Landslide susceptibility ensemble learning evaluation method considering spatial heterogeneity partitioning and factor feature screening

Abstract

The invention belongs to the technical field of landslide susceptibility analysis, and relates to a landslide susceptibility ensemble learning evaluation method considering spatial heterogeneity partitioning and factor feature screening, which comprises the following steps: generating a landslide sample based on historical landslide catalog data, and selecting a non-landslide sample through environmental factor frequency ratio analysis; the method comprises the following steps of: extracting static and dynamic environment factor data sets, realizing factor space interpretation force transformation by utilizing a t-SNE-ISO clustering algorithm and a feature screening strategy, eliminating high-correlation factors through a Pearson correlation coefficient method, quantifying interpretation force of each factor on landslide space differentiation by combining a geographic detector, screening optimal feature combinations under global and partition frameworks respectively, and performing landslide space differentiation on the landslide space. According to the method, a Stacking integrated learning framework is combined with CNN, DNN, MLP-based learners and LR element learners, a landslide susceptibility probability prediction model is formed, the generalization ability and prediction accuracy of the model are improved, and the method is especially suitable for landslide high-incidence areas with severe topographic relief and complex geological conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of landslide susceptibility analysis, and in particular to an integrated learning evaluation method for landslide susceptibility that takes into account spatial heterogeneity partitioning and factor feature screening. Background Art

[0002] Landslide susceptibility assessment (LSA) is a core technical requirement for disaster prevention and mitigation. Traditional assessment methods fall into two main categories: qualitative analysis based on expert knowledge, which delineates landslide-prone areas through geological surveys and empirical judgment, but is highly subjective and difficult to quantify. Data-driven quantitative models, such as logistic regression and support vector machines, rely on statistical relationships between historical disaster data and environmental factors, but suffer from two major technical bottlenecks. First, traditional models assume that the study area is "spatially homogeneous," assuming that landslide development mechanisms are uniformly distributed across the entire region. However, in reality, factors such as topography, geological structure, and climate often exhibit significant spatial heterogeneity, leading to "averaging" errors when using a single model for prediction in complex regions. Second, feature selection strategies are crude. Existing technologies often use global correlation analysis or single indicators (such as the variance inflation factor) to select factors, ignoring the dynamic modulation of factor influence by spatial heterogeneity. For example, the same factor may have completely different triggering mechanisms for landslides in mountainous areas and plains, and traditional methods cannot capture this regional variability.

[0003] The shortcomings of existing technologies are concentrated in two aspects: First, the lack of a mechanism to handle spatial heterogeneity leads to a "southern orange tree, northern tangerine" phenomenon in cross-regional predictions, that is, the performance of the model optimized in the training area drops sharply in the validation area; second, feature screening is disconnected from spatial partitioning. Traditional methods first screen factors and then divide regions (or vice versa), forming a serial "screening-partitioning" logic. In fact, the importance of factors themselves is spatially dependent. For example, rainfall factors may dominate landslide occurrence in humid areas, but degenerate to a secondary factor in arid areas. This technical approach leads to two core problems: first, the combination of factors uniformly screened across the entire region may contain variables that are redundant or even misleading for some subregions; second, after partitioning, each subregion still uses the global screening results, which cannot achieve spatially adaptive expression of factor influence. These shortcomings directly limit the accuracy and generalization ability of landslide susceptibility assessment models, especially in complex geological environments, where prediction reliability needs to be improved. Summary of the Invention

[0004] The purpose of the present invention is to provide an integrated learning evaluation method for landslide susceptibility that takes into account spatial heterogeneity zoning and factor feature screening, which solves the problems of insufficient prediction accuracy and rigid factor screening strategy caused by ignoring spatial heterogeneity in existing landslide susceptibility evaluation methods.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions: The present invention proposes an integrated learning evaluation method for landslide susceptibility that takes into account spatial heterogeneity partitioning and factor feature screening. The method comprises: Generate landslide samples based on historical landslide catalog data in the target area, and select non-landslide samples through environmental factor frequency ratio analysis; Extract the factor data set that affects the development of landslides in the target area from landslide samples and non-landslide samples. The factor data set includes static environmental factors and dynamic environmental factors. Based on the landslide catalog data and the factor dataset, the target area is divided into spatially heterogeneous sub-areas; By combining the correlation coefficient method with the feature screening method of geographic detectors, the factor feature combinations of spatially heterogeneous sub-regions in the global framework and the partitioned framework are determined. At least two heterogeneous deep learning models are constructed as base learners, and the base learners are combined based on the Stacking ensemble learning framework to generate landslide susceptibility evaluation results corresponding to the factor feature combination.

[0006] Furthermore, the specific steps of determining the landslide sample and the non-landslide sample include: The landslide boundary polygon vector data in the historical landslide inventory of the target area are converted into raster cells as landslide samples; The frequency ratio FR value of each environmental factor classification was calculated, and in the area with the lowest FR value, the grid cells with the same number as the landslide samples were selected as non-landslide samples.

[0007] Furthermore, after extracting the data set of factors affecting landslide development in the target area from the landslide samples and non-landslide samples, the method further includes data preprocessing: All static environmental factor data were unified in coordinate system and resampled to the same resolution; The frequency ratio method was used to normalize the values ​​of environmental factors.

