological maps, and the faults in 729 geological maps at the scale of 1:250,000 were extracted. The influence of faults on landslides is related to the distance, which forms the distance to the fault map, as shown in Figure 9.

2.4.6. Distance to Streams

The scouring action of water induces landslides. The scouring effect of water refers to the river, ditch, lake, and seawater currents scouring the slope, eroding the foot of the slope, weakening the support, and making the slope higher, steeper, and unstable. The scouring effect of water is reflected in the distribution of streams. By downloading the Chinese stream data through OpenStreetMap (OSM) (https://www.openstreetmap.org/, accessed on 8 July 2023), information on the distribution of streams in China was extracted, and a distance-to-stream map was formed, as shown in Figure 10.

2.4.7. Average Annual Precipitation

Rainfall is a key factor in inducing landslides. When rainfall is high, the rain penetrates the rock and soil layer on the surface of the slope, which leads to an increase in the water content of the rock and soil layer until it is saturated. When rainfall continues to increase, it leads to an increase in the water content in the water-barrier layer in the lower part of the slope, which in turn increases the weight of the slope, resulting in landslide disasters. Average annual precipitation expresses the intensity of rainfall [60]. Data on precipitation obtained from 2400 meteorological stations nationwide from 2000 to 2020 were obtained from the China Meteorological Data Network (https://data.cma.cn/, accessed on 8 July 2023). The monthly average precipitation was computed through interpolation to derive the mean monthly precipitation, subsequently resulting in the creation of the average annual precipitation map. The map was then resampled to a spatial resolution of 40 m, as illustrated in Figure 11.

2.4.8. Distance to Epicenter

Earthquakes have a significant impact on landslides, triggering uncoordinated movements of slopes and slides along the slip surface, and making slopes unstable [61]. Following a major earthquake, which is typically succeeded by a series of aftershocks, geotechnical structures or materials initially endure without damage; however, with each subsequent aftershock of comparable intensity, these geotechnical elements progressively weaken and may eventually fail.

In this study, earthquake data for China from 1950 to 2021, totaling 6331 records, were obtained from the China Seismic Network (http://www.ceic.ac.cn/history, accessed on 8 July 2023). The earthquake sources were projected onto the map, and the distance to the epicenter map was formed, as shown in Figure 12.

2.4.9. Distance to Roads

Engineering construction can affect rock stability. In the process of slope excavation, landslides caused by engineering construction often damage the original slope, and these should be appropriately represented in terms of their distribution and scale for comparison with the landslides caused by natural factors. Most landslides induced by engineering construction are mainly caused by changing the shape of slopes or allowing water to enter the underlying material, especially in mountainous areas where road construction tends to destabilize slopes; the impact of engineering construction in this study is reflected by the road network.

China’s road network data were downloaded from OSM (https://www.openstreetmap.org/, accessed on 8 July 2023). The distribution of China’s road network was extracted to form the distance to road map, as shown in Figure 13.

2.5. LSA Approach

Existing approaches to LSA are based on statistical analysis and mathematical modeling methods [62]. Statistical analysis methods include information quantity, hierarchical analysis and weight of evidence methods [63]. Mathematical modeling methods include random forest (RF) [64,65], artificial neural network (ANN) [66], support vector machine (SVM) [67] methods, and so on.

Each method has its advantages and disadvantages. The statistical analysis method can intuitively reflect the role of each assessment factor in the calculation process, but it is easily affected by subjective factors; therefore, it is suitable for regions with a single environment and with unchanging assessment factors over the long term. The mathematical model method is adaptive, self-learning, and has a strong nonlinear mapping ability; however, it is limited by the quality of sample data, and it is suitable for regions with complex environments consisting of nonlinear assessment factors.

Many researchers have compared the advantages and disadvantages of different LSA methods and have proven that RF is the most suitable [68,69,70]. Therefore, in this study, the RF method was selected to build a model to complete the assessment.

RF is an integrated machine learning method that utilizes random resampling techniques and random splitting of nodes to construct multiple decision trees. The final classification results are obtained through voting [71]. The steps for constructing an RF model are as follows: (1) historical landslides and non-landslides were used as the sample set, from which n samples were generated using random sampling; (2) assessment factors were used as features, and k of them were selected for n samples to build a decision tree; (3) the process was repeated m times to produce m decision trees; and (4) the majority voting mechanism made predictions.

Each tree has the same distribution, and the classification error depends on the classification ability of each tree and the correlations between them. The error in the out-of-bag data is an unbiased estimation that can be used to verify the performance of the model to prevent overfitting.

In this study, we used the scikit-learn package in a Python 3.8 environment to implement RF algorithm training, parameter tuning, and prediction. The package integrates a range of typical machine learning algorithms, including RF, SVM, and logistic regression.

