Geological disaster multi-scale environment control factor identification system based on machine learning

Abstract

The application discloses a geological disaster multi-scale environment control factor identification system based on machine learning and relates to the technical field of geological disaster identification.The system comprises a multi-scale environment factor collection module, which acquires historical geological disaster data of a target region, extracts multi-scale environment characteristics of geological disasters and constructs a geological disaster-environment factor correlation characteristic set; a quantitative evaluation module, which establishes a geological disaster risk probability prediction model, quantitatively evaluates the marginal contribution degree of each environment factor at different spatial scales, identifies environment factors that dominate the development of geological disasters in the multi-scale of regions, river basins and slope surfaces and establishes a target region geological disaster risk probability surface and an environment factor contribution spatial distribution map; and a prevention and control module, which introduces a target region human activity intensity index and an ecological system state index, establishes an environment factor dynamic coupling risk propagation model, realizes real-time multi-scale risk transmission paths and generates a dynamic risk classification early warning atlas. The application improves the precision of geological disaster risk assessment.

Description

Technical Field

[0001] This invention relates to the field of geological hazard identification technology, specifically to a multi-scale environmental control factor identification system for geological hazards based on machine learning. Background Technology

[0002] Current geological hazard risk assessment methods mainly rely on static geological environmental factor analysis, which fails to adequately consider the dynamic interactions between environmental factors and the real-time impact of human activities and ecosystem changes. Existing technologies often use single-scale or fixed models for assessment, making it difficult to accurately quantify the synergistic control effects of multi-scale environmental factors such as regional, watershed, and slope factors on disasters. This results in risk assessment models exhibiting lag and limitations in predicting spatiotemporal changes, and failing to effectively adapt to the dynamic risk early warning needs under complex environmental changes. Summary of the Invention

[0003] To address the aforementioned technical issues, a machine learning-based multi-scale environmental control factor identification system for geological disasters is provided. This technical solution resolves the problems of early warning lag and insufficient assessment accuracy caused by the synergistic effects of the aforementioned multi-scale environmental factors.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A machine learning-based multi-scale environmental control factor identification system for geological hazards includes:

[0006] Multi-scale environmental factor acquisition module, quantitative assessment module, and prevention and control module;

[0007] Among them, the quantitative assessment module is electrically connected to the multi-scale environmental factor acquisition module, and the prevention and control module is electrically connected to the quantitative assessment module;

[0008] The multi-scale environmental factor acquisition module acquires historical geological disaster data of the target area, extracts multi-scale environmental features of geological disaster areas, watersheds and slopes, filters multi-scale environmental factors that are strongly correlated with the occurrence of geological disasters, and constructs a geological disaster-environmental factor correlation feature set.

[0009] The quantitative assessment module establishes a geological disaster risk probability prediction model based on the geological disaster-environmental factor correlation feature set, quantitatively assesses the marginal contribution of each environmental factor at different spatial scales, identifies the environmental factors that dominate the development of geological disasters at multiple scales such as region, watershed and slope, and establishes a geological disaster risk probability surface and spatial distribution map of environmental factor contribution in the target area.

[0010] The prevention and control module, based on the geological disaster risk probability surface and the spatial distribution map of environmental factor contributions in the target area, introduces the human activity intensity index and ecosystem state index in the target area as dynamic adjustment variables, establishes a dynamic coupling risk transmission model of environmental factors, analyzes the multi-scale risk transmission path under changes in human activities and ecological state in real time, generates a dynamic risk classification and early warning map with spatiotemporal dimensions, and realizes intelligent geological disaster risk prevention and control.

[0011] Preferably, the multi-scale environmental factor acquisition module specifically includes:

[0012] The data preprocessing unit, based on geological disaster research institutions, acquires historical geological disaster location data of the target area, obtains vector data of landslide and debris flow events, collects topographic data, remote sensing data, soil data and meteorological data of the target area during the same period of geological disaster occurrence, performs unified map projection transformation on all data, resamples to a fixed spatial resolution, converts to the same raster cell, and continues data preprocessing on the data.

[0013] The regional scale feature extraction unit, based on GIS software, uses the DEM digital elevation model to calculate the elevation standard deviation and mean of each raster unit, obtain the elevation variation coefficient, count the area proportion of various rock group types within the unit, calculate the average rainfall within the unit, and associate the elevation variation coefficient, the area proportion of various rock group types, the average rainfall with historical geological disaster points in the target area to establish an attribute table with regional unit-attribute feature structure, and extract the regional scale features of historical geological disasters in the target area;

[0014] The watershed-scale feature extraction unit calculates the channel density of each grid cell, analyzes the distribution of rivers within the watershed, calculates the topographic humidity index of the grid cell, calculates the catchment area of ​​each watershed based on hydrological analysis, quantitatively assesses the impact of watershed hydrological conditions on historical geological disasters, and extracts watershed-scale features of historical geological disasters in the target area.

[0015] The slope-scale feature extraction unit utilizes the DEM digital elevation model to extract slope, aspect, and curvature features within the raster unit. It calculates the NDVI (Normalized Difference Vegetation Index) using multispectral remote sensing imagery. Based on fault line vector data, it calculates the Euclidean distance from each raster unit to the nearest fault, generating a distance raster map. This extracts the slope-scale features of historical geological disasters in the target area. Combining these features with the regional and watershed-scale features of historical geological disasters in the target area, the unit obtains multi-scale environmental characteristics of the geological disaster area, watershed, and slope.