[0008] Furthermore, the static environmental factors include elevation, slope, aspect, plan curvature, profile curvature, terrain moisture index, runoff intensity index, distance from faults, stratum lithology, distance from rivers, distance from roads, land use, average annual rainfall, normalized vegetation index, and kernel density of points of interest; the dynamic environmental factors include the ascending orbit surface deformation rate and the descending orbit surface deformation rate.

[0009] Furthermore, the dividing the target area into spatially heterogeneous sub-areas based on the landslide catalog data and the factor dataset includes: determining a spatial distribution of a plurality of landslide points based on the landslide samples, wherein each grid cell represents a potential landslide occurrence location; Performing cluster analysis on the plurality of landslide points according to a preset distance metric criterion, and determining the number of clusters based on the CH index; The t-SNE algorithm is used to perform nonlinear dimensionality reduction on static and dynamic environmental factors to form low-dimensional spatial data after dimensionality reduction. Performing cluster analysis on the low-dimensional spatial data using an ISO clustering algorithm to form a plurality of spatially heterogeneous sub-regions; A corresponding Thiessen polygon is constructed for each spatially heterogeneous sub-region of the cluster, and the boundary of the Thiessen polygon is used as the dividing line of the spatially heterogeneous sub-region to form multiple spatially heterogeneous sub-regions with similar environmental conditions.

[0010] Furthermore, the step of determining the factor feature combination includes: Calculate the correlation coefficients of all factors in the dataset within all spatially heterogeneous sub-regions, and determine the screened spatially heterogeneous sub-regions based on the relationship between the set coefficient threshold and the correlation coefficient, including the screened static environmental factors and dynamic environmental factors; The factor detector in the geographic detector was used to analyze the selected spatially heterogeneous sub-regions to determine the explanatory power of each environmental factor on the spatial differentiation of landslide susceptibility and to characterize it through the q value. The factor feature combination is determined based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region.

[0011] Furthermore, determining the factor feature combination based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region includes: In the global spatial heterogeneous sub-region, a global analysis of the environmental factors of the global spatial heterogeneous sub-region is conducted based on the geographic detector, and the factors whose q-value ranking and significance level meet the preset standards in the global spatial heterogeneous sub-region are selected to form a target factor feature combination applicable to the entire region; In each spatially heterogeneous sub-region, geographic detector analysis is performed independently. Based on the q-value ranking and significance level within each sub-region, the optimal feature combination unique to each sub-region is selected to form independent analysis units with different landslide development characteristics.

[0012] Furthermore, in the partitioning framework analysis, the optimal feature combination of each spatially heterogeneous sub-region is determined by the following steps: In each independent analysis unit, the q value of the environmental factor is calculated separately, and the candidate factors whose q value exceeds the preset threshold are screened out; In each independent analysis unit, the q value of the environmental factor is calculated separately, and the candidate factors whose q value exceeds the preset threshold are screened out; Based on the geographic detector, the candidate factors of each sub-region are sorted by q value and significance test, and the factors with strong explanatory power and statistical significance are retained to form the optimal feature combination of the sub-region; The optimal feature combination of each sub-region is compared with the global target factor feature combination to verify the influence of spatial heterogeneity on the factor influence.

[0013] Furthermore, the calculation formulas for the q values ​​of each environmental factor in the global spatial heterogeneous sub-region and each spatial heterogeneous sub-region are as follows: Where h is the number of layers of environmental factor X, h=1,2,3,…, L, N and N h is the total number of samples and the number of samples in the hth layer, and are the total variance of the dependent variable landslide susceptibility Y and the variance of Y within the hth layer, respectively.

[0014] Furthermore, the stacking ensemble learning framework is used to combine base learners to generate landslide susceptibility evaluation results corresponding to the factor feature combination, including: Construct convolutional neural network, deep neural network and multi-layer perceptron models as base learners, and construct logistic regression model as meta-learner; A five-fold cross-validation method is used to input the prediction results of the base learner as features into the meta-learner to generate the final landslide susceptibility probability prediction results; The prediction results are divided into five susceptibility levels of "very low", "low", "medium", "high" and "very high" through the natural break method, and the landslide susceptibility evaluation mapping is completed.

