A method for assessing debris flow susceptibility in watershed units based on multi-model stacking integration

Abstract

The present invention discloses a method for assessing the debris flow susceptibility of a watershed unit based on multi-model Stacking integration, comprising the steps of: constructing a debris flow disaster data set; building a Stacking integrated learning model, including four base models, a feature stacking layer, an attention layer, a mixed feature layer, a feature splicing layer and a meta-model; training the Stacking integrated learning model; predicting the probability of debris flow disasters in the watershed unit to be identified; and assessing the debris flow susceptibility. The present invention can efficiently fuse multi-source data, and creatively splice the mixed probability vector with the original sample, so that the meta-model can simultaneously utilize the intrinsic features of the original data and the prediction information of the base model during the training process, can reduce the interference of irrelevant information and the dependence on the quality of the base model, can more comprehensively capture the key information in the data, and enhance the model's ability to interpret complex data and the prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of disaster identification, and in particular to a method for assessing the susceptibility of debris flow in a watershed unit based on multi-model Stacking integration. Background Art

[0002] At present, there are several main methods for debris flow risk assessment: statistical and empirical model methods, numerical simulation and physical modeling methods, multi-factor comprehensive analysis methods and integrated learning methods.

[0003] Statistical and empirical model methods are mainly based on historical data and expert experience. Their inherent linear assumptions make it difficult to capture the complexity and high uncertainty of the debris flow mechanism, resulting in limited accuracy and comprehensiveness of the assessment results.

[0004] Based on numerical simulation methods, mathematical models of debris flow movement are constructed or physical process simulation software is used to simulate the occurrence, development and evolution of debris flows. However, due to the complexity of the modeling process and the high demand for computing resources, these methods are time-consuming, labor-intensive and costly in actual operations, and their universality and efficiency need to be further improved.

[0005] The multi-factor comprehensive analysis method considers multiple influencing factors and evaluates these factors through a certain algorithm or model. Its main drawbacks are that it is affected by subjective factors, resulting in inaccurate weight distribution, and the simplified processing of the complex interactions between factors may reduce the accuracy of the evaluation.

[0006] The Stacking algorithm is an integrated learning method that improves the generalization ability and prediction performance of the model by combining the prediction results of multiple base learners. The basic idea is to first train multiple different base learners, then use the prediction results of these base learners as new training sets, and then train a meta-learner to synthesize these prediction results to obtain the final prediction output. However, in the traditional Stacking algorithm, the meta-model usually only receives the prediction results of the base model as input. This approach ignores the rich information in the original data, especially when there are correlated features in the original data, which may limit the prediction performance of the meta-model. Summary of the invention

[0007] The purpose of the present invention is to provide a method for assessing debris flow susceptibility in a watershed unit based on multi-model Stacking integration, which solves the problems of inaccurate weight distribution of multiple influencing factors and excessive reliance on the quality of the base model in the traditional Stacking algorithm.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows: a method for assessing the susceptibility of debris flow in a watershed unit based on multi-model Stacking integration, comprising the following steps;

[0009] S1, construct debris flow disaster data set, including S11~S13;

[0010] S11, define several influencing factors that cause debris flow disasters, select a target area, count the total number of debris flow occurrences M in the target area within a preset time period, divide the target area into several watershed units, obtain the value of each influencing factor corresponding to each watershed unit, and mark the probability of debris flow disaster b, b=m / M, m is the number of debris flow occurrences in the watershed unit within the preset time period;

[0011] S12, perform collinearity screening on the influencing factors, remove the influencing factors with high collinearity, and if there are N influencing factors left, mark the remaining influencing factors as y1~y N ;

[0012] S13, generate samples based on watershed units, and form a debris flow disaster dataset with all samples, where the sample generated by the i-th watershed unit in the target area is u i , ,y in is the influencing factor y corresponding to the i-th watershed unit n The value of b, 1≤n≤N, i is the probability of debris flow disaster in the ith watershed unit;

[0013] S2, builds a Stacking ensemble learning model, including four base models, feature stacking layer, attention layer, mixed feature layer, feature concatenation layer and meta-model;