[0016] Preferably, the multi-scale environmental factor acquisition module also includes:

[0017] The initial feature matrix is ​​established by spatially connecting historical geological disaster points in the target area with their corresponding slope-scale raster units. If a raster unit overlaps with a geological disaster point or disaster polygon, the disaster label of that raster unit is marked as 1; otherwise, it is marked as 0. The regional, watershed, and slope-scale features corresponding to all raster units are extracted and combined with the disaster label and raster unit ID. The raster unit is used as the row, and the raster unit ID, disaster label, and the regional, watershed, and slope-scale features corresponding to all raster units are used as the column to establish the initial feature matrix.

[0018] Preferably, the multi-scale environmental factor acquisition module also includes:

[0019] The screening unit takes the initial feature matrix as input, the disaster label as the target variable, and the regional, watershed, and slope-scale features corresponding to all grid cells as environmental factors. Using the Bagging ensemble method, the initial feature matrix is ​​sampled with replacement to generate multiple different training subsets. Corresponding decision trees are constructed, and the total reduction in Gini impurity brought by each feature when splitting at all decision tree nodes is evaluated. After normalization, the importance score of each feature is obtained. According to actual business needs, an importance score threshold h is set, and features below the threshold are removed to obtain the screened disaster label - regional, watershed, and slope-scale features.

[0020] For each pair of disaster labels—region, watershed, and slope scale features—the ratio of the covariance to the standard deviation of each pair is calculated to obtain the Pearson correlation coefficient. The closer the absolute value of each pair of Pearson correlation coefficients is to 1, the stronger the linear correlation between the two features.

[0021] The feature set construction unit uses regional, watershed, and slope-scale features as dependent variables and disaster labels as independent variables. It calculates their coefficients of determination and combines them with the absolute values ​​of each pair of Pearson correlation coefficients. If the absolute values ​​of the coefficients of determination and their Pearson correlation coefficients are both greater than 0.7, it proves that there is a strong correlation between disaster labels and regional, watershed, and slope-scale features, and constructs a geological disaster-environmental factor correlation feature set.

[0022] Preferably, the quantitative evaluation module specifically includes:

[0023] The model building unit takes the geological hazard-environmental factor correlation feature set as input, extracts environmental factors and corresponding hazard labels at the regional, watershed and slope scales of the target area, divides the geological hazard-environmental factor correlation feature set into training set and validation set, performs data preprocessing, establishes a geological hazard risk probability prediction model, optimizes the hyperparameters of the number of trees, learning rate, maximum depth and minimum number of sample splits through grid search and K-fold cross-validation, introduces cross-validation for model evaluation, and takes the geological hazard risk probability prediction of any grid cell in the target area as output.

[0024] Preferably, the quantitative evaluation module also includes:

[0025] The dominant factor acquisition unit calculates the SAP value of each environmental factor for each grid cell geological hazard risk probability prediction sample, quantifies its local marginal contribution, calculates the SAP value of each environmental factor across all samples, and calculates the mean of the absolute values. The larger the mean, the greater the overall impact of the environmental factor on the model evaluation results. All environmental factors are ranked by importance to identify the dominant environmental factors for geological hazard development at multiple scales, including regional, watershed, and slope levels.

[0026] Preferably, the quantitative evaluation module also includes:

[0027] The geological hazard risk probability surface raster map acquisition unit extracts environmental factors of regional, watershed and slope multi-scale environmental characteristics of all raster units in the study area, generates a whole-area raster feature dataset with the same structure as the geological hazard-environmental factor association feature set, substitutes it into the geological hazard probability prediction model, obtains the geological hazard risk prediction probability corresponding to each unit, maps it according to its original spatial location, and generates a geological hazard risk probability surface raster map covering the whole area and continuous.

[0028] The contribution intensity spatial distribution map generation unit is based on the environmental factors that dominate the development of geological hazards at multiple scales, including regions, watersheds, and slopes. For each grid cell in the whole area, the SHAP value of the dominant environmental factor under the specific combination of environmental characteristics in the cell is calculated to obtain the marginal contribution of the dominant environmental factor to the final geological hazard risk probability prediction, and a spatial distribution map of the contribution intensity of the dominant environmental factor is generated.

[0029] Preferably, the prevention and control module specifically includes:

[0030] The Human Activity Intensity Index Acquisition Unit, based on the Atmospheric Administration, acquires monthly / annual nighttime light intensity data for the target area. Based on the local transportation department, it acquires the vector road network for the target area, calculates the road network density by calculating the road length per unit area of ​​the target area, and calculates the rate of construction land expansion and the rate of change of cultivated land in the target area based on multi-period land use-cover classification data. The human activity intensity index for the target area is obtained by weighted summation of light intensity data, road network density, and land area change rate.

[0031] The ecosystem state index acquisition unit, based on long-term NDVI normalized vegetation index data, uses trend analysis to calculate the long-term trend of vegetation greenness in the target area, defines the negative trend area as the vegetation degradation area, uses a known soil loss equation model, dynamically inputs rainfall erosion force and vegetation cover management factors, calculates the soil erosion modulus in the target area, and calculates the ecosystem state index of the target area by combining the long-term trend of vegetation greenness and the soil erosion modulus.

