Landslide susceptibility assessment method based on dynamic adaptive weighted ensemble model

Abstract

The present application relates to the technical field of geological disaster risk assessment, and more particularly to a landslide susceptibility evaluation method based on a dynamic self-adaptive weighted integrated model. The steps are as follows: a plurality of base learners are constructed, and based on cross-validation, the meta-features are generated, and the meta-learners are trained to obtain the static prior weights of each base model; a dynamic weight generation network is designed, and the landslide influence factors are input to generate exclusive dynamic weights for each sample; a two-level weight fusion module is constructed, and the static and dynamic weights are adaptively balanced through the learnable parameter λ to obtain the final integrated weight at the sample level; and the gradient descent method is used to perform end-to-end joint optimization on the dynamic weight generation network and the fusion coefficient with the validation set loss as the target. The present application realizes sample-level adaptive regulation and control of the base model weights, significantly improves the precision, robustness and interpretability of landslide susceptibility prediction, and has a wide application prospect in the fine evaluation of landslide disasters.

Description

Technical Field

[0001] This invention relates to the field of geological hazard risk assessment technology, and in particular to a landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model. Background Technology

[0002] Landslides are among the most widespread, frequent, and devastating natural disasters, posing a significant threat to people's lives and property, infrastructure, and the ecological environment. Landslide susceptibility assessment aims to quantitatively predict the probability of landslides in different spatial units of a region based on historical landslide sites and disaster-prone environmental factors. It is a core foundation for landslide disaster risk management and territorial spatial planning.

[0003] Early landslide susceptibility assessments primarily relied on statistical analysis methods such as expert scoring, information content models, and the weight of evidence method. While these methods offer interpretability, they struggle to characterize the complex nonlinear relationships between landslides and influencing factors. With the development of machine learning techniques, models such as support vector machines, random forests, and artificial neural networks have been widely adopted in susceptibility assessments, significantly improving prediction accuracy. However, single models inherently suffer from limitations in generalization ability, sensitivity to data distribution, and susceptibility to local optima.

[0004] To overcome the limitations of single models, ensemble learning strategies have gradually become a research hotspot. Existing techniques typically employ two ensemble paradigms: homogeneous ensembles, such as Random Forest, XGBoost, and LightGBM, which improve overall performance through the diversity of base learners; and heterogeneous ensembles, such as Stacking and Blending, which fuse the outputs of multiple heterogeneous base models through meta-learners. The latter, by integrating the complementary advantages of different models, has achieved better results than single models in landslide susceptibility assessment.

[0005] Although heterogeneous integration methods have shown great potential, existing solutions still face the following technical bottlenecks:

[0006] (1) The weight allocation mode is singular and lacks sample-level adaptive capability. Most ensemble methods use global fixed weights in the fusion stage—either calculating static weights based on the overall performance of the validation set or relying on a meta-learner to train global mapping relationships. This "one weight applies to all samples" mode ignores the heterogeneity of landslide influencing factors in different spatial units, resulting in improper weight allocation under local complex geological conditions, and the model is unable to accurately capture the spatial variability of landslide occurrence conditions.

[0007] (2) The separation between weight generation and optimization objective makes it difficult to guarantee global optimum. Some studies have attempted to introduce dynamic weight mechanism, such as estimating weight based on sample similarity, clustering results or local accuracy. However, the weight generation process is usually independent of the final prediction loss of the model. The weight assignment lacks clear optimization objective guidance and cannot minimize evaluation error in an end-to-end manner, which limits the effectiveness of adaptive weight control.

[0008] (3) Static priors and dynamic adjustment lack a synergistic mechanism. Existing technologies either rely entirely on global static weights, ignoring sample specificity, or adopt data-driven dynamic weights, which are susceptible to noise samples and discard the stable contribution information of the base model that has been verified on the overall distribution. How to organically integrate global stability and local adaptability to achieve synergistic gains is a key issue that ensemble learning urgently needs to address in the field of landslide susceptibility.

