Landslide susceptibility assessment method based on heterogeneous ensemble learning considering spatial heterogeneity

Abstract

This disclosure relates to a landslide susceptibility assessment method based on heterogeneous ensemble learning considering spatial heterogeneity, and belongs to the field of risk prediction for landslide disasters. According to the method, a zoning strategy is adopted to fully explore the spatial heterogeneity of landslide assessment factors; besides, dominant influencing factors of landslide development in different zoned regions and the whole region are revealed; and on this basis, a heterogeneous ensemble learning model is constructed to assess the landslide susceptibility. This disclosure can realize accurate landslide prediction.

Description

TECHNICAL FIELD

[0001] This disclosure belongs to the eld of risk prediction for landslide disasters, and specically relates to a spatial heterogeneity-considered landslide susceptibility assessment method for heterogeneous ensemble learning. BACKGROUND

[0002] With the frequent occurrence of landslide disasters, the social and economic development has been seriously impacted. As a key technical link in a landslide risk assessment system, landslide susceptibility mapping has been widely applied in elds such as prevention and control of geological disasters. However, there are still two obvious limitations in the existing landslide susceptibility assessment: (1) most studies fail to pay enough attention to the spatial heterogeneity of landslide assessment factors in spatial distribution, and the ignorance of spatial differences of these factors often leads to the decline of adaptability between assessment results and actual geological environments, making it difcult to accurately reect landslide development potential in different regions; and (2) the performance of a single assessment model is highly dependent on its own inherent algorithm structure and assumptions, which limits the applicability of the model in complex geological scenarios and signicantly reduces the room for performance improvement, making it difcult to meet the actual needs of high-precision landslide risk assessment.

[0003] Current landslide susceptibility assessment models can be divided into a qualitative method and a quantitative method. (1) The qualitative method is characterized by experience- driven, mainly including empirical methods. The assessment process is mainly based on the knowledge and experience of domain experts for judgment. The widely used methods include a fuzzy theory analysis method and an analytic hierarchy process. (2) The quantitative method is characterized by data-driven, mainly including mathematical statistics methods and machine learning methods. In the mathematical statistics methods, the information method and certainty factor method are commonly used techniques. However, there is often a complex nonlinear coupling relationship between assessment factors, and conventional mathematical statistical models are not sensitive enough to complex interaction between factors. (3) In order to solve the shortcomings of mathematical statistical models, machine learning models are widely used in landslide susceptibility assessment due to its strong nonlinear tting ability. Logistic regression (LR) model, support vector machine (SVM), random forest (RF), extreme gradient boosting (XGBoost) and other models are included. However, the performance of a single machine learning model is often limited by its own structural characteristics, and the overtting of specic data distribution is easily produced in the training process. (4) An ensemble learning model is used to carry out landslide susceptibility modeling through the fusion of multiple models, which improves the reliability and accuracy of prediction results. There are two types of ensemble learning: homogeneous and heterogeneous ensembles. The homogeneous ensemble is realized by combining multiple base classiers with a same structure. Although the stability of the model is enhanced, the inherent defects of the base classiers may be amplied in the ensemble process, thus increasing the potential risk of over-tting. However, the heterogeneous ensemble optimizes the overall prediction effect of the model by fusing different types of classiers and relying on the complementary characteristics of various classiers.

[0004] Currently, many scholars have studied the methods of landslide susceptibility, but there are still the following shortcomings: l. The performance of a single model depends on the inherent structural characteristics of the model itself, and the room for performance improvement is obviously limited, which makes it difcult to adapt to complex and changeable geological environments. The current study on the heterogeneous ensemble learning model of landslide susceptibility is still limited. 2. The current susceptibility assessment model does not rely enough on the spatial heterogeneity of assessment factors, which greatly restricts the improvement of the prediction accuracy of the model. SUMMARY

[0005] A purpose of this disclosure is to provide a . A zoning strategy is adopted to fully explore the spatial heterogeneity characteristics of landslide assessment factors; besides, dominant inuencing factors of landslide development in different zoned region and the whole region are revealed; and on this basis, a heterogeneous ensemble learning model is constructed to assess landslide susceptibility.