[0032] The unified unit resamples the human activity intensity index and the ecosystem status index of the target area to the same spatial resolution and geographic coordinate system, performs data normalization processing, and aligns the timestamps.

[0033] The dynamic adjustment variable generation unit establishes an m×n evaluation matrix for each grid cell, calculates the proportion of the i-th sample value under the j-th indicator in the evaluation matrix, calculates the entropy value of the j-th indicator, calculates the difference coefficient of each indicator, obtains the final weight of each indicator, normalizes it, and weights and sums the values ​​of each indicator and their corresponding weights to obtain two spatial grid layers: the comprehensive index of human activities and the comprehensive index of ecosystem status, which serve as dynamic adjustment variables.

[0034] m represents a grid cell, and n represents sub-indices within the target area's human activity intensity index and ecosystem state index.

[0035] Preferably, the prevention and control module also includes:

[0036] The multi-scale risk transmission path generation unit sets the watershed risk level as an endogenous latent variable, and human activity pressure, ecological status and slope stability as exogenous latent variables. It takes all watershed units in the target area as inputs, establishes a dynamic coupling risk transmission model of environmental factors, and combines the geological disaster risk probability prediction of any grid unit in the target area to calculate the influence coefficient of human activities and ecological status on risk probability, and analyzes the multi-scale risk transmission path under changes in human activities and ecological status in real time.

[0037] The watershed risk level is obtained by collecting the characteristics of landslide density and debris flow frequency in the target area; human activity pressure is obtained by calculating the comprehensive index of human activities and road density characteristics in the target area; ecological status is obtained by calculating the comprehensive index of ecosystem status and vegetation degradation index in the target area; and slope stability is obtained by calculating the static risk probability and topographic relief characteristics in the target area.

[0038] Preferably, the prevention and control module also includes:

[0039] The risk probability raster dataset generation unit collects real-time rainfall data and future rainfall forecast data for the target area, substitutes them into the environmental factor dynamic coupling risk propagation model, calculates the disaster risk probability sequence of the target area at different future time points, and generates a risk probability raster dataset updated according to the time series.

[0040] The hierarchical unit uses the natural breakpoint method to classify the disaster risk probability sequence of the target area at different future time points, and sets low, medium, high and extremely high risk probability threshold intervals to obtain multiple levels. The disaster risk probability sequence at different future time points is divided into the corresponding threshold intervals, generating a dynamic risk classification and early warning map with spatiotemporal dimensions, so as to realize intelligent geological disaster risk prevention and control.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] This invention proposes a machine learning-based multi-scale environmental control factor identification scheme for geological hazards. This scheme integrates multi-scale environmental factor acquisition, quantitative assessment, and dynamic prevention and control modules, constructing a complete technology chain from feature identification to risk early warning. This achieves a leap from static single-scale to dynamic multi-scale geological hazard risk assessment. The system quantifies the marginal contribution of environmental factors at each scale through machine learning, accurately identifying the dominant risk factors for regions, watersheds, and slopes. It introduces human activities and ecosystem status as dynamic regulating variables, and combines real-time data-driven risk propagation models to dynamically analyze the cascading transmission effects of human activities, ecological status, slope stability, and watershed risk. The generated spatiotemporal dynamic risk classification and early warning map significantly improves the accuracy, timeliness, and predictability of geological hazard risk assessment, providing scientific decision support for intelligent and refined risk prevention and control. Attached Figure Description

[0043] Figure 1 This is a framework diagram of a machine learning-based multi-scale environmental control factor identification system for geological hazards. Detailed Implementation

[0044] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0045] Reference Figure 1 As shown, the machine learning-based multi-scale environmental control factor identification system for geological hazards includes:

[0046] Multi-scale environmental factor acquisition module, quantitative assessment module, and prevention and control module;

[0047] Among them, the quantitative assessment module is electrically connected to the multi-scale environmental factor acquisition module, and the prevention and control module is electrically connected to the quantitative assessment module;

[0048] The multi-scale environmental factor acquisition module acquires historical geological disaster data of the target area, extracts multi-scale environmental features of geological disaster areas, watersheds and slopes, filters multi-scale environmental factors that are strongly correlated with the occurrence of geological disasters, and constructs a geological disaster-environmental factor correlation feature set.

[0049] The quantitative assessment module establishes a geological disaster risk probability prediction model based on the geological disaster-environmental factor correlation feature set, quantitatively assesses the marginal contribution of each environmental factor at different spatial scales, identifies the environmental factors that dominate the development of geological disasters at multiple scales such as region, watershed and slope, and establishes a geological disaster risk probability surface and spatial distribution map of environmental factor contribution in the target area.

[0050] The prevention and control module, based on the geological disaster risk probability surface and the spatial distribution map of environmental factor contributions in the target area, introduces the human activity intensity index and ecosystem state index in the target area as dynamic adjustment variables, establishes a dynamic coupling risk transmission model of environmental factors, analyzes the multi-scale risk transmission path under changes in human activities and ecological state in real time, generates a dynamic risk classification and early warning map with spatiotemporal dimensions, and realizes intelligent geological disaster risk prevention and control.

