Landslide hazard result evaluation method based on improved Stacking fusion model

Abstract

The invention discloses a landslide hazard result evaluation method based on an improved Stacking fusion model, and the method comprises: S1, constructing a landslide feature sample data set, extracting a landslide influencing factor set and storing it in association with a landslide hazard result; S2, constructing a landslide hazard evaluation model, the landslide hazard evaluation model comprises a primary learner layer and a secondary learner layer, the secondary learner layer comprises a support vector machine SVM, the support vector machine SVM uses the output and weight of the primary learner layer as input and outputs a landslide hazard classification result using a kernel function; S3, collecting landslide data in a study area and inputting it into the landslide hazard evaluation model, and then outputting a landslide hazard result. The invention has the advantages of high model prediction accuracy and high comprehensiveness through the training of multiple models of the primary learner layer of the landslide hazard evaluation model and the model training of the secondary learner layer, and is of great significance for the prevention of geological disaster hazards.

Description

Technical Field

[0001] The invention relates to the field of geological disasters, and in particular to a landslide hazard result evaluation method based on an improved Stacking fusion model. Background Art

[0002] Landslide disasters are the most common type of geological disasters. According to statistics, 7,840 geological disasters occurred in China in 2020, of which landslides accounted for 4,810, exceeding 60% of the total number of geological disasters. Landslides are likely to occur all over the country, causing serious impacts on human life and production methods. Therefore, landslide disasters have always been one of the current research hotspots. Research on landslide disasters can clearly understand the distribution, development and occurrence laws of landslide disasters, evaluate the danger of landslide disasters, and enable people to have a clear understanding of the threats that landslides may bring in order to prevent them in advance, which can reduce the loss of personnel and property to a certain extent. Therefore, it is of great significance to carry out landslide hazard assessment. Summary of the invention

[0003] The purpose of the present invention is to overcome the technical problems pointed out by the background technology, provide a landslide hazard result evaluation method based on an improved Stacking fusion model, construct an improved autoencoder model to preprocess landslide feature sample data, extract a set of landslide influencing factors and store them in association with landslide hazard results, obtain the weight relationship of the model through the training of multiple models of the primary learner layer of the landslide hazard evaluation model and the model training of the secondary learner layer, and also perform weight verification of the influencing factors based on multiple models. The method has the advantages of high model prediction accuracy and high comprehensiveness, can help geological disaster departments to formulate corresponding preventive measures, and is of great significance for the prevention of geological disaster hazards.

[0004] The purpose of the present invention is achieved through the following technical solutions:

[0005] A landslide hazard result evaluation method based on an improved Stacking fusion model includes:

[0006] S1. Construct a landslide feature sample data set including landslide feature sample data and corresponding landslide hazard results, construct an improved autoencoder model to preprocess the landslide feature sample data, extract a landslide influencing factor set and store it in association with the landslide hazard results. The landslide influencing factor set includes natural environment factors, topographic factors, geological factors, human activity factors and other classified influencing factors.

[0007] S2. Construct a landslide hazard assessment model based on the improved Stacking fusion model structure. The landslide hazard assessment model includes a primary learner layer and a secondary learner layer. The primary learner layer includes a random forest model, an AdaBoost model and a gradient boosting algorithm model. The random forest model combines multiple decision trees for classification and regression processing, and uses a landslide influencing factor set and landslide hazard results to train the random forest model; the AdaBoost model uses a series of weak classifiers to form a strong classifier. Each round of weak classifier processing includes weighted processing of sample data. After several rounds of processing, the classification results of the previous round adjust the weight of sample data of the next round, and the landslide influencing factor set and landslide hazard results are used to train the AdaBoost model; the gradient boosting algorithm model iteratively trains multiple weak learners, and each iteration adjusts the weight of the weak learner and gradually reduces the loss of the model. The gradient boosting algorithm model uses a combination of grid search and cross-validation to search for the best parameter combination, and uses a landslide influencing factor set and landslide hazard results to train the gradient boosting algorithm model;

[0008] The landslide hazard assessment model sets the confidence and weight of the random forest model, AdaBoost model and gradient boosting algorithm model in the primary learner layer based on the improved Stacking algorithm; the secondary learner layer includes a support vector machine SVM, which uses the output and weight of the primary learner layer as input and uses the kernel function to output the landslide hazard classification result;

[0009] S3. Determine the study area, collect landslide data in the study area and input it into the landslide hazard assessment model, and then output the landslide hazard results.