[0009] To address this issue, a landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model was designed to provide a technical solution for the aforementioned technical problems. Through a complete technical chain of "static prior guidance + dynamic sample adaptation + gradient optimization closed loop", it achieves for the first time in the field of landslide susceptibility assessment sample-level fine-tuning of integrated weights and optimization goal-oriented adaptive learning, significantly improving the accuracy, robustness and spatial adaptability of the assessment results. It has outstanding substantive features and significant technical progress. Summary of the Invention

[0010] Therefore, it is necessary to provide a landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model to address the aforementioned technical problems.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0012] The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model comprises the following steps:

[0013] S1. Obtain the landslide impact factor dataset and landslide hazard catalog data for the target area, and construct labeled training and test sample sets after data preprocessing;

[0014] S2. Construct a base model pool consisting of multiple heterogeneous base learners, and train each base learner independently using the training sample set to obtain multiple base prediction models;

[0015] S3. Using a K-fold cross-validation strategy, generate meta-features for each base prediction model on the training samples, and concatenate the meta-features of all base models into a meta-feature matrix. With the landslide true label as supervision, train the meta-learner and extract the weight coefficients of the meta-learner as the static global weights of each base model.

[0016] S4. Construct a dynamic weight generation network that takes sample features as input;

[0017] S5. Introduce a learnable fusion coefficient λ, and perform linear weighted fusion of the static global weights obtained in step S3 and the sample-level dynamic weights obtained in step S4 to obtain the final integrated weights of each sample, and normalize the fused weights.

[0018] S6. Input the test samples into each base prediction model to obtain the predicted landslide occurrence probability value of each model. Use the final integrated weights obtained in step S5 to perform a weighted summation of the predicted probabilities to obtain the landslide susceptibility index of the samples.

[0019] S7. Divide the training sample set into a validation subset, use the prediction loss on the validation subset as the objective function, perform joint optimization to complete the training of the entire ensemble model;

[0020] S8. After the landslide impact factor data of the area to be evaluated is processed in the same way, it is input into the trained ensemble model, and the landslide susceptibility index of each spatial unit is calculated through step S6.

[0021] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, the landslide influencing factors in step S1 include one or more of the following: elevation, slope, aspect, curvature, stratigraphic lithology, fault distance, water system distance, road distance, land use type, normalized vegetation index, and rainfall intensity; the data preprocessing includes missing value imputation, outlier correction, and normalization or standardization.

[0022] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, the heterogeneous base learners in step S2 include, but are not limited to, any four or more combinations of Ultimate Gradient Boosting Tree (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Random Forest (RandomForest), Gradient Boosting Decision Tree (GBDT), and Deep Neural Network (DNN). Each base learner adopts differentiated hyperparameter configurations to enhance ensemble diversity.

[0023] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, the meta-learner in step S3 adopts a logistic regression model with L2 regularization, and its weight coefficients are used as static global weights after sparse constraints; if the meta-learner does not have explicit weight coefficients, the normalized proportion of the average AUC value of each base model in cross-validation is used as an alternative estimate of the static weights.

[0024] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, the dynamic weight generation network in step S4 is a three-layer fully connected neural network with 32 and 16 neurons in the hidden layer, respectively. The activation function is ReLU, and a batch normalization (BatchNorm) and dropout layer are introduced after the first hidden layer. The output layer uses the Softmax function, and the output dimension is the same as the number of base models.

[0025] As a preferred embodiment of the landslide susceptibility evaluation method based on the dynamic adaptive weighted ensemble model provided by the present invention, the fusion coefficient λ in step S5 is initialized to 0.5, and in the optimization process of step S7, it is mapped to the (0,1) interval through the Sigmoid function to ensure the boundedness of the fusion weight.