[0014] The four base models M1-M4 are obtained by training the logistic regression model, random forest model, support vector machine model and deep neural network model with the debris flow disaster data set, respectively. i , and output u respectively i The predicted probability of debris flow disaster p i1 、p i2 、p i3 、p i4 ;

[0015] The feature stacking layer is used according to the formula , get u i The predicted probability vector A1 i ;

[0016] The attention layer is used according to the formula Calculate u iAttention score i , and normalize it to get the attention weight A2 i , where T is the transposition operation;

[0017] The hybrid feature layer is used according to the formula Get the mixed probability vector A3 i ;

[0018] The feature concatenation layer is used to transform u i With A3 i Splice to generate u i Corresponding splicing samples , ;

[0019] The meta-model is screened from four base models by a three-fold cross validation method;

[0020] S3, training the Stacking ensemble learning model;

[0021] The debris flow disaster dataset is input into the Stacking ensemble learning model. Each sample is processed through four base models, feature stacking layer, attention layer, mixed feature layer and feature splicing layer to generate a spliced ​​sample, forming a spliced ​​sample dataset.

[0022] The spliced ​​sample data set is input into the meta-model, and the probability of debris flow disaster in the spliced ​​samples is used as the expected output. The meta-model is trained until convergence to obtain a trained meta-model;

[0023] S4, predict the probability of debris flow disaster in the watershed unit to be identified;

[0024] Get the watershed unit to be identified and construct the corresponding sample u according to step S13 d , input into the Stacking ensemble learning model to generate splicing samples , and then sent to the trained meta-model output to predict the probability of debris flow disaster ;

[0025] S5, debris flow susceptibility assessment;

[0026] Divide the [0,1] interval into several subintervals, corresponding to different levels from small to large. The sub-interval level is Debris flow susceptibility level.

[0027] Preferably: in S11, the watershed unit is divided by GIS software or QGIS software, and the influencing factors include the watershed area, elevation difference, elevation value, slope, aspect, plane curvature, annual rainfall, normalized vegetation index, land use type, watershed density, distance from water system, distance from road, soil type, geological structure, and / or debris flow record of the watershed unit.

[0028] As a preferred method: the collinearity screening of the influencing factors is specifically as follows;

[0029] Sa1, preset Pearson correlation coefficient threshold r, tolerance threshold TOL, variance inflation factor threshold VIF;

[0030] Sa2, group the influencing factors in pairs, calculate the Pearson correlation coefficient of each group, and calculate the tolerance and variance inflation factor of each influencing factor;

[0031] Sa3, delete the influencing factors whose Pearson correlation coefficient is greater than r, tolerance is less than TOL, or variance inflation factor is greater than VIF.

[0032] Preferably, r=0.5, TOL=0.1, and VIF=5.

[0033] As a preference: In the attention layer, normalization is performed according to the formula We get: is the softmax function.

[0034] As a preferred method: when screening the meta-model, a three-fold cross-validation method is used to screen out the base model with the smallest error between the predicted probability of debris flow disaster and the probability of debris flow disaster as the meta-model from the four base models.

[0035] As a preferred embodiment: the sub-intervals divided in S5 are specifically:

[0036] The interval [0,1] is divided into five levels using the equidistant segmentation method: extremely low incidence area [0,0.2), low incidence area [0.2,0.4), medium incidence area [0.4,0.6), high incidence area [0.6,0.8) and extremely high incidence area [0.8,1].

[0037] The idea of ​​the present invention is: first divide the target area into watershed units; then count the various influencing factors of the watershed units, and delete some influencing factors through collinearity screening; then construct samples based on the remaining influencing factors, train the base model, and screen the meta-model; then process the samples through feature stacking layer, attention layer, mixed feature layer, and feature splicing layer to generate spliced ​​samples for meta-model training.

[0038] When generating spliced ​​samples, the present invention uses an attention layer to generate an attention weight for each sample, and the mixed feature layer uses the attention weight to mix the four base models to generate a mixed probability vector. The processed mixed probability vector reflects the importance and contribution of different base models in the prediction process. This step enables the meta-model to focus more on the base models with excellent performance during prediction and reduce the impact of low-precision base models.