[0051] The multi-scale environmental factor acquisition module specifically includes:

[0052] The data preprocessing unit, based on geological disaster research institutions, acquires historical geological disaster location data of the target area, obtains vector data of landslide and debris flow events, collects topographic data, remote sensing data, soil data and meteorological data of the target area during the same period of geological disaster occurrence, performs unified map projection transformation on all data, resamples to a fixed spatial resolution, converts to the same raster cell, and continues data preprocessing on the data.

[0053] The regional scale feature extraction unit, based on GIS software, uses the DEM digital elevation model to calculate the elevation standard deviation and mean of each raster unit, obtain the elevation variation coefficient, count the area proportion of various rock group types within the unit, calculate the average rainfall within the unit, and associate the elevation variation coefficient, the area proportion of various rock group types, the average rainfall with historical geological disaster points in the target area to establish an attribute table with regional unit-attribute feature structure, and extract the regional scale features of historical geological disasters in the target area;

[0054] The watershed-scale feature extraction unit calculates the channel density of each grid cell, analyzes the distribution of rivers within the watershed, calculates the topographic humidity index of the grid cell, calculates the catchment area of ​​each watershed based on hydrological analysis, quantitatively assesses the impact of watershed hydrological conditions on historical geological disasters, and extracts watershed-scale features of historical geological disasters in the target area.

[0055] The slope-scale feature extraction unit utilizes the DEM digital elevation model to extract slope, aspect, and curvature features within the raster unit. It calculates the NDVI (Normalized Difference Vegetation Index) using multispectral remote sensing imagery. Based on fault line vector data, it calculates the Euclidean distance from each raster unit to the nearest fault, generating a distance raster map. This extracts the slope-scale features of historical geological disasters in the target area. Combining these features with the regional and watershed-scale features of historical geological disasters in the target area, the unit obtains multi-scale environmental characteristics of the geological disaster area, watershed, and slope.

[0056] The multi-scale environmental factor acquisition module also includes:

[0057] The initial feature matrix is ​​established by spatially connecting historical geological disaster points in the target area with their corresponding slope-scale raster units. If a raster unit overlaps with a geological disaster point or disaster polygon, the disaster label of that raster unit is marked as 1; otherwise, it is marked as 0. The regional, watershed, and slope-scale features corresponding to all raster units are extracted and combined with the disaster label and raster unit ID. The raster unit is used as the row, and the raster unit ID, disaster label, and the regional, watershed, and slope-scale features corresponding to all raster units are used as the column to establish the initial feature matrix.

[0058] The multi-scale environmental factor acquisition module also includes:

[0059] The screening unit takes the initial feature matrix as input, the disaster label as the target variable, and the regional, watershed, and slope-scale features corresponding to all grid cells as environmental factors. Using the Bagging ensemble method, the initial feature matrix is ​​sampled with replacement to generate multiple different training subsets. Corresponding decision trees are constructed, and the total reduction in Gini impurity brought by each feature when splitting at all decision tree nodes is evaluated. After normalization, the importance score of each feature is obtained. According to actual business needs, an importance score threshold h is set, and features below the threshold are removed to obtain the screened disaster label - regional, watershed, and slope-scale features.

[0060] For each pair of disaster labels—region, watershed, and slope scale features—the ratio of the covariance to the standard deviation of each pair is calculated to obtain the Pearson correlation coefficient. The closer the absolute value of each pair of Pearson correlation coefficients is to 1, the stronger the linear correlation between the two features.

[0061] The feature set construction unit uses regional, watershed, and slope-scale features as dependent variables and disaster labels as independent variables. It calculates their coefficients of determination and combines them with the absolute values ​​of each pair of Pearson correlation coefficients. If the absolute values ​​of the coefficients of determination and their Pearson correlation coefficients are both greater than 0.7, it proves that there is a strong correlation between disaster labels and regional, watershed, and slope-scale features, and constructs a geological disaster-environmental factor correlation feature set.

[0062] When using it, combine it with the above module units:

[0063] Existing geological hazard assessment methods often rely on single-scale or limited environmental factors, and feature selection is frequently based on experience or simple statistics. They fail to systematically integrate multi-scale features at the regional, watershed, and slope levels, and do not fully consider multicollinearity among features and their nonlinear correlation with hazards, resulting in weak model interpretability and limited prediction accuracy. This method, through multi-source data fusion and multi-scale feature extraction, combined with Bagging integrated feature selection and Pearson correlation coefficient and coefficient of determination quantitative verification, constructs a set of features strongly correlated with geological hazards and environmental factors. This improves the objectivity and scientific rigor of feature selection, enhances the model's generalization ability and the depth of hazard mechanism explanation, and provides a more reliable and multi-dimensional data foundation for geological hazard risk assessment.

[0064] The quantitative evaluation module specifically includes:

[0065] The model building unit takes the geological hazard-environmental factor correlation feature set as input, extracts environmental factors and corresponding hazard labels at the regional, watershed and slope scales of the target area, divides the geological hazard-environmental factor correlation feature set into training set and validation set, performs data preprocessing, establishes a geological hazard risk probability prediction model, optimizes the hyperparameters of the number of trees, learning rate, maximum depth and minimum number of sample splits through grid search and K-fold cross-validation, introduces cross-validation for model evaluation, and takes the geological hazard risk probability prediction of any grid cell in the target area as output.