3. Results

3.1. LSM

Considering the resolution and scale of multisource data collected in this study, a grid cell with a side length of 40 m was used as the spatial scale for the LSM in China. Certain positive and negative sample data were selected as the dataset, where the positive sample data were landslides that occurred. A total of 219,351 historical landslides were selected covering an equal number of raster cells. The selection rule for negative sample data was as follows: (1) establishment of buffer zones on landslides that have occurred, with a buffer zone distance of 1 km, (2) extraction of the area outside the overlap between the landslide site buffer and study area, and (3) For each geomorphic district, we select the same number of negative samples as positive samples in that geomorphic district to ensure the balance of positive and negative samples in each geomorphic district. The total number of positive and negative samples was 438,702, and the dataset was classified into training and test data at a ratio of 7:3. Training data were used to generate the model, and the test dataset was used to test the accuracy of the model.

Because the main controlling factors of landslides vary in different geomorphological districts, it is necessary to complete geomorphological regionalization before carrying out LSA in China. Adopting China’s geomorphological regionalization scheme, the country was classified into 37 geomorphological districts. Considering the dynamic characteristics of landslides and environmental factors, InSAR surface deformation, elevation, slope, aspect, lithology, distance to faults, distance to streams, average annual precipitation, distance to epicenter, and distance to roads were used as assessment factors. Utilizing the RF, the landslide susceptibility model of each geomorphological district was constructed individually. The probability value of landslide occurrence within each raster cell was obtained by modeling, and the value was considered the landslide susceptibility index with a value domain of [0,1]. The landslide susceptibility index was classified into five intervals, including 0.0–0.2, 0.2–0.4, 0.4–0.6, 0.6–0.8 and 0.8–1 representing very low, low, medium, high, and very high susceptibility zones, respectively, which were stitched together to form the Chinese LSM, as shown in Figure 14.

3.2. Map Validation

In this study, receiver operating characteristic (ROC) curves and area under the ROC (AUC) values were used to reflect the accuracy of the model [72]. Additionally, a comprehensive evaluation was conducted using Precision, Recall, F1-score and Matthews correlation coefficient (MCC). The confusion matrix is shown in

Table 3. Confusion matrix.

Ground Truth	Predicted as Positive Sample	Predicted as Negative Sample
Positive sample	Ture Positive (TP)	False Negative (FN)
Negative sample	False Positive (FP)	True Negative (TN)

.

The ROC curve was plotted as a point pair of coordinates formed by the false-positive rate (FPR) as the horizontal coordinate and the true-positive rate (TPR).

$$FPR={\displaystyle \frac{FP}{FP+TN}}$$

$$TPR={\displaystyle \frac{TP}{TP+FN}}$$

Precision, recall, F1-score and MCC, as common metrics in machine learning, also demonstrate the accuracy of LSA. The calculation formulas are as follows:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

$$Recall=TPR$$

$$F1=2·{\displaystyle \frac{Precision·Recall}{Precision+Recall}}$$

$$MCC={\displaystyle \frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}}$$

As China is classified into 37 geomorphological districts in this study, 37 models were formed in the LSA, and each geomorphological district will have an ROC curve and AUC value; therefore, the AUC value of all geomorphological districts were averaged as the accuracy of the LSM in China. The ROC curves of each geomorphological region were placed into the same coordinate system. The ROC curves of each geomorphological district in different geomorphological regions are shown in Figure 15a–f. The AUC, precision, recall, and F1-score for different geomorphic zones are presented in

Table 4. Evaluation indicators of geomorphologic districts.

Geomorphologic Districts	AUC	Precision	Recall	F1-Score	MCC	Geomorphologic Districts	AUC	Precision	Recall	F1-Score	MCC
I-A	0.948	0.868	0.865	0.865	0.796	IV-A	0.964	0.940	0.780	0.838	0.816
I-B	0.917	0.830	0.830	0.830	0.692	IV-B	0.928	0.845	0.787	0.805	0.805
I-C	0.930	0.847	0.847	0.847	0.747	IV-C	0.961	0.879	0.843	0.857	0.847
I-D	0.946	0.908	0.771	0.812	0.704	IV-D	0.921	0.832	0.826	0.825	0.792
I-E	0.972	0.916	0.832	0.863	0.837	IV-E	0.981	0.938	0.800	0.851	0.862
I-F	0.881	0.783	0.752	0.753	0.652	V-A	0.878	0.792	0.792	0.792	0.683
I-G	0.973	0.899	0.883	0.890	0.843	V-B	0.793	0.702	0.708	0.707	0.631
I-H	0.972	0.912	0.912	0.912	0.826	V-C	0.933	0.891	0.876	0.878	0.811
II-A	0.852	0.803	0.774	0.778	0.678	V-D	0.846	0.751	0.747	0.744	0.648
II-B	0.911	0.847	0.837	0.834	0.713	V-E	0.835	0.741	0.683	0.684	0.632
II-C	0.860	0.783	0.774	0.777	0.658	VI-A	0.968	0.907	0.848	0.871	0.844
II-D	0.839	0.757	0.751	0.753	0.663	VI-B	0.978	0.926	0.919	0.922	0.853
II-E	0.975	0.910	0.884	0.886	0.875	VI-C	0.976	0.932	0.874	0.897	0.848
III-A	0.936	0.860	0.788	0.806	0.802	VI-D	0.956	0.889	0.881	0.883	0.817
III-B	0.857	0.765	0.768	0.764	0.704	VI-E	0.964	0.898	0.867	0.880	0.820
III-C	0.975	0.932	0.718	0.782	0.864	VI-F	0.972	0.915	0.829	0.861	0.841
III-D	0.960	0.895	0.883	0.888	0.831	VI-G	0.961	0.930	0.730	0.788	0.823
III-E	0.817	0.724	0.712	0.703	0.622	VI-H	0.963	0.900	0.893	0.896	0.850
						VI-I	0.988	0.952	0.914	0.932	0.881