[0006] To achieve the above purpose, this disclosure provides the following technical solution: a , including:

[0007] step SOI: obtaining historical landslide data and landslide assessment factor data;

[0008] step S02: preprocessing the obtained historical landslide data and landslide assessment factor data, to construct a landslide assessment factor database;

[0009] step S03: constructing a heterogeneous ensemble landslide susceptibility assessment model considering spatial heterogeneity and training the model based on the landslide assessment factor database; and

[0010] step 804: using the landslide susceptibility assessment model trained in step S03 to predict landslide susceptibility in a region.

[0011] Further, in step 803, when constructing the heterogeneous ensemble landslide susceptibility assessment model considering spatial heterogeneity, spatial heterogeneity of a study region is explored and landslide assessment factors of the region are screened to construct the heterogeneous ensemble landslide susceptibility assessment model.

[0012] Further, in step S03, the constructing a heterogeneous ensemble landslide susceptibility assessment model considering spatial heterogeneity specically includes:

[0013] SO31: performing spatial heterogeneity modeling and regional division on the landslide susceptibility based on a combined manner of FR, GWR, and AGNES clustering methods;

[0014] S032: screening the landslide assessment factors by using a geographic detector (GeoDetector); and

[0015] SO33: using a stacking strategy to construct the heterogeneous ensemble landslide susceptibility assessment model.

[0016] Further, step S031 is realized as follows: the FR is used to preliminarily assess a probability of landslide occurrence, a calculated FR value is taken as a dependent variable, and a local regression coefcient of the landslide assessment factors is calculated by the GWR model; and then landslide points are clustered by the AGNES, and a Thiessen polygon is constructed by the clustered landslide points to zone the study region.

[0017] Further, the FR is used to calculate a probability of landslide occurrence of landslide assessment factors in different classication intervals, and a calculation formula is as follows: N- N FR = à

[0018] in the formula: FR represents a frequency ratio of the j-th landslide assessment factor in the ith classication interval; Ni represents a number of landslide occurrence in the ith classication interval of the landslide assessment factors; N represents a total number of landslides in the whole study region; Ai represents an area of the i-th classication interval; and A represents a total area of the study region;

[0019] the GWR model quanties an inuence coefcient of a single driving factor in different spaces by establishing a local regression equation of each spatial unit within a study range, thus reecting spatial heterogeneity and non-stationarity between study objects and the landslide assessment factors; and an equation of the GWR model is as follows: Q y = ßommwm) + Z ßk(umvm>xmk + s k=1

[0020] in the equation: ym represents a dependent variable of the m-th sample, that is, a probability of landslide occurred at the m-th landslide point, [30(um,vm) represents a local intercept term of the m-th sample, (uwv) represents spatial coordinates of the m-th sample, ßk represents a coefcient of the k-th independent variable, xmk represents the kth independent variable at the m-th sample, and represents a random error of a neighborhood at the m-th sample, and Q is a number of independent variables; and

[0021] the AGNES is used to cluster regression coefcients of the landslide points, similarity of spatial units is given priority in the process of clustering, landslide points with a similar driving mechanism are classied into a same cluster, the Thiessen polygon is constructed according to clustering results and the landslide points to divide the study region into several homogeneous sub- regions, wherein each sub-region represents a relatively landslide driving mechanism- homogeneous region.

[0022] Further, in step SO32, the geographic detector (GeoDetector) quanties spatial stratied heterogeneity of geographical elements through variance decomposition and identies its driving factors; and specically, q statistic is an index to assess explanatory power of the landslide assessment factors, the closer the q value is to l, the stronger the explanatory power of the spatial heterogeneity of corresponding landslide assessment factors, and a calculation formula of the q value is: EE N1m of, wss q = 1 T = 1 È s WSS = Z Nm ofn m=1 TSS = No2

[0023] in the formula: m is a layer of a variable Y or a factor X, wherein Y is a probability of landslide occurrence calculated by the frequency ratio, and X refers to a landslide assessment factor; Nm is a number of units in the m layer, and N is a number of units in the whole region; oËn is a method of the Y value in the m layer, 02 is a variance of the Y value in the whole region, WSS is a sum of intra-layer variances, and TSS is a total variance of the whole region.