[0010] In order to better realize the present invention, the preprocessing method of the improved autoencoder model includes: directly extracting the numerical value of the numerical content of the landslide feature sample data and using the surrounding mean to perform missing interpolation processing; converting the category and attribute data of the landslide feature sample data into corresponding feature values ​​according to the rules and using the surrounding mean to perform missing feature value interpolation processing.

[0011] A further technical solution is: the training processing method of the improved autoencoder model is as follows: the improved autoencoder model loads a landslide feature sample data set, uses a polynomial attenuation pruning strategy for the improved autoencoder model, sets different pruning ratios to perform model parameter pruning processing; the improved autoencoder model after the model parameter pruning processing is loaded with a landslide feature sample data set for multiple retraining and fine-tuning.

[0012] Preferably, parameters with Gaussian noise are defined, and the landslide characteristic sample data are added with Gaussian noise and then input into the improved autoencoder model for denoising training.

[0013] Preferably, the decision tree parameter tuning within the random forest model includes: using the N_estimators parameter to limit the maximum number of weak learners, using the Max_depth parameter to limit the maximum depth of the decision tree, and using the Min_samples_split parameter to limit the conditions for further division of the decision tree; using the Min_samples_leaf parameter to limit the minimum number of samples of the leaf nodes of the decision tree. When the number of leaf nodes is less than the number of samples, the leaf node is pruned together with its sibling nodes.

[0014] Preferably, the training process of the AdaBoost model is as follows:

[0015] The weights of each landslide feature sample data set are initialized to an equal value, and the AdaBoost model is used for iterative training. The iterative training includes: A11, using the current weight to train a weak classifier to classify a single feature; applying the weak classifier to the landslide feature sample data set, and calculating the classification error rate, and calculating the weight of the current weak classifier according to the classification error rate; increasing the weight of the sample misclassified by the weak classifier, and reducing the weight of the sample correctly classified; then normalizing the sample weights so that the sum of the sample weights is 1; A12, training a new weak classifier again according to the sample weight updated in step A11, and obtaining the weight of the new weak classifier and the sample weight according to the method of step A11; A13, obtaining several weak classifiers according to the method of step A12, performing weighted processing on each weak classifier based on the weak classifier weight, and combining the results of all weak classifiers to obtain the trained AdaBoost model; the AdaBoost model uses a combination of grid search and cross-validation to search for the best parameter combination.

[0016] Preferably, the training processing method of the gradient boosting algorithm model is as follows: B11, constructing an initial predictor of the model, the initial predictor expression is as follows: Where γ is a constant term; y i is an element in the sample i label set, N is the number of samples; B12, the iterative training method of the gradient boosting algorithm model is as follows:

[0017] The residual r of sample i mi The expression of is as follows: mi =f m-1 (x), f m-1 (x) is the m-1th weak learner, and the m-1th weak learner is recorded as the m-1th decision tree, m = 1, 2, ..., M; M represents the total number of weak learners; the leaf node region R of the jth node of the mth decision tree mj , j = 1, 2, ..., J, J is the number of leaf nodes, calculate R mj The corresponding output value fm-1 (x i ) is the output value of sample i of the m-1th decision tree, and L is the loss function of the model; then the regression tree is updated, and the expression is as follows: x∈R mj , I(x) is the indicator function; the classification expression of the gradient boosting algorithm model is obtained: x∈R mj .

[0018] Preferably, the kernel function of the support vector machine SVM adopts the Gaussian kernel function K(x, y), which is expressed as follows: K(x, y)=exp(-Y*||xy|| 2 ), where parameter Y controls the width of the Gaussian kernel function, x and y represent input samples respectively; ||xy|| represents the norm operation of the input samples; the parameters of the support vector machine SVM also include a penalty factor C parameter greater than 0; the support vector machine SVM uses a 5-fold cross-validation method and grid search to obtain the best parameter combination.