. The AUC, prevision, recall, F1-score and MCC of LSA in China are 0.927, 0.859, 0.815, 0.828 and 0.773, respectively, indicating high accuracy.

It can be seen that the accuracy of the landslide susceptibility model is different in each geomorphological region. The overall accuracy of landslide susceptibility is the highest in region VI (Tibetan Plateau), with average AUC, precision, recall, F1-score and MCC of 0.969, 0.916, 0.861, 0.881 and 0.842, respectively, and the accuracy of the assessment in region V (Southwest Asia Alpine Mesopotamia) is the lowest with average AUC, precision, recall, F1-score and MCC of 0.857, 0.775, 0.761, 0.761 and 0.681, respectively.

3.3. Spatial Rationality Analysis

The AUC value reflects the performance of the model but does not reflect whether the spatial distribution of the LSA results is reasonable [73]. To verify the spatial rationality of the LSM, we have statistically analyzed the area of highly susceptible regions under each susceptibility level, the proportion of the area of highly susceptible regions, the number of landslide points, and the density of landslide points, as shown in

Table 5. Statistics of LSM.

Susceptibility	Area (million km2)	Percentage	Recorded Landslide	Landslide Density (nos./km2)
Very low	5.17	53.9%	11626	0.002
Low	1.76	18.3%	17767	0.01
Moderate	1.63	17%	22154	0.013
High	0.73	7.6%	35315	0.048
Very high	0.31	3.2%	132489	0.427

. We found that the area and its proportion of regions with susceptibility levels increasing from low to high gradually decrease. The transition between regions of each susceptibility level is gentle, indicating a reasonable spatial distribution. Meanwhile, it is obtained that the density of landslide points that have occurred in the extremely low susceptibility regions is 0.002 per square kilometer, and the density of landslide points that have occurred in the extremely high susceptibility regions reaches 0.427 per square kilometer. This demonstrates that the spatial locations of historical disaster points are consistent with the distribution of the extremely high susceptibility regions in the LSM.

3.4. Comparison with LSM of CIGEM

In 2020, the China Institute of Geological and Environmental Monitoring (CIGEM) released an LSM for China at a scale of 1:5 million, as shown in Figure 16a. Because there are only four susceptibility classes for LSM published by the CIGEM, namely, very low, low, medium, and high susceptibility, the very high and high susceptibility levels were combined in our results (Figure 16b). Differences in the area and spatial distribution of susceptibility classes were found between the two results (Figure 16c), which delineated very low-susceptibility areas (53.9% versus 23.1%) and fewer high-vulnerability areas (10.8% versus 13.6%) compared with the CIGEM results. The distribution of areas of the various susceptibility classes was more refined. The reason for this difference is the spatial resolution. The result of the CIGEM is at a scale of 1:5 million, and the different levels of susceptibility are continuously distributed, which makes the result more suitable for macroscopic presentation. Our results have a spatial resolution of 40 m, which is suitable for mapping at a scale of 1:250,000 or larger, allowing a more accurate distinction between various susceptibility levels and providing data for engineering and construction.

4. Discussion

4.1. Importance of Assessment Factors

The importance of the LSA factors in each geomorphological district can be obtained using the RF model, which can be represented in the form of radar charts, as shown in Figure 17. In the assessment of the 37 geomorphological districts, 27 geomorphological districts showed the highest importance for elevation or slope, and the other geomorphological zones also had high scores for these factors. The average importance values of elevation and slope were 0.243 and 0.198, respectively; thus, topographic conditions had the highest degree of correlation with landslide development and played a dominant role in the assessment process. The average annual precipitation had the second-highest degree of importance, and in some mountain-dominated areas, it was the main factor that induces landslides. It is worth mentioning that the highest importance of mean annual precipitation is 0.239 in the LSA of Loess Plateau (III-E), and the distance to streams and distance to roads are also highly important in this region. Therefore, precipitation and human engineering activities are the most active triggering factors in the study of landslides on the Loess Plateau [74]. The remaining factors had different importance values in the assessment process, with distance to roads being the most important in districts VI-E and VI-F, distance to earthquake sources being the most important in district II-E, and lithology being the most important in district I-G.

Although, the introduced dynamic feature, InSAR surface deformation does not have an outstanding importance, its value is higher in some geomorphological districts. Its contribution value is >0.1 in districts IV-D, V-A, and V-B. Some large deformation fields in these districts are directly related to landslides, and the degree of contribution of surface deformation is lower in some plain-dominated geomorphological districts.

