Landslide susceptibility prediction method based on multi-scale geographically weighted regression and spatial heterogeneity partitioning

Abstract

The present application relates to the technical field of geological disaster risk assessment, and discloses a landslide-prone prediction method based on multi-scale geographic weighted regression and spatial heterogeneity partitioning, which solves the problems of low local accuracy, poor area efficiency and insufficient prediction stability caused by single spatial scale, uniform modeling and single evaluation system in traditional landslide prediction models. The present application organically combines multi-scale geographic weighted regression with spatial heterogeneity partitioning, first adaptively assigns a dedicated optimal bandwidth to each disaster-inducing factor through a multi-scale regression model to accurately depict the spatial non-stationary characteristics of landslide disaster-inducing relationships, then reconstructs the discrete regression coefficients into a spatial continuous surface with the help of Kriging interpolation, dynamically extracts the dominant factors with the optimal discrete and natural discontinuity method to realize the heterogeneity partitioning of the study area, takes each partition as an independent unit to carry out landslide-prone deduction, and constructs a multi-dimensional comprehensive evaluation system from the dimensions of prediction accuracy, area rationality and uncertainty robustness.

Description

Technical Field

[0001] This invention relates to the field of geological hazard risk assessment technology, specifically to a landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning. Background Technology

[0002] Landslides are a widespread and destructive geological hazard globally. Accurate prediction of regional landslide susceptibility is a core and fundamental task for geological disaster prevention and mitigation, national land space safety planning, and risk management of major projects. From a technological evolution perspective, landslide susceptibility assessment has shifted from early expert experience-based judgments and traditional bivariate statistical analysis to an intelligent quantitative prediction paradigm driven by multi-source geospatial data and centered on machine learning and spatial regression models.

[0003] With the development of high-resolution remote sensing, geographic information systems (GIS), and geological exploration technologies, global machine learning models such as logistic regression, random forests, and gradient boosting trees have become mainstream techniques for landslide susceptibility prediction due to their strong fitting capabilities. However, these global models assume that the relationship between landslides and disaster-causing factors remains spatially homogeneous, neglecting the spatial non-stationarity of the landslide formation process and failing to characterize the differences in local disaster-causing mechanisms in different regions.

[0004] To compensate for the shortcomings of global models, existing technologies have introduced the traditional geographically weighted regression (GWR) model. By embedding spatial coordinates into the regression framework and using the distance decay kernel function to estimate the local regression coefficients that vary with location, the spatial heterogeneity of landslide occurrence has been revealed to some extent, achieving a technological advancement from global uniform assessment to local analysis.

[0005] However, traditional geographically weighted regression models suffer from fundamental theoretical and engineering flaws: they presuppose that all disaster-causing factors operate at the same spatial scale, and can only determine a single, fixed spatial search bandwidth through cross-validation. In reality, during landslide formation, macroscopic geomorphology and regional tectonics typically control the basic development conditions of landslides at a large regional scale, while lithological changes and hydrological upheavals often trigger disasters at a more refined local scale. Therefore, the scales at which different factors act differ fundamentally. A single bandwidth can lead to over-smoothing of key local disaster-causing features or ineffective fragmentation of macroscopic patterns, resulting in modeling results that severely deviate from the actual physical mechanisms.

[0006] Meanwhile, at the engineering implementation level, existing solutions generally treat the highly geologically variable, wide-area research areas as a homogeneous whole for probabilistic extrapolation. Lacking a mechanism for intelligent spatial partitioning based on differences in disaster-prone coupling, this approach not only results in computational redundancy but also dilutes the prediction accuracy of local micro-regions. Furthermore, the existing evaluation system is too simplistic, relying solely on the receiver operating characteristic (ROC) curve to measure overall accuracy. It fails to quantify the efficiency of engineering screening for high-risk prediction ranges or examine the impact of accidental disturbances in underlying data on the stability of prediction results, creating a significant reliability blind spot in practical applications.

[0007] The aforementioned deficiencies, compounded in engineering practice, constitute a systemic technical bottleneck restricting accurate early warning of geological disasters: the coupling of modeling constraints at a single spatial scale with the assumption of homogenized space means that prediction models can only output a severely compromised "average response relationship" when facing complex terrain, thus masking the strong local heterogeneity of specific areas and triggering a chain reaction of degradation where "global statistical accuracy is acceptable, but local micro-risks are out of control." Attempting to compensate for the homogenization deficiency by artificially dividing the study area within the existing framework not only lacks geostatistical basis but also artificially severs the physical continuity of natural disaster processes, leading to a technical dilemma of neglecting one aspect for another. More significantly, the lack of an evaluation system with area efficiency and robustness constraints often results in high-scoring models outputting broad and unstable high-risk prediction areas, leading to a substantial increase in actual geological investigation costs. This deep mismatch between underlying theoretical assumptions and the need for cost reduction and efficiency improvement in engineering makes existing solutions inadequate for high-reliability disaster prevention resource allocation tasks.

[0008] Therefore, how to break through the constraints of a single spatial scale, realize multi-factor adaptive scale modeling, conduct objective spatial heterogeneity partitioning based on disaster-causing mechanisms, and construct a multi-dimensional evaluation system that takes into account accuracy, efficiency, and robustness are key technical problems that urgently need to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning, which solves the problems of low local accuracy, poor area efficiency and insufficient prediction stability caused by the single spatial scale, uniform modeling and single evaluation system of traditional landslide prediction models.

[0010] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0011] In a first aspect, the present invention provides a landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning, comprising the following steps:

[0012] S1. Multi-source pregnancy and malformation data are fused and processed, and the study area is discretized by a regular grid space. Feature optimization is completed through collinearity diagnosis to construct a multi-dimensional pregnancy and malformation feature space.

[0013] S2. A multi-scale geographically weighted regression model is used to adaptively allocate independent bandwidth to each disaster-prone factor, and the local regression coefficients are obtained by fitting.

[0014] S3. Perform Kriging interpolation on the local regression coefficients to generate a spatially continuous regression coefficient surface;

[0015] S4. Based on the regression coefficient surface, perform optimal discretization and dominant factor extraction to complete spatial heterogeneity partitioning;

[0016] S5. Landslide susceptibility is extrapolated using spatially heterogeneous zoning as the unit, and the prediction quality is evaluated through multi-dimensional indicators.

[0017] In this scheme, multi-scale geographical weighted regression is used to adapt to the real spatial scale of different disaster-causing factors, avoiding the smoothing distortion of local disaster-causing information caused by a single scale; spatial heterogeneity partitioning is used to divide the study area into units with homogeneous disaster-causing mechanisms, thereby eliminating the local accuracy dilution caused by uniform modeling across the entire region; and multi-dimensional indicators are used to simultaneously measure prediction accuracy, area efficiency and robustness, making up for the shortcomings of traditional single accuracy indicators, and improving the overall accuracy, stability and engineering practicality of landslide prediction.

[0018] Furthermore, in step S1, the multi-source disaster-prone data includes remote sensing satellite image interpretation results, historical geological disaster cataloging data, digital elevation models, and regional meteorological and hydrological time-series data;

[0019] The methods for fusing and processing multi-source pregnancy and malpregnancy data include:

[0020] By performing unified coordinate benchmark transformation and projection transformation on all data through a Geographic Information System (GIS), a structured continuous spatial layer with spatial matching and unified format is formed.

[0021] In this scheme, the data types comprehensively cover the core landslide-causing elements such as topography, geology, hydrology, meteorology, and surface cover, which can fully characterize the environmental conditions for landslide formation in the region and provide a comprehensive and reliable data foundation for subsequent high-precision modeling. By unifying coordinates and projections, spatial offsets, misalignments, and matching errors between multi-source heterogeneous data can be completely eliminated, ensuring that all layers participate in the calculation within the same spatial framework, and ensuring the consistency and accuracy of subsequent spatial analysis and regression calculations.