[0066] The quantitative evaluation module also includes:

[0067] The dominant factor acquisition unit calculates the SAP value of each environmental factor for each grid cell geological hazard risk probability prediction sample, quantifies its local marginal contribution, calculates the SAP value of each environmental factor across all samples, and calculates the mean of the absolute values. The larger the mean, the greater the overall impact of the environmental factor on the model evaluation results. All environmental factors are ranked by importance to identify the dominant environmental factors for geological hazard development at multiple scales, including regional, watershed, and slope levels.

[0068] The quantitative evaluation module also includes:

[0069] The geological hazard risk probability surface raster map acquisition unit extracts environmental factors of regional, watershed and slope multi-scale environmental characteristics of all raster units in the study area, generates a whole-area raster feature dataset with the same structure as the geological hazard-environmental factor association feature set, substitutes it into the geological hazard probability prediction model, obtains the geological hazard risk prediction probability corresponding to each unit, maps it according to its original spatial location, and generates a geological hazard risk probability surface raster map covering the whole area and continuous.

[0070] The contribution intensity spatial distribution map generation unit is based on the environmental factors that dominate the development of geological hazards at multiple scales, including regions, watersheds, and slopes. For each grid cell in the whole area, the SHAP value of the dominant environmental factor under the specific combination of environmental characteristics in the cell is calculated to obtain the marginal contribution of the dominant environmental factor to the final geological hazard risk probability prediction, and a spatial distribution map of the contribution intensity of the dominant environmental factor is generated.

[0071] When using it, combine it with the above module units:

[0072] Existing studies often employ single-scale environmental factors for geological hazard risk modeling, failing to fully integrate the synergistic mechanisms of regional, watershed, and slope-level multi-scale interactions. Furthermore, model hyperparameter optimization frequently relies on empirical settings, resulting in insufficient spatial continuity and physical interpretability of prediction results. Traditional importance assessment methods struggle to quantify the marginal contribution of environmental factors to individual predictions, limiting the accurate identification of multi-scale dominant factors and the analysis of spatial heterogeneity. This step improves the accuracy and robustness of geological hazard risk probability prediction through multi-scale environmental factor fusion and hyperparameter optimization cross-validation. By introducing SHAP marginal contribution analysis, the importance of environmental factors is quantitatively ranked from global to local levels, and a spatial distribution map of the contribution intensity of dominant factors is generated, enhancing the model's interpretability and providing a more refined decision-making basis for geological hazard risk zoning and prevention.

[0073] The prevention and control module specifically includes:

[0074] The Human Activity Intensity Index Acquisition Unit, based on the Atmospheric Administration, acquires monthly / annual nighttime light intensity data for the target area. Based on the local transportation department, it acquires the vector road network for the target area, calculates the road network density by calculating the road length per unit area of ​​the target area, and calculates the rate of construction land expansion and the rate of change of cultivated land in the target area based on multi-period land use-cover classification data. The human activity intensity index for the target area is obtained by weighted summation of light intensity data, road network density, and land area change rate.

[0075] The ecosystem state index acquisition unit, based on long-term NDVI normalized vegetation index data, uses trend analysis to calculate the long-term trend of vegetation greenness in the target area, defines the negative trend area as the vegetation degradation area, uses a known soil loss equation model, dynamically inputs rainfall erosion force and vegetation cover management factors, calculates the soil erosion modulus in the target area, and calculates the ecosystem state index of the target area by combining the long-term trend of vegetation greenness and the soil erosion modulus.

[0076] The unified unit resamples the human activity intensity index and the ecosystem status index of the target area to the same spatial resolution and geographic coordinate system, performs data normalization processing, and aligns the timestamps.

[0077] The dynamic adjustment variable generation unit establishes an m×n evaluation matrix for each grid cell, calculates the proportion of the i-th sample value under the j-th indicator in the evaluation matrix, calculates the entropy value of the j-th indicator, calculates the difference coefficient of each indicator, obtains the final weight of each indicator, normalizes it, and weights and sums the values ​​of each indicator and their corresponding weights to obtain two spatial grid layers: the comprehensive index of human activities and the comprehensive index of ecosystem status, which serve as dynamic adjustment variables.

[0078] m represents a grid cell, and n represents a sub-indicator within the target area's human activity intensity index and ecosystem state index.

[0079] The prevention and control module also includes:

[0080] The multi-scale risk transmission path generation unit sets the watershed risk level as an endogenous latent variable, and human activity pressure, ecological status and slope stability as exogenous latent variables. It takes all watershed units in the target area as inputs, establishes a dynamic coupling risk transmission model of environmental factors, and combines the geological disaster risk probability prediction of any grid unit in the target area to calculate the influence coefficient of human activities and ecological status on risk probability, and analyzes the multi-scale risk transmission path under changes in human activities and ecological status in real time.

[0081] The watershed risk level is obtained by collecting the characteristics of landslide density and debris flow frequency in the target area; human activity pressure is obtained by calculating the comprehensive index of human activities and road density characteristics in the target area; ecological status is obtained by calculating the comprehensive index of ecosystem status and vegetation degradation index in the target area; and slope stability is obtained by calculating the static risk probability and topographic relief characteristics in the target area.

[0082] Multi-scale risk transmission pathways: human activity pressure → ecological status, ecological status → slope stability, slope stability → watershed risk level.