4.2. Comparison with LSM Without Geomorphologic Regionalization

To compare the effect of LSA through geomorphological regionalization in China, the geomorphological region V was selected and the same assessment factors were applied to produce LSM, as shown in Figure 18. The accuracy of the results was measured using the AUC values. The accuracies of the LSMs without and with geomorphological regionalization were 0.806 and 0.857, respectively, and the accuracy improved by 6.3% with geomorphological regionalization.

To compare the rationality of the spatial distribution of the two LSMs, the validated landslide points in geomorphological region V were projected onto the LSMs. There were 64,811 landslides recorded in geomorphological region V, of which the number of validated points was 19,443. The area and proportion of each susceptibility class in the LSM without geomorphological regionalization and the LSM based on geomorphological regionalization, as well as the number of validated landslides and the landslide density in the subregion, were counted separately, as shown in

Table 6. Statistics of LSMs generated based on geomorphological regionalization.

LSM	AUC	Susceptibility	Area (thousand km2)	Percentage	Recorded Landslide	Landslide Density
LSM without geomorphological regionalization	0.806	Very low	105.8	8.8%	410	3.9
		Low	274.6	23.0%	1889	6.9
		Moderate	358.9	30.0%	3813	10.2
		High	241.0	20.1%	4247	17.6
		Very high	216.1	18.1%	9084	42.0
LSM based on geomorphological regionalization	0.857	Very low	348.1	29.1%	1588	4.6
		Low	182.1	15.2%	1714	9.4
		Moderate	253.2	21.2%	3269	12.9
		High	232.3	19.4%	4470	19.2
		Very high	180.7	15.1%	8402	46.5

, These statistics shows the reasonableness of the LSM results from real data, and can intuitively show whether or not to use geomorphological regionalization LSM changes in the area of each susceptibility class.

Based on the statistical results, the density of the verified historical landslides increased with an increase in susceptibility class. However, the spatial distribution of susceptibility classes in the LSM without geomorphological regionalization was not reasonable, with the areas of very low, very high, and medium susceptibility zones accounting for 8.8, 18.1, and 30%, respectively. This indicates that the LSM without geomorphological regionalization has a low accuracy and fails to extract very-low and very high-susceptibility zones. In comparison, the area distribution of each susceptibility class in the LSM based on geomorphological regionalization is reasonable, and the percentage of area decreases with an increase in the susceptibility class, in which the areas of high and very high susceptibility zones are smaller than those of the LSM without geomorphological regionalization, while the density of verified landslides is higher.

To further verify the reliability of the results based on geomorphological regionalization, two regions, R1 and R2, in geomorphological region V were selected for comparison, and the locations of R1 and R2 are labelled in Figure 18b,c, respectively. Figure 19 shows the results of the remotely sensed optical imagery and landslide susceptibility assessment in areas R1 and R2. R1 is located in the central part of the Sichuan Basin and passes through the region; the geomorphology is a river alluvial plain, which does not have an environment for landslides. Results based on geomorphological regionalization have considered slope as the main controlling factor of landslides, so that all the areas where there are topographical undulations are classified at very high susceptibility levels (Figure 19a). With geomorphological regionalization, R1 can be more rationally extracted from high-susceptibility areas, reducing the false alarm rate in the assessment (Figure 19b). R2 is located in the central part of the Yunnan–Guizhou Plateau, which is a typical V-shaped river valley affected by river erosion. The terrain on both sides of the river is steep, and the very high susceptibility class of the LSM without geomorphological regionalization is observed in R2 (Figure 19d), which is centrally distributed in a certain range of buffer zones. The distance to streams dominates the susceptibility to landslides. Based on geomorphological regionalization (Figure 19e), the very high susceptibility classes were unevenly distributed on both sides of the river, and the distribution was highly correlated with topography and geological conditions. A comparison of the results of R1 and R2 show that the LSM based on geomorphological regionalization was more in line with the actual situation.

4.3. Comparison with LSM Using Environmental Factors

To compare the effect of introducing dynamic features on LSA, the geomorphological districts IV-D and I-C with higher and lower importance of the InSAR deformation factor, respectively, were selected. The LSA of these two geomorphological zones were completed using only the environmental factor and introducing dynamic features, as shown in Figure 20.

The AUC values for each LSM were compared, and the area and percentage of each susceptibility class, as well as the number of validation points and density of landslides in each subarea, are shown in

Table 7. Statistics of LSMs generated using different methods.