[0015] The beneficial effects of the present invention are: The present invention effectively solves the problem of insufficient prediction accuracy caused by ignoring spatial heterogeneity in traditional methods by taking into account the dual optimization strategy of spatial heterogeneity partitioning and factor feature screening. The study area is divided into spatially heterogeneous sub-regions by using t-SNE dimensionality reduction and ISO clustering algorithm, and the explanatory power of factors is quantitatively measured by geographic detectors to accurately capture the landslide development mechanism in different regions. Through the global-partitioned dual framework comparative analysis, the optimal feature combination of each sub-region is dynamically screened, avoiding the rigid problem of factor screening in traditional methods. On this basis, the Stacking integrated learning framework is used to fuse heterogeneous deep learning models such as CNN, DNN and MLP, giving full play to the complementary advantages of each model, significantly improving the generalization ability and prediction accuracy of the model, and is particularly suitable for high-incidence landslide areas with severe terrain fluctuations and complex geological conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of the main body of the landslide susceptibility integrated learning evaluation method considering spatial heterogeneity zoning and factor feature screening proposed in a specific embodiment of the present invention; Figure 2 A flowchart for the specific implementation of the landslide susceptibility ensemble learning evaluation method considering spatial heterogeneity zoning and factor feature screening proposed in a specific embodiment of the present invention; Figure 3 Schematic diagram of CH values ​​of various cluster numbers under different distance measurement criteria in a case study of a specific embodiment of the present invention; Figure 4 This is a spatial heterogeneity zoning map of the landslide monitoring demonstration area in Honghe Prefecture in a case study of a specific embodiment of the present invention; Figure 5 This is a heat map of the Pearson correlation coefficient of the LCFs in the case study of the specific embodiment of the present invention; Figure 6 This is a diagram of GeoDetector factor detection results in a case study of a specific embodiment of the present invention; Figure 7 This is a graph showing the landslide susceptibility prediction results based on the Stacking model in a case study of a specific embodiment of the present invention; Figure 8 This is a graph showing the landslide susceptibility prediction results based on the F-Stacking model in a case study of a specific embodiment of the present invention; Figure 9 This is a graph showing the landslide susceptibility prediction results based on the H-Stacking model in a case study of a specific embodiment of the present invention; Figure 10 This is a graph showing the landslide susceptibility prediction results based on the HF-Stacking model in a case study of a specific embodiment of the present invention; Figure 11 This is a multi-index radar chart of the landslide susceptibility evaluation model in different sub-regions in the study case of the specific embodiment of the present invention. DETAILED DESCRIPTION

[0017] The present application is described in further detail below in conjunction with the accompanying drawings. It is necessary to point out that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technicians in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0018] Combine Figure 1 and Figure 2 As shown, in a specific embodiment, the present application proposes an integrated learning evaluation method for landslide susceptibility that takes into account spatial heterogeneity partitioning and factor feature screening, and the method includes the following steps: S1. Based on historical landslide data from the target area, polygonal vector data of landslide boundaries were converted into raster cells as landslide samples. To balance the sample distribution, non-landslide samples were selected using the environmental factor frequency ratio (FR) analysis method. The FR values ​​for each environmental factor classification were calculated, and raster cells with the lowest FR values ​​were randomly selected as non-landslide samples, equal to the number of landslide samples. This method ensured the representativeness of the non-landslide samples and prevented sample bias from interfering with model training.

[0019] S2. Extract a dataset of factors influencing landslide development from both landslide and non-landslide samples, including static environmental factors (such as elevation, slope, and stratum lithology) and dynamic environmental factors (such as surface deformation rate). Data preprocessing includes unifying the coordinate system and resolution (using bilinear interpolation and resampling), and normalizing the factor values ​​using the frequency ratio method to eliminate the impact of dimensional differences on the model.

[0020] S3. Based on the landslide catalog and factor datasets, the t-SNE algorithm was used to perform nonlinear dimensionality reduction on the high-dimensional environmental factor data, preserving the data's local structural characteristics. The ISO clustering algorithm was then used to cluster the reduced data. The optimal number of clusters (e.g., three) was determined using the CH index, and spatially heterogeneous subregions were delineated using Thiessen polygons. This method effectively addressed the low sample discrimination of traditional clustering methods in high-dimensional data and accurately characterized the spatial heterogeneity of the study area.

[0021] S4. Redundant factors (e.g., factors with r > 0.45) were initially eliminated using the Pearson correlation coefficient method. GeoDetector was then used to quantify the explanatory power (q-value) of each factor in landslide spatial variation. Globally dominant factors (e.g., distance to roads, NDVI, etc.) were screened within the global framework. Within the subregional framework, each subregion was independently analyzed, and the optimal regional feature combination was determined using recursive feature elimination. For example, the q-value of the surface deformation rate in one subregion was significantly higher than that of the global level, indicating that dynamic factors play a key role in the landslide triggering mechanism in this region.

[0022] S5. Convolutional neural networks, deep neural networks, and multilayer perceptron models were constructed as heterogeneous base learners. Their prediction results were fused using a stacking ensemble framework, and logistic regression (LR) was used as a meta-learner to generate the final landslide susceptibility probability. Model performance was optimized through five-fold cross-validation, and the prediction results were categorized into five susceptibility levels, ranging from "very low" to "very high."

[0023] In a preferred embodiment, the specific steps of determining the landslide sample and the non-landslide sample include: converting the landslide boundary polygon vector data in the historical landslide list of the target area into grid cells as the landslide sample; calculating the frequency ratio FR value of each environmental factor classification, and selecting the same number of grid cells as the landslide sample in the area with the lowest FR value as the non-landslide sample.

[0024] It can be understood that the landslide list is constructed based on historical landslide data, combined with remote sensing images and field investigation data; the environmental factors include static environmental factors and dynamic environmental factors, and in the embodiment, the static environmental factors include elevation, slope, slope direction, plan curvature, profile curvature, terrain moisture index, runoff intensity index, distance from fault, stratum lithology, distance from river, distance from road, land use, annual average rainfall, normalized vegetation index, and point of interest kernel density; and the dynamic environmental factors include ascending track surface deformation rate and descending track surface deformation rate.

[0025] In a preferred embodiment, after extracting the factor data set affecting the development of landslides in the target area from the landslide sample and the non-landslide sample, the method further includes data preprocessing: unifying the coordinate systems of all static environmental factor data and resampling to the same resolution; and using the frequency ratio method to normalize the environmental factor values.