[0019] Preferably, the confidence of the random forest model, AdaBoost model and gradient boosting algorithm model is obtained by using a softmax function and undergoing standardization; the weight calculation formulas of the random forest model, AdaBoost model and gradient boosting algorithm model are as follows:

[0020] Where R a is the prediction result of model a, P a is the confidence corresponding to model a, and A represents the total number of input test samples i.

[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0022] (1) The present invention constructs a landslide feature sample data set, constructs an improved autoencoder model to preprocess the landslide feature sample data, extracts a landslide influencing factor set and stores it in association with the landslide hazard result, obtains the weight relationship of the model through the training of multiple models of the primary learner layer of the landslide hazard assessment model and the model training of the secondary learner layer, and also verifies the weight of the influencing factor based on multiple models. The model has the advantages of high prediction accuracy and high comprehensiveness, can help geological disaster departments formulate corresponding preventive measures, and is of great significance for the prevention of geological disaster hazards.

[0023] (2) The present invention uses an improved autoencoder model to preprocess the landslide feature sample data set, fully considering the correlation and redundancy between the feature factors and reducing the information redundancy of the feature factors, making the feature factors in the landslide hazard assessment model more representative and improving the robustness and computational efficiency of the model.

[0024] (3) The present invention determines the weights of different models in the primary learner by obtaining confidence, so that the model with better training results can play a better data advantage in the secondary learner. Adding the original training set data to the training set of the secondary learner can effectively avoid the loss of part of the information in the original training set, making the training results more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the landslide hazard result evaluation method of the present invention;

[0026] Figure 2 It is a schematic diagram of the principle of the improved Stacking fusion model in the embodiment;

[0027] Figure 3 This is a schematic diagram of the principle of the landslide hazard result evaluation method;

[0028] Figure 4 The algorithm diagram of the random forest model in the embodiment;

[0029] Figure 5 A schematic diagram of the principle of the AdaBoost model in the embodiment;

[0030] Figure 6 Schematic diagram of the principle of the support vector machine SVM in the embodiment. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with embodiments:

[0032] Example

[0033] like Figure 1 to Figure 3 As shown, a landslide hazard result evaluation method based on an improved Stacking fusion model includes:

[0034] S1. Construct a landslide feature sample data set including landslide feature sample data and corresponding landslide hazard results, construct an improved autoencoder model to preprocess the landslide feature sample data, the landslide feature sample data in the landslide feature sample data set includes various factors in the landslide influencing factor set, extract the landslide influencing factor set and store it in association with the landslide hazard result, and the landslide hazard result can be quantitatively characterized by a value of 0 to 100 (for a specific and detailed evaluation of the landslide feature sample data, the landslide hazard is quantitatively described according to a value of 0 to 100. If a landslide almost never occurs, the probability of the landslide hazard result is 0, and if a landslide occurs 100%, the probability of the landslide hazard result is 100); of course, the landslide hazard result can also be classified and expressed by hazard level, and the landslide hazard result is low hazard, medium hazard, and high hazard levels. The landslide influencing factor set includes natural environment factors, topographic factors, geological factors, human activity factors, and other classified influencing factors. The natural environment factors include earthquake, rainfall (annual average rainfall can be used), hydrology (groundwater and the relative position of the landslide site to the river can be used), soil, and vegetation (NDVI can be used as a substitute); topographic factors include DEM, slope, aspect, and curvature; geological factors include geological structure and stratum lithology; human activity factors include buildings (which can be expressed by building type, distance from the landslide, and area), roads (which can be expressed by road grade and location), and cultivated land; other classified influencing factors (referred to as other influencing factors) include landslide shape, landslide rate (which can be expressed by maximum rate and average rate), and crack conditions.