Geomorphologic District	LSM	AUC	Susceptibility	Area (thousand km2)	Area (%)	Number of Recorded Landslides	Landslide Density
IV-D	LSM using environmental factors	0.902	Very low	86.5	37.8	30	0.3
			Low	51.8	22.6	48	0.9
			Moderate	42.8	18.7	123	2.9
			High	29.8	13.0	224	7.5
			Very high	18.2	7.9	466	25.6
	LSM with dynamic features	0.921	Very low	91.2	39.8	26	0.3
			Low	44	19.2	33	0.8
			Moderate	43.8	19.1	104	2.4
			High	30.8	13.4	209	6.8
			Very high	19.4	8.5	519	26.8
I-C	LSM using environmental factors	0.924	Very low	43.6	43.8	33	0.8
			Low	15.4	15.5	14	0.9
			Moderate	15	15.1	63	4.2
			High	12.8	12.9	109	8.5
			Very high	12.6	12.7	281	22.3
	LSM with dynamic features	0.930	Very low	44.1	44.4	33	0.7
			Low	14	14.1	13	0.9
			Moderate	13.3	13.4	41	3.1
			High	14.1	14.2	99	7.0
			Very high	13.8	13.9	314	22.8

. There were 891 IV-D validation points and 500 I-C validation points. When the dynamic feature factor is introduced, the AUC value of the IV-D increased by 0.019, the area of very high susceptibility class increased from 7.9 to 8.5%, and the density of validation landslides increased from 25.6 to 26.8. Similarly, the AUC value of the I-C increased by 0.006, the area of very high susceptibility class increased from 12.7 to 13.9%, and the density of validation landslides increased from 22.3 to 22.8; After the introduction of dynamic features, the LSM prediction effects were all improved to different degrees. As the importance of the InSAR deformation factor in IV-D was higher than that of I-C, the enhancement effect of the LSM after the introduction of dynamic features in IV-D was better than that of I-C. The accuracy of the LSM can be further improved by introducing dynamic features.

To further validate the reliability of the LSA when dynamic features were introduced, the typical landslide-prone high-risk areas were selected from zones IV-D and I-C, respectively. Figure 21 shows the optical remote sensing imagery, LSM, and InSAR deformation for regions L1 and L2 (labelled in Figure 20). The InSAR time series results (Figure 21e,j) show that during the 5-year period, the cumulative deformation of L1 reached 35 cm and that of L2 reached 50 cm, and both landslide risk zones were in the stage of rapid creep. However, without considering InSAR deformation, most of the areas within these two landslide risk zones were categorized as high susceptibility zones (Figure 21b,g), and their landslide susceptibility was underestimated. With the introduction of InSAR deformation (Figure 21c,h), many of the areas in L1 and L2 were correctly classified as very high susceptibility zones.

In summary, the dynamic characterization factor of landslides can highlight areas of potential hazards that are not captured by environmental factors and are of great significance in evaluation.

4.4. Typical High Landslide Susceptibility Zones in China

By analyzing the LSM throughout China, we found that landslide susceptibility areas are distributed in the Tianshan Mountains, southern Tibetan Valley, Western Sichuan Basin, Qinling–Dabashan Mountains, mountainous areas of west-central Yunnan, mountainous areas of southern Zhejiang, middle reaches of the Yangtze River, middle part of the Qilian Mountains, Taihang Mountain, Liaodong Peninsula, and mountainous areas of central Shandong. We chose seven typical areas for analysis, and describe them in Section 4.4.1, Section 4.4.2, Section 4.4.3, Section 4.4.4, Section 4.4.5, Section 4.4.6 and Section 4.4.7.

4.4.1. Tianshan Mountains

The Tianshan Mountains are located in the northern part of Xinjiang, China (Figure 22a). The landslide susceptibility of the Tianshan Mountains region is affected by a combination of topographic, geological, and meteorological factors, with extremely high landslide susceptibility zones distributed in a belt-like pattern (Figure 22b). The northern slopes of the Tianshan Mountain range are rich in precipitation, whereas the southern slopes are affected by a desert climate and receive less precipitation, making the northern slopes more prone to landslides than the southern slopes. The tectonic activity is strong in the Tianshan Mountains with more faults because of the collision between the Indian Plate with the Asian and European Plate. Tianshan north and south have reentered the state of extrusion, which is the main manifestation of the thrust-nappe structure and undergo slip movement and differential uplift. In addition, the lithology of sedimentary clastic rocks and Quaternary slope deposits are soft and facilitate slip. Highly susceptible areas are strongly related to the complexity of the geological conditions.

On both sides of the Ili River Valley in the Tianshan Mountains, we observed a concentrated distribution of landslides, and >80% of which were loess-type soil landslides. The Ili River Valley is rich in pasture and has a developed animal husbandry industry, and the grazing activities are mainly concentrated in the middle and lower mountainous areas from May to September each year. The destruction of vegetation caused by overgrazing and the excavation of herbs in the region has a greater impact on the stability of slopes, which, combined with rainfall and snowmelt, are highly susceptible to landslides.

