Landslide disaster susceptibility intelligent evaluation system and method
By combining GBDT, SVM, and Adaboost algorithms with feature fusion of deep learning models, the problem of insufficient model generalization ability in landslide hazard susceptibility assessment is solved, and high-precision landslide susceptibility assessment and prediction map generation are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-27
- Publication Date
- 2026-03-27
AI Technical Summary
Existing technologies for assessing landslide hazard susceptibility neglect the spatiotemporal heterogeneity of data during model training, resulting in insufficient model generalization ability and inadequate feature extraction and fusion among multiple models. This makes it difficult to accurately capture the complex interactions between geological factors, leading to low prediction accuracy.
Gradient boosting tree (GBDT), support vector machine (SVM) and Adaboost algorithm are used as base models. Deep learning models are combined for feature fusion. Features are learned through multilayer perceptron structure. The natural breakpoint method is used to divide the susceptibility results into five levels to generate a landslide susceptibility prediction map.
It significantly improved the accuracy of landslide susceptibility assessment, with an AUC value of 0.9056, and generated an intuitive susceptibility prediction map, providing reliable decision support for geological disaster prevention and control.
Smart Images
Figure CN121745360A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological disaster prediction and assessment technology, specifically to an intelligent assessment system and method for landslide disaster susceptibility. Background Technology
[0002] Landslides, as a common geological disaster, pose a serious threat to human life and property safety and socio-economic development.
[0003] With global climate change and increased human activity, the frequency and impact of landslides are expanding. Traditional landslide susceptibility assessment methods mainly rely on statistical models and empirical formulas, such as the analytic hierarchy process (AHP), logistic regression, and information content method.
[0004] However, these methods have significant limitations when dealing with nonlinear relationships and high-dimensional data, making it difficult to accurately capture the complex interactions between geological factors. In recent years, machine learning techniques such as GBDT, Support Vector Machine (SVM), and Adaboost have been widely used in landslide prediction. Although these algorithms have improved prediction accuracy to some extent, they still have shortcomings in feature extraction and model fusion, are prone to overfitting, and cannot fully utilize the advantages of multiple models. In addition, existing methods often ignore the spatiotemporal heterogeneity of data during model training, resulting in insufficient model generalization ability.
[0005] Therefore, there is an urgent need for a landslide hazard susceptibility assessment method that can effectively integrate multi-source geological data, fully utilize the advantages of multiple models, and improve prediction accuracy and reliability. Summary of the Invention
[0006] The purpose of this invention is to provide an intelligent assessment system and method for landslide hazard susceptibility to solve the technical problems mentioned in the background art.
[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent assessment method for landslide hazard susceptibility, comprising at least the following steps:
[0008] S1: Data preprocessing and collection;
[0009] S2: Factor extraction, extracting multiple factor layers that affect landslide occurrence, with the original geological factor data directly used for subsequent analysis;
[0010] S3: Base model construction, constructing base models for the GBDT model, Support Vector Machine (SVM) model, and Adaboost model, and training each base model;
[0011] S4: Feature Fusion: The prediction results of the base model are used as feature inputs to the deep learning fusion model, which fuses and learns features through a multilayer perceptron structure;
[0012] S5: Model Training: Train the deep learning fusion model using historical landslide data and optimize the model parameters;
[0013] S6: Susceptibility Assessment: Apply the model to the target area to generate a landslide susceptibility prediction map, and classify the susceptibility results into five levels: very low, low, medium, high, and very high using the natural breakpoint method.
[0014] Furthermore, the data preprocessing in S1 includes at least coordinate system 1, spatial resolution resampling, and data normalization;
[0015] Coordinate System 1: Convert all data sources into a unified coordinate system, using the WGS1984UTM projected coordinate system to ensure spatial consistency between different data.