[0039] Compared with the prior art, the advantages of the present invention are:

[0040] (1) High efficiency of multi-source data fusion: The research of this invention is based on the watershed unit, and deeply studies multiple influencing factors including terrain such as slope and elevation, climate such as rainfall and drought, land characteristics such as normalized vegetation index and land use type. Through collinearity analysis, this invention realizes the precise integration of these complex and heterogeneous data, effectively reduces the subjective bias common in traditional multi-factor analysis methods, and ensures the objectivity and accuracy of the analysis results. This comprehensive utilization of multi-source data provides more comprehensive and in-depth data support for the risk assessment of debris flow disasters.

[0041] (2) Significant improvement in prediction accuracy: The Stacking integration method designed by the present invention combines multiple different base models, avoiding the possible bias and limitations of a single model, and creatively splicing the mixed probability vector with the original sample, so that the model can focus more on the key features or information that have an important impact on the prediction results. This method ensures that the meta-model can simultaneously utilize the intrinsic characteristics of the original data and the prediction information of the base model during the training process, which can not only reduce the interference of irrelevant information and reduce the dependence on the quality of the base model, but also more comprehensively capture the key information in the data, enhance the model's ability to interpret complex data and prediction accuracy, and provide a new and feasible method for reducing and preventing debris flow disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a structural diagram of the Stacking ensemble learning model in the present invention;

[0043] Figure 2 It is a comparison chart of performance parameters on the five model training sets in Example 3;

[0044] Figure 3 This is a comparison chart of performance parameters on the test set of the five models in Example 3. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings.

[0046] Example 1: See Figure 1, a method for assessing debris flow susceptibility of a watershed unit based on multi-model Stacking integration, comprising the following steps;

[0047] S1, construct debris flow disaster data set, including S11~S13;

[0048] S11, define several influencing factors that cause debris flow disasters, select a target area, count the total number of debris flow occurrences M in the target area within a preset time period, divide the target area into several watershed units, obtain the value of each influencing factor corresponding to each watershed unit, and mark the probability of debris flow disaster b, b=m / M, m is the number of debris flow occurrences in the watershed unit within the preset time period;

[0049] S12, perform collinearity screening on the influencing factors, remove the influencing factors with high collinearity, and if there are N influencing factors left, mark the remaining influencing factors as y1~y N ;

[0050] S13, generate samples based on watershed units, and form a debris flow disaster dataset with all samples, where the sample generated by the i-th watershed unit in the target area is u i , ,y in is the influencing factor y corresponding to the i-th watershed unit n The value of b, 1≤n≤N, i is the probability of debris flow disaster in the ith watershed unit;

[0051] S2, builds a Stacking ensemble learning model, including four base models, feature stacking layer, attention layer, mixed feature layer, feature concatenation layer and meta-model;

[0052] The four base models M1-M4 are obtained by training the logistic regression model, random forest model, support vector machine model and deep neural network model with the debris flow disaster data set, respectively. i , and output u respectively i The predicted probability of debris flow disaster p i1 、p i2 、p i3 、p i4 ;

[0053] The feature stacking layer is used according to the formula , get u i The predicted probability vector A1 i ;

[0054] The attention layer is used according to the formula Calculate u i Attention score i, and normalize it to get the attention weight A2 i , where T is the transposition operation;

[0055] The hybrid feature layer is used according to the formula Get the mixed probability vector A3 i ;

[0056] The feature concatenation layer is used to transform u i With A3 i Splice to generate u i Corresponding splicing samples , ;

[0057] The meta-model is screened from four base models by a three-fold cross validation method;

[0058] S3, training the Stacking ensemble learning model;

[0059] The debris flow disaster dataset is input into the Stacking ensemble learning model. Each sample is processed through four base models, feature stacking layer, attention layer, mixed feature layer and feature splicing layer to generate a spliced ​​sample, forming a spliced ​​sample dataset.