Figure 22c shows a remote sensing image of a typical landslide in the Ili Valley area of the Tianshan Mountains, located on the bank of the Tekes River, where the substrate is loess with a loose structure, low shear strength, and weathering resistance. From February to March each year, the soil structure is damaged due to freezing and thawing, which results in the gradual development of sliding surfaces. The sliding surface develops gradually, and the slope body is infiltrated by snowmelt and rainfall in June and September, which results in the wetting and creeping of the loess, which occurs when the stability of the body of the landslide reaches the critical threshold value. Figure 22d shows the InSAR time series deformation results from 2018 to 2022 for the location of point A of this landslide, from June to September every year. The surface of the landslide at this location shows a decreasing trend, while the surface is stable in other months, and the correlation between the surface subsidence trend and precipitation during the flood season is relatively high. Thus, it is inferred that water action is the main triggering factor of landslides at this location. And this area is located in geomorphological region IV-D, in Section 4.1 we analyzed the importance of evaluation factors in different geomorphological zones. The main controlling factors of landslide susceptibility in IV-D region are elevation, slope, distance to the water system, and InSAR surface deformation, and hence, topographic conditions and river erosion play a role in the development of landslides in this region. Figure 22e shows the LSM of the area around this landslide, which is classified as a very high susceptibility zone, while some of the surrounding areas, especially the steep slopes along the river, have the same hazard-bearing environment as that of the landslide at this location and have the same potential risk of landslides.

4.4.2. Southern Tibetan Valley

The Southern Tibetan Valley is located in the southern part of the Tibetan Plateau in China (Figure 23a), which is high in the northwest and low in the southeast, bounded in the north by the Nyingchi–Tanggula Mountains, and in the south by the Himalayas, with the Yarlung Tsangpo River flowing through the area that consists of a well-developed water system and numerous tributaries within the region. The combination of the region’s deep rivers, steep valley slopes, fragile geological and environmental conditions, such as complex geological formations and broken slopes, and abundant rainfall, ice, and snowmelt water infiltration have resulted in a wide distribution and variety of landslides in the region. This area is located in the Himalayan mountain region, and the main factors controlling landslide susceptibility in this geomorphological region are elevation, distance to faults, distance to streams, and distance to roads. The development of landslides is highly correlated with the linear evaluation factors, and hence, very high landslide susceptibility areas are mainly distributed linearly (Figure 23b).

Figure 23c shows the remote sensing image of the landslide at the tributary of the Yarlung Tsangpo River, which is located next to the river and highway, comprising granite with a slope of 33–42°. Although the rock is hard, the high and steep slopes of the mountain provide potential energy for the rock body, and the influence of prolonged precipitation facilitates rockslides along the jointing surface. Figure 23d shows the results of the InSAR time series deformation from 2018 to 2022 at point B. Because the landslide has no obvious surface deformation characteristics during this period, the development of the landslide is mainly affected by environmental characteristics such as steep slope, effects of river scouring, and road construction, resulting in the deterioration of stability at point B. Moreover, the Yarlung Zangbo River Valley usually has a high water table and high soil water content, and in the rainy season, the groundwater table increases further, increasing the weight of earth and rock bodies, triggering landslides. Figure 23e shows the LSM for the area around point B. The landslide occurred along the river in a very-high susceptibility zone.

4.4.3. Western Sichuan Basin

The Western Sichuan Basin is located on the eastern edge of the Tibetan Plateau (Figure 24a). This geomorphic evolution is closely tied to the region’s complex and diverse geological history, with a particular emphasis on the influence of the Himalayan movement, especially during the Cenozoic era. The topography and geomorphology of the Western Sichuan Basin have been significantly shaped by the dynamic external forces acting on the Earth’s crust. These forces have led to the creation of striking features, such as valley incisions and deep canyons, in the local landscape. This geomorphic evolution is closely tied to the region’s complex and diverse geological history, with a particular emphasis on the influence of the Himalayan movement, especially during the Cenozoic era. The Western Sichuan Basin is one of the most earthquake-prone regions in China, with three earthquakes of magnitude 7.0 and above occurring since 2008, all of which have triggered a large number of geological hazards, resulting in numerous casualties and property damage. Our LSM show that the very high landslide susceptibility zones in the region are mainly distributed in a linear pattern along the river (Figure 24b).

On 12 May 2008, an earthquake of magnitude 8 struck Wenchuan, China. Figure 24c shows a remote sensing optical image of a mountainous area after the earthquake. The area, situated in Sichuan Province east of the Minjiang River’s source, is characterized by a steep and varied terrain with an average slope of 35°. The predominant lithology consists of mudstone and soft debris-dominated sandstone, creating an environment conducive to landslide development. This, in turn, threatens the stability of the region’s rock and soil structures, leading to the formation of landslide clusters. The source of the earthquake was in the southwest direction at a distance of 78 km. A total of 10 earthquakes with magnitudes >5 occurred within the 300-km buffer zone centered on point C between January 2018 and December 2022. Frequent earthquake damage made it impossible to use InSAR technology to measure surface deformation, and InSAR time-series data for this region are unavailable. Figure 24d shows the LSM around point C, which is located in geomorphological district V-E. In the analysis in Section 4.1, we found that the distance to epicenter is a major controlling factor for landslide susceptibility in the V-E geomorphological district. Because of the close proximity to the source of the earthquake, the entire mountainous area is classified as above the medium susceptibility class and is at a higher risk of landslides.