[0016] Spatial resolution resampling: Resample data of different resolutions to a uniform resolution, typically set to 30 meters, to match the resolution of most remote sensing data and digital elevation models (DEMs);
[0017] Data normalization: Normalizing data maps values to the interval [0, 1]. The formula is as follows:
[0018] x′=
[0019] Where x is the original data value; and These are the minimum and maximum values of the data, respectively; x′ is the normalized data value.
[0020] The data collection in S1 is multi-source geospatial data, which includes at least elevation, slope, geological lithology, vegetation cover and precipitation, and is standardized using geographic information system (GIS) technology.
[0021] Furthermore, the multiple factor layers are extracted from the preprocessed data;
[0022] The factor layer includes at least elevation, slope, aspect, and distance from the river.
[0023] Furthermore, S3 includes at least the following steps:
[0024] Building the GBDT model:
[0025] Multiple decision trees are constructed, and each decision tree is trained using a subset of the dataset sampled from Bootstrap. The computational expression for GBDT is as follows:
[0026]
[0027] Among them, Ly i F(x) i )] is a measure of the difference between the predicted and the true values (such as mean squared error); Ω is a regularization term used to prevent overfitting.
[0028] Construct a Support Vector Machine (SVM) model, which is the same as the SVM model:
[0029] By finding the optimal classifying hyperplane that maximizes the classification margin, the decision function of the SVM model is:
[0030] )
[0031] in, They are Lagrange multipliers; These are the sample labels; K is the kernel function. b is the support vector; b is the bias.
[0032] Building the Adaboost model:
[0033] By employing an exponential loss function minimization strategy, a linear combination of weak classifiers is constructed. This progressively focuses on misclassified samples in the feature space. The AdaBoost model's prediction formula is as follows:
[0034]
[0035] in, It is the t-th weak classifier. It is its weight.
[0036] Furthermore, S4 includes at least the following steps:
[0037] Feature fusion: The predicted probabilities of GBDT, SVM and Adaboost models are used as input features to construct a fused feature vector;
[0038] Deep learning model: A multilayer perceptron (MLP) structure is constructed, comprising three fully connected layers with 1024, 512, and 256 neurons in each layer, respectively, and employing the ReLU activation function and Layer Normalization.
[0039] The first layer is the input layer, which receives the fused feature vector;
[0040] The second layer is a fully connected layer with 1024 neurons and ReLU activation function:
[0041] ReLU(x) = max(0,x)
[0042] The third layer is a fully connected layer with 512 neurons and the activation function is ReLU.
[0043] The fourth layer is a fully connected layer with 256 neurons and the activation function is ReLU;
[0044] The output layer outputs the probability of a landslide occurring.
[0045] Model training: The Adam optimization algorithm is used to update model parameters through backpropagation, minimizing the cross-entropy loss function. The expression for the cross-entropy loss function is:
[0046]
[0047] in, This represents the result of the loss function; Indicates the number of samples; This represents the true label of the i-th sample; This represents the probability that the i-th sample predicted by the model is a landslide point.
[0048] Furthermore, S5 includes at least the following steps:
[0049] Model validation involves evaluating model performance using a test dataset, plotting the ROC curve, and calculating the AUC value. A higher AUC value indicates better predictive performance of the model.
[0050] Landslide susceptibility prediction: The model is applied to the target area to generate a landslide susceptibility prediction map;
[0051] Susceptibility partitioning uses the natural breakpoint method to classify susceptibility results into five levels: very low, low, medium, high, and very high.
[0052] A landslide hazard susceptibility intelligent assessment system includes a data processing module, a factor extraction module, a model training module, and an assessment and mapping module;
[0053] The data processing module is used to collect, store, and preprocess geospatial data;
[0054] The factor extraction module is used to extract multiple factor layers that affect the occurrence of landslides;
[0055] The model training module is used to build and train base models and deep learning fusion models, and the model training module supports distributed computing.
[0056] The assessment and mapping module is used to generate landslide susceptibility prediction maps and perform susceptibility zoning.