[0026] Regarding the data preprocessing, in specific implementation, first, the coordinate systems of all the basic data are unified to ensure that spatial data from different sources can be correctly connected and overlaid. Since there are differences in spatial resolution of different types of geographic data, bilinear interpolation method can be used for resampling to unify the spatial resolution, so as to ensure that the number of rows and columns of LCFs (hereinafter referred to as landslide environmental factors) is consistent. In addition, the differences in dimensions and value ranges of LCFs will affect the evaluation results of the model, and the frequency ratio method (Frequency Ratio, FR) is selected to eliminate the unbalanced influence between LCFs. At the same time, the frequency ratio method is used to select the non-landslide sample for model training, to calculate the FR value of each classification state of LCFs, and to superimpose and sum the grid data of LCFs to obtain the FR value of all grids in the study area. In the area with low FR value, the same number of units as the landslide sample are randomly selected as the non-landslide sample. Finally, the landslide sample and the non-landslide sample are combined as the sample data set.

[0027] In a preferred embodiment, based on the landslide catalog data and the factor dataset, the target area is divided into spatially heterogeneous sub-areas, which includes: determining the spatial distribution of multiple landslide points based on landslide samples, wherein each grid cell represents a potential landslide location; performing cluster analysis on the multiple landslide points according to a preset distance metric, and determining the number of clusters based on the CH index (Calinski-Harabasz Index); performing linear dimensionality reduction on static environmental factors and dynamic environmental factors using the t-SNE algorithm to form low-dimensional spatial data after dimensionality reduction; performing cluster analysis on the low-dimensional spatial data using the ISO clustering algorithm to form multiple spatially heterogeneous sub-areas; constructing corresponding Thiessen polygons for each spatially heterogeneous sub-area of ​​the cluster (to facilitate visualization and further analysis in a GIS platform, and the partitioning results can be directly used in regional disaster prevention planning), and using the boundaries of the Thiessen polygons as the boundaries of the spatially heterogeneous sub-areas to form multiple spatially heterogeneous sub-areas with similar environmental conditions (such as homogeneous landslide categories).

[0028] To address the issue of spatial heterogeneity, based on the above-mentioned embodiments, this application proposes a combined method (t-SNE-ISO) that incorporates the t-SNE dimensionality reduction algorithm and the ISO clustering algorithm to mitigate the adverse effects of spatial heterogeneity on the construction of landslide susceptibility assessment models. Specifically, the optimal distance metric and number of clusters are first determined based on the CH index, where the distance metrics include Euclidean distance, Manhattan distance, and cosine distance. The nonlinear dimensionality reduction properties of the t-SNE algorithm are then utilized to map the high-dimensional and complex LCFs data into a low-dimensional space, preserving the data's local structure and characteristic information to the greatest extent possible. Finally, the ISO clustering algorithm is used to cluster the landslide points containing the dimensionality reduction information and ultimately construct Thiessen polygons to achieve spatial heterogeneity partitioning.

[0029] As can be seen, the present invention, through the combination of t-SNE dimensionality reduction and ISO clustering, can accurately identify the spatial distribution characteristics of different landslide development mechanisms within a target area. For example, in mountainous areas with dramatic terrain, clustering results can clearly distinguish between landslide areas driven by tectonic activity and those induced by rainfall, resolving the problem of fuzzy zoning in complex geological environments caused by traditional methods.

[0030] In a preferred embodiment, the step of determining the factor feature combination includes: calculating the correlation coefficient of all factors in the data set in all spatially heterogeneous sub-regions, and determining the screened spatially heterogeneous sub-regions based on the relationship between the set coefficient threshold and the correlation coefficient, including the screened static environmental factors and dynamic environmental factors; using the factor detector in the geographic detector to analyze the screened spatially heterogeneous sub-regions, determine the explanatory power of each environmental factor on the spatial differentiation of landslide susceptibility, and characterize it through the q value; determine the factor feature combination based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region.

[0031] It should be noted that landslide development mechanisms and inducing factors vary significantly across subregions. Indiscriminately inputting all features into a machine learning model would reduce data validity and, in turn, affect model accuracy. This application uses a PG feature screening strategy (combining the Pearson correlation coefficient method and geographic detectors for feature screening) to determine the optimal feature combination.

[0032] First, calculate the Pearson correlation coefficient r value between the global feature factors. If the r value is greater than the threshold (0.45), the feature is deleted, otherwise it is retained. The calculation method is as follows: Where x i with y i are the observed values ​​of the two characteristic factors; and is the sample mean of x and y; n is the sample size.

[0033] Then, for the retained features, the factor detector in the geographic detector is used to screen out the most significant features in each sub-region as the final optimal feature combination. The factor detection q value is used to measure the degree to which the characteristic factor X explains the spatial variation of landslide susceptibility Y, and the value range is [0,1]. The larger the q value, the stronger the explanatory power of X on Y. Its calculation formula is as follows: Where h=1,2,3,…, L is the number of layers of the characteristic factor X, N and Nh are the total number of samples and the number of samples in the hth layer, and are the total variance of the dependent variable Y and the variance of Y within the hth layer, respectively.