[0035] In some embodiments, the preprocessing method of the improved autoencoder model includes: directly extracting the numerical value for the numerical content of the landslide feature sample data and using the surrounding mean to perform missing interpolation processing; converting the data of the category and attribute of the landslide feature sample data into the corresponding eigenvalue according to the rules and using the surrounding mean to perform interpolation processing for missing eigenvalues. In the collected data, some data are continuous data, and some are discrete data; for continuous data, the numerical value can be directly extracted, while discrete data is usually data representing categories and attributes, and it is necessary to convert it into a numerical value to replace its eigenvalue (different types of data are dimensionless, all features are converted to the same dimension, and the data is converted to [0, 1]). In the acquired data, not every landslide can extract all the required data. Usually, there will be some landslides with missing data. At this time, it is necessary to supplement the actual data of the landslide here through spatial interpolation processing, and the mean of the adjacent landslides or the mean of the remaining landslides is often used for interpolation processing. For the acquired landslide feature sample data, the improved autoencoder model should select a part of the data as a training sample and another part of the data as a test sample. Through the above data preprocessing, data with unified standards are obtained, and the landslide influencing factor feature set NTGHO (Natural, Topography, Geology, Human, Other) is constructed according to different influencing factor categories.

[0036] S2, Fusion model structure based on improved Stacking (such as Figure 2 The landslide hazard assessment model is constructed as shown in FIG. 1 . The landslide hazard assessment model includes a primary learner layer and a secondary learner layer. The primary learner layer includes a random forest model, an AdaBoost model, and a gradient boosting algorithm model. The random forest model combines multiple decision trees for classification and regression processing. The landslide influencing factor set and the landslide hazard results are used to train the random forest model. The AdaBoost model uses a series of weak classifiers to form a strong classifier, such as Figure 5 As shown in the figure, each round of weak classifier processing includes weighted processing of sample data. Then, after several rounds of processing, the classification results of the previous round adjust the weight of sample data of the next round, and the landslide influence factor set and landslide hazard results are used to train the AdaBoost model; the gradient boosting algorithm model iteratively trains multiple weak learners, and each iteration adjusts the weight of the weak learner and gradually reduces the loss of the model. The gradient boosting algorithm model uses a combination of grid search and cross-validation to search for the best parameter combination, and uses the landslide influence factor set and landslide hazard results to train the gradient boosting algorithm model.

[0037] Preferably, the fusion model structure of Stacking includes a primary learner layer (first layer) and a secondary learner layer (second layer). The primary learner layer includes multiple primary learners, and the second layer usually uses a secondary learner. The training data is input into the model of the first layer, and the prediction results output by the model of the first layer are used as the training data of the model of the second layer. Finally, after passing through the model of the second layer, the final prediction results are obtained. The following is the pseudo code of the Stacking algorithm:

[0038]

[0039]

[0040] The main steps of the Stacking algorithm are as follows: First, prepare the original training data set, which includes feature x and feature y. Then, build the first-layer basic model and the second-layer meta-model, and input the training data set into the primary learner for training and prediction. Next, input the new training set for training and prediction to obtain the final model.

[0041] The landslide hazard assessment model sets the confidence and weight of the random forest model, AdaBoost model and gradient boosting algorithm model in the primary learner layer based on the improved Stacking algorithm. In some embodiments, the confidence of the random forest model, AdaBoost model and gradient boosting algorithm model is obtained by using a softmax function (the confidence or probability distribution can be obtained by using a softmax function) and normalization (using a minimum-maximum normalization process); the weight calculation formula of the random forest model, AdaBoost model and gradient boosting algorithm model is as follows:

[0042] Where R a is the prediction result of model a, P a is the confidence corresponding to model a, and A represents the total number of input test samples i (corresponding to the landslide influencing factors in the landslide influencing factor set). The landslide hazard assessment model of the present invention sets weights for the random forest model, the AdaBoost model, and the gradient boosting algorithm model, and can also sort out the weights of the landslide influencing factors in the landslide influencing factor set in each model.