[0083] The prevention and control module also includes:

[0084] The risk probability raster dataset generation unit collects real-time rainfall data and future rainfall forecast data for the target area, substitutes them into the environmental factor dynamic coupling risk propagation model, calculates the disaster risk probability sequence of the target area at different future time points, and generates a risk probability raster dataset updated according to the time series.

[0085] The hierarchical unit uses the natural breakpoint method to classify the disaster risk probability sequence of the target area at different future time points, and sets low, medium, high and extremely high risk probability threshold intervals to obtain multiple levels. The disaster risk probability sequence at different future time points is divided into the corresponding threshold intervals, generating a dynamic risk classification and early warning map with spatiotemporal dimensions, so as to realize intelligent geological disaster risk prevention and control.

[0086] When using it, combine it with the above module units:

[0087] Existing technologies for regional geological hazard risk assessment typically rely on static or single-factor models, which struggle to dynamically couple the interactions between human activities and ecosystems. Furthermore, they fail to adequately consider multi-scale risk transmission mechanisms, resulting in limited timeliness and spatial targeting of early warnings. This approach constructs a human activity intensity index integrating nighttime light, road network density, and land use change, along with an ecosystem state index encompassing vegetation trends and soil erosion. Based on a dynamic weighted integration using the entropy method, an environmental factor-coupled risk propagation model is established, quantifying multi-scale risk pathways involving human activities, ecological status, and slope stability. By combining real-time and forecasted rainfall data to drive model iteration, a spatiotemporal dynamic risk probability sequence and tiered early warning map are generated, significantly improving the timeliness, systematicity, and spatial accuracy of risk assessments. This provides intelligent and forward-looking decision support for geological hazard prevention and control.

[0088] Based on the above module units, the specific implementation is as follows:

[0089] In its application in the Wumeng Mountains of southern Sichuan, this system covers a study area of ​​2,973 square kilometers. It comprehensively integrates multi-source geographic information data. The multi-scale environmental factor acquisition module first integrates 327 historical geological disaster sites, high-precision DEMs, multi-period remote sensing images, geological maps, and meteorological data from 2005 to 2025. After preprocessing with a unified coordinate system and 30-meter resolution resampling, the system extracts three types of scale features: at the regional scale (1km grid), it calculates the coefficient of variation of elevation, rock mass area ratio, and average annual rainfall; at the watershed scale, it extracts river density, topographic humidity index, and catchment area; and at the slope scale, it interprets slope, aspect, NDVI, and distance from faults.

[0090] The module uses the Bagging ensemble method to rank the importance of the initial 11 features, and selects 6 core factors such as slope, elevation variation coefficient, and annual average rainfall based on thresholds. The Pearson correlation coefficient and determination coefficient between these factors and the disaster label are calculated (both > 0.7), thus constructing a set of features with strong correlation between geological disasters and environmental factors, laying a reliable data foundation for risk assessment.

[0091] Based on the constructed feature set, the quantitative assessment module established a geological disaster risk probability prediction model. The module divided the sample data into training and validation sets in a 7:3 ratio and optimized the hyperparameters through grid search and 5-fold cross-validation (finally determined the number of trees = 200 and the learning rate = 0.05). The model achieved an AUC value of 0.91 on the validation set, demonstrating excellent prediction performance. To quantify the contribution of each environmental factor, the module introduced SHAP value analysis to calculate the mean marginal contribution of each factor to the prediction of all samples.

[0092] The results showed that slope (mean SHAP = 0.253) and elevation variation coefficient (mean SHAP = 0.189) were the most influential dominant factors. The module input the environmental characteristics of all grids in the study area into the trained model, generated a geological hazard risk probability surface grid map covering the entire area, and simultaneously output the spatial distribution map of the contribution intensity of the dominant factors, intuitively revealing the spatial coupling relationship between high-risk areas and key control factors, and realizing the visualization and mechanism explanation of the risk pattern.

[0093] To achieve dynamic risk prevention and control, the prevention and control module first integrates data on nighttime lights, road network density, and land use change, and generates a human activity intensity index by weighting the data using the entropy method.

[0094] Based on long-term NDVI and soil erosion models, an ecosystem state index was calculated. These two dynamic variables were used as exogenous moderating factors. A structural equation model was constructed to quantify the cross-scale risk transmission path of "human activity pressure → ecological state → slope stability → watershed risk level" (all path coefficients p < 0.01).

[0095] The module leverages real-time and 72-hour rainfall forecast data for the study area to drive model iteration, outputting hourly updated risk probability raster sequences. Using the natural breakpoint method, the probability values ​​are categorized into four risk levels: low, medium, high, and extremely high. This results in the generation of a spatiotemporal dynamic risk classification and early warning map. This map successfully identifies river valley sections affected by the combined effects of road excavation and vegetation degradation as extremely high-risk areas under heavy rainfall scenarios, providing local disaster prevention departments with precise and forward-looking decision support for evacuation control and engineering management.