[0024] Further, in step SO33, the using a stacking strategy to construct the heterogeneous ensemble landslide susceptibility assessment model is that three complementary machine learning algorithms, RF, SVM, and XGBoost, are used as base models and LR as a meta model to construct the heterogeneous ensemble landslide susceptibility assessment model.

[0025] Further, in step 804, the using the landslide susceptibility assessment model trained in step 803 to predict landslide susceptibility in a region is that the landslide susceptibility assessment model trained in step 803 is used to calculate a landslide susceptibility index of a grid cell in the region, so as to predict the landslide susceptibility in the region.

[0026] This disclosure further provides a landslide susceptibility assessment system based on heterogeneous ensemble learning considering spatial heterogeneity, including a memory, a processor, and a computer program instruction that is stored on the memory and can be executed by the processor, where when the processor executes the computer program instruction, the steps of any of the methods as described above can be implemented.

[0027] This disclosure further provides a computerreadable storage medium storing a computer program instruction that can be executed by a processor, where when the processor executes the computer program instruction, the steps of any of the methods as described above can be implemented.

[0028] Compared with the prior art, the present invention has the following benecial effects:

[0029] (1) The study region is divided through the combination of FR, GWR and AGNES, which effectively explores the spatial heterogeneity of the landslide assessment factors. (2) The dominant factors in each subregion screened by the GeoDetector reduce the number of assessment factors, while without reducing the performance of the model, thus reducing data redundancy through factor screening. (3) The heterogeneous ensemble learning model considering spatial heterogeneity proposed in this disclosure is superior to other models in performance, providing more accurate landslide prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] FIG. 1 is an implementation owchart of an embodiment of this disclosure;

[0031] FIG. 2 is a owchart of a landslide susceptibility model;

[0032] FIG. 3 shows ROC curve results of four models; and

[0033] FIG. 4 shows classication statistics of landslide susceptibility assessment results, where FIG. 4 (a) shows a susceptibility classication area; FIG. 4 (b) shows a number of landslides in each classication; and FIG. 4 (c) shows a landslide density. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0034] The technical solution of this disclosure will be described below in detail with reference to the accompanying drawings.

[0035] This disclosure provides a . First, historical landslide data and assessment factor data are integrated, landslide assessment factors are preliminarily screened by adopting correlation analysis and multicollinearity analysis, and a landslide assessment factor database is constructed. A buffer zone is then established around a historical landslide point, and non-landslide samples are randomly extracted outside the buffer zone to create training set and test set data. Second, spatial division of a study region is carried out to reveal a spatial heterogeneity law of landslide assessment factors; and dominant factors of the whole region and each zoned region are then screened, and on this basis, a heterogeneous ensemble landslide susceptibility assessment model is constructed.

[0036] A specic implementation process of this disclosure is as follows.

[0037] Referring to FIG. 1 and FIG. 2, this disclosure further provides a heterogeneous ensemble landslide susceptibility assessment method considering spatial heterogeneity, including the following steps.

[0038] Step SOI: Obtain historical landslide data and landslide assessment factor data. [003 9] The historical landslide data used in this disclosure is from the national geological disaster point spatial distribution data set released by the Resources and Environmental Science Data Center of China Academy of Sciences. The digital elevation model (DEM) used in this disclosure is from the geospatial data cloud, and a data set with 30 m resolution provided by the ASTER satellite is downloaded in this disclosure. Lithological data is from ISRIC World Soil Information Database; fault data is from the GMT Chinese community; and roads and water systems come from OpenStreetMap. Soil type data is provided by the Resources and Environmental Science Data Platform of China Academy of Sciences; and rainfall data is based on site observation data provided by the National Centers for Environmental Information (NCEI) under the National Oceanic and Atmospheric Administration (NOAA). Land use data is from China Land Use Data Set of Wuhan University; and the normalized difference vegetation index (NDVI) is from the National Tibetan Plateau Scientic Data Center.