[0060] The spliced ​​sample data set is input into the meta-model, and the probability of debris flow disaster in the spliced ​​samples is used as the expected output. The meta-model is trained until convergence to obtain a trained meta-model;

[0061] S4, predict the probability of debris flow disaster in the watershed unit to be identified;

[0062] Get the watershed unit to be identified and construct the corresponding sample u according to step S13 d , input into the Stacking ensemble learning model to generate splicing samples , and then sent to the trained meta-model output to predict the probability of debris flow disaster ;

[0063] S5, debris flow susceptibility assessment;

[0064] Divide the [0,1] interval into several subintervals, corresponding to different levels from small to large. The sub-interval level is Debris flow susceptibility level.

[0065] In this embodiment, in S11, the watershed unit is divided by GIS software or QGIS software, and the influencing factors include the watershed area, elevation difference, elevation value, slope, slope aspect, plane curvature, annual rainfall, normalized vegetation index, land use type, watershed density, distance from the water system, distance from the road, soil type, geological structure, and / or debris flow record of the watershed unit.

[0066] The specific collinearity screening of influencing factors is as follows;

[0067] Sa1, preset Pearson correlation coefficient threshold r, tolerance threshold TOL, variance inflation factor threshold VIF;

[0068] Sa2, group the influencing factors in pairs, calculate the Pearson correlation coefficient of each group, and calculate the tolerance and variance inflation factor of each influencing factor;

[0069] Sa3, delete the influencing factors corresponding to the Pearson correlation coefficient greater than r, the tolerance less than TOL, or the variance inflation factor greater than VIF. And r=0.5, TOL=0.1, VIF=5.

[0070] In the attention layer, normalization is performed according to the formula We get: is the softmax function.

[0071] When screening the meta-model, the three-fold cross-validation method is used to screen out the base model with the smallest error between the predicted probability of debris flow disaster and the probability of debris flow disaster as the meta-model from the four base models.

[0072] The sub-intervals in S5 are specifically divided as follows:

[0073] The interval [0,1] is divided into five levels using the equidistant segmentation method: extremely low incidence area [0,0.2), low incidence area [0.2,0.4), medium incidence area [0.4,0.6), high incidence area [0.6,0.8) and extremely high incidence area [0.8,1].

[0074] Example 2: See Figure 1 ,Based on Example 1, we provide a more detailed method for assessing the susceptibility of debris flow in a watershed unit based on multi-model Stacking integration, including the following steps;

[0075] S1, construct debris flow disaster data set, including S11~S13;

[0076] S11, defines 12 influencing factors that cause debris flow disasters, namely, watershed area, elevation difference, elevation value, slope, slope aspect, plane curvature, annual rainfall, normalized difference vegetation index, land use type, watershed density, distance from water system, and distance from road;

[0077] Select the target area. The target area of ​​the present invention contains 921 debris flow disaster points. The total number of debris flow disasters in the target area in the past year is M=300. These data come from geological environment monitoring departments, scientific research institutions, and historical documents, and are recorded in the format of longitude and latitude. In the susceptibility assessment of geological disasters, the spatial impact area of ​​landslides, avalanches, floods and other phenomena can usually be simplified to one point for analysis, because their impact range is relatively concentrated and easy to define.

[0078] The watershed units were divided using GIS software, including (1) watershed unit extraction: 0.5 km 2 The watershed units in the target area are divided and extracted as the waterlogged area threshold, and the watershed area within 1km is screened and retained. 2 For watershed units above 0.5 km 2 , 1km 2 The value of can be adjusted according to actual needs. (2) Extract the value of each influencing factor in the basin unit: When analyzing and processing data, for each influencing factor, use the mean as the influencing factor value representing the basin unit to ensure the accuracy and representativeness of the data; (3) Mark the debris flow disaster probability of each basin unit. Assuming that the number of debris flow disasters in a basin unit in the past year is 10, then its debris flow disaster probability is 10 / 300; (4) Divide the training set and the test set. 70% of the data is used as the training set for model training, and 30% of the data is used as the test set for model verification. Both the training data set and the test data set contain equal proportions of positive samples (debris flow basins) and negative samples (non-debris flow basins).