[0034] In a preferred embodiment, determining a factor feature combination based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region includes: In the global spatial heterogeneous sub-region, a global analysis of the environmental factors of the global spatial heterogeneous sub-region is conducted based on the geographic detector, and the factors whose q-value ranking and significance level meet the preset standards in the global spatial heterogeneous sub-region are selected to form a target factor feature combination applicable to the entire region; In each spatially heterogeneous sub-region, geographic detector analysis is performed independently. Based on the q-value ranking and significance level within each sub-region, the optimal feature combination unique to each sub-region is selected to form independent analysis units with different landslide development characteristics.

[0035] In a preferred embodiment, in the partitioning framework analysis, the optimal feature combination of each spatially heterogeneous sub-region is determined by the following steps: in each independent analysis unit, the q value of the environmental factor is calculated separately, and the candidate factors whose q value exceeds the preset threshold are screened out; in each independent analysis unit, the q value of the environmental factor is calculated separately, and the candidate factors whose q value exceeds the preset threshold are screened out; based on the geographic detector, the candidate factors of each sub-region are sorted by q value and significance test, and the factors with strong explanatory power and statistical significance are retained to form the optimal feature combination of the sub-region; the optimal feature combination of each sub-region is compared with the feature combination of the global target factor to verify the degree of influence of spatial heterogeneity on the factor influence.

[0036] It is understood that the present invention achieves optimized screening of global and regional feature combinations through the dual-frame analysis mechanism of the geodetector. In the global framework, all environmental factors are first subjected to global q-value calculation and significance test (p < 0.05), and dominant factors with strong explanatory power and statistical significance are screened to form a universal feature combination applicable to the entire region. In the regional framework, geodetector analysis is performed independently for each spatially heterogeneous sub-region delineated by the t-SNE-ISO algorithm, and the feature selection criteria are dynamically adjusted. For example, for sub-regions dominated by tectonic activity, surface deformation rate (q = 0.351) and fault distance (q = 0.298) are included in the feature combination; while in rainfall-induced sub-regions, average annual rainfall (q = 0.412) and topographic wetness index (q = 0.327) become key factors. This dual-frame design not only retains the overall consistency of the global model, but also captures the variation in factor influence caused by spatial heterogeneity through partition-specific screening. For example, in a certain sub-region, the aspect factor, which originally ranked 7th in the global q value, jumped to 2nd place (q=0.364) in the partition due to its coupling with the local structural trend, significantly improving the explanatory accuracy of the partition model.

[0037] In a preferred embodiment, the present invention utilizes a stacking ensemble learning framework to construct a landslide susceptibility prediction model. Three deep learning models with complementary characteristics are first selected as base learners: a convolutional neural network (CNN) is used to extract local spatial features, a deep neural network (DNN) processes high-dimensional nonlinear relationships, and a multilayer perceptron (MLP) captures global patterns. Using a five-fold cross-validation method, each base learner is trained separately in both a global and partitioned framework, with its predicted probability being fed as a new feature into a logistic regression (LR) meta-learner for secondary modeling. This heterogeneous ensemble strategy effectively combines the strengths of different models, such as the CNN's ability to extract terrain texture features with the DNN's superior processing of dynamic deformation data. The final output of the landslide susceptibility probability (ranging from 0 to 1) is divided into five levels using the natural breaks method, generating an intuitive susceptibility zoning map on a GIS platform.

[0038] Based on the above embodiments, the present application scheme is further elaborated and illustrated in conjunction with a research example below.

[0039] 1. Overview of the study area: The Honghe Landslide Monitoring Demonstration Area is located in the Honghe Hani and Yi Autonomous Prefecture, Yunnan Province, China, encompassing Jinping, Pingbian, and Hekou Counties. The area lies between 102°31′33″-104°16′42″ east longitude and 22°26′40″-23°23′25″ north latitude, covering approximately 6,915 square kilometers. Located in the low-latitude region within the Tropic of Cancer, the demonstration area experiences a low-latitude subtropical humid mountain monsoon climate, with an average annual rainfall of 2,000.9 mm and an average annual temperature of 19.1°C. The rainy season, from June to August, is characterized by frequent geological disasters. The demonstration area is a typical mountainous area with an altitude range of 76.4m-3074m, of which the area with an altitude above 800m accounts for about 70.49%; the regional water system is highly developed and dense, which leads to strong river erosion, resulting in the widespread distribution of steep slopes and active gravity geological effects; in recent years, human engineering activities have become increasingly intense, and various infrastructures need to be built by cutting slopes. These factors provide the material source basis and breeding environment for the development of landslide geological disasters.

[0040] 2. Slippery Slope List The landslide inventory, constructed based on historical landslide data and combined with remote sensing imagery and field survey data, includes 514 landslides in the study area, covering a total area of ​​8.569 km2. The largest landslide covers an area of ​​1.238 km2, and the smallest covers an area of ​​40 m2. Landslides are categorized by volume: 395 are small, accounting for approximately 76.85% of the total; 99 are medium, accounting for approximately 19.26%; 18 are large, accounting for approximately 3.50%; and two are extremely large, accounting for approximately 0.39%. Most landslides in the study area are small and primarily located near valley slopes.