[0026] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, in step S7, a validation subset is divided from the training sample set. The prediction loss on the validation subset is used as the objective function, and the gradient descent method is used to jointly optimize the parameters and fusion coefficients λ of the dynamic weight generation network, so that the dynamic weight generation network learns the weight mapping relationship that minimizes the validation error, and the training of the entire ensemble model is completed.

[0027] As a preferred embodiment of the landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by the present invention, it further includes a feature importance screening step, which is used to calculate the importance score of each influencing factor using a random forest model before step S2, and screen out the top N dominant factors as model input, thereby reducing feature redundancy and improving computational efficiency.

[0028] A dynamic adaptive weighted integrated landslide susceptibility assessment system based on dual gradient guidance includes:

[0029] The data acquisition and preprocessing module is used to execute step S1 in the method;

[0030] The base model training module is used to execute step S2 in the method;

[0031] The static weight learning module is used to execute step S3 in the method;

[0032] The dynamic weight generation module is used to execute step S4 in the method, and it integrates a dynamic weight generation network.

[0033] The two-level weight fusion module is used to execute step S5 in the method;

[0034] An integrated prediction module is used to execute step S6 in the method;

[0035] The gradient-guided optimization module is used to execute step S7 in the method, and to perform end-to-end optimization of the dynamic weight generation network and fusion coefficients with the goal of verifying the loss.

[0036] The output and visualization module is used to generate landslide susceptibility zoning maps, model performance evaluation indicators, ROC curves, feature importance ranking maps, and output evaluation reports.

[0037] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described herein.

[0038] It is clear without a doubt that the technical solution described above in this application can solve the technical problem that this application aims to address.

[0039] Meanwhile, through the above technical solutions, the present invention has at least the following beneficial effects:

[0040] 1. The landslide susceptibility assessment method based on the dynamic adaptive weighted ensemble model provided by this invention breaks through the static paradigm of the traditional ensemble method of "one weight applies to all". By constructing a dynamic weight generation network with sample features as input, the base model weights are independently assigned to each evaluation unit. This mechanism can adaptively adjust the model contribution according to the local disaster-prone environment characteristics, which significantly improves the model's ability to accurately characterize complex terrain and heterogeneous geological conditions.

[0041] 2. This invention organically couples static global weights (derived from the cross-validation gradient of the meta-learner) with dynamic weight networks (derived from the end-to-end gradient of the validation loss) to form a dual guidance mechanism of "prior gradient + dynamic gradient". Static weights provide global stability constraints, while dynamic weights enable local fine-tuning. The two are jointly optimized through learnable fusion coefficients, avoiding the overfitting risk of purely data-driven dynamic weights, while overcoming the insufficient adaptability of purely static weights.

[0042] 3. This invention constructs the entire process of base model prediction, dynamic weight generation, two-level fusion, and loss calculation into a differentiable computational graph, achieving for the first time joint optimization of integrated weight allocation and model prediction objectives in the field of landslide susceptibility. The dynamic weight network updates parameters with the goal of directly minimizing the validation loss, giving weight generation a clear optimization objective and resulting in a substantial improvement in accuracy compared to dynamic weight methods based on heuristic rules or independent clustering.

[0043] 4. The present invention has been verified by actual landslide area data. The method proposed in this invention is significantly better than the existing mainstream ensemble models (XGBoost, LightGBM, Random Forest, GBDT and traditional Stacking ensemble) in terms of AUC, F1 score, accuracy and Matthews correlation coefficient. The average AUC improvement is 2.5% to 5.8%, and it shows stronger generalization stability in cross-regional transfer tests.

[0044] 5. This invention reveals the global importance ranking of the base model through static weights, and identifies the advantageous areas of different models in different terrain regions through the spatial distribution visualization of dynamic weights, providing support for the analysis of landslide controlling factors and the evaluation of model decision credibility. The modular design of the system facilitates deployment and can be seamlessly integrated into existing geological disaster early warning platforms, showing broad prospects for industrial application. Attached Figure Description

[0045] 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 accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is the overall flowchart of the present invention;

[0047] Figure 2 This is a schematic diagram of the ROC curve of the present invention. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0049] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0050] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0051] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0052] Example 1

[0053] Reference Figures 1-2The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model has the following steps:

[0054] S1: Data Acquisition and Standardization Processing

[0055] S101: Landslide logging and sample point generation.