[0022] Furthermore, in step S1, the process of performing regular grid spatial discretization on the study area and completing feature optimization through collinearity diagnosis includes:

[0023] Using a set resolution regular grid as the basic mapping unit, the spatial kernel density estimation algorithm is used to calculate the landslide density as a continuous response variable. Collinearity diagnosis is performed by Pearson correlation test and variance inflation factor, and redundant factors are eliminated to construct a multidimensional disaster-prone feature space.

[0024] The variance inflation factor is calculated as follows:

[0025] ;

[0026] in, For the first The variance inflation factor of each pregnancy factor; To make the first The coefficient of determination obtained when performing multiple linear regression with one factor as the dependent variable and all other factors as independent variables.

[0027] The method for diagnosing collinearity using Pearson correlation test and variance inflation factor is as follows:

[0028] Factors with a variance inflation factor greater than 10 or a Pearson correlation coefficient greater than 0.7 are considered as redundant factors.

[0029] In this scheme, collinearity diagnosis can effectively avoid matrix singularity, computational instability and overfitting caused by linear overlap of high-dimensional factors, thereby improving model stability and parameter interpretability.

[0030] Furthermore, in step S1, the constructed multidimensional disaster-prone feature space includes four types of features: terrain category, engineering geological rock group, distance from river, and melting index.

[0031] In this scheme, by eliminating multicollinearity and redundant factors, four types of core disaster-causing features covering macro-geomorphology, geological conditions, hydrological erosion, and thermal conditions are retained. This not only fully characterizes the key environmental driving factors of landslide formation, but also avoids computational redundancy and model bias caused by overlapping factor information, providing a stable feature foundation with simplified dimensions and clear physical meaning for subsequent multi-scale regression modeling.

[0032] Furthermore, in step S2, the adoption of a multi-scale geographically weighted regression model to adaptively allocate independent bandwidths to each disaster-causing factor and fit local regression coefficients includes:

[0033] Construct a multi-scale geographically weighted regression baseline equation to achieve independent scale modeling of each factor:

[0034] ;

[0035] in, Indicates the first spatial grid location The response variable on; Indicates the first spatial grid location The first Explanatory variables (such as engineering geological rock groups); For the first The optimal spatial bandwidth corresponding to each explanatory variable; For the first The explanatory variable in the th... spatial grid location The local regression coefficient at a given location, which is based on the optimal spatial bandwidth of the corresponding explanatory variable. Obtained through fitting; For spatial intercept; To obtain white noise residuals that conform to a Gaussian distribution;

[0036] Separating partial residuals of a single factor using a backfit algorithm:

[0037] ;

[0038] in, For the first The partial residuals of each variable; The actual observed values ​​of the response variable; No. The current regression summation term of all variables other than the one variable; For the first A regression summation term for each variable; For residuals;

[0039] Construct a distance-decay spatial weight kernel function for local weighted calculation:

[0040] ;

[0041] in, For observation point For target point Spatial weights; The Euclidean distance between the two points; For the current trial bandwidth, when hour, ;

[0042] With the goal of minimizing the modified Akaike Information Criterion (AICc), a golden section search is used to continuously update the trial bandwidth. This yields the local regression coefficients of the factor at the current scale.

[0043] In the iterative optimization process, the ratio of the sum of squared changes is introduced as a criterion for model convergence:

[0044] ;

[0045] in, The ratio of the sum of squares of the changes; This represents the total number of pregnancy-related factors. This represents the total number of spatial grid cells; and Representing the first In the current iteration and the previous iteration, the factor is located at the th . Local regression coefficient values ​​for each grid location; The preset convergence tolerance threshold is used;

[0046] when When the value is below the preset convergence tolerance threshold, the model is deemed to have converged, and the optimal bandwidth and local regression coefficients of each factor are output.

[0047] In this scheme, by independently allocating the optimal spatial bandwidth to each landslide-causing factor and purifying the fitted signal factor by factor, the spatial non-stationary characteristics of landslide-causing factors can be realistically restored. This avoids the local detail smoothing distortion caused by a single scale and ensures the stability and reliability of the coefficients through strict convergence judgment, ultimately obtaining the local regression coefficients of each factor.

[0048] Furthermore, in step S3, the step of performing kriging interpolation on the local regression coefficients to generate a spatially continuous regression coefficient surface includes:

[0049] For any unmodeled grid location within the study area to be estimated First, based on the discrete local regression coefficients, the experimental semivariogram is calculated and the theoretical variogram is obtained by fitting.

[0050] The experimental semivariance value is calculated as follows:

[0051] ;

[0052] in, The lag distance is The experimental semivariance value at time; The spatial distance lag step; Spatial distance interval equal to The total number of valid sample point pairs; and Spatial location and its distance The regression coefficient value at the location;

[0053] Construct and solve the Kriging equations to obtain the weight coefficients that satisfy the unbiased and optimal conditions; the Kriging equations are expressed as:

[0054] ;

[0055] in, For known sample points and The theoretical semivariance between them; For known sample points And the point to be estimated The theoretical semivariance between them; It is a Lagrange multiplier;

[0056] Based on the calculated weight coefficients, the regression coefficients of the surrounding known sample points are weighted and combined to obtain the predicted regression coefficient values ​​for the location to be estimated.

[0057] ;

[0058] in, Location to be estimated The predicted values ​​of the regression coefficients; This represents the number of known neighboring sample points used in the estimation. To be assigned to the Weight coefficients for each known sample point;

[0059] By traversing all grid locations to be estimated within the study area, regression coefficient prediction is completed point by point, generating a spatially continuous regression coefficient surface.

[0060] In this scheme, discrete local regression coefficients are reconstructed into a spatially continuous surface through Kriging interpolation. This fully utilizes the spatial autocorrelation of the regression coefficients, and while ensuring unbiasedness and optimal estimation, it completely fills in the coefficient values ​​of the unmodeled grid, eliminates spatial expression discontinuities caused by discrete points, and makes the global regression coefficient distribution smooth, continuous, and consistent with the real geographic spatial variation patterns. This provides high-precision, continuous, and unified basic data support for subsequent optimal coefficient discretization, dominant factor identification, and spatial heterogeneity partitioning.

[0061] Furthermore, in step S4, the optimal discretization and dominant factor extraction based on the regression coefficient surface to complete the spatial heterogeneity partitioning includes:

[0062] First, perform optimal clustering and discretization of the attribute data on the regression coefficient surface, and calculate the sum of squares of the global data deviations from the mean for a single factor layer:

[0063] ;

[0064] in, The sum of squared deviations from the mean of global data for a single factor layer; The total number of pixels involved in the classification; For the first The regression coefficient values ​​for each pixel; This is the global mean.

[0065] Simultaneously, calculate the sum of squared deviations from the mean within each subclass:

[0066] ;

[0067] in, The sum of squared deviations from the mean within each subclass; For the first Subclasses; For the first The within-class mean of the regression coefficients of all pixels within each subclass; This represents the total number of subclasses. For the first Within each subclass The regression coefficient values ​​for each pixel;

[0068] Introducing the Goodness-of-Class (GVF) index as the optimal discrete evaluation metric:

[0069] ;

[0070] in, For classification excellence index;

[0071] By continuously adjusting the combination of discontinuities through enumeration or heuristic search, the optimal discontinuity scheme that makes GVF achieve the global maximum value is locked, thus completing the optimal discretization of the continuous regression coefficient surface.

[0072] Subsequently, a GIS spatial topology overlay operation was performed on all the discretized factor layers to cut the study area into subdivided polygonal patches with single attributes. Within each patch, the mean of the absolute values ​​of the regression coefficients of each disaster-causing factor was calculated, and the factor with the largest mean absolute value was extracted as the dominant disaster-causing factor of that patch.