[0040] Step 802: Preprocess the historical landslide data and landslide assessment factor data, to construct a landslide assessment factor database. The following steps are mainly carried out:

[0041] 8021: After a series of corrections and processing on the historical landslide data set, 6,115 landslide points were nally screened in the study region, and these landslide points were used as landslide samples for subsequent model construction. Considering that there is still a high risk of landslide occurred in the adjacent area of the landslide, a buffer zone of 1000 rn is established around the known landslide disaster point, and nonlandslide samples are randomly selected outside the buffer zone according to a ratio of landslide points to non-landslide points of 1:1. The data set is divided into 70% of training set and 30% of test set.

[0042] 8022: Processing such as splicing and projection cutting is performed on DEM data; terrain factors such as slope, slope direction, plane curvature, prole curvature, terrain undulation, terrain surface roughness, and TRI are obtained through a series of operations on DEM; based on the Standard for Engineering Classication of Rock Mass GBT502l8-2014, original multiclass lithology data is reclassied according to the hardness of rock mass, and nally the lithology is divided into six categories: extremely soft rock, soft rock, relatively soft rock, relatively hard rock, hard rock, and water body; Euclidean distance rasterization processing is carried out on the faults, roads, and water systems to respectively generate three assessment factors: a distance from faults, a distance from rivers, and a distance from roads; rasterization and re-classication are carried out on the soil type data; and the rainfall factor is used to calculate the average rainfall during 10 years from 2014 to 2023, which is then interpolated to 30 m resolution by an inverse distance weighting method.

[0043] Step 803: Construct a heterogeneous ensemble landslide susceptibility assessment model considering spatial heterogeneity.

[0044] When constructing the heterogeneous ensemble landslide susceptibility assessment model considering spatial heterogeneity, the study region is rst zoned, and dominant factors are then screened in the zoned region. An ensemble learning model constructed using a stacking strategy uses three machine learning algorithms, RF, SVM, and XGBoost, as a base model and LR as a meta model. Specically, it is implemented through the following steps:

[0045] SO31: Perform spatial heterogeneity modeling and regional division on the landslide susceptibility based on a combined manner of FR, GWR, and AGNES clustering methods. The FR is used to preliminarily assess a probability of landslide occurrence, a calculated FR value is taken as a dependent variable, and a local regression coefcient of the landslide assessment factors is calculated by the GWR model. The FR is used to calculate a probability of landslide occurrence of landslide assessment factors in different classication intervals, and a calculation formula is as follows: N- N FR = Ü

[0046] In the formula: FR represents a frequency ratio of the j -th assessment factor in the i-th classication interval; Ni represents a number of landslide occurrence in the i-th classication interval of the assessment factors; N represents a total number of landslides in the whole study region; Ai represents an area of the ith classication interval; and A represents a total area of the study region.

[0047] The GWR model quanties an inuence coefcient of a single driving factor in different spaces by establishing a local regression equation of each spatial unit within a study range, thus reecting spatial heterogeneity and nonstationarity between study objects and the assessment factors. In this disclosure, the FR is used to preliminarily assess a probability of landslide occurrence, a calculated FR value is taken as a dependent variable, and a local regression coefcient of the landslide assessment factors in each landslide point is calculated by the GWR model, so as to reveal the spatial heterogeneity of the landslide assessment factors. An equation of the GWR model is as follows: Q y = son) + Z ßk(umvm)xmk + s k=1

[0048] In the equation: ym represents a dependent variable of the m-th sample, that is, a probability of landslide occurred at the m-th landslide point, 80(um,vm) represents a local intercept term of the m-th sample, (um,vm) represents spatial coordinates of the m-th sample, ßk represents a coefcient of the k-th independent variable, xmk represents the k-th independent variable at the m-th sample, and sm represents a random error of a neighborhood at the m-th sample, and Q is a number of independent variables.