[0096] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A machine learning-based multi-scale environmental control factor identification system for geological hazards, characterized in that, include: Multi-scale environmental factor acquisition module, quantitative assessment module, and prevention and control module; Among them, the quantitative assessment module is electrically connected to the multi-scale environmental factor acquisition module, and the prevention and control module is electrically connected to the quantitative assessment module; The multi-scale environmental factor acquisition module acquires historical geological disaster data of the target area, extracts multi-scale environmental features of geological disaster areas, watersheds and slopes, filters multi-scale environmental factors that are strongly correlated with the occurrence of geological disasters, and constructs a geological disaster-environmental factor correlation feature set. The quantitative assessment module establishes a geological disaster risk probability prediction model based on the geological disaster-environmental factor correlation feature set, quantitatively assesses the marginal contribution of each environmental factor at different spatial scales, identifies the environmental factors that dominate the development of geological disasters at multiple scales such as region, watershed and slope, and establishes a geological disaster risk probability surface and spatial distribution map of environmental factor contribution in the target area. The prevention and control module, based on the geological disaster risk probability surface and the spatial distribution map of environmental factor contributions in the target area, introduces the human activity intensity index and ecosystem state index in the target area as dynamic adjustment variables, establishes a dynamic coupling risk transmission model of environmental factors, analyzes the multi-scale risk transmission path under changes in human activities and ecological state in real time, generates a dynamic risk classification and early warning map with spatiotemporal dimensions, and realizes intelligent geological disaster risk prevention and control. The multi-scale environmental factor acquisition module specifically includes: The data preprocessing unit, based on geological disaster research institutions, acquires historical geological disaster location data of the target area, obtains vector data of landslide and debris flow events, collects topographic data, remote sensing data, soil data and meteorological data of the target area during the same period of geological disaster occurrence, performs unified map projection transformation on all data, resamples to a fixed spatial resolution, converts to the same raster cell, and continues data preprocessing on the data. The regional scale feature extraction unit, based on GIS software, uses the DEM digital elevation model to calculate the elevation standard deviation and mean of each raster unit, obtain the elevation variation coefficient, count the area proportion of various rock group types within the unit, calculate the average rainfall within the unit, and associate the elevation variation coefficient, the area proportion of various rock group types, the average rainfall with historical geological disaster points in the target area to establish an attribute table with regional unit-attribute feature structure, and extract the regional scale features of historical geological disasters in the target area; The watershed-scale feature extraction unit calculates the channel density of each grid cell, analyzes the distribution of rivers within the watershed, calculates the topographic humidity index of the grid cell, calculates the catchment area of ​​each watershed based on hydrological analysis, quantitatively assesses the impact of watershed hydrological conditions on historical geological disasters, and extracts watershed-scale features of historical geological disasters in the target area. The slope-scale feature extraction unit utilizes the DEM digital elevation model to extract slope, aspect, and curvature features within the raster unit. It calculates the NDVI normalized vegetation index using multispectral remote sensing imagery, and calculates the Euclidean distance from each raster unit to the nearest fault based on fault line vector data, generating a distance raster map. This extracts the slope-scale features of historical geological disasters in the target area. Combining the regional and watershed-scale features of historical geological disasters in the target area, the unit obtains multi-scale environmental characteristics of the geological disaster area, watershed, and slope. The prevention and control module specifically includes: The Human Activity Intensity Index Acquisition Unit, based on the Atmospheric Administration, acquires monthly / annual nighttime light intensity data for the target area. Based on the local transportation department, it acquires the vector road network for the target area, calculates the road network density by calculating the road length per unit area of ​​the target area, and calculates the rate of construction land expansion and the rate of change of cultivated land in the target area based on multi-period land use-cover classification data. The human activity intensity index for the target area is obtained by weighted summation of light intensity data, road network density, and land area change rate. The ecosystem state index acquisition unit, based on long-term NDVI normalized vegetation index data, uses trend analysis to calculate the long-term trend of vegetation greenness in the target area, defines the negative trend area as the vegetation degradation area, uses a known soil loss equation model, dynamically inputs rainfall erosion force and vegetation cover management factors, calculates the soil erosion modulus in the target area, and calculates the ecosystem state index of the target area by combining the long-term trend of vegetation greenness and the soil erosion modulus. The unified unit resamples the human activity intensity index and the ecosystem status index of the target area to the same spatial resolution and geographic coordinate system, performs data normalization processing, and aligns the timestamps. The dynamic adjustment variable generation unit establishes an m×n evaluation matrix for each grid cell, calculates the proportion of the i-th sample value under the j-th indicator in the evaluation matrix, calculates the entropy value of the j-th indicator, calculates the difference coefficient of each indicator, obtains the final weight of each indicator, normalizes it, and weights and sums the values ​​of each indicator and their corresponding weights to obtain two spatial grid layers: the comprehensive index of human activities and the comprehensive index of ecosystem status, which serve as dynamic adjustment variables. m is a grid cell, and n is a sub-indicator within the target area human activity intensity index and the target area ecosystem state index. The prevention and control module also includes: The multi-scale risk transmission path generation unit sets the watershed risk level as an endogenous latent variable, and human activity pressure, ecological status and slope stability as exogenous latent variables. It takes all watershed units in the target area as inputs, establishes a dynamic coupling risk transmission model of environmental factors, and combines the geological disaster risk probability prediction of any grid unit in the target area to calculate the influence coefficient of human activities and ecological status on risk probability, and analyzes the multi-scale risk transmission path under changes in human activities and ecological status in real time. The watershed risk level is obtained by collecting the characteristics of landslide density and debris flow frequency in the target area; human activity pressure is obtained by calculating the comprehensive index of human activities and road density characteristics in the target area; ecological status is obtained by calculating the comprehensive index of ecosystem status and vegetation degradation index in the target area; and slope stability is obtained by calculating the static risk probability and topographic relief characteristics in the target area.

2. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 1, characterized in that, The multi-scale environmental factor acquisition module also includes: The initial feature matrix is ​​established by spatially connecting historical geological disaster points in the target area with their corresponding slope-scale raster units. If a raster unit overlaps with a geological disaster point or disaster polygon, the disaster label of that raster unit is marked as 1; otherwise, it is marked as 0. The regional, watershed, and slope-scale features corresponding to all raster units are extracted and combined with the disaster label and raster unit ID. The raster unit is used as the row, and the raster unit ID, disaster label, and the regional, watershed, and slope-scale features corresponding to all raster units are used as the column to establish the initial feature matrix.

3. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 2, characterized in that, The multi-scale environmental factor acquisition module also includes: The screening unit takes the initial feature matrix as input, the disaster label as the target variable, and the regional, watershed, and slope-scale features corresponding to all grid cells as environmental factors. Using the Bagging ensemble method, the initial feature matrix is ​​sampled with replacement to generate multiple different training subsets. Corresponding decision trees are constructed, and the total reduction in Gini impurity brought by each feature when splitting at all decision tree nodes is evaluated. After normalization, the importance score of each feature is obtained. According to actual business needs, an importance score threshold h is set, and features below the threshold are removed to obtain the screened disaster label - regional, watershed, and slope-scale features. For each pair of disaster labels—region, watershed, and slope scale features—the ratio of the covariance to the standard deviation of each pair is calculated to obtain the Pearson correlation coefficient. The closer the absolute value of each pair of Pearson correlation coefficients is to 1, the stronger the linear correlation between the two features. The feature set construction unit uses regional, watershed, and slope-scale features as dependent variables and disaster labels as independent variables. It calculates their coefficients of determination and combines them with the absolute values ​​of each pair of Pearson correlation coefficients. If the absolute values ​​of the coefficients of determination and their Pearson correlation coefficients are both greater than 0.7, it proves that there is a strong correlation between disaster labels and regional, watershed, and slope-scale features, and constructs a geological disaster-environmental factor correlation feature set.

4. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 3, characterized in that, The quantitative evaluation module specifically includes: The model building unit takes the geological hazard-environmental factor correlation feature set as input, extracts environmental factors and corresponding hazard labels at the target area, watershed and slope scale, divides the geological hazard-environmental factor correlation feature set into training set and validation set, performs data preprocessing, establishes a geological hazard risk probability prediction model, optimizes the hyperparameters of the number of trees, learning rate, maximum depth and minimum number of sample splits through grid search and K-fold cross-validation, introduces cross-validation for model evaluation, and takes the geological hazard risk probability prediction of any grid cell in the target area as output.

5. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 4, characterized in that, The quantitative evaluation module also includes: The dominant factor acquisition unit calculates the SAP value of each environmental factor for each grid cell geological hazard risk probability prediction sample, quantifies its local marginal contribution, calculates the SAP value of each environmental factor across all samples, and calculates the mean of the absolute values. The larger the mean, the greater the overall impact of the environmental factor on the model evaluation results. All environmental factors are ranked by importance to identify the dominant environmental factors for geological hazard development at multiple scales, including regional, watershed, and slope levels.

6. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 5, characterized in that, The quantitative evaluation module also includes: The geological hazard risk probability surface raster map acquisition unit extracts environmental factors of regional, watershed and slope multi-scale environmental characteristics of all raster units in the study area, generates a whole-area raster feature dataset with the same structure as the geological hazard-environmental factor association feature set, substitutes it into the geological hazard probability prediction model, obtains the geological hazard risk prediction probability corresponding to each unit, maps it according to its original spatial location, and generates a geological hazard risk probability surface raster map covering the whole area and continuous. The contribution intensity spatial distribution map generation unit is based on the environmental factors that dominate the development of geological hazards at multiple scales, including regions, watersheds, and slopes. For each grid cell in the whole area, the SHAP value of the dominant environmental factor under the specific combination of environmental characteristics in the cell is calculated to obtain the marginal contribution of the dominant environmental factor to the final geological hazard risk probability prediction, and a spatial distribution map of the contribution intensity of the dominant environmental factor is generated.

7. The machine learning-based multi-scale environmental control factor identification system for geological hazards according to claim 1, characterized in that, The prevention and control module also includes: The risk probability raster dataset generation unit collects real-time rainfall data and future rainfall forecast data for the target area, substitutes them into the environmental factor dynamic coupling risk propagation model, calculates the disaster risk probability sequence of the target area at different future time points, and generates a risk probability raster dataset updated according to the time series. The hierarchical unit uses the natural breakpoint method to classify the disaster risk probability sequence of the target area at different future time points, and sets low, medium, high and extremely high risk probability threshold intervals to obtain multiple levels. The disaster risk probability sequence at different future time points is divided into the corresponding threshold intervals, generating a dynamic risk classification and early warning map with spatiotemporal dimensions, so as to realize intelligent geological disaster risk prevention and control.

Images