[0079] S12, collinearity screening of influencing factors, including Sa1~Sa3;

[0080] Sa1, preset Pearson correlation coefficient threshold r=0.5, tolerance threshold TOL=0.1, variance inflation factor threshold VIF=5;

[0081] Sa2, group the influencing factors in pairs, calculate the Pearson correlation coefficient of each group, and get the following Table 1;

[0082] Table 1 Pearson correlation coefficient analysis table

[0083] Influencing factors Basin area Elevation difference Elevation value slope Slope aspect Plane curvature Annual rainfall Normalized Difference Vegetation Index Land use type Watershed density Distance to water system Distance from road Basin area 1 - - - - - - - - - - - Elevation difference 0.192 1 - - - - - - - - - - Elevation value 0.248 -0.138 1 - - - - - - - - - slope -0.091 0.849 -0.287 1 - - - - - - - - Slope aspect -0.418 -0.362 -0.38 -0.033 1 - - - - - - - Plane curvature 0.206 0.441 -0.065 0.26 -0.335 1 - - - - - - Annual rainfall 0.006 -0.040 0.125 -0.055 -0.116 -0.005 1 - - - - - Normalized Difference Vegetation Index 0.11 -0.129 0.229 -0.361 -0.286 0.022 0.16 1 - - - - Land use type -0.096 -0.068 -0.327 0.004 0.042 0.115 0.029 -0.057 1 - - - Watershed density 0.14 -0.582 0.506 -0.685 -0.073 -0.037 0.030 0.161 -0.030 1 - - Distance to water system -0.16 -0.498 0.31 -0.425 0.126 -0.19 -0.077 0.033 -0.127 0.381 1 - Distance from road -0.155 -0.475 0.334 -0.412 0.09 -0.192 -0.066 0.030 -0.132 0.361 0.854 1

[0084] Based on Table 1, we deleted 3 influencing factors, namely elevation difference, land use type, and distance from the road, leaving only 9 influencing factors. We then calculated the tolerance and variance inflation coefficient of these 9 influencing factors and obtained the following Table 2;

[0085] Table 2 Tolerance and tolerance analysis table

[0086] Influencing factors Tolerance Variance Inflation Factor Basin area 0.737 1.356 Elevation value 0.593 1.686 slope 0.614 1.630 Slope aspect 0.944 1.060 Plane curvature 0.749 1.336 Annual rainfall 0.784 1.275 Normalized Difference Vegetation Index 0.850 1.177 Watershed density 0.595 1.682 Distance to water system 0.679 1.474

[0087] Sa3, as can be seen from Table 2, these influencing factors just meet the conditions of not being screened, so the remaining 9 influencing factors are directly used to construct the sample.

[0088] S13, generate samples based on watershed units, and form a debris flow disaster dataset with all samples, where the sample generated by the i-th watershed unit in the target area is u i , ;

[0089] S2, builds a Stacking ensemble learning model, including four base models, feature stacking layer, attention layer, mixed feature layer, feature concatenation layer and meta-model;

[0090] The four base models M1-M4 are obtained by training the logistic regression model, random forest model, support vector machine model and deep neural network model with the debris flow disaster data set, respectively. i , and output u respectively i The predicted probability of debris flow disaster p i1 、p i2 、p i3 、p i4 Among them, the logistic regression model, random forest model, support vector machine model, and deep neural network model are all models in the existing technology, and the parameters are optimized using random search and gradient descent method when training with debris flow disaster data.

[0091] The feature stacking layer, attention layer, mixed feature layer, feature concatenation layer, and meta-model are the same as those in Example 1. The meta-model is selected by fitting four base models using a three-fold cross-validation method, and the average performance index of the four base models in three iterations is evaluated using a validation set to screen out the optimal model.

[0092] S3, training the Stacking ensemble learning model, the same as step S3 in Example 1.

[0093] S4, predict the probability of debris flow disaster in the watershed unit to be identified; first obtain the DEM data of the area containing the watershed to be identified or import the data points with geographic coordinates into GIS software for watershed unit division. 2 The watershed area threshold was used to divide the study area into watersheds, and the watershed area within 1km was selected and retained. 2 Extract the characteristic value of the watershed unit and obtain the 9 influencing factor values ​​of the watershed unit to form the sample u to be identified d . Then u dIn the stacking ensemble learning model, the base model first obtains 4 predicted debris flow disaster probabilities, and then goes through the feature stacking layer, attention layer, mixed feature layer, feature splicing layer and generates splicing samples. ,Will Input the trained meta-model output to predict the probability of debris flow disaster .