[0057] Compared with the prior art, the beneficial effects of the present invention are:
[0058] 1. This invention improves the accuracy of landslide susceptibility assessment by fusing prediction results from Gradient Boosting Tree (GBDT), Support Vector Machine (SVM), and Adaboost algorithms, and utilizing a deep learning model for feature fusion and optimization. Experiments show that the AUC value of this fusion model can reach 0.9056, a significant improvement compared to single models, providing a more reliable basis for the prevention and management of geological disasters.
[0059] 2. This invention avoids subjective weighting by directly using raw geological factor data, ensuring the authenticity of the data. Simultaneously, it utilizes the natural discontinuity method to classify susceptibility results into five levels (extremely low, low, medium, high, and extremely high), generating an intuitive landslide susceptibility prediction map, providing clear and easy-to-understand decision support for geological disaster prevention and control. Attached Figure Description
[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0061] Figure 1 This is the overall flowchart of the present invention;
[0062] Figure 2 This is a schematic diagram of the ROC curve of the present invention;
[0063] Figure 3 This is a schematic diagram of the four performance indicators of the model of the present invention. Detailed Implementation
[0064] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0065] Example 1:
[0066] Please see Figures 1-3 A method for intelligent assessment of landslide hazard susceptibility, comprising at least the following steps:
[0067] S1: Data preprocessing and collection;
[0068] S2: Factor extraction, extracting multiple factor layers that affect landslide occurrence, with the original geological factor data directly used for subsequent analysis;
[0069] S3: Base model construction, constructing base models for the GBDT model, Support Vector Machine (SVM) model, and Adaboost model, and training each base model;
[0070] S4: Feature Fusion: The prediction results of the base model are used as feature inputs to the deep learning fusion model, which fuses and learns features through a multilayer perceptron structure;
[0071] S5: Model Training: Train the deep learning fusion model using historical landslide data and optimize the model parameters;
[0072] S6: Susceptibility Assessment: Apply the model to the target area to generate a landslide susceptibility prediction map, and classify the susceptibility results into five levels: very low, low, medium, high, and very high using the natural breakpoint method.
[0073] 1. Data Preprocessing
[0074] Multi-source geospatial data of the study area were collected, including elevation, slope, geological lithology, vegetation cover, precipitation, river distribution, and historical landslide data. Geographic Information System (GIS) technology was used to standardize the data; the specific steps are as follows:
[0075] 1.1 Coordinate System One: Convert all data sources to a unified coordinate system, typically WGS1984UTM projected coordinate system, to ensure spatial consistency between different data. The following are the data sources.
[0076] Digital Elevation Model (DEM): Sourced from geospatial data cloud, with a resolution of 30 meters.
[0077] Lithology and distance from faults: sourced from the Geological Science Publishing System.
[0078] Road and hydrological data: sourced from geospatial data cloud.
[0079] Rainfall data: sourced from the Qinghai-Tibet Plateau Data Center, with a resolution of 1 kilometer.
[0080] Normalized Difference Vegetation Index (NDVI): Sourced from the Qinghai-Tibet Plateau Data Center.
[0081] Land use type: Sourced from Professor Huang Xin of Wuhan University, resolution 30 meters.
[0082] 1.2 Spatial resolution resampling: Resample data of different resolutions to a uniform resolution, generally set to 30 meters, to match the resolution of most remote sensing data and digital elevation models (DEMs).
[0083] 1.3 Data Normalization: The data is normalized to map the values to the [0,1] interval. The formula is as follows:
[0084] x′=
[0085] Where x is the original data value, x min and x max These are the minimum and maximum values of the data, respectively, and x′ is the normalized data value.
[0086] 2. Factor Extraction
[0087] Extract multiple factor layers that influence landslide occurrence from the preprocessed data, including but not limited to the following factors:
[0088] Elevation: Extracted from the digital elevation model (DEM), it is usually divided into multiple intervals, such as <600 meters, 600-800 meters, 800-1000 meters, etc.
[0089] Slope: Calculated using DEM and divided into multiple levels, such as <10°, 10°-20°, 20°-30°, etc.