[0073] Finally, adjacent patches with the same dominant disaster-causing factors are spatially stitched together to form spatially heterogeneous zones with homogeneous internal physical disaster-causing patterns and independent zones.

[0074] This scheme achieves objective optimal discretization of continuous coefficient surfaces through a classification goodness index, avoiding subjective bias caused by manually set thresholds and ensuring that the partition boundaries conform to the spatial variation law of regression coefficients. Through spatial topological overlay and intelligent extraction of dominant factors, it accurately identifies the core driving factors of landslide formation in each region, making the disaster-causing mechanisms within each partition highly homogeneous and the differences between partitions significant. Finally, it reconstructs the complex study area into multiple independent evaluation units, fundamentally avoiding the problem of local accuracy dilution caused by uniform modeling across the entire region, and greatly improving the pertinence and accuracy of subsequent landslide susceptibility prediction.

[0075] Furthermore, in step S5, the landslide susceptibility simulation based on spatial heterogeneity partitions and the prediction quality evaluation using multi-dimensional indicators include:

[0076] Each spatially heterogeneous zone is treated as an independent unit, and the probability of landslide susceptibility is extrapolated separately.

[0077] The simulation results are sequentially subjected to accuracy benchmark diagnosis, area rationality diagnosis and prediction uncertainty robustness mapping to complete the multidimensional prediction quality evaluation.

[0078] The accuracy benchmark diagnosis method is as follows: construct the receiver operating characteristic (ROC) curve, calculate the area under the curve (AUC), evaluate the local prediction accuracy within a single partition and the global comprehensive accuracy after splicing the probabilities of each partition, and establish the accuracy benchmark for susceptibility prediction.

[0079] The method for diagnosing the rationality of the area is as follows: a prediction efficiency curve is constructed with the cumulative proportion of the area of ​​the high-risk interval as the horizontal axis and the cumulative proportion of the corresponding historical landslides as the vertical axis. The rationality of the predicted area distribution is quantitatively evaluated based on the criteria that the high-risk interval is small and the area covers more historical landslide points.

[0080] The method for mapping the robustness of prediction uncertainty is as follows: multiple independent datasets are generated by stratified random sampling within each partition, and the training set and validation set are divided proportionally. Multiple landslide susceptibility probability values ​​are obtained by repeatedly running the classifier. The mean and standard deviation of the prediction probability for each spatial location are calculated to achieve robustness assessment of the prediction results.

[0081] In this scheme, spatial heterogeneous partitions are used as units for independent simulation, which can avoid the problem of local accuracy dilution caused by uniform modeling of the whole domain. The three-dimensional quality evaluation system breaks through the limitations of traditional single accuracy indicators. While ensuring the global prediction accuracy, it improves the efficiency of high-risk area locking and the stability of prediction results, making the model output more in line with the actual disaster prevention and mitigation engineering application needs.

[0082] Secondly, the present invention also provides an electronic device, including a processor and a memory; the memory stores a computer program; when the processor executes the computer program, it implements the steps in the above-mentioned landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning.

[0083] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon; when the computer program is executed by a processor, it implements the steps in the above-mentioned landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning.

[0084] The beneficial effects of this invention are:

[0085] (1) It overcomes the feature smoothing defect caused by a single spatial scale and improves the quantification accuracy of disaster-causing mechanisms:

[0086] Regarding the quantitative accuracy of disaster-causing mechanisms, this invention introduces a multi-scale geographically weighted regression model with a backfit correction mechanism. This model adaptively allocates an independent spatial search bandwidth to each disaster-causing factor, dynamically separating macro-background factors from micro-activating factors based on the physical influence range of different environmental factors. Existing traditional geographically weighted regression models use a fixed and unique spatial scale, inevitably leading to over-smoothing of local features or ineffective fragmentation of global features. In contrast, this invention significantly improves the global goodness of fit and substantially reduces the sum of squared residuals, accurately reproducing spatially non-stationary disaster mechanisms that conform to the true attributes of nature, laying a solid algorithmic foundation for subsequent high-precision risk prediction.

[0087] (2) It eliminates the subjective arbitrariness of feature space partitioning and realizes unbiased reconstruction of discrete data into a continuous surface:

[0088] Regarding the objective reconstruction of spatial features, this invention combines ordinary kriging spatial interpolation with the natural discontinuity method to continuously reconstruct discrete local regression coefficients and classify optimal attributes. This mechanism first generates an unbiased, continuous feature surface by fitting a theoretical semivariance model and solving the kriging equations. Then, it automatically determines the optimal discrete discontinuity point of the attribute, targeting the extreme value of the classification goodness index. Existing technologies rely on manual experience to set classification thresholds or use simple equidistant divisions, inevitably introducing subjective uncertainty at the cognitive level. In contrast, this invention fundamentally eliminates spatial boundary distortion caused by human intervention, significantly reducing the engineering complexity of disaster prevention system data processing and providing a clear, objective, and rigorous physical boundary basis for refined geological management.

[0089] (3) It avoids the local precision dilution caused by the homogenization spatial assumption and improves the spatial identification of high-risk and prone areas:

[0090] Regarding the spatial identification accuracy of local high-risk areas, this invention constructs a disaster scene spatial reconstruction mechanism centered on the extraction of "dominant disaster-causing factors." Through spatial topological overlay, a large-scale watershed is deconstructed into multiple independent evaluation zones. This mechanism dynamically extracts the feature with the largest absolute value of the average regression coefficient from subdivided patches as the dominant factor, stitching together areas with highly homogeneous internal landslide coupling processes into independent working units, and assigning differentiated prediction benchmarks to different mapping units. Existing technologies treat the entire complex study area as a uniform space, leading to a continuous dilution of local prediction accuracy in low-correlation areas. In contrast, this invention effectively avoids this dilution effect, ensuring that the accuracy of both local and global susceptibility predictions remains stable at a high level, significantly enhancing the practical applicability of the assessment model under complex terrain and geological conditions.

[0091] (4) A three-dimensional evaluation system that takes into account both area efficiency and anti-interference ability has been established, which enhances the engineering application value of the prediction results:

[0092] Regarding the engineering application value of the prediction results, this invention establishes a three-dimensional comprehensive diagnostic system encompassing the receiver operating characteristic curve, prediction efficiency curve, and Monte Carlo random sampling robustness mapping. This system not only quantitatively verifies the effectiveness of locking onto historical real disaster points under conditions of minimal high-risk area proportion, but also spatially and explicitly isolates and identifies prediction fluctuations caused by underlying data disturbances by tracking the variance and range of single pixels through multiple random sampling. Existing technologies rely solely on a single global accuracy indicator, which cannot measure the utilization efficiency of disaster prevention resources, nor can it assess the risk of accidental misjudgments by the model. In contrast, this invention achieves a substantial improvement in prediction area efficiency and strictly controls global uncertainty within a safe threshold, thereby breaking down the key barrier between pure theoretical deduction and engineering practice, and significantly enhancing the guiding value of susceptibility zoning maps in on-site geological investigation and disaster prevention resource allocation. Attached Figure Description

[0093] Figure 1 This is a technical roadmap for the landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning in this invention.

[0094] Figure 2 This is a logic diagram of the multi-scale geographically weighted regression backfit correction algorithm in this invention.

[0095] Figure 3 This is a schematic diagram of the polarization phenomenon of the regression coefficient of the pregnancy factor and the spatial heterogeneity partitioning and recombination in Embodiment 1 of the present invention.

[0096] Figure 4 This is a line graph comparing the prediction efficiency curves output in Embodiment 1 of the present invention.

[0097] Figure 5 This is a comprehensive diagnostic diagram of the multi-scale model for high mountain and deep canyon areas in Embodiment 1 of the present invention, wherein (a) is a bar chart of the scale polarization of the spatial effect of disaster-causing factors; (b) is a line chart of the dual effects of global prediction accuracy and area efficiency; and (c) is a scatter plot of the Monte Carlo random sampling uncertainty robustness.