[0094] S5, Debris flow susceptibility assessment, if =0.56, it is located in the medium-prone area [0.4, 0.6), and its debris flow susceptibility level is medium-prone. =0.04, it is located in the extremely low susceptibility zone [0, 0.2), and its debris flow susceptibility level is extremely low.

[0095] Example 3: See Figures 1 to 3 In order to illustrate the effect of the present invention, based on Example 2, we select a certain area in Sichuan Province as the target area to construct a debris flow disaster dataset. The target area is divided into 489 watershed units, so it contains 489 corresponding samples. The debris flow disaster dataset is divided into a training set and a test set, and four base models are used to compare multiple performance parameters with the model of the present invention in the training set and the test set, as follows Figure 2 , Figure 3 . Figure 2 , Figure 3 In the figure, the five models are from left to right:

[0096] SVM: Support Vector Machine, support vector machine model;

[0097] LR: Logistic Regression, logistic regression model;

[0098] RF: RandomForest, random forest model;

[0099] The present invention: Stacking ensemble learning model;

[0100] DNN: Deep Neural Network, deep neural network.

[0101] In the figure, precision is precision, recall is recall, F1 score is f1-score, and accuracy is accuracy.

[0102] In addition, the ROC curve and the area under the curve AUC are also important indicators for evaluating model performance. The ROC curve is drawn with the false positive rate as the X-axis and the true positive rate (TPR) as the Y-axis, and the AUC value is the area under the ROC curve. The value of AUC ranges from 0.5 to 1. The closer the value is to 1, the better the performance of the model. We plotted the ROC curves of the test sets of 5 models, and found that the AUC values ​​of LR, RF, SVM, the model of the present invention, and DNN were 0.940, 0.948, 0.829, 0.987, and 0.708, respectively. The AUC values ​​of all 5 models are greater than 0.7, indicating that the prediction accuracy is high. Among them, the model of the present invention showed the highest prediction accuracy, followed by the RF model.

[0103] Example 4: We use the five models in Example 3 to predict the probability of debris flow disasters, and then evaluate the susceptibility of debris flow according to the five levels divided in step S5 of the present invention: extremely low susceptibility area [0, 0.2), low susceptibility area [0.2, 0.4), medium susceptibility area [0.4, 0.6), high susceptibility area [0.6, 0.8), and extremely high susceptibility area [0.8, 1]. The number of watershed units at different levels in different models is shown in Table 3, and the proportion of watershed units at different levels in different models is shown in Table 4; Table 3 Statistical table of the number of watershed units at different levels in different models

[0104] ,

[0105] Table 4 Statistics of the proportion of watershed units at different levels in different models

[0106] ,

[0107] According to Tables 3 and 4, the LR model classifies 58.8% of the watershed units as extremely low susceptibility, and the RF model classifies 65% of the watershed units as extremely low susceptibility, which is the highest proportion among the five models. The SVM model has the lowest proportion at this level, only 9.6%, and the present invention classifies 60.5% of the watershed units as extremely low susceptibility. This shows that the RF model and the present invention are more conservative in identifying extremely low susceptibility areas, while the SVM model tends to classify these areas as higher susceptibility levels. The proportions of the LR model, RF model, SVM model, DNN model and the present invention in the high risk level are 1.2%, 1.6%, 21.8%, 9.6% and 0.8%, respectively. The proportion of the DNN model at this level is also relatively high, while the proportion of the present invention is the lowest. The LR model and the present invention classified about one-third of the basin units (35.8% and 37.2%, respectively) as extremely high susceptibility, the RF model accounted for 28.1% of the extremely high susceptibility level, slightly lower than the LR model and the present invention, and the SVM model had the lowest proportion of 13.5% at this level, indicating that the LR model and the present invention were relatively consistent in identifying extremely high susceptibility areas, while the RF model and the SVM model were relatively conservative. In summary, the target area has a high susceptibility to debris flows.