[0049] The AGNES regards each data point as an independent cluster, adopts a bottom-up agglomerative strategy, and builds a tree-like clustering structure by iteratively merging the most similar clusters until a termination condition is met, which is used to realize landslide point clustering in this disclosure. The AGNES clustering method is used to cluster regression coefcients of the landslide points according to the regression coefcient of each landslide point, and the AGNES method gives priority to similarity of spatial units in the process of clustering, and classies landslide points with a similar driving mechanism into a same cluster. Then, a Thiessen polygon is constructed according to clustering results and the landslide points to divide the study region into several homogeneous sub-regions, where each sub-region represents a relatively landslide driving mechanism-homogeneous region.

[0050] 8032: Screen the landslide assessment factors by using a geographic detector (GeoDetector).

[0051] GeoDetector is a statistical method based on the spatial heterogeneity theory, which quanties spatial stratied heterogeneity of geographical elements through variance decomposition and identies its driving factors. q statistic is an index to assess explanatory power of the landslide assessment factors, the closer the q value is to 1, the stronger the explanatory power of the spatial heterogeneity of the assessment factor, and a calculation formula of the q value is: EE Nm sân wss q = 1 = T = 1 = TSS s wss = 2 Nm og] m=1 TSS = No2

[0052] In the formula: m is a layer of a variable Y or a factor X, where Y is a probability of landslide occurrence calculated by the frequency ratio, and X refers to a landslide assessment factor; Nm is a number of units in the rn layer, and N is a number of units in the whole region; 612,, is a method of the Y value in the m layer, and oz is a variance of the Y value in the whole region. WSS is a sum of intra-layer variances, and TSS is a total variance of the whole region.

[0053] S033: Construct the heterogeneous ensemble landslide susceptibility assessment model.

[0054] The ensemble learning model is constructed using the stacking strategy. Specically, three complementary machine learning algorithms, RF, SVM, and XGBoost, are used as base models and LR as a meta model. The RF can effectively alleviate the imbalance of landslide samples through a Bootstrap sampling and feature random selection mechanism; the SVM has certain advantages in dealing with a complex boundary of a discrete assessment factor such as lithology; and the XGBoost can effectively capture a nonlinear relationship between assessment factors and landslides through a gradient boosting tree-structure. Taking data sets of landslide assessment factors as the inputs of the three base models, and taking the respective prediction results of the three base models as the input features of the meta model, linear combination characteristics of the meta model LR can be automatically learned and relative contribution weight of each base model can be quantied. The original input is a landslide assessment factor data set X = [x1 x2, ...,xp] , FR, SVM, and XGBoost respectively predict the X, and f1,f2f3 are prediction results of the three base models respectively.

[0055] The RF constructs m decision trees through subsets of different data, votes the results of the multiple decision trees, and nally obtains output results. The core idea of Random Forest is Bagging and random feature subspaces, which aims to reduce the model variance and avoid over- tting. 1 m f1(X)= 52300

[0056] where Tt (X ) is a prediction result of the t-th tree. [005 7] The SVM is a supervised learning algorithm based on the statistical learning theory, which is essentially a nonlinear data processing method. This method maps input low-dimensional nonlinear data into a highdimensional space, and performs linear regression in the high dimensional space, thus achieving a nonlinear regression effect in the original space. In the landslide susceptibility modeling process, a radial basis function is usually used as a sum function of the SVM, which can effectively describe a complex nonlinear relationship between landslide assessment factors and landslides. KOC) = eX10(-1 / ||x - X'IIZ) MX) = 2:109 Mom) + b

[0058] where K (x, x') is the radial basis function, ocl- is a support vector coefcient, and b is a bias term.

[0059] The XGBoost algorithm is an improvement of the gradient boosting decision tree. Through iterative ensemble learning and regularization optimization, the algorithm signicantly improves the computational efciency and prediction performance of the traditional GBDT, and is widely used in classication, regression, and sorting tasks. f3(X) = íîzlhtm

[0060] where [? is a learning rate, and ht (X) is a prediction value of the t-th tree.

[0061] The LR is a statistical learning method widely used in classication tasks, especially in binary classication problems. The model maps linear regression results to (O, 1) by a Sigmoid function, thus realizing the landslide probability prediction. The LR function can be expressed as: 1 Y(X) = m m = WO + w1f1 + w2f2+...+wnfn

[0062] where y(x) represents a probability of landslide, wl, wz, ..., and wn represent relative contribution of RF, SVM, and XGBoost, and fl, fz, ..., frl are prediction results of the three base models respectively.