[0056] Through high-resolution remote sensing image interpretation (such as GF-2 satellite imagery) and verification through three years of field investigation, 203 historical landslide disaster sites were accurately identified and delineated within Cili County. To address the requirement of combining landslide areal features with point-based model input, and to avoid the problem of insufficient representativeness of single pixels, a buffer circular area with a radius of 30 meters was generated, centered on the geometric center of each landslide body. This buffer effectively covers the main landslide area, and sampling points were systematically deployed within this buffer, ultimately generating 609 positive sample points representing "landslide occurrence". To ensure sample balance, in areas confirmed to have not experienced landslides and with representative geological and geographical conditions (such as flat, intact bedrock areas and stable floodplains), a spatially constrained random sampling method was used to select an equal number of 609 negative sample points. Thus, a model training and validation dataset of a total of 1218 sample points was constructed, with each sample point associated with a binary label (Y: landslide = 1, non-landslide = 0).

[0057] S102: Landslide Influencing Factor Extraction and Database Construction. Based on multi-source data including digital elevation model (DEM, 12.5m resolution), geological map (1:200,000), land use dataset, and annual precipitation interpolation map, 16 environmental factors closely related to landslide development were extracted in ArcGIS and Python environments, forming an initial feature set. Mainly includes:

[0058] Topographic factors: elevation, slope, aspect, plane curvature, profile curvature, topographic relief, and topographic humidity index.

[0059] Geological condition factors: stratigraphic lithology (coded as a categorical variable), distance from fault.

[0060] Hydrological factors: distance from the river, multi-year average rainfall.

[0061] Land cover and human activity factors: Normalized Difference Vegetation Index (NDVI), land use type, and distance from roads.

[0062] Spatial registration and rasterization were performed on the landslide influencing factor data to form a unified raster unit; the factor value of each raster unit was extracted, and an initial feature matrix X∈ℝ^(n×p) was constructed, where n is the total number of raster units and p is the number of influencing factors;

[0063] right Missing values ​​are checked, and numerical features are imputed using the median. All numerical features are normalized and scaled to the [0,1] interval, as shown in the following formula:

[0064] ;

[0065] in, The value of the j-th feature of the i-th sample after normalization. Let j be the feature value of the i-th sample. and Let X be the minimum and maximum values ​​of feature j, respectively. Let X be the processed feature set.

[0066] S103: Data Preprocessing and Partitioning. Continuous factors were checked for missing values ​​(no missing values ​​in this case) and normalized using Min-Max standardization. Categorical variables (e.g., lithology) were uniquely coded. The processed standardized feature set is denoted as X. Stratified sampling was used to randomly partition the 1218 samples into a training set of 852 samples and an independent test set of 366 samples at a 7:3 ratio. , This ensures that the ratio of positive to negative samples remains consistent in the training and test sets (approximately 1:1).

[0067] S2. Base Model Pool Construction and Training

[0068] S201 selects M heterogeneous machine learning models as base learners to form a base model pool. (1) Heterogeneous algorithm principles: It includes at least three of the following: tree ensemble models (such as XGBoost, LightGBM), bagged ensemble models (such as random forest), gradient boosting models (such as GBDT), and neural network models.

[0069] (2) Heterogeneous feature space: Different feature subsets or feature transformation methods can be used for each base model, such as principal component analysis, dimensionality reduction features of autoencoders, original features, etc.

[0070] (3) Heterogeneous hyperparameter configuration: By setting hyperparameters such as the number of trees, maximum depth, learning rate, and subsampling ratio in a differentiated manner, the diversity of the model decision boundary can be expanded.