[0098] Figure 6 This is a hardware architecture and logical topology diagram of the wide-area distributed landslide prediction system in Embodiment 2 of the present invention. Detailed Implementation

[0099] This invention aims to provide a landslide susceptibility prediction method based on multi-scale geographically weighted regression and spatial heterogeneous zoning, addressing the problems of low local accuracy, poor area efficiency, and insufficient prediction stability caused by traditional landslide prediction models due to their single spatial scale, homogeneous modeling, and singular evaluation system. The core idea is to organically combine multi-scale geographically weighted regression with spatial heterogeneous zoning. First, a multi-scale regression model adaptively allocates optimal bandwidth to each landslide-prone factor, accurately characterizing the spatial non-stationary characteristics of landslide-prone relationships. Then, Kriging interpolation is used to reconstruct the discrete regression coefficients into a spatially continuous surface. The dominant factors are dynamically extracted using optimal discretization and natural discontinuity methods to achieve heterogeneous zoning of the study area. Landslide susceptibility is extrapolated using each zone as an independent unit. A multi-dimensional comprehensive evaluation system is constructed from three dimensions: prediction accuracy, area rationality, and uncertainty robustness. This comprehensively improves the accuracy of landslide prediction, the efficiency of high-risk area identification, and the stability of results, forming a landslide susceptibility prediction scheme that conforms to real geological laws and possesses both scientific rigor and engineering practicality.

[0100] Based on the above core ideas, the core innovations of this invention include at least the following points:

[0101] (1) Spatial nonstationarity adaptive quantization architecture based on multi-scale geographic weighted regression (MGWR):

[0102] This architecture introduces a computational process with a backfit correction mechanism. By dynamically solving for the optimal parameters of the spatial kernel function, it adaptively allocates independent spatial search bandwidth to each disaster-causing factor, accurately distinguishing between globally homogeneous control factors and locally heterogeneous triggering factors from the underlying computing power architecture. This design fundamentally overcomes the technical bottleneck of traditional assessment models, which use a single fixed spatial scale, leading to the smooth masking of local disaster-causing mechanisms.

[0103] (2) Disaster scenario spatial reorganization mechanism driven by the extraction of "dominant disaster-causing factors":

[0104] This mechanism proposes an intelligent spatial heterogeneity partitioning method: by performing spatial topological overlay on continuous layers of various disaster-causing factors, it dynamically calculates and extracts the feature with the largest absolute value of the average regression coefficient in the subdivided patches as the dominant factor of the region, and then stitches adjacent patches with the same dominant factor into independent evaluation units. This approach completely abandons the inherent assumptions of traditional large-scale watershed homogeneous modeling, deconstructing the complex study area into a microscopic working space with highly homogeneous internal landslide coupling processes, thereby significantly improving the model's spatial identification accuracy of local high-risk areas.

[0105] (3) Unbiased continuous surface reconstruction method combining Kriging interpolation and natural discontinuity method:

[0106] This method constructs a complete feature reconstruction algorithm for discrete spatial regression coefficients: by establishing a theoretical semivariance model including range and sill values ​​to solve the Kriging equations, a statistically unbiased continuous coefficient surface is generated; subsequently, the natural discontinuity method is seamlessly nested to achieve automatic discretization of attribute boundaries with the goal of maximizing the classification goodness index. This processing method accurately transforms discrete mathematical parameters into a continuous feature layer with clear physical boundaries, fundamentally eliminating the subjective arbitrariness of manually setting classification thresholds and providing rigorous underlying data support for the aforementioned spatial heterogeneity partitioning.

[0107] (4) A three-dimensional disaster assessment system covering "accuracy, area efficiency, and uncertainty robustness":

[0108] This system innovatively introduces a prediction efficiency curve, building upon traditional global and local subject operating characteristic curves, to quantitatively characterize the nonlinear coupling relationship between the cumulative area ratio of the prediction probability interval and the historical disaster lock-in rate. Simultaneously, based on multiple Monte Carlo random samplings, it achieves spatial explicit mapping of robustness by calculating twice the standard deviation of a single pixel. This design not only quantitatively verifies the model's disaster interception effectiveness under extremely small control areas for the first time, but also effectively eliminates the risk of misjudgment of disaster prevention levels caused by accidental disturbances in the underlying data through spatialization, thus breaking down the key barrier between pure theoretical deduction and practical engineering resource scheduling.

[0109] The technical approach of the landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning provided by this invention is as follows: Figure 1 As shown, it is divided into five layers: data fusion and processing layer, multi-scale regression modeling layer, spatial heterogeneity intelligent partitioning layer, closed-loop prediction and three-dimensional evaluation layer, and engineering output layer. Each layer is progressive, and the specific process is as follows:

[0110] I. Data Fusion Processing Layer:

[0111] Using high-resolution remote sensing images, historical disaster catalogs, digital elevation models (DEMs), and regional meteorological and hydrological time-series data as inputs, spatial gridding and response variable (landslide density) extraction are first performed. Then, core disaster-causing factors are screened through Pearson test and VIF collinearity diagnosis. Finally, a high-dimensional feature space matrix is ​​constructed to provide a high-quality and non-redundant input foundation for subsequent model building.

[0112] II. Multi-scale regression optimization layer:

[0113] The MGWR multi-scale quantization architecture is adopted. First, a multi-scale geographically weighted regression (MGWR) model is constructed based on the feature space. By introducing a backfit algorithm, each disaster-causing factor is adaptively allocated its own optimal spatial bandwidth to achieve accurate identification of the scale of action of disaster-causing factors. Then, the local regression coefficient array is iteratively updated until the model converges, thus completing a high-precision quantitative characterization of the spatial nonstationarity of landslide disaster-causing mechanisms.

[0114] III. Spatial Heterogeneity Intelligent Partitioning Layer:

[0115] For the discrete local regression coefficients output by the regression model, the semivariance function is first constructed using the Kriging interpolation algorithm, and the optimal unbiased estimation weights are solved to generate a spatially continuous, smooth, and unbiased global regression coefficient surface. Then, based on the continuous coefficient surface, spatial heterogeneity intelligent partitioning is performed. The optimal discretization is achieved through Jenks' natural discontinuity method, and the "dominant disaster-causing factor" is extracted using GSI classification index and overlay analysis. Finally, adjacent patches with consistent dominant factors are integrated to form a spatially heterogeneous partitioning result with homogeneous internal disaster-causing mechanisms and clear boundaries.

[0116] IV. Closed-loop prediction and three-dimensional evaluation layer:

[0117] Using each spatially heterogeneous zone as an independent evaluation unit, landslide susceptibility probability extrapolation is carried out separately, and a three-dimensional evaluation system is constructed to complete the model performance diagnosis. This includes using ROC curves and AUC values ​​to achieve global and local accuracy benchmark diagnosis, conducting area rationality diagnosis based on prediction efficiency curves, and using Monte Carlo sampling to achieve prediction uncertainty robustness mapping, ultimately forming a multi-dimensional, all-angle model performance evaluation closed loop.

[0118] V. Engineering Output Layer:

[0119] Based on the multidimensional evaluation results, a high-precision regional landslide susceptibility prediction map is output. Combined with the risk warning threshold, a landslide susceptibility prediction map and a risk warning alarm database are constructed, which directly serves the practical engineering needs of regional landslide disaster prevention and mitigation, risk management and emergency command.