[0090] Slope aspect: Indicates the orientation of a slope, which affects the distribution of water and vegetation.
[0091] Distance from the river: Calculated based on river system data, divided into multiple intervals, such as 0-100 meters, 100-200 meters, etc.
[0092] Distance from road: Calculated based on road network data, divided into multiple intervals, such as 0-10 meters, 10-20 meters, etc.
[0093] 3. Base Model Construction and Training
[0094] Base models for GBDT, Support Vector Machine (SVM), and Adaboost algorithms are constructed respectively, and each base model is trained:
[0095] GBDT model: Construct multiple decision trees, each of which is trained using a subset of the dataset obtained by Bootstrap sampling.
[0096] Parameter settings: Learning rate = 0.1, Number of decision trees = 100, Maximum depth = 5, Minimum split = 2, Minimum leaf node = 1. A training dataset was constructed using the same historical landslide and non-landslide data. The training dataset was split into training and validation sets in a 7:3 ratio. The GBDT model was built using the GradientBoostingClassifier class from Python's scikit-learn library. The final model achieved an AUC of 0.8925 on the validation set.
[0097] The expression for GBDT is as follows:
[0098]
[0099] Among them, Ly i ,F(x i )] is a measure of the difference between the predicted and the true values (such as mean squared error); Ω is a regularization term used to prevent overfitting.
[0100] Support Vector Machine (SVM) model: finds the optimal classification hyperplane that maximizes the classification margin.
[0101] Parameter settings: Penalty parameter C = 1.0, kernel function = radial basis function (RBF), RBF parameter γ = 0.1. A training dataset was constructed using the same historical landslide and non-landslide data. The training dataset was split into training and validation sets in a 7:3 ratio. An SVM model was built using the SVC class from Python's scikit-learn library. The final model achieved an AUC of 0.8190 on the validation set.
[0102] The decision function of SVM is:
[0103] )
[0104] in It is a Lagrange multiplier. These are the sample labels, and K is the kernel function. 'b' is the support vector, and 'b' is the bias.
[0105] The Adaboost model constructs a linear combination of weak classifiers through an exponential loss function minimization strategy, progressively focusing on misclassified samples in the feature space.
[0106] Parameter settings: Number of weak classifiers = 100, learning rate = 0.1, weak classifier type = decision stump. A training dataset was constructed using the same historical landslide and non-landslide data. The training dataset was split into training and validation sets in a 7:3 ratio. An AdaBoost model was built using the AdaBoostClassifier class from Python's scikit-learn library. The final model achieved an AUC of 0.8554 on the validation set.
[0107] Its prediction formula is:
[0108]
[0109] in It is the t-th weak classifier. It is its weight.
[0110] 4. Feature fusion and deep learning model training
[0111] The prediction results of the base model are used as features to input into the deep learning fusion model. The specific steps are as follows:
[0112] 4.1 Feature Fusion: The predicted probabilities of GBDT, SVM, and Adaboost are used as input features to construct a fused feature vector. The predicted probability of each base model is calculated using the following formula:
[0113]
[0114] in It is the output logits of the base model.
[0115] 4.2 Deep Learning Model: A multilayer perceptron (MLP) structure is constructed, containing three fully connected layers with 1024, 512, and 256 neurons in each layer, respectively. ReLU activation function and Layer Normalization are employed.
[0116] First layer: Input layer, which receives the fused feature vector.
[0117] The second layer is a fully connected layer with 1024 neurons and the activation function is ReLU: ReLU(x) = max(0,x).
[0118] The third layer is a fully connected layer with 512 neurons and the activation function is ReLU.
[0119] Fourth layer: Fully connected layer, with 256 neurons and ReLU activation function.