[0071] In this embodiment, M=4, and four models are selected: XGBoost, LightGBM, RandomForest, and GBDT.

[0072] S202 utilizes the training set ( , Each base learner is trained independently to obtain M trained base prediction models. Each base model can output the probability value that a sample belongs to the positive class. After training, the AUC values ​​of each base model are calculated on the validation set to quantify its global discriminative ability. In this embodiment, the AUC values ​​of each base model on the validation set are: XGBoost 0.855, LightGBM 0.848, RandomForest 0.860, and GBDT 0.849 (see Appendix). Figure 2 This indicator will serve as the basis for the static weight substitution estimation in S303 and provide a benchmark for subsequent dynamic weight optimization.

[0073] S3. Static Global Weight Learning

[0074] S301 employs a K-fold cross-validation strategy for each base model. Generate its meta-feature vector across all training samples: Randomly divide the training set into K subsets. For the k-th fold, train the model using the remaining K-1 subsets and predict the probability on the k-th subset. After traversing all folds, obtain the predicted probability vector on the complete training set. ;where P m is the predicted probability vector generated by the m-th base model through cross-validation on the training set. train This represents the number of training samples. R is the real number field, indicating that each element in the vector has a real probability value.

[0075] The predicted probability vectors of the M base models are horizontally concatenated to form the meta-feature matrix. ;

[0076] S302 takes the meta-feature matrix Z as input and the true labels as input. For supervision, a meta-learner g is trained; the meta-learner is a logistic regression model with L2 regularization, and its decision function is:

[0077] y ;

[0078] Where σ(·) is the Sigmoid function, and β∈ℝ^M are the weight coefficients corresponding to the base model. For bias terms;

[0079] S303 solves for the meta-learner parameters by minimizing the cross-entropy loss function:

[0080]

[0081] ;

[0082] in The loss function of the meta-learner consists of cross-entropy loss and L2 regularization term. : The true label of the i-th sample (1 for landslides, 0 for non-landslides). : The prediction probability of the meta-learner for the i-th sample. λ: L2 regularization coefficient, controlling the sparsity of the weights. β: The weight coefficient vector assigned by the meta-learner to each base model. The L2 norm squared of the weight vector is used to prevent overfitting.

[0083] The optimal weight coefficient β^ is obtained; after normalizing β^, it is used as the static global weight vector for each base model:

[0084] ;

[0085] in : Static global weight vector, with dimension M, representing the fixed contribution of each base model to the global dataset.

[0086] The meta-learner (logistic regression with L2 regularization) is the optimal weight coefficient vector learned by each base model.

[0087] : The weight coefficients of the j-th base model are taken as positive values, and the negative weights are truncated to 0 to ensure the non-negativity of the static weights.

[0088] When the meta-learner does not have explicit weight coefficients, the normalized proportion of the average AUC value of each base model in cross-validation is used as an alternative estimate of the static weights.

[0089] S4. Sample-level dynamic weight generation

[0090] S401 constructs a dynamic weight generation network Ψ(·;θ), whose input is the original feature vector x∈ℝ^p of the sample, and the output is an M-dimensional weight vector that is non-negative and sums to 1; the network structure is a three-layer fully connected neural network.

[0091] ;

[0092] ;

[0093] ;

[0094] Where θ={ , , , , , } represents the trainable parameters of the network; the Dropout ratio is set to 0.2.

[0095] x: Feature vector of landslide impact factors for the input sample.

[0096] W1, W2, W3: Weight matrices of the fully connected layers.

[0097] b1, b2, b3: Bias vectors of the fully connected layer.

[0098] BatchNorm: Batch normalization operation, which accelerates training and stabilizes the distribution.

[0099] ReLU: An activation function that introduces non-linearity.

[0100] Dropout: Random deactivation operation to prevent overfitting.

[0101] Softmax: The activation function of the output layer, which makes the generated dynamic weights non-negative and sum to 1.