[0120] In practical implementation, the landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning provided by this invention includes the following implementation process:

[0121] S1. Fusion processing and feature space construction of multi-source environmental disaster-prone data:

[0122] The purpose of this step is to integrate multi-source heterogeneous geospatial data, complete data standardization, redundancy removal, and feature purification, and construct a core disaster-causing factor feature space that is non-collinear and highly representative, providing a high-quality, unbiased input foundation for subsequent multi-scale regression modeling. In one exemplary implementation scheme, the following implementation methods are included:

[0123] (1) Multi-source heterogeneous data acquisition and spatial gridded discretization:

[0124] In terms of data acquisition and spatial gridding, the system first calls the geospatial data acquisition interface to integrate multi-source heterogeneous inputs, including high-resolution remote sensing satellite image interpretation results, historical geological disaster cataloging data, digital elevation models, and regional meteorological and hydrological time-series data. To establish the basic framework for spatial computation, the system constructs a regular grid with a fixed resolution (preferably 30m×30m) as the basic mapping unit based on the geomorphic span of the study area, and uses a spatial kernel density estimation algorithm to calculate the landslide density value (unit: landslides / square kilometer) within each mapping unit, which is then used as a continuous response variable for system simulation.

[0125] (2) Collinearity diagnosis and dimensionality reduction of high-dimensional feature space:

[0126] In terms of collinearity diagnosis and feature space dimensionality reduction, to avoid the matrix singularity problem in high-dimensional data, the system sequentially performs Pearson correlation test and variance inflation factor (VIF) collinearity diagnosis on the initially extracted environmental factors. The VIF is calculated as follows:

[0127] (1);

[0128] in, For the first The variance inflation factor of each pregnancy factor; To make the first The coefficient of determination obtained when performing multiple linear regression with one factor as the dependent variable and all other candidate factors as independent variables.

[0129] The system sets strict cutoff thresholds and automatically excludes [illegible characters]. or Pearson correlation coefficient Redundant factors. After the above data cleaning, the system finally constructs a landslide feature space composed of core disaster-causing factors, and outputs a four-dimensional feature vector: terrain category, engineering geological rock group, distance from river, and melting index.

[0130] S2. Multi-scale geographically weighted regression modeling based on backfitting algorithm:

[0131] The purpose of this step is to overcome the limitations of traditional global regression and single-scale regression, to adaptively match the optimal spatial scale of action for different disaster-causing factors, and to accurately quantify the spatial heterogeneity between disaster-causing factors and landslide evolution. In one exemplary implementation scheme, the logic of the multi-scale geographically weighted regression backfit algorithm is as follows: Figure 2 As shown, it includes:

[0132] (1) Global initialization of the MGWR model:

[0133] For global model initialization, the system constructs the following baseline mathematical equations to analyze the differential effects of each explanatory variable on the response variable at local, regional, and global scales:

[0134] (2);

[0135] in, Indicates the first spatial grid location The response variable (landslide density) on the surface. Indicates the first position at that location Explanatory variables (such as engineering geological rock groups); For the first The optimal spatial bandwidth corresponding to each explanatory variable; For the first The explanatory variable in the th... spatial grid location The local regression coefficient at a given location, which is based on the optimal spatial bandwidth of the corresponding explanatory variable. Obtained through fitting; For spatial intercept; To obtain white noise residuals that conform to a Gaussian distribution.

[0136] (2) Partial residual separation and spatial weight kernel function optimization:

[0137] In terms of partial residual separation and spatial weight kernel function optimization, since the equation contains Each bandwidth is independently variable. The system embeds a Generalized Additive Model (GAM) logic and executes a dynamic iterative backfit algorithm to solve the problem.

[0138] The system first uses the classic geographically weighted regression (GWR) algorithm or ordinary least squares (OLS) to complete the baseline fitting and obtain the initial regression summation term. Then, it sequentially isolates the partial residuals of individual target variables. :

[0139] (3);

[0140] in, For the first The partial residual of the nth variable, which purifies the system only with respect to the nth variable. The residual variance associated with each variable; The actual observed values ​​of the response variable; No. The current regression summation term of all variables other than the one variable; For the first A regression summation term for each variable; It represents the residual.

[0141] Construct a distance-decay spatial weight kernel function for local weighted calculation:

[0142] (4);

[0143] in, For observation point For target point Spatial weights; The Euclidean distance between the two points; This represents the current trial bandwidth. When... hour, .

[0144] The system uses the modified Akaike Information Criterion (AICc) as the fitness evaluation function, and iteratively searches using the golden section method until it locks the exclusive optimal bandwidth that minimizes AICc, and then updates the regression coefficient matrix.

[0145] (3) Parameter array update and convergence determination of multidimensional matrix:

[0146] The system iterates through all factors in the feature space to complete a single full iteration. To prevent model divergence, the system monitors the magnitude of change in the local coefficient array between consecutive iterations in real time, introducing the sum of squared changes ratio (SOC) as a hard criterion for model convergence.

[0147] (5);

[0148] in, This represents the total number of pregnancy-related factors. This represents the total number of spatial grid cells; and Representing the first In the current iteration and the previous iteration, the factor is located at the th . The local regression coefficient values ​​at each grid location.

[0149] When the calculated SOC value is lower than the system's preset minimum tolerance threshold (preferred) When the regression algorithm converges, the system determines that the regression algorithm has converged and outputs a set of local regression coefficients with the optimal bandwidth for each factor.

[0150] S3. Spatial continuity of regression coefficients and modeling of variation structure based on ordinary Kriging interpolation:

[0151] The purpose of this step is to reconstruct the discretely distributed local regression coefficients into a globally continuous, smooth, and unbiased regression coefficient surface, fully characterizing the spatial variation structure of the regression coefficients, eliminating spatial expression discontinuities at discrete points, and providing continuous and unified basic data for subsequent heterogeneous partitioning. In one exemplary implementation scheme, the following implementation means are included:

[0152] (1) Construction of the experimental semivariance function:

[0153] The system extracts the set of regression coefficients for any pregnancy-related factors exhibiting local heterogeneity. The spatial autocorrelation structure was detected by calculating the experimental semivariance function:

[0154] (6);

[0155] in, The lag distance is The experimental semivariance value at time; The spatial distance lag step; Spatial distance interval equal to The total number of valid sample point pairs; and Spatial location and its distance The regression coefficient value at the location.

[0156] Based on the fitted theoretical semivariance function, the system targets any unmodeled grid location within the study area. The regression coefficient value is predicted by using a linear combination of known surrounding sample points. :

[0157] (7);

[0158] in, Location to be estimated The predicted values ​​of the regression coefficients; This represents the number of known neighboring sample points used in the estimation. To be assigned to the Unbiased estimated weights for known sample points.

[0159] (2) Fitting the theoretical variation model and assigning unbiased weights:

[0160] In terms of fitting the theoretical variability model and allocating unbiased weights, the system employs the least squares method, using theoretical variability functions such as spherical and exponential models to fit the experimental semivariogram data, extracting three core spatial variability parameters: nugget value, sill value, and range. For any grid location to be estimated, the system constructs and solves the Kriging equations to allocate the optimal unbiased estimation weights.

[0161] (8);

[0162] in, For known sample points and The theoretical semivariance between them; For known sample points And the point to be estimated The theoretical semivariance between them; is the Lagrange multiplier, used to ensure the unbiased estimation condition that the sum of the weights is 1.

[0163] After system calculation, a smooth, unbiased continuous layer of regression coefficients is generated.

[0164] S4. Spatial heterogeneity partitioning based on the "dominant factor" discrete attribute optimization:

[0165] The purpose of this step is to achieve intelligent and objective partitioning of the study area based on the continuous regression coefficient surface, ensuring that the disaster-causing mechanisms within each partition are homogeneous and the differences between partitions are significant, thereby fundamentally avoiding the accuracy dilution problem of uniform modeling across the entire domain. An exemplary implementation scheme includes the following implementation methods:

[0166] (1) Attribute data optimization clustering discretization:

[0167] The system calculates the sum of squares of the deviations from the mean of global data in a single factor layer. Sum of squared deviations from the mean within subclasses :

[0168] (9);

[0169] (10);

[0170] in, The total number of pixels involved in the classification; For the first The regression coefficient values ​​for each pixel; This is the global mean. For the first Subclasses; Its within-class mean; For the first Within each subclass The regression coefficient value of each pixel.