4.4.4. Qinling–Dabashan Mountains

The Qinling–Dabashan Mountains are located in the mountainous region of southern Shaanxi (Figure 25a), exhibiting two mountains sandwiched by a river, with wide differences in topographic elevation and dense gullies and valleys. The lithology is dominated by weak metamorphic rocks, such as slate, kyanite, and schist, which are loose, jointed, and strongly weathered, providing materials that can cause the occurrence of landslides. The complex and variable climatic conditions in the region and the concentration of rainfall, particularly heavy and torrential rainfall in summer, can exacerbate soil erosion and trigger landslides. Landslides triggered by anthropogenic factors outnumber those induced by natural factors in this region because of intense economic developmental activities and construction of infrastructure such as roads, railways, and hydroelectric power stations, leading to damages to the ecological environment at varying degrees. Areas with high susceptibility to landslides are located in areas with high slopes and many access roads (Figure 25b).

We observed that landslides are more often found on the sides of roads, and that when roads are built, actions such as excavation at the foot of slopes and accumulation of waste soil destabilize the lower parts of slopes that trigger landslides. Figure 25c shows a typical landslide optical remote sensing image at a tributary of the Han River, adjoining a highway. It is assumed that the slope is destabilized owing to the combined effect of river erosion and highway construction, which leads to the occurrence of landslides. Figure 25d shows the InSAR time series deformation that occurred from 2018 to 2022 at point D of the landslide. During the 5-year period, the cumulative deformation at point D has reached 40 cm, the ground is in a downward trend, and there is a possibility of a recurring landslide. The LSM (Figure 25e) classifies both sides of the highway as highly susceptible to landslides.

4.4.5. Mountainous Areas of West-Central Yunnan

Mountainous areas of west-central Yunnan are located at the southeastern edge of the Tibetan Plateau and adjoining the Yunnan–Guizhou Plateau (Figure 26a), and is a typical low-latitude and high-altitude inland area. It predominantly has an elevation of >1000 m, with a high relief. The highest elevation is >5000 m, and the lowest is <500 m. The topography is complex, prone to changes, and changes in the ecological environment with elevation are obvious. The region is located at the junction of the Asian and Indian Plates, and consists of the most complex geological structures in the Asian continent, mainly fractured at a relatively large scale. The region is located in the transition zone between the Mediterranean and Himalayan and Pacific seismic belts, and is subject to frequent seismic activity. Landslide development is widespread owing to the complex natural environment. Landslides in the region are characterized by a certain periodicity, intensity, and clustering, depending mainly on the periodicity of rainfall and seismic activity. Areas with very high landslide susceptibility are concentrated along the river (Figure 26b).

Figure 26c shows a localized remote sensing image of a high landslide susceptibility area. The area consists of a steep river valley, and the lithology mainly consists of Quaternary clayey sand, sandstone, siltstone, and other crushed and weathered sedimentary rocks that are soft, thereby allowing the development of frequent landslides. Figure 26d shows the LSM of the region, indicating that the landslide susceptibility was higher with an increase in proximity to the river. As seismic effects cannot be measured using InSAR surface deformation in this region, landslide susceptibility was primarily determined based on environmental features.

4.4.6. Southern Mountains of Zhejiang

The southern mountains of Zhejiang are located on the southeast coast of China (Figure 27a), and the region is mostly mountainous (>1000 m above sea level). Hills and basins are present in the transitional areas, with complex topography and diverse geomorphological patterns. The lithology of the region is complex and varied, and the distribution of fractures is widespread, which are conducive for the occurrence of landslides. In addition, the region lies in a subtropical monsoon belt with abundant precipitation that often experiences extreme rainstorms. Hydrological and climatic conditions facilitate the occurrence of landslides, and hence, the middle and lower mountainous areas of southern Zhejiang experience the most frequent landslide disasters in China. The main landslide type that occurs in this region is rainfall induced. The region has very high landslide susceptibility zones scattered in the mountainous areas (Figure 27b).

Figure 27c shows a localized remote sensing image of the high-susceptibility landslide area. No landslides were detected in the optical images, but our evaluation results delineated very high landslide susceptibility zones based on the geomorphological setting, geological background, and meteorological and hydrological factors of the area, which are distributed on both sides of the river (Figure 27d). Landslides in the lower and middle mountainous areas of southern Zhejiang are highly susceptible to landslide hazards during extreme rainfall events. Bao et al. (2016) analyzed rainfall data from 1592 landslide records obtained from 2457 rainfall stations in Zhejiang Province between 1990 and 2013 to determine the rainfall intensity thresholds for landslide disasters [75]. They specified landslide warnings for areas when the maximum rainfall reached 53, 97, 142, 207, and 304 mm in 1, 3, 6, 12, and 24 h, respectively. Information on rainfall thresholds for inducing landslides at different periods, combined with the spatial characteristics of very high landslide susceptibility zones provided by the LSM, allows for more accurate forecasting of landslide hazards.