[0043] The secondary learner layer includes support vector machine SVM (the main parameters include penalty factor C and kernel function), such as Figure 6As shown, the support vector machine SVM uses the output and weight of the primary learner layer as input and uses the kernel function to output the landslide hazard classification result. The kernel function can use a linear kernel function, a Sigmoid kernel function, a Gaussian kernel function or a polynomial kernel function. In this embodiment, the kernel function of the support vector machine SVM uses a Gaussian kernel function K(x, y), which is expressed as follows: K(x, y)=exp(-Y*||xy|| 2 ); wherein parameter Y controls the width of the Gaussian kernel function, x and y represent input samples respectively; ||xy|| represents the norm operation of the input sample; the parameters of the support vector machine SVM also include a penalty factor C parameter greater than 0; the support vector machine SVM of the present invention adopts a 5-fold cross-validation method and a grid search to obtain the best parameter combination. Preferably, the original training set data is added to the training set of the support vector machine SVM of the secondary learner layer as a data supplement for the original data and factors, which can effectively avoid the loss of part of the information in the original training set and make the training results more accurate.

[0044] S3. Determine the study area, collect landslide data in the study area and input it into the landslide hazard assessment model, and then output the landslide hazard results.

[0045] In some embodiments, the training processing method of the improved autoencoder model (which can reduce the dimension of the features in the landslide feature sample data set) is as follows: the improved autoencoder model loads the landslide feature sample data set, and the improved autoencoder model uses a polynomial decay pruning strategy and sets different pruning ratios to perform model parameter pruning processing; the improved autoencoder model after the model parameter pruning processing loads the landslide feature sample data set for multiple retraining and fine-tuning until the compression requirements are met. In the training process of the improved autoencoder model, preferably, a parameter with Gaussian noise is defined, and the landslide feature sample data is added with Gaussian noise (the method is: generate a Gaussian distribution random number with a mean of 0 and a standard deviation of 1, and add the generated random number to the original data to obtain the data after adding Gaussian noise) and then input into the improved autoencoder model for denoising training, so that the improved autoencoder model can learn the robustness of the noise data, thereby improving its ability to reconstruct the original data. The improved autoencoder model only adds Gaussian noise learning during model training, and after the model training is completed, there is no need to add Gaussian noise.

[0046] In some embodiments, Figure 4As shown in the figure, the random forest model combines multiple decision trees for classification and regression tasks. The decision tree constructs a tree structure by recursively dividing the features, so that the samples on each node belong to the same category as much as possible or minimize the regression error. The basis for dividing each node is to select a feature and a threshold of the feature to divide the sample into two sub-nodes. The random forest model combines multiple decision trees together through ensemble learning to form a more powerful and stable prediction model. By integrating the prediction results of multiple decision trees, the random forest can improve the robustness and generalization ability of the model and reduce the risk of overfitting. The Bagging framework parameters of the random forest model mainly include N_estimators and Oob_score. The value of the Oob_score parameter is True or False, and the default value is False. The Oob_score parameter indicates whether to use out-of-bag samples to evaluate the quality of the model. The decision tree parameter tuning within the random forest model includes: using the N_estimators parameter to limit the maximum number of weak learners, using the Max_depth parameter to limit the maximum depth of the decision tree, and using the Min_samples_split parameter to limit the conditions for continued division of the decision tree (when the number of samples at a node is less than the Min_samples_split value, the optimal feature will no longer be selected for division); using the Min_samples_leaf parameter to limit the minimum number of samples of the leaf nodes of the decision tree (enter the integer of the minimum number of samples). When the number of leaf nodes is less than the number of samples, the leaf node is pruned together with its sibling nodes.

[0047] In some embodiments, the training process of the AdaBoost model is as follows:

[0048] Initialize the weights of each landslide feature sample data set to equal values, and use the AdaBoost model for iterative training. The iterative training includes: A11, use the current weight to train a weak classifier (equivalent to a decision stump) to classify a single feature; apply the weak classifier to the landslide feature sample data set, and calculate the classification error rate, and calculate the weight of the current weak classifier according to the classification error rate (the weight is negatively correlated with the classification error rate, the smaller the classification error rate, the larger the weight of the classifier); increase the weight of the samples misclassified by the weak classifier, and reduce the weight of the samples correctly classified; then normalize the samples A12, train a new weak classifier again according to the sample weight updated in step A11, and obtain the weight and sample weight of the new weak classifier according to the method in step A11; A13, obtain several weak classifiers according to the method in step A12, perform weighted processing on each weak classifier based on the weak classifier weight (using weighted voting method, and determining the importance of voting according to its classifier weight), and obtain the trained AdaBoost model by combining the results of all weak classifiers; the AdaBoost model uses a combination of grid search and cross-validation to search for the best parameter combination.