[0102] S402: For any input sample x, the dynamic weight generation network generates a unique weight vector for it.

[0103]

[0104] ;

[0105] Where Ψ(x;θ) is the dynamic weight generation network with parameter θ.

[0106] M: The number of base models.

[0107] The network assigns dynamic weights to the m-th base model for the current sample x.

[0108] This weight vector characterizes the contribution ratio that each base model should be assigned under the current sample features, achieving a leap from "global uniformity" to "sample adaptation".

[0109] S5. Two-level weighted adaptive fusion

[0110] S501 introduces learnable fusion coefficients λ and constrains them to the (0,1) interval using the Sigmoid function to ensure the boundedness of the fusion weights:

[0111] ;

[0112] λ is initialized to 0.5 (i.e., =0), indicating that the static weights and dynamic weights each contribute equally initially. λ is updated with the gradient during subsequent optimization, automatically learning the optimal balance point.

[0113] S502 performs a linear weighted fusion of the static global weights and the dynamically generated weights for each sample x to obtain the final sample-level integrated weights:

[0114] ;

[0115] in : The final integrated weight vector of sample x.

[0116] λ: Learnable fusion coefficient, constrained to the (0,1) interval by the Sigmoid function.

[0117] : Static global weight vector, extracted from the meta-learner.

[0118] The dynamic weight generation network outputs a dynamic weight vector for sample x.

[0119] S503 normalizes the fused weight vector to ensure that the sum of the weights of each base model is 1.

[0120] ;

[0121] : The final integrated weight vector of sample x (unnormalized after fusion).

[0122] : The weight component corresponding to the m-th base model in this vector.

[0123] M: Total number of base models

[0124] The final sample-level weight vector used for integration is obtained. This vector incorporates the prior advantage of the base model on the global dataset (through...). It also incorporates dynamic adjustment based on the current sample features (through...). This enables the collaboration between prior and subsequent knowledge.

[0125] S6. Integrated Prediction and Susceptibility Index Generation

[0126] S601 for test samples The data is then input into M basic prediction models to obtain predicted landslide occurrence probabilities:

[0127] ;

[0128] in The m-th base model for the test sample x test The predicted probability of a landslide occurring.

[0129] fm: The m-th basis prediction model

[0130] S602 uses the final ensemble weights at the sample level calculated in step S5 to perform a weighted summation of the predicted probabilities of each base model, thus obtaining the landslide susceptibility index for that sample:

[0131] ;

[0132] in Test sample x test The landslide susceptibility index indicates that the higher the value, the greater the likelihood of a landslide.

[0133] Sample x test The final ensemble weights assigned to the m-th base model.

[0134] The higher the LSI value, the greater the likelihood of a landslide occurring in that grid cell;

[0135] S7. Gradient-guided end-to-end joint optimization

[0136] S701 randomly selects 10% to 20% of the samples in the training set as a validation subset. These samples do not participate in parameter updates of the dynamic weight generation network; they are only used to calculate the target loss. The remaining samples are used as a subset for training the dynamic network. In this embodiment, 10% of the samples (85) are extracted for validation, and the remaining 767 are used for training.

[0137] S702 uses the prediction loss on the validation subset as the objective function and employs gradient descent to generate network parameters θ and fusion coefficients for dynamic weights. Perform joint optimization; the objective function is defined as the binary cross-entropy loss:

[0138] ;

[0139] in : The objective function for the dynamic optimization phase is the binary cross-entropy loss on the validation set.

[0140] θ: All trainable parameters of the dynamic weight generation network.

[0141] : The original fusion coefficients without Sigmoid constraints, used to control the balance between static and dynamic weights.

[0142] n val : The number of samples in the validation subset.

[0143] y i : The true label of the i-th validation sample (1 for landslides, 0 for non-landslides).

[0144] Sample x iThe landslide susceptibility index (i.e., the final predicted probability of the ensemble model).