[0120] Output layer: Outputs the probability of a landslide occurring, using the Sigmoid activation function: σ(x) =
[0121] 4.3 Model Training: The Adam optimization algorithm is used to update model parameters through backpropagation, minimizing the cross-entropy loss function. The expression for the cross-entropy loss function is:
[0122]
[0123] in, This represents the result of the loss function. Indicates the number of samples. This represents the true label of the i-th sample. This represents the probability that the i-th sample predicted by the model is a landslide point. The final model achieved an AUC of 0.9056 on the validation set.
[0124] 5. Susceptibility assessment and mapping
[0125] 5.1 Model Validation: The model performance was evaluated using an independent test dataset. The test dataset had no overlap with the training dataset to ensure the objectivity of the evaluation results. ROC curves were plotted and AUC values were calculated. The fusion model achieved an AUC value of 0.9056, indicating that the model has good predictive performance.
[0126] 5.2 Landslide Susceptibility Prediction: Apply the model to the target area to generate a landslide susceptibility prediction map.
[0127] 5.3 Susceptibility Zoning: Susceptibility results are divided into five levels using the natural breakpoint method: very low, low, medium, high, and very high. The specific steps are as follows:
[0128] 5.3.1 Calculate the susceptibility index for each grid cell.
[0129] 5.3.2 The classification threshold was determined using the natural breakpoint method, and the susceptibility index was divided into five intervals.
[0130] 5.3.3 Generate a landslide susceptibility zoning map, with different colors representing different susceptibility levels, providing intuitive decision support for geological disaster prevention and control.
[0131] Example 2:
[0132] This embodiment proposes an intelligent assessment system for landslide hazard susceptibility based on the above embodiment one, including a data processing module, a factor extraction module, a model training module, and an assessment and mapping module;
[0133] The data processing module is used to collect, store, and preprocess geospatial data;
[0134] The factor extraction module is used to extract multiple factor layers that influence landslide occurrence;
[0135] The model training module is used to build and train base models and deep learning fusion models. The model training module supports distributed computing and can process large-scale datasets in parallel to improve training efficiency.
[0136] The assessment and mapping module is used to generate landslide susceptibility prediction maps and perform susceptibility zoning.
[0137] In summary:
[0138] This invention utilizes Geographic Information System (GIS) to preprocess multi-source geospatial data, extracting multiple factor layers that influence landslide occurrence; it directly uses raw geological factor data without weighting, and fuses and learns the geological factor features through a deep learning network; it constructs a base model incorporating GBDT, Support Vector Machine (SVM), and Adaboost algorithms, and uses the prediction results of the base model as feature inputs to train the deep learning fusion model; it uses the trained model to perform landslide susceptibility analysis on the target area, and divides the susceptibility results into five levels (extremely low, low, medium, high, and extremely high) using the natural breakpoint method, generating a landslide susceptibility prediction map.
[0139] This invention effectively improves the accuracy and reliability of landslide susceptibility assessment through multi-model feature fusion and deep learning technology, with an AUC value of up to 0.9056, providing a scientific basis for geological disaster prevention and control.
[0140] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for intelligent assessment of landslide hazard susceptibility, characterized in that: At least the following steps are included: S1: Data preprocessing and collection; S2: Factor extraction, extracting multiple factor layers that affect landslide occurrence, with the original geological factor data directly used for subsequent analysis; S3: Base model construction, constructing base models for the GBDT model, support vector machine model, and Adaboost model, and training each base model; S4: Feature Fusion: The prediction results of the base model are used as feature inputs to the deep learning fusion model, which fuses and learns features through a multilayer perceptron structure; S5: Model Training: Train the deep learning fusion model using historical landslide data and optimize the model parameters; S6: Susceptibility Assessment: Apply the model to the target area to generate a landslide susceptibility prediction map, and classify the susceptibility results into five levels: very low, low, medium, high, and very high using the natural breakpoint method.