4.4.7. Central Liaodong Peninsula

The Liaodong Peninsula is located in the northeastern part of China, bordering the Yellow Sea and the Bohai Sea (Figure 28a). It is situated in the transitional zone between the Northeast Plain and the Liaodong Mountains. The terrain is mainly characterized by mountains and hills, with the landform sloping from the northeast to the southwest. There is a significant variation in elevation on the peninsula, and there are numerous mountain ranges. The highest peak is Buyun Mountain, with an altitude of 1130 m. The elevation difference between the mountains and the plain is quite distinct, and the terrain undulates greatly. Such complex topography provides certain topographical conditions for the occurrence of landslides. The geological structure of the Liaodong Peninsula is rather complex. Fault structures are well-developed within the region. These fault structures have undermined the integrity of the rocks, causing them to be fragmented and increasing the unstable factors of the mountain masses. As observed from the LSM, the high landslide-prone areas are mainly distributed in the river valley areas in the central part of the Liaodong Peninsula (Figure 28b). The remote sensing image of the area with landslide hazards is shown in Figure 28c. This area is located at the boundary between the human activity area and the mountainous area. Figure 28d is a partial map of the LSM of this region. The terrain slope in this area is relatively steep. The lithology within the region mainly consists of granite, sandstone, shale, etc. Under the long-term weathering effect, the rocks are fragmented, and the weathered layer is relatively thick, resulting in poor stability of the rock and soil masses. Moreover, this area has abundant precipitation in summer, with frequent heavy rains. In addition, human activities such as road construction and real estate development have caused significant disturbances to the mountain bodies. Therefore, this area exhibits a high susceptibility to landslides.

The distribution of landslide susceptibility zones in the Tianshan Mountains is mainly influenced by topographic factors and seasonal freeze-thaw and precipitation effects. Landslide susceptibility areas in the Southern Tibetan Valley are controlled by topographic factors and some linear characteristic factors (e.g., rivers and faults). Most landslides in the Western Sichuan Basin and mountainous areas of central and western Yunnan were induced by seismic activity. Landslide susceptibility areas in the Qinling–Dabashan Mountains and the Central Liaodong Peninsula are closely related to human engineering construction, and strongly correlated with road networks. Landslides in the southern mountains of Zhejiang were mainly rainfall induced.

The above seven high landslide-prone areas are located in different major geomorphological regions, and their environmental characteristics vary significantly, leading to different controlling factors for the occurrence of landslides. Therefore, during the process of landslide susceptibility assessment, especially at the national scale, it is a necessary technical approach to conduct the assessment separately using geomorphological zoning.

4.5. Limitations

This study has some limitations.

(1)

In this study, the representation form of the stream data used is linear. However, due to the lack of attribute information about the river width, it is impossible to shield the rivers. As a result, in the LSM, some areas on the water surface have been identified as high landslide-prone areas, which does not conform to the actual situation. In future research, we hope to fully consider the attribute of the river width to improve the accuracy of the LSM.

(2)

the LSA was completed using a geomorphic regionalization scheme, which can suffer from joining discontinuities at the divisions of geomorphic districts because training samples and the main controlling factors of landslides can be different in each geomorphic district. Filtering techniques should be combined to deal with the splicing areas in different geomorphological districts, so as to solve the problem of poor continuity in the spliced areas.

(3)

the RF model was applied in the assessment, which is a black-box model. Although the importance of assessment factors was analyzed during the evaluation, the interpretive nature of landslide susceptibility in each geomorphological area is still lacking. Therefore, during the assessment process, it is necessary to fully consider the development mechanism of landslides, and combine physical models to increase the interpretability in the assessment process.

5. Conclusions

In this study, China was classified into 37 geomorphic districts using an existing geomorphic zoning scheme. The surface deformation rate map of China for five years from 2018 to 2022 was prepared and included as the landslide dynamic feature. Nine environmental factors related to topography, geology, meteorology and hydrology, seismicity, and engineering construction were analyzed to evaluate landslide susceptibility in different geomorphic districts. An LSM of the whole territory of China with a spatial resolution of 40 m was produced. The results were analyzed, and the following conclusions were drawn:

(1)

Using AUC, precision, recall, F1-score and MCC as evaluation metrics for the LSA in China, the obtained values are 0.927, 0.859, 0.815, 0.828, and 0.773, respectively. These results indicate that the model performed accurately.

(2)

The analysis of the importance of assessment factors reveals that the main factors controlling landslides are different in each geomorphological district, among which elevation and slope have the highest average importance in LSA, followed by average annual precipitation. The InSAR deformation factor, a dynamic feature, is highly important in some geomorphological areas, indicating that large deformation fields in these zones are directly related to landslides.

(3)

Through comparative experiments in typical areas, it is proved that the introduction of geomorphic regionalization and dynamic characteristics in LSA can improve the accuracy of the result.

(4)

This paper summarizes the distribution of very high landslide susceptibility areas in China and analyzes in detail the distribution characteristics of landslides and the main factors that control the development of landslides in seven typical areas: the Tianshan Mountain range, Southern Tibetan Valley, Western Sichuan Basin, Qinling–Dabashan Mountains, mountainous areas in west-central Yunnan, mountainous areas in southern Zhejiang, and Central Liaodong Peninsula. The LSM is expected to serve as a valuable resource for engineering projects, disaster management, and mitigation efforts in China.