[0171] The system introduces the Goodness-of-Class (GVF) index as an evaluation metric:

[0172] (11);

[0173] The system continuously adjusts the combination of discontinuities through enumeration or heuristic search, locks the optimal discontinuity scheme that makes the GVF reach the global maximum, and completes the optimal discretization of continuous variables.

[0174] (2) Spatial topological overlay and intelligent extraction of dominant factors:

[0175] In terms of spatial topological overlay and intelligent extraction of dominant factors, the system performs GIS spatial topological overlay operations on all discretized factor layers, dividing the study area into subdivided polygonal patches with single attributes. Within each patch, the system calculates the mean of the absolute values ​​of the regression coefficients of each relevant factor, and automatically extracts the factor with the largest mean absolute value as the "dominant disaster-causing factor" for that patch. Finally, the system spatially stitches adjacent patches with the same dominant factor characteristics, forming spatially heterogeneous partitions with homogeneous internal physical disaster-causing laws and independent operation between different areas.

[0176] S5. Landslide Susceptibility Deduction and Multi-Dimensional Prediction Quality Assessment:

[0177] The purpose of this step is to conduct accurate landslide susceptibility simulations using spatially heterogeneous zones as independent units, and to construct a multi-dimensional evaluation system to comprehensively verify the model's prediction accuracy, area rationality, and robustness, ultimately outputting landslide susceptibility prediction results that meet engineering application requirements. An exemplary implementation plan includes the following implementation methods:

[0178] The system treats each spatially heterogeneous partition as an independent working unit, performing landslide susceptibility probability inferences separately. To overcome the limitations of conventional single-indicator verification, the system embeds a three-dimensional quality assessment center to comprehensively diagnose model convergence and prediction reliability.

[0179] (1) Global and local accuracy benchmark diagnosis:

[0180] In terms of accuracy benchmark diagnosis, the system constructs receiver operating characteristic (ROC) curves and calculates the area under the curve (AUC). It evaluates the local AUC value within a single partition and the global comprehensive AUC value after splicing the probabilities of each partition, thereby establishing a benchmark accuracy standard for susceptibility prediction.

[0181] (2) Reasonableness diagnosis based on the area of ​​the predicted efficiency curve:

[0182] In diagnosing the reasonableness of the predicted area, the system introduces the Prediction Efficiency Curve (PEC) as a supplementary evaluation tool. Its core logic is: given a relatively low proportion of high-susceptibility landslide areas, the system should cover as many historical landslide points as possible; simultaneously, low-susceptibility areas should occupy a large proportion of the area and ideally not include any historical landslide points, thus ensuring a good match between the spatial distribution of high- and low-susceptibility areas and the actual historical landslide pattern. To quantitatively characterize the evolutionary relationship between the proportion of susceptibility areas and the number of historical landslides, the system constructs a prediction efficiency curve with the cumulative proportion of each area on the horizontal axis and the corresponding cumulative proportion of historical landslides on the vertical axis, and calculates the area under the curve. The larger the area under the curve, the more reasonable the predicted area distribution is, indicating that the model can still cover more landslide points while maintaining a relatively small proportion of high-susceptibility areas.

[0183] (3) Robustness mapping of prediction uncertainty based on Monte Carlo sampling:

[0184] Regarding the robustness mapping of prediction uncertainty, to quantitatively evaluate the random variation characteristics of landslide susceptibility prediction results, the system designed the following quantification process: First, stratified random sampling was used to generate 20 independent datasets in each spatially heterogeneous partition. Each dataset was randomly matched with landslide and non-landslide samples at a 1:1 ratio, and the training and validation sets were divided at a 7:3 ratio. Then, the 20 datasets were independently and repeatedly run using the Extra Trees classifier algorithm to obtain 20 landslide susceptibility probability values ​​at the same spatial location. Finally, the system plotted an uncertainty scatter plot and a fitting curve by statistically analyzing the mean and standard deviation of the predicted probability at each location. When the scatter plot shows an overall flat distribution, it indicates that the prediction results have good robustness and a strong ability to suppress random bias.

[0185] Example 1:

[0186] This embodiment takes the Yuqu high mountain deep canyon watershed in Tibet as the study area and fully demonstrates the landslide susceptibility prediction process based on multi-scale geographical weighted regression and spatial heterogeneity zoning.

[0187] In terms of system composition and structure, this embodiment relies on a single-node workstation equipped with a high-performance multi-core processor and large-capacity memory as the hardware execution carrier. The internal logic of the system is divided into a multi-source data fusion module, an MGWR regression optimization engine, a Kriging space reconstruction module, a Jenks partitioning module, and a multi-dimensional quality evaluation center. Each functional module relies on the workstation's internal high-speed bus and shared memory space to achieve low-latency transmission of high-dimensional matrix data and operation instructions.

[0188] The system first activates the multi-source data fusion module, loads historical landslide locations in the Yuqu River basin, and then... High-resolution regular grids cover approximately [amount missing] of the total area. The study area is spatially discretized into individual grids. After rigorous screening for correlation and VIF collinearity, the system finally outputs a four-dimensional core feature vector consisting of terrain category, engineering geological rock group (EGR), distance from river (DR), and melting index (MC).

[0189] Subsequently, the system's MGWR regression optimization engine initiates a backfit correction algorithm. During this process, the system sets a minimum convergence tolerance threshold. The calculation results show that the terrain category is identified as a globally homogeneous factor with a bandwidth ratio as high as 0.64; the engineering geological rock group, distance from the river, and melting index are accurately identified as strongly local heterogeneous factors with bandwidth ratios between 0.04 and 0.09, thus accurately distinguishing the differences in the scale of action of different disaster-causing factors.

[0190] Next, the Kriging space reconstruction module extracts discrete regression coefficients and performs theoretical semivariance fitting to generate an unbiased continuous surface. The Jenks partitioning module sets the target number of classifications. The system automatically optimizes and finds the breakpoint with the highest goodness-of-class (GVF) index, dividing the continuous layer into three physical regions: strongly negatively correlated, weakly correlated, and strongly positively correlated. Through spatial topological overlay, the system finely reorganizes the entire watershed into 18 independent spatially heterogeneous partitions. For example... Figure 3 The diagram shown is a schematic diagram of the polarization phenomenon of the regression coefficients of disaster-causing factors and the spatial heterogeneity of the partitions in this embodiment. It intuitively presents the spatial differences and scale polarization characteristics of the regression coefficients of each disaster-causing factor, as well as the spatial pattern after partitioning and reorganization based on the extraction of the dominant factor. Different partitions correspond to different dominant disaster-causing factors and disaster-causing mechanisms.

[0191] This embodiment strictly adheres to the core technical approach of extracting dominant disaster-causing factors: through bottom-level grid overlay and coefficient statistics, the average regression coefficient of the engineering geological rock group in partition 1 has the largest absolute value, reaching 0.61, and is identified as the EGR dominant area; the average regression coefficient of the distance from the river in partition 7 has the largest absolute value, reaching 0.23, and is identified as the DR dominant area. This cuts off the statistical interference of different physical driving mechanisms and ensures the homogeneity of disaster logic within each partition.

[0192] After partitioning, the system uses the 18 spatially heterogeneous partitions as independent units, calls the ExtraTrees Classifier algorithm to perform landslide susceptibility probability inference, and verifies the model effectiveness through multi-dimensional indicators.