[0145] The S703 uses the Adam optimizer with an initial learning rate of 0.001 and a weight decay of 1 × 10⁻⁶. -4 , for θ and The process involves iterative updates; in each iteration, the dynamic weights, fusion weights, ensemble prediction, and loss are first calculated via forward propagation, followed by backpropagation to calculate the gradient and update the parameters.

[0146] ;

[0147] ;

[0148] S704 sets the maximum number of iterations Emax (80 iterations in this example) and monitors the validation loss. If the validation loss does not decrease for P consecutive iterations (P=12), training is terminated, and the value of θ that minimizes the validation loss is saved. After optimization, the fusion coefficient... The parameters θ of the dynamic weight generation network are fixed.

[0149] S8. Model Application and Landslide Susceptibility Zoning

[0150] S801 inputs the landslide impact factor data of the area to be evaluated into the trained ensemble model after the same preprocessing, and calculates the landslide susceptibility index of each spatial unit through step S6.

[0151] S802 uses the natural discontinuity method, quantile method or equal interval method to classify the susceptibility index and generate a landslide susceptibility zoning map.

[0152] The S803 outputs evaluation results, including a susceptibility index raster plot, a grading plot, statistical distribution of weights for each base model, ROC curve (Receiving Receiver Operating Characteristic), AUC value (area under the curve), and feature importance ranking. The method of this invention achieves an AUC of 0.8847 on the test set (after feature selection and optimization), significantly outperforming single base models (maximum 0.860) and traditional ensemble methods (maximum 0.872). (Appendix) Figure 2 The ROC curves of the method of this invention are compared with those of various base models, intuitively demonstrating the advantages of AUC.

[0153] Example 2

[0154] Based on the above embodiment one, the following is disclosed:

[0155] A dynamic adaptive weighted integrated landslide susceptibility assessment system based on dual gradient guidance includes:

[0156] The data acquisition and preprocessing module is used to perform step S1 in Example 1 to collect landslide influencing factor data, spatial registration, normalization and training / test set division;

[0157] The base model training module is used to execute step S2 in Example 1, integrates multiple heterogeneous machine learning algorithms, and completes the independent training and saving of the base model.

[0158] The static weight learning module is used to execute step S3 in Example 1, which trains the meta-learner based on cross-validation meta-features and extracts and normalizes static global weights.

[0159] The dynamic weight generation module is used to execute step S4 in Embodiment 1. It has a built-in dynamic weight generation neural network that takes sample features as input and outputs a sample-level dynamic weight vector.

[0160] The two-level weight fusion module is used to execute step S5 in Example 1, and realizes adaptive weighted fusion of static and dynamic weights through learnable fusion coefficient λ.

[0161] The integrated prediction module is used to execute step S6 in Example 1, receive test samples, call the base model and dynamic weight generation module, and calculate the landslide susceptibility index of each sample.

[0162] The gradient-guided optimization module is used to execute step S7 in Example 1, and to perform end-to-end optimization of the dynamic weight generation network and fusion coefficients with the goal of verifying the loss.

[0163] Feature selection module: used to perform step S8 in Example 1, calculate feature importance using random forest or mutual information method, and select dominant factors;

[0164] The output and visualization module is used to generate landslide susceptibility zoning maps, model performance evaluation indicators, ROC curves, feature importance ranking maps, and output evaluation reports.