2. The intelligent assessment method for landslide hazard susceptibility according to claim 1, characterized in that: The data preprocessing in S1 includes at least coordinate system one, spatial resolution resampling, and data normalization. Coordinate System 1: Convert all data sources into a unified coordinate system, using the WGS1984UTM projected coordinate system to ensure spatial consistency between different data. Spatial resolution resampling: resampling data at different resolutions to a uniform resolution to match the resolution of most remote sensing data and digital elevation models; Data normalization: Normalizing data maps values to the interval [0, 1]. The formula is as follows: x′= Where x is the original data value; and These are the minimum and maximum values of the data, respectively; x′ is the normalized data value. The data collection in S1 is multi-source geospatial data, which includes at least elevation, slope, geological lithology, vegetation cover and precipitation, and is standardized using geographic information system technology.
3. The intelligent assessment method for landslide hazard susceptibility according to claim 2, characterized in that: The multiple factor layers are extracted from the preprocessed data; The factor layer includes at least elevation, slope, aspect, and distance from the river.
4. The intelligent assessment method for landslide hazard susceptibility according to claim 3, characterized in that: The S3 includes at least the following steps: Building the GBDT model: Multiple decision trees are constructed, and each decision tree is trained using a subset of the dataset sampled from Bootstrap. The computational expression for GBDT is as follows: Among them, Ly i F(x) i )] is a measure of the difference between the predicted and the true values; Ω is a regularization term used to prevent overfitting; Construct a Support Vector Machine (SVM) model, which is the same as the SVM model: By finding the optimal classifying hyperplane that maximizes the classification margin, the decision function of the SVM model is: ) in, They are Lagrange multipliers; These are the sample labels; K is the kernel function. b is the support vector; b is the bias. Building the Adaboost model: By employing an exponential loss function minimization strategy, a linear combination of weak classifiers is constructed. This progressively focuses on misclassified samples in the feature space. The AdaBoost model's prediction formula is as follows: in, It is the t-th weak classifier. It is its weight.
5. The intelligent assessment method for landslide hazard susceptibility according to claim 4, characterized in that: The S4 includes at least the following steps: Feature fusion: The predicted probabilities of GBDT, SVM and Adaboost models are used as input features to construct a fused feature vector; Deep learning model: A multilayer perceptron structure is constructed, comprising three fully connected layers with 1024, 512, and 256 neurons in each layer, respectively, and employing ReLU activation function and Layer Normalization. The first layer is the input layer, which receives the fused feature vector; The second layer is a fully connected layer with 1024 neurons and ReLU activation function: ReLU(x) = max(0,x) The third layer is a fully connected layer with 512 neurons and the activation function is ReLU. The fourth layer is a fully connected layer with 256 neurons and the activation function is ReLU; The output layer outputs the probability of a landslide occurring. Model training: The Adam optimization algorithm is used to update model parameters through backpropagation, minimizing the cross-entropy loss function. The expression for the cross-entropy loss function is: in, This represents the result of the loss function; Indicates the number of samples; This represents the true label of the i-th sample; This represents the probability that the i-th sample predicted by the model is a landslide point.
6. The intelligent assessment method for landslide hazard susceptibility according to claim 5, characterized in that: The S5 includes at least the following steps: Model validation involves evaluating model performance using a test dataset, plotting the ROC curve, and calculating the AUC value. A higher AUC value indicates better predictive performance of the model. Landslide susceptibility prediction: The model is applied to the target area to generate a landslide susceptibility prediction map; Susceptibility partitioning uses the natural breakpoint method to classify susceptibility results into five levels: very low, low, medium, high, and very high.
7. A landslide hazard susceptibility intelligent assessment system, used in the landslide hazard susceptibility intelligent assessment method according to any one of claims 1-6, characterized in that: It includes a data processing module, a factor extraction module, a model training module, and an evaluation and graphing module; The data processing module is used to collect, store, and preprocess geospatial data; The factor extraction module is used to extract multiple factor layers that affect the occurrence of landslides; The model training module is used to build and train base models and deep learning fusion models, and the model training module supports distributed computing. The assessment and mapping module is used to generate landslide susceptibility prediction maps and perform susceptibility zoning.