[0193] like Figure 4 As shown, this is a line graph comparing the prediction efficiency curves output in this embodiment. It quantitatively depicts the matching relationship between the area proportion of high-risk zones and the coverage rate of historical landslide points, intuitively demonstrating the area rationality advantage of this invention compared to traditional methods. Figure 5The diagram shown is a comprehensive diagnostic chart of the multi-scale model's multi-dimensional performance in this embodiment, where: Figure 5 (a) is a scale polarization bar chart of the spatial effects of disaster-causing factors, which clearly shows the scale differences between topographical categories as global factors and engineering geological rock groups as local factors; Figure 5 (b) is a line graph showing the combined effect of global prediction accuracy and area efficiency, which simultaneously reflects the synergistic improvement in model accuracy and investigation efficiency; Figure 5 (c) is a scatter plot of Monte Carlo random sampling uncertainty robustness, used to characterize the ability of the prediction results to resist random disturbances.

[0194] Technical performance verification shows that the proposed solution achieves significant technical improvements in the Yuqu River Basin: In terms of accuracy diagnosis, the global AUC value after reconstruction increases from 0.93 in the traditional baseline model to over 0.95; in terms of area rationality diagnosis, combined with… Figure 4 and Figure 5 (b) It is evident that for some optimized zoning, only the top 20% of the high-risk red areas need to be investigated to identify 100% of the historical landslide points, improving area accuracy by 9% compared to traditional methods; regarding uncertainty robustness, combined with Figure 5 (c) shows that after 20 independent Monte Carlo sampling iterations, the quadratic standard deviation and range of over 94% of the regions were strictly suppressed within the safe threshold of 0.2. The above quantitative data fully demonstrates the progress this case study has made in improving the explanatory power of disaster-causing mechanisms and the efficiency of disaster prevention resource allocation.

[0195] Example 2:

[0196] This embodiment is an extended application for large-scale regions such as the southeastern edge of the Qinghai-Tibet Plateau and cross-provincial natural zones. It aims to demonstrate the hardware cluster adaptability and algorithm generalization ability of the present invention when dealing with ultra-large-scale spatial data throughput and multi-source dynamic factor access.

[0197] In terms of system composition and structure, such as Figure 6 The diagram shows the hardware architecture and logical topology of the wide-area distributed landslide prediction system used in this embodiment. Completely different from the single-machine workstation architecture of Embodiment 1, this embodiment uses a wide-area distributed cloud server cluster as the physical hardware foundation. The overall technical link progresses sequentially along the external data input layer → data preprocessing layer → static balancing foundation & load layer → distributed response core computing layer → collaborative inference & output layer. Each layer collaborates to complete the entire process of wide-area landslide susceptibility prediction.

[0198] External data input layer: The system accesses two types of core data sources. One type is conventional static environmental feature sources, covering traditional static geospatial data such as topography, geology, hydrology, and meteorology. The other type is time-series InSAR dynamic deformation maps, which are used to obtain the temporal dynamic features of surface micro-deformation, forming a multi-source data input system that combines static and dynamic data.

[0199] Data preprocessing layer: Both types of data sources are uniformly integrated into the edge computing aggregation node to complete front-end data cleaning, format standardization, coordinate registration and preliminary aggregation, providing a unified and standardized data entry point for subsequent distributed processing.

[0200] Static balancing base & load layer: The aggregated multi-source data is dynamically scheduled through the load balancing gateway to distribute data tasks evenly to the backend distributed computing cluster, avoiding overload of single-node computing power and ensuring low-latency and high-stability processing of ultra-large-scale data.

[0201] Distributed Response Core Computing Layer: The cluster is divided into three core functional modules that work together to complete core computations:

[0202] A. Distributed Fusion Module for Streaming Data: It analyzes and accesses a five-dimensional environmental response matrix that combines static and dynamic data. For a wide-area grid scale of tens of millions of grid cells, it adaptively lowers the spatial discretization accuracy and uses a 90m×90m regular grid to complete the discretization of the entire study area.

[0203] B.MGWR Parallel Solving Engine: Executes a dynamic step-size backfit algorithm, appropriately relaxing the convergence tolerance threshold. It achieves an optimal balance between prediction accuracy and computing power consumption, and adaptively allocates dedicated optimal spatial bandwidth to each disaster-prone factor.

[0204] C. Multi-core concurrent spatial heterogeneity recombination module: Generates a continuous regression coefficient surface through Kriging unbiased interpolation, adjusts the target number of Jenks' natural discontinuity method to K=5, achieves a delicate transition division of 5 attributes, and then extracts dozens or even hundreds of macroscopic heterogeneous working bands through high-density spatial topology overlay.

[0205] Collaborative Inference & Output Layer: The core computation results are input into the distributed ensemble learning inference center, which calls the Extreme Gradient Boosting Tree (XGBoost) algorithm to independently perform landslide susceptibility probability inference on each heterogeneous working zone in a multi-process synchronous manner on distributed nodes; the inference results are input into the wide-area three-dimensional evaluation system to complete the full-process performance verification.

[0206] In terms of differentiated design of working principles and implementation steps, this embodiment is fully adapted to the needs of wide-area distributed scenarios: For the massive data volume of tens of millions of grids, the system adaptively reduces the resolution of the basic mapping unit to a 90m×90m regular grid, significantly reducing computing power consumption while ensuring regional prediction accuracy; in the MGWR model backfit calculation stage, a dynamic step size adjustment strategy is adopted and the convergence tolerance threshold is relaxed to balance model accuracy and computing power efficiency; in the spatial heterogeneity partitioning stage, the number of target classifications is expanded to 5 levels to capture the attribute transition characteristics of the complex geological environment in a more delicate way, and dozens to hundreds of macroscopic heterogeneous working zones containing dynamic deformation dominant factors are formed through high-density topological overlay; finally, the XGBoost algorithm with higher parallel efficiency is adopted, and the susceptibility probability inference of each partition is completed synchronously with the distributed cloud cluster, fully releasing the computing power advantages of the distributed architecture.

[0207] In terms of performance verification, this embodiment relies on a three-dimensional evaluation system to achieve closed-loop performance verification in a wide-area scenario: such as Figure 6 As shown, the PEC efficiency verification results indicate that by investigating only the top 15% of high-risk areas across the entire region, over 90% of historical disaster hazard points can be identified, significantly reducing the ineffective costs of wide-area disaster prevention and investigation. The global accuracy verification results show that the model's global AUC value remains consistently at an excellent level of AUC≥0.92. Meanwhile, relying on the distributed architecture and optimized parallel solution strategy, the overall optimization time is reduced by >60% compared to traditional single-machine full computation, achieving a leapfrog improvement in computing power efficiency.

[0208] The above results fully verify that the technical architecture of this invention, which includes "unbiased continuous surface reconstruction" and "dominant factor partitioning", still has extremely high technical stability and engineering reuse value when crossing different hardware platforms, different mesh accuracies and different algorithm classifiers.

[0209] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning, characterized in that, Includes the following steps: S1. Multi-source pregnancy and malformation data are fused and processed, and the study area is discretized by a regular grid space. Feature optimization is completed through collinearity diagnosis to construct a multi-dimensional pregnancy and malformation feature space. S2. A multi-scale geographically weighted regression model is used to adaptively allocate independent bandwidth to each disaster-prone factor, and the local regression coefficients are obtained by fitting. S3. Perform Kriging interpolation on the local regression coefficients to generate a spatially continuous regression coefficient surface; S4. Based on the regression coefficient surface, perform optimal discretization and dominant factor extraction to complete spatial heterogeneity partitioning; S5. Landslide susceptibility is extrapolated using spatially heterogeneous zoning as the unit, and the prediction quality is evaluated through multi-dimensional indicators.

2. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 1, characterized in that, In step S1, the multi-source disaster-prone data includes remote sensing satellite image interpretation results, historical geological disaster cataloging data, digital elevation models, and regional meteorological and hydrological time-series data; The methods for fusing and processing multi-source pregnancy and malpregnancy data include: By performing unified coordinate benchmark transformation and projection transformation on all data through a geographic information system, a structured continuous spatial layer with spatial matching and unified format is formed.

3. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 1, characterized in that, In step S1, the process of performing regular grid spatial discretization on the study area and completing feature optimization through collinearity diagnosis includes: Using a set resolution regular grid as the basic mapping unit, the spatial kernel density estimation algorithm is used to calculate the landslide density as a continuous response variable. Collinearity diagnosis is performed by Pearson correlation test and variance inflation factor, and redundant factors are eliminated to construct a multidimensional disaster-prone feature space. The variance inflation factor is calculated as follows: ； in, For the first The variance inflation factor of each pregnancy factor; To make the first The coefficient of determination obtained when performing multiple linear regression with one factor as the dependent variable and all other factors as independent variables. The method for diagnosing collinearity using Pearson correlation test and variance inflation factor is as follows: Factors with a variance inflation factor greater than 10 or a Pearson correlation coefficient greater than 0.7 are considered as redundant factors.

4. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 3, characterized in that, In step S1, the constructed multidimensional disaster-prone feature space includes four types of features: terrain category, engineering geological rock group, distance from river, and melting index.

5. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 1, characterized in that, In step S2, the adoption of a multi-scale geographically weighted regression model to adaptively allocate independent bandwidths to each disaster risk factor and fit local regression coefficients includes: Construct a multi-scale geographically weighted regression baseline equation to achieve independent scale modeling of each factor: ； in, Indicates the first spatial grid location The response variable on; Indicates the first spatial grid location The first One explanatory variable; For the first The optimal spatial bandwidth corresponding to each explanatory variable; For the first The explanatory variable at the th ... spatial grid location The local regression coefficient at a given location, which is based on the optimal spatial bandwidth of the corresponding explanatory variable. Obtained through fitting; For spatial intercept; To obtain white noise residuals that conform to a Gaussian distribution; Separating partial residuals of a single factor using a backfit algorithm: ； in, For the first The partial residuals of each variable; The actual observed values ​​of the response variable; No. The current regression summation term of all variables other than the one variable; For the first A regression summation term for each variable; For residuals; Construct a distance-decay spatial weight kernel function for local weighted calculation: ； in, For observation point For target point Spatial weights; The Euclidean distance between the two points; For the current trial bandwidth, when hour, ; With the goal of minimizing the modified Akaike information criterion, a golden section search is used to continuously update and test the bandwidth. This yields the local regression coefficients of the factor at the current scale. In the iterative optimization process, the ratio of the sum of squared changes is introduced as a criterion for model convergence: ； in, The ratio of the sum of squares of the changes; The total number of pregnancy-related factors; This represents the total number of spatial grid cells; and Representing the first In the current iteration and the previous iteration, the factor is located at the th . Local regression coefficient values ​​for each grid location; The preset convergence tolerance threshold is used; when When the value is below the preset convergence tolerance threshold, the model is deemed to have converged, and the optimal bandwidth and local regression coefficients of each factor are output.

6. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 1, characterized in that, In step S3, performing kriging interpolation on the local regression coefficients to generate a spatially continuous regression coefficient surface includes: For any unmodeled grid location within the study area to be estimated First, based on the discrete local regression coefficients, the experimental semivariogram is calculated and the theoretical variogram is obtained by fitting. The experimental semivariance value is calculated as follows: ； in, The lag distance is The experimental semivariance value at time; The spatial distance lag step; Spatial distance interval equal to The total number of valid sample point pairs; and Spatial location and its distance The regression coefficient value at the location; Construct and solve the Kriging equations to obtain the weight coefficients that satisfy the unbiased and optimal conditions; the Kriging equations are expressed as: ； in, For known sample points and The theoretical semivariance between them; For known sample points And the point to be estimated The theoretical semivariance between them; It is a Lagrange multiplier; Based on the calculated weight coefficients, the regression coefficients of the surrounding known sample points are weighted and combined to obtain the predicted regression coefficient values ​​for the location to be estimated. ； in, Location to be estimated The predicted values ​​of the regression coefficients; This represents the number of known neighboring sample points used in the estimation. To be assigned to the Weight coefficients for each known sample point; By traversing all grid locations to be estimated within the study area, regression coefficient prediction is completed point by point, generating a spatially continuous regression coefficient surface.

7. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in claim 6, characterized in that, In step S4, the optimal discretization and dominant factor extraction based on the regression coefficient surface to complete the spatial heterogeneity partitioning includes: First, perform optimal clustering and discretization of the attribute data on the regression coefficient surface, and calculate the sum of squares of the global data deviations from the mean for a single factor layer: ； in, The sum of squared deviations from the mean of global data for a single factor layer; The total number of pixels involved in the classification; For the first The regression coefficient values ​​for each pixel; This is the global mean. Simultaneously, calculate the sum of squared deviations from the mean within each subclass: ； in, The sum of squared deviations from the mean within each subclass; For the first Subclasses; For the first The within-class mean of the regression coefficients of all pixels within each subclass; This represents the total number of subclasses. For the first Within each subclass The regression coefficient values ​​for each pixel; Introducing the classification goodness index as the optimal discrete evaluation index: ； in, For classification excellence index; By continuously adjusting the combination of discontinuities through enumeration or heuristic search, the optimal discontinuity scheme that makes GVF achieve the global maximum value is locked, thus completing the optimal discretization of the continuous regression coefficient surface. Subsequently, a GIS spatial topology overlay operation was performed on all the discretized factor layers to cut the study area into subdivided polygonal patches with single attributes. Within each patch, the mean of the absolute values ​​of the regression coefficients of each disaster-causing factor was calculated, and the factor with the largest mean absolute value was extracted as the dominant disaster-causing factor of that patch. Finally, adjacent patches with the same dominant disaster-causing factors are spatially stitched together to form spatially heterogeneous zones with homogeneous internal physical disaster-causing patterns and independent zones.

8. The landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in any one of claims 1 to 7, characterized in that, In step S5, the landslide susceptibility simulation is performed using spatially heterogeneous partitions as units, and the prediction quality assessment is completed through multi-dimensional indicators, including: Each spatially heterogeneous zone is treated as an independent unit, and the probability of landslide susceptibility is extrapolated separately. The simulation results are sequentially subjected to accuracy benchmark diagnosis, area rationality diagnosis and prediction uncertainty robustness mapping to complete the multidimensional prediction quality evaluation. The accuracy benchmark diagnosis method is as follows: construct the receiver operating characteristic curve, calculate the area under the curve, evaluate the local prediction accuracy within a single partition and the global comprehensive accuracy after splicing the probabilities of each partition, and establish the accuracy benchmark for susceptibility prediction. The method for diagnosing the rationality of the area is as follows: a prediction efficiency curve is constructed with the cumulative proportion of the area of ​​the high-risk interval as the horizontal axis and the cumulative proportion of the corresponding historical landslides as the vertical axis. The rationality of the predicted area distribution is quantitatively evaluated based on the criteria that the high-risk interval is small and the area covers more historical landslide points. The method for mapping the robustness of prediction uncertainty is as follows: multiple independent datasets are generated by stratified random sampling within each partition, and the training set and validation set are divided proportionally. Multiple landslide susceptibility probability values ​​are obtained by repeatedly running the classifier. The mean and standard deviation of the prediction probability for each spatial location are calculated to achieve robustness assessment of the prediction results.

9. An electronic device comprising a processor and a memory; wherein the memory stores a computer program; characterized in that, When the processor executes the computer program, it implements the steps in the landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneous zoning as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon; characterized in that, When the computer program is executed by the processor, it implements the steps in the landslide susceptibility prediction method based on multi-scale geographical weighted regression and spatial heterogeneity zoning as described in any one of claims 1 to 8.

Images