[0165] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model, characterized in that, The steps are as follows: S1. Obtain the landslide impact factor dataset and landslide hazard catalog data for the target area, and construct labeled training and test sample sets after data preprocessing; S2. Construct a base model pool consisting of multiple heterogeneous base learners, and train each base learner independently using the training sample set to obtain multiple base prediction models; S3. Using a K-fold cross-validation strategy, generate meta-features for each base prediction model on the training samples, and concatenate the meta-features of all base models into a meta-feature matrix. With the landslide true label as supervision, train the meta-learner and extract the weight coefficients of the meta-learner as the static global weights of each base model. S4. Construct a dynamic weight generation network that takes sample features as input; S5. Introduce a learnable fusion coefficient λ, and perform linear weighted fusion of the static global weights obtained in step S3 and the sample-level dynamic weights obtained in step S4 to obtain the final integrated weights of each sample, and normalize the fused weights. S6. Input the test samples into each base prediction model to obtain the predicted landslide occurrence probability value of each model. Use the final integrated weights obtained in step S5 to perform a weighted summation of the predicted probabilities to obtain the landslide susceptibility index of the samples. S7. Divide the training sample set into a validation subset, use the prediction loss on the validation subset as the objective function, perform joint optimization to complete the training of the entire ensemble model; S8. After the landslide impact factor data of the area to be evaluated is processed in the same way, it is input into the trained ensemble model, and the landslide susceptibility index of each spatial unit is calculated through step S6.

2. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, The landslide influencing factors mentioned in step S1 include one or more of the following: elevation, slope, aspect, curvature, lithology, fault distance, water system distance, road distance, land use type, normalized vegetation index, and rainfall intensity; the data preprocessing includes missing value imputation, outlier correction, and normalization or standardization.

3. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, The heterogeneous base learners mentioned in step S2 include, but are not limited to, any four or more combinations of Extreme Gradient Boosting Tree (XGBoost), Lightweight Gradient Boosting Machine (LightGBM), Random Forest (RandomForest), Gradient Boosting Decision Tree (GBDT), and Deep Neural Network (DNN). Each base learner employs differentiated hyperparameter configurations to enhance ensemble diversity.

4. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, The meta-learner mentioned in step S3 adopts a logistic regression model with L2 regularization, and its weight coefficients are used as static global weights after being subject to sparse constraints. If the meta-learner does not have explicit weight coefficients, the normalized proportion of the average AUC value of each base model in cross-validation is used as an alternative estimate of the static weights.

5. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, The dynamic weight generation network described in step S4 is a three-layer fully connected neural network with 32 and 16 neurons in the hidden layers, respectively. The activation function is ReLU, and a batch normalization (BatchNorm) and dropout layer are introduced after the first hidden layer. The output layer uses the Softmax function, and the output dimension is the same as the number of base models.

6. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, In step S5, the fusion coefficient λ is initialized to 0.5, and in the optimization process of step S7, it is mapped to the (0,1) interval through the Sigmoid function to ensure the boundedness of the fusion weight.

7. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, In step S7, a validation subset is divided from the training sample set. The prediction loss on the validation subset is used as the objective function. The gradient descent method is used to jointly optimize the parameters and fusion coefficients λ of the dynamic weight generation network, so that the dynamic weight generation network learns the weight mapping relationship that minimizes the validation error, thus completing the training of the entire ensemble model.

8. The landslide susceptibility assessment method based on a dynamic adaptive weighted ensemble model according to claim 1, characterized in that, It also includes a feature importance screening step, which is used to calculate the importance score of each influencing factor using a random forest model before step S2, and screen out the top N dominant factors as model input, thereby reducing feature redundancy and improving computational efficiency.

9. A dynamic adaptive weighted integrated landslide susceptibility assessment system based on dual gradient guidance, characterized in that, include: The data acquisition and preprocessing module is used to perform step S1 in claim 1; A base model training module for performing step S2 in claim 1; A static weight learning module is used to perform step S3 in claim 1; The dynamic weight generation module is used to execute step S4 in claim 1, and has an integrated dynamic weight generation network. A two-level weight fusion module is used to perform step S5 in claim 1; An integrated prediction module is used to perform step S6 in claim 1; The gradient-guided optimization module is used to execute step S7 in claim 1, and to perform end-to-end optimization of the dynamic weight generation network and fusion coefficients with the goal of verifying the loss. The output and visualization module is used to generate landslide susceptibility zoning maps, model performance evaluation indicators, ROC curves, feature importance ranking maps, and output evaluation reports.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 8.

Images