[0041] 3. Environmental factors Landslides are the result of the combined effects of multiple internal and external environmental factors. Scientifically selecting landslide environmental factors (LCFs) is crucial for improving the reliability and accuracy of LSMs. Drawing on previous research and expert knowledge, this study initially selected 15 static factors (elevation, slope, aspect, plan curvature (PLC), profile curvature (PRC), terrain wetness index (TWI), runoff intensity index (SPI), distance to fault (DFFA), stratigraphic lithology, distance to river (DFRI), distance to road (DFRO), land use (LULC), average annual rainfall, normalized difference vegetation index (NDVI), and point of interest (POI) kernel density) and two dynamic factors (ascending surface deformation rate (ASC) and descending surface deformation rate (DES)).

[0042] Based on the landslide susceptibility ensemble learning evaluation method proposed in the above embodiment, which takes into account spatial heterogeneity zoning and factor feature screening, the results are analyzed.

[0043] 3. Results Analysis 3.1 Spatial heterogeneity analysis like Figure 3 As shown, in order to determine the optimal distance measurement criterion and the number of clusters for clustering, the study tested the CH values ​​of each number of clusters under different distance measurement criteria. The results show that the clustering effect of Euclidean distance is the best, and the optimal number of clusters is 3, and the corresponding CH value is 14.1; the clustering effect of Manhattan distance is second, the number of clusters is 4, and the corresponding CH value is 10.3; the clustering effect of cosine distance is the worst, the number of clusters is 3, and the corresponding CH value is 1.2. In summary, this application finally uses Euclidean distance as the distance measurement criterion, and uses the t-SNE-ISO algorithm to cluster the landslide points containing LCFs information into 3 categories. The spatial heterogeneity zoning results of the study area are shown in Figure 2. Figure 4 shown.

[0044] 3.2 Factor feature screening Figure 5 is the Pearson correlation coefficient result of the global LCFs. The results show that the r value of PLC and slope is -0.56, and the r value of TWI and SPI is 0.45, both greater than 0.45. It is worth noting that the correlation between TWI and slope reaches -0.41, showing a linear correlation trend. Therefore, TWI and PLC are finally eliminated in this paper. Subsequently, GeoDetector was used to further evaluate the influence of the remaining 15 LCFs on the spatial distribution of landslides in the global and sub-regions. The results are shown in Figure 6 , Figure 6 p<0.05 is “significant” and is indicated by ★★; p≥0.05 is “not significant” and is indicated by ★.

[0045] The results show that the q values ​​of LCFs vary across different regions, but DFRO and NDVI are the two most significant factors in both the entire region and each subregion. The impact of DFRO on landslides is essentially the result of the interaction between human engineering activities and the natural geological environment. The construction of roads changes the geological conditions of the region, destroys the stability of slopes, and thus increases the probability of landslides. The closer to the road, the higher the landslide risk. NDVI indirectly reflects regional surface stability and hydrological conditions by quantifying the state of vegetation cover and is an important environmental factor in landslide susceptibility assessment. In areas with good vegetation cover, the soil is fixed by vegetation roots, which enhances stability and reduces landslide risk.

[0046] For the global framework, DFRO had an explanatory power of 40.2%, with 75.9% of landslides occurring within a 750-m buffer zone from roads. Within the zonal framework, DFRO had the greatest impact on landslides in Regions I and II, with explanatory powers of 41.6% and 38.4%, respectively. 81% and 74% of landslides occurred within the 750-m buffer zone from roads, respectively. The number of landslide events decreased with increasing distance. For Region III, NDVI was the most significant factor, with a q-value of 37.5%. Landslides in this region occurred primarily in areas with an NDVI less than 0.7. Furthermore, Lithology failed the 5% significance test in any region, indicating weak explanatory power. In summary, based on the results of the GeoDetector, insignificant factors were eliminated from each region, and the significantly influential LCFs were retained as the optimal feature combination. The optimal feature combination for the entire domain is: DFRO, NDVI, elevation, precipitation, slope, LULC, POI, DFRI, ASC, aspect, DES, PRC and DFFA; the optimal feature combination for region I is: DFRO, NDVI, elevation, slope, DFFA, precipitation, DES and ASC; the optimal feature combination for region II is: DFRO, NDVI, slope, elevation, ASC, LULC, DFRI, aspect, precipitation, POI, SPI and PRC; the optimal feature combination for region III is: NDVI, DFRO, POI, elevation, slope, DFRI and PRC.

[0047] 3.3 Landslide susceptibility assessment based on ensemble learning To effectively mitigate the impact of spatial heterogeneity on landslide susceptibility research and verify the effectiveness of feature screening strategies, this study employed CNN, DNN, and MLP as base learners, and LR as a meta-learner. Four comparative models were constructed using stacking ensemble learning techniques: 1) a global framework stacking model (without feature screening); 2) a global framework F-Stacking model (with feature screening); 3) a partitioned framework H-Stacking model (with reduced spatial heterogeneity); and 4) a partitioned framework HF-Stacking model (with reduced spatial heterogeneity and feature screening). After constructing the ensemble models, landslide susceptibility probability prediction was performed for the target area using LCFs as input variables. A quantitative landslide susceptibility index (LSI) was obtained for each evaluation unit in the study area, ranging from 0 to 1, with higher values ​​indicating greater landslide susceptibility. Finally, visualization was performed in ArcGIS 10.7, and the LSI was categorized into five levels: very low, low, medium, high, and very high using the natural breaks method. Figure 7-10 The LSMs predicted by the four models are shown, all of which have similar landslide susceptibility distribution patterns.