[0049] In some embodiments, it is assumed that T = {(x1, y1), (x2, y2), ..., (x N ,y N )} is the training data set input to the gradient boosting algorithm model, x=(x1,x2,...,x N ) is the sample attribute set (corresponding to each landslide influencing factor in the landslide characteristic sample data), y = (y1, y2, ..., y N ) is the sample label set (corresponding to the landslide hazard result), N is the number of samples; f(x) is the objective function, and L is the loss function. The training processing method of the gradient boosting algorithm model is as follows: B11. Construct the initial predictor of the model. The initial predictor expression is as follows: Where γ is a constant term; y i is an element in the sample i label set, N is the number of samples; B12, the iterative training method of the gradient boosting algorithm model is as follows:

[0050] The residual r of sample i mi The expression of is as follows: mi =f m-1 (x), f m-1 (x) is the m-1th weak learner, and the m-1th weak learner is recorded as the m-1th decision tree, m = 1, 2, ..., M; M represents the total number of weak learners (i.e., decision trees); the leaf node region R of the jth node of the mth decision tree mj , j = 1, 2, ..., J, J is the number of leaf nodes, calculate Rmj The corresponding output value f m-1 (x i ) is the output value of the m-1th decision tree instance i; then the regression tree is updated, and the expression is as follows: x∈R mj , I(x) is the indicator function; the classification expression of the gradient boosting algorithm model is obtained: x∈R mj For some variable parameters, the gradient boosting algorithm model uses a combination of grid search and cross-validation to search for the best parameter combination. When the classification accuracy of the model training sample is the highest, it is the optimal value, and the final gradient boosting algorithm model is obtained at this time.