[0063] Step 804: Use the trained landslide susceptibility assessment model to predict landslide susceptibility in a region.

[0064] A landslide susceptibility index of a grid cell in the region is calculated according to the landslide susceptibility assessment model trained in step 803.

[0065] This disclosure has been tested and proved to be feasible, with the accuracy higher than that of the existing methods.

[0066] In order to assess the performance of the model created by this disclosure, the following comparative tests are set: adopting four models, specically including a Stacking ensemble learning model, GWR-S, S-Geo, and GWR-S-Geo models, where the GWR-S model only considers spatial heterogeneity; the SGeo model only uses the GeoDetector to screen dominant factors of the whole region; and the GWR-SGeo model not only considers the spatial heterogeneity, but also screens the dominant factors of each different subregion separately. Accuracy assessment results are shown in FIG. 3.

[0067] The GWR-S-Geo model created by this disclosure can more accurately predict the probability of landslide occurrence, and AUC values of the GWR-S, S-Geo, and GWR-SGeo models are 0.836, 0.815, and 0.838 respectively. It can be seen that, compared with the Stacking model, the performance of the GWR-S-Geo model considering the spatial heterogeneity and using the GeoDetector for factor screening is better than other models.

[0068] This disclosure also counts landslide susceptibility zoning areas of different models proposed by this disclosure and number and density of landslides of this zoning respectively. As shown in FIG. 4, landslide density values of the four models all gradually increase with the increase of susceptibility level, which indicates that all four models can correctly reect the spatial distribution of landslide points. The sum of landslide densities with extremely high and high susceptibility levels obtained by the GWR-S-Geo model proposed by this disclosure is 0.311, which is the highest compared with the GWR-S (0.269), S-Geo (0.260), and Stacking (0.271) models. The zoning area with extremely high and high susceptibility levels predicted by the GWR SGeo model is the smallest, which indicates that the model predicts more landslides in the relatively small high susceptibility zoning area, providing more accurate prediction results.

[0069] This disclosure further provides a landslide susceptibility assessment system based on heterogeneous ensemble learning considering spatial heterogeneity, including a memory, a processor, and a computer program instruction that is stored on the memory and can be executed by the processor, where when the processor executes the computer program instruction, the steps of any of the methods as described above can be implemented.

[0070] This disclosure further provides a computerreadable storage medium storing a computer program instruction that can be executed by a processor, where when the processor executes the computer program instruction, the steps of any of the methods as described above can be implemented.

[0071] The abovedescribed embodiments are exemplary embodiments of the present invention, and any changes made in accordance with the technical solutions of the present invention and the resulting functional effects without departing from the scope of the technical solutions of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for assessing landslide susceptibility that takes into account taking into account spatial heterogeneity in heterogeneous integrative learning methods, characterized that: Step SOl, obtaining historical landslide data and landslide assessment factor data; Step 802: Preprocessing the historical landslide data and the landslide assessment factor data and building a landslide assessment factor database; Step SO3, Building a Heterogeneous Integration Model for Assessing landslide sensitivities that take into account spatial heterogeneity and training based on the landslide assessment factor database; step 804, predicting the landslide susceptibility in an area with the step SO3 trained model for landslide susceptibility assessment.

2. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 1, characterized by exploring the spatial heterogeneity of a research area in step SO3 in building the heterogeneous integration model for assessing the landslide sensitivities that take into account spatial heterogeneity, the selecting the landslide assessment factors in the area and building the heterogeneous integration model for landslide susceptibility assessment.

3. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 1, characterized by building the heterogeneous integration model for assessing the landslide sensitivities that take into account the spatial heterogeneity in the step SO3, where the specific includes the following: SO31, modeling the spatial heterogeneity and dividing the areas into based on the combination of the FR, GWR and AGNES clustering methods; SO32, selecting the assessment factors for the landslides with a geographic detector GeoDetector; 8033, Building the Heterogeneous Integration Model for Assessing the landslide sensitivities with a Stacking strategy.

4. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 3, characterized in that step SO31 is realized by: preliminary assessment of the landslide probability with FR, calculating local regression coefficients of the landslide assessment factors using a GWR model and the calculated FR value as a dependent variable, and clustering landslide points with AGNES, and building a Thiessen polygon to divide the research area.

5. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 4, characterized in that the FR is used to estimate the probability of landslide occurrence within different classification intervals, the formula is as follows: Ni / N FR = m where: FR denotes a frequency ratio of the j th value of the landslide assessment factor in the ith classification interval; Ni indicates the number landslides of the landslide assessment factor that occurs within the ide classification interval; N gives the total number of landslides in the entire research area; Ai denotes the area of ​​the ith classification interval; A denotes the total area of ​​the study area; The GWR model quantifies the influence factors of individual control factors in different spaces by building local regression equations for each spatial unity within the study area, resulting in spatial heterogeneity and non-stationarity characteristics between the research object and the landslide assessment factors are reflected; the equation of the GWR model is as follows: Q y = God) + Z ßk(umvm)xmk + 8 k=1 where: ym denotes a dependent variable of the mth sample point, i.e., the probability that a landslide occurs at the mth landslide point, ß0(um vm) gives a local intercept term of the m-th sample, (u, v) denotes a spatial coordinate of the m-th sample, ßk denotes a coefficient of the k-th independent variable, xmk denotes the kth independent variable of the mth sample, denotes the random error in the neighborhood of the m-th sample, where Q is the number of independent variables; the clustering of clusters the regression factors of the different landslide points with the AGNES and giving priority to similarity in the clustering process of the spatial units; classifying the landslide points with similar control mechanisms in the same cluster; building the Thiessen polygon based on the clustering results and the landslide points to divide the study area into different to divide homogeneous areas, with each subzone representing a relatively homogeneous area of ​​the landslide mechanisms.

6. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 3, characterized in that in step 8032 the geographical detector GeoDetector measures the spatial layering of geographical elements quantified by means of variance distribution and identifies the driving factors; specifically, the q-statistic is an indicator for assessing of the explanatory power of the landslide assessment factors; the closer the The q-value is close to 1, the stronger the explanatory power of the spatial differentiation characteristics of the associated landslide assessment factors, where the calculation formula for the q-value is: q: l_EÊnzlNmO-à : 1_Ë N02 TSS s WSS = Z Nm 0121, m=1 TSS = N62 where: m is the layer of variable Y or factor X, where Y is the landslide probability which is calculated via the frequency ratio, and X refers to the rating factors for landslides; Nm is the number of units in layer m, N is the total number of units in the entire area; an is the method for the Y-value of the layer m, oz is the variance of the Y- values ​​in the whole area, WSS is the sum of the variances within the layer, and TSS is the total variance of the entire area.

7. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 3, characterized in that in step SO33 the Stacking strategy is applied to obtain a heterogeneous to build an integration model for landslide susceptibility assessment, i.e. the heterogeneous integration model for landslide susceptibility assessment built with three complementary machine learning algorithms, RF, SVM and XGBoost, as basic models and LR as meta-model.

8. The method for assessing landslide susceptibility taking into account taking into account spatial heterogeneity in heterogeneous integrative learning methods according to conclusion 1, characterized in that in step SO4 a model trained in step SO3 for assessing The landslide sensitivities are used to estimate the landslide sensitivities in a to predict a certain area, i.e. the landslide susceptibility index of the grid units within the area are calculated using the model trained in step SO3 for the assessing landslide sensitivities in order to determine the landslide sensitivities of to predict the area.

9. A landslide susceptibility assessment system that takes into account taking into account spatial heterogeneity in heterogeneous integrative learning methods, characterized that it has a memory, a processor, and computer program instructions stored in memory comprises, wherein the steps of the method as described in any of claims 1-8 are realized can be performed when the processor executes the computer program instructions.

10. A computer-readable storage medium on which the computer software instructions are stored stored that can be executed by the processor, where the steps of the method as described in any of claims 1 to 8 can be realized when the processor executes computer program instructions. FIG. 1