[0048] The study conducted a comprehensive analysis using an integrated learning model and came to the following conclusions: The extremely high susceptibility zone is mainly distributed in the central part of the study area and on both sides of the main traffic arteries. Among them, the Jinhe Town area has a very high LSI in the prediction results of the four models, and the possibility of landslides is the greatest. The geomorphological characteristics of this area are large terrain undulations and a developed river network. The front edge of the slope is eroded by the river, causing the rock and soil at the foot of the slope to soften, easily collapse, and reduce the anti-sliding force. In addition, the stratum rock type in this area is mainly quartzite, which has high hardness and high compressive strength and can support steep terrain. However, it is significantly brittle. Under gravity or tectonic stress, the bottom is easily damaged by tensile cracks, forming a potential sliding surface, which provides direct conditions for the occurrence of geological disasters.

[0049] 3.4 Comparison of Evaluation Performance between Single Deep Learning Model and Ensemble Learning Model To accurately reflect the performance differences among CNN, DNN, MLP, and Stacking models, this study first compared the predictive performance of the four models under global conditions. The global dataset was divided into training and test sets in a 7:3 ratio. Each model was trained using the training set, and the test set was used to test model performance. The corresponding EMs for each model were calculated (see Table 1). The results show that, given the same test samples, the Stacking model achieved the highest AUC value of 0.9167, followed by the CNN model (0.9143). The MLP model (0.8988) had the lowest AUC value. Furthermore, the Stacking model achieved the highest values ​​for OA (0.8350), Recall (0.7453), F1-score (0.8247), and Kappa (0.6721), demonstrating the superior predictive performance of the Stacking model and providing reliable support for the predictions of subsequent partitioning models.

[0050] Table 1 EMs of the four models under the whole region (the bold and underlined values ​​indicate the highest EMs in that column) ; 3.5 Optimizing the Evaluation Performance of Ensemble Learning Models by Spatial Heterogeneity Partitioning and Feature Screening In order to explore the optimization effect of spatial heterogeneity zoning modeling and feature screening strategy on landslide susceptibility evaluation model, the prediction performance of four integrated models including Stacking, F-Stacking, H-Stacking and HF-Stacking was systematically analyzed using six evaluation indicators including OA, Precision, Recall, F1-score, Kappa and MCC (see Figure 11(Figure: (a) Subregion I; (b) Subregion II; (c) Subregion III; (d) Average index of subregions). The results show that within the spatial heterogeneity partitioning framework, the HF-Stacking model exhibits the best performance in all subregions. Its OA, Recall, F1-score, Kappa, and MCC values ​​all rank first, and it exhibits the largest coverage area in radar chart analysis, fully verifying the superiority of this model. Furthermore, the study optimized the model's performance by introducing a feature screening strategy. Results show that within the global analysis framework, the F-Stacking model exhibits significant performance improvements over the traditional Stacking model: OA reaches 0.8544 (a relative increase of 1.94%), Recall increases to 0.7888 (a 4.35%), and F1-score increases to 0.8495 (a 2.48%). Furthermore, the Kappa coefficient (0.7100) and MCC (0.7180) significantly increase by 3.79% and 3.21%, respectively. It is worth noting that under the partition analysis framework, the HF-Stacking model also shows comprehensive improvements compared to the H-Stacking model, and all evaluation indicators show significant improvements.

[0051] Figure 11 (d) A radar chart showing the mean values ​​of subregional indicators shows that compared to the global stacking model, HF-Stacking achieves significant improvements in key indicators: OA (0.9141) increases by 7.91%, F1-score (0.9021) increases by 7.74%, Kappa (0.7979) increases by 12.58%, and MCC (0.8012) increases by 11.53%. These results demonstrate that by constructing a spatial heterogeneity partitioning framework and implementing a regional optimal feature screening strategy, the interference of spatial heterogeneity in global modeling can be effectively reduced, significantly improving the generalization ability and accuracy of the landslide susceptibility prediction model.

[0052] According to the analysis of the above research case results, the present invention significantly improves the accuracy and reliability of landslide susceptibility assessment by taking into account the method of spatial heterogeneity zoning and factor feature screening. The study first used the t-SNE dimensionality reduction and ISO clustering algorithm to divide the study area into spatially heterogeneous sub-regions with different landslide development characteristics, effectively solving the problem of spatial heterogeneity. Subsequently, the environmental factors of the entire region and each sub-region were screened by combining the Pearson correlation coefficient method and geographic detectors to determine the optimal feature combination. Finally, based on the Stacking ensemble learning framework, deep learning models such as CNN, DNN and MLP were integrated to construct a landslide susceptibility prediction model. The experimental results show that this method exhibits excellent prediction performance in both the global and partitioned frameworks. In particular, in the partitioned framework, by weakening spatial heterogeneity and performing feature screening, the generalization ability and prediction accuracy of the model are further improved, providing strong support for regional disaster prevention planning.

[0053] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A landslide susceptibility ensemble learning evaluation method that takes into account spatial heterogeneity partitioning and factor feature screening, characterized by: The method comprises: Generate landslide samples based on historical landslide catalog data in the target area, and select non-landslide samples through environmental factor frequency ratio analysis; Extract the factor data set that affects the development of landslides in the target area from landslide samples and non-landslide samples. The factor data set includes static environmental factors and dynamic environmental factors. Based on the landslide catalog data and the factor dataset, the target area is divided into spatially heterogeneous sub-areas; By combining the correlation coefficient method with the feature screening method of geographic detectors, the factor feature combinations of spatially heterogeneous sub-regions in the global framework and the partitioned framework are determined. At least two heterogeneous deep learning models are constructed as base learners, and the base learners are combined based on the Stacking ensemble learning framework to generate landslide susceptibility evaluation results corresponding to the factor feature combination.

2. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 1 is characterized in that: The specific steps of determining the landslide sample and the non-landslide sample include: The landslide boundary polygon vector data in the historical landslide inventory of the target area are converted into raster cells as landslide samples; The frequency ratio FR value of each environmental factor classification was calculated, and in the area with the lowest FR value, the grid cells with the same number as the landslide samples were selected as non-landslide samples.

3. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 1 is characterized in that: After extracting the data set of factors affecting landslide development in the target area from the landslide samples and non-landslide samples, the method further includes data preprocessing: All static environmental factor data were unified in coordinate system and resampled to the same resolution; The frequency ratio method was used to normalize the values ​​of environmental factors.

4. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 3 is characterized in that: The static environmental factors include elevation, slope, aspect, plan curvature, profile curvature, terrain moisture index, runoff intensity index, distance from faults, stratum lithology, distance from rivers, distance from roads, land use, average annual rainfall, normalized difference vegetation index, and kernel density of points of interest; the dynamic environmental factors include the ascending orbit surface deformation rate and the descending orbit surface deformation rate.

5. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 2 is characterized in that: The dividing of the target area into spatially heterogeneous sub-areas based on the landslide catalog data and the factor dataset includes: determining a spatial distribution of a plurality of landslide points based on the landslide samples, wherein each grid cell represents a potential landslide occurrence location; Performing cluster analysis on the plurality of landslide points according to a preset distance metric criterion, and determining the number of clusters based on the CH index; The t-SNE algorithm is used to perform nonlinear dimensionality reduction on static and dynamic environmental factors to form low-dimensional spatial data after dimensionality reduction. Performing cluster analysis on the low-dimensional spatial data using an ISO clustering algorithm to form a plurality of spatially heterogeneous sub-regions; A corresponding Thiessen polygon is constructed for each spatially heterogeneous sub-region of the cluster, and the boundary of the Thiessen polygon is used as the dividing line of the spatially heterogeneous sub-region to form multiple spatially heterogeneous sub-regions with similar environmental conditions.

6. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 5 is characterized in that: The step of determining the factor feature combination includes: Calculate the correlation coefficients of all factors in the dataset within all spatially heterogeneous sub-regions, and determine the screened spatially heterogeneous sub-regions based on the relationship between the set coefficient threshold and the correlation coefficient, including the screened static environmental factors and dynamic environmental factors; The factor detector in the geographic detector was used to analyze the selected spatially heterogeneous sub-regions to determine the explanatory power of each environmental factor on the spatial differentiation of landslide susceptibility and to characterize it through the q value. The factor feature combination is determined based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region.

7. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 6 is characterized in that: Determining the factor feature combination based on the relationship between the q value and the preset threshold in the spatially heterogeneous sub-region includes: In the global spatial heterogeneous sub-region, a global analysis of the environmental factors of the global spatial heterogeneous sub-region is conducted based on the geographic detector, and the factors whose q-value ranking and significance level meet the preset standards in the global spatial heterogeneous sub-region are selected to form a target factor feature combination applicable to the entire region; In each spatially heterogeneous sub-region, geographic detector analysis is performed independently. Based on the q-value ranking and significance level within each sub-region, the optimal feature combination unique to each sub-region is selected to form independent analysis units with different landslide development characteristics.

8. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 7 is characterized in that: In the partitioning framework analysis, the optimal feature combination for each spatially heterogeneous sub-region is determined by the following steps: In each independent analysis unit, the q value of the environmental factor is calculated separately, and the candidate factors whose q value exceeds the preset threshold are screened out; Based on the geographic detector, the candidate factors of each sub-region are sorted by q value and significance test, and the factors with strong explanatory power and statistical significance are retained to form the optimal feature combination of the sub-region; The optimal feature combination of each sub-region is compared with the global target factor feature combination to verify the influence of spatial heterogeneity on the factor influence.

9. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 8 is characterized in that: The calculation formula for the q value of each environmental factor in the global spatial heterogeneous sub-region and each spatial heterogeneous sub-region is as follows: Where h is the number of layers of environmental factor X, h=1,2,3,…, L, N and N h is the total number of samples and the number of samples in the hth layer, and are the total variance of the dependent variable landslide susceptibility Y and the variance of Y within the hth layer, respectively.

10. The landslide susceptibility ensemble learning evaluation method taking into account spatial heterogeneity partitioning and factor feature screening according to claim 1 is characterized in that: The stacking ensemble learning framework is based on combining base learners to generate a landslide susceptibility evaluation result corresponding to the factor feature combination, including: Construct convolutional neural network, deep neural network and multi-layer perceptron models as base learners, and construct logistic regression model as meta-learner; A five-fold cross-validation method is used to input the prediction results of the base learner as features into the meta-learner to generate the final landslide susceptibility probability prediction results; The prediction results are divided into five susceptibility levels of "very low", "low", "medium", "high" and "very high" using the natural break method, and the landslide susceptibility evaluation mapping is completed.

Images