[0051] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A landslide hazard result evaluation method based on an improved Stacking fusion model, characterized by: The methods include: S1. Construct a landslide feature sample data set including landslide feature sample data and corresponding landslide hazard results, construct an improved autoencoder model to preprocess the landslide feature sample data, and the training processing method of the improved autoencoder model is as follows: load the landslide feature sample data set into the improved autoencoder model, use a polynomial decay pruning strategy for the improved autoencoder model, set different pruning ratios to perform model parameter pruning processing; load the landslide feature sample data set into the improved autoencoder model after the model parameter pruning processing, and retrain and fine-tune it multiple times; extract a landslide influencing factor set and store it in association with the landslide hazard results, wherein the landslide influencing factor set includes natural environment factors, topographic factors, geological factors, human activity factors and other classified influencing factors; S2. A landslide hazard assessment model is constructed based on the improved Stacking fusion model structure. The landslide hazard assessment model includes a primary learner layer and a secondary learner layer. The primary learner layer includes a random forest model, an AdaBoost model and a gradient boosting algorithm model. The random forest model combines multiple decision trees for classification and regression processing. The landslide influencing factor set and the landslide hazard results are used to train the random forest model. The AdaBoost model uses a series of weak classifiers to form a strong classifier. Each round of weak classifier processing includes weighted processing of sample data. After several rounds of processing, the classification results of the previous round adjust the weight of the sample data of the next round. The landslide influencing factor set and the landslide hazard results are used to train the AdaBoost model. The training processing method of the AdaBoost model is as follows: Initialize the weights of each landslide feature sample data set to an equal value, and use the AdaBoost model for iterative training, the iterative training includes: A11, use the current weight to train a weak classifier for single feature classification; apply the weak classifier to the landslide feature sample data set, and calculate the classification error rate, and calculate the weight of the current weak classifier according to the classification error rate; increase the weight of the sample misclassified by the weak classifier, and reduce the weight of the sample correctly classified; then normalize the sample weights so that the sum of the sample weights is 1; A12, train a new weak classifier again according to the sample weight updated in step A11, and obtain the weight and sample weight of the new weak classifier according to the method of step A11; A13, obtain several weak classifiers according to the method of step A12, perform weighted processing on each weak classifier based on the weak classifier weight, and obtain the trained AdaBoost model by combining the results of all weak classifiers; the AdaBoost model uses a combination of grid search and cross-validation to search for the best parameter combination; The gradient boosting algorithm model iteratively trains multiple weak learners. Each iteration adjusts the weights of the weak learners and gradually reduces the loss of the model. The gradient boosting algorithm model uses a combination of grid search and cross-validation to search for the best parameter combination. The landslide influencing factor set and landslide hazard results are used to train the gradient boosting algorithm model. The training processing method of the gradient boosting algorithm model is as follows: B11. Construct the initial predictor of the model. The initial predictor expression is as follows: Where Υ is a constant term; y i is an element in the sample i label set, N is the number of samples; B12, the iterative training method of the gradient boosting algorithm model is as follows: The residual r of sample i mi The expression of is as follows: mi =f m-1 (x), f m-1 (x) is the m-1th weak learner, and the m-1th weak learner is recorded as the m-1th decision tree, m = 1, 2, ..., M; M represents the total number of weak learners; the leaf node region R of the jth node of the mth decision tree mj ,j=1,2,...,J, J is the number of leaf nodes, calculate R mj The corresponding output value c mj , f m-1 (x i ) is the output value of sample i of the m-1th decision tree, and L is the loss function of the model; then the regression tree is updated, and the expression is as follows: I(x) is the indicator function; the classification expression of the gradient boosting algorithm model is obtained: The landslide hazard assessment model sets the confidence and weight of the random forest model, AdaBoost model and gradient boosting algorithm model in the primary learner layer based on the improved Stacking algorithm; the confidence of the random forest model, AdaBoost model and gradient boosting algorithm model is obtained by using the softmax function and normalization; the weight calculation formula of the random forest model, AdaBoost model and gradient boosting algorithm model is as follows: Where R a is the prediction result of model a, P a is the confidence corresponding to model a, A represents the total number of input test samples i; the secondary learner layer includes support vector machine SVM, which uses the output and weight of the primary learner layer as input and outputs the landslide hazard classification result using the kernel function; the kernel function of the support vector machine SVM adopts the Gaussian kernel function K(x,y), which is expressed as follows: K(x,y)=exp(-Y*||xy 2 ); wherein the parameter Y controls the width of the Gaussian kernel function, x and y represent input samples respectively; ||xy|| represents the norm operation of the input sample; the parameters of the support vector machine SVM also include a penalty factor C parameter greater than 0; the support vector machine SVM uses a 5-fold cross validation method and a grid search to obtain the best parameter combination; S3. Determine the study area, collect landslide data in the study area and input it into the landslide hazard assessment model, and then output the landslide hazard results.

2. The landslide hazard result evaluation method based on the improved Stacking fusion model according to claim 1 is characterized in that: The preprocessing method of the improved autoencoder model includes: directly extracting the numerical content of the landslide characteristic sample data and using the surrounding mean to perform missing interpolation processing; converting the category and attribute data of the landslide characteristic sample data into corresponding eigenvalues ​​according to rules and using the surrounding mean to perform missing eigenvalue interpolation processing.

3. The landslide hazard result evaluation method based on the improved Stacking fusion model according to claim 1 is characterized in that: The parameters of Gaussian noise are defined, and the landslide characteristic sample data are added with Gaussian noise and then input into the improved autoencoder model for denoising training.

4. The landslide hazard result evaluation method based on the improved Stacking fusion model according to claim 1 is characterized in that: The decision tree parameter tuning within the random forest model includes: using the N_estimators parameter to limit the maximum number of weak learners, using the Max_depth parameter to limit the maximum depth of the decision tree, using the Min_samples_split parameter to limit the conditions for further division of the decision tree; using the Min_samples_leaf parameter to limit the minimum number of samples of the leaf nodes of the decision tree. When the number of leaf nodes is less than the number of samples, the leaf node and its sibling nodes are pruned together.

Images