[0108] 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 method for assessing the susceptibility of debris flow in a watershed unit based on multi-model stacking integration, characterized by: The steps include: S1, construct debris flow disaster data set, including S11~S13; S11, define several influencing factors that cause debris flow disasters, select a target area, count the total number of debris flow occurrences M in the target area within a preset time period, divide the target area into several watershed units, obtain the value of each influencing factor corresponding to each watershed unit, and mark the probability of debris flow disaster b, b=m / M, m is the number of debris flow occurrences in the watershed unit within the preset time period; S12, perform collinearity screening on the influencing factors, remove the influencing factors with high collinearity, and if there are N influencing factors left, mark the remaining influencing factors as y1~y N ; S13, generate samples based on watershed units, and form a debris flow disaster dataset with all samples, where the sample generated by the i-th watershed unit in the target area is u i , ,y in is the influencing factor y corresponding to the i-th watershed unit n The value of b, 1≤n≤N, i is the probability of debris flow disaster in the ith watershed unit; S2, builds a Stacking ensemble learning model, including four base models, feature stacking layer, attention layer, mixed feature layer, feature concatenation layer and meta-model; The four base models M1-M4 are obtained by training the logistic regression model, random forest model, support vector machine model and deep neural network model with the debris flow disaster data set, respectively. i , and output u respectively i The predicted probability of debris flow disaster p i1 、p i2 、p i3 、p i4 ; The feature stacking layer is used according to the formula , get u i The predicted probability vector A1 i ; The attention layer is used according to the formula Calculate u i Attention score i , and normalize it to get the attention weight A2 i , where T is the transposition operation; The hybrid feature layer is used according to the formula Get the mixed probability vector A3 i ; The feature concatenation layer is used to transform u i With A3 i Splice to generate u i Corresponding splicing samples , ; The meta-model is screened from four base models by a three-fold cross validation method; S3, training the Stacking ensemble learning model; The debris flow disaster dataset is input into the Stacking ensemble learning model. Each sample is processed through four base models, feature stacking layer, attention layer, mixed feature layer and feature splicing layer to generate a spliced ​​sample, forming a spliced ​​sample dataset. The spliced ​​sample data set is input into the meta-model, and the probability of debris flow disaster in the spliced ​​samples is used as the expected output. The meta-model is trained until convergence to obtain a trained meta-model; S4, predict the probability of debris flow disaster in the watershed unit to be identified; Get the watershed unit to be identified and construct the corresponding sample u according to step S13 d , input into the Stacking ensemble learning model to generate splicing samples , and then sent to the trained meta-model output to predict the probability of debris flow disaster ; S5, debris flow susceptibility assessment; Divide the [0,1] interval into several subintervals, corresponding to different levels from small to large. The sub-interval level is Debris flow susceptibility level.

2. The method for assessing debris flow susceptibility of a watershed unit based on multi-model stacking integration according to claim 1 is characterized by: In S11, the watershed unit is divided by GIS software or QGIS software, and the influencing factors include the watershed area, elevation difference, elevation value, slope, slope aspect, plane curvature, annual rainfall, normalized difference vegetation index, land use type, watershed density, distance from water system, distance from road, soil type, geological structure, and / or debris flow record of the watershed unit.

3. The method for assessing debris flow susceptibility of a watershed unit based on multi-model stacking integration according to claim 1 is characterized by: The specific collinearity screening of influencing factors is as follows; Sa1, preset Pearson correlation coefficient threshold r, tolerance threshold TOL, variance inflation factor threshold VIF; Sa2, group the influencing factors in pairs, calculate the Pearson correlation coefficient of each group, and calculate the tolerance and variance inflation factor of each influencing factor; Sa3, delete the influencing factors whose Pearson correlation coefficient is greater than r, tolerance is less than TOL, or variance inflation factor is greater than VIF.

4. The method for assessing debris flow susceptibility in a watershed unit based on multi-model stacking integration according to claim 3 is characterized by: The r=0.5, TOL=0.1, VIF=5.

5. The method for assessing debris flow susceptibility in a watershed unit based on multi-model stacking integration according to claim 1, characterized in that: In the attention layer, normalization is performed according to the formula We get: is the softmax function.

6. The method for assessing debris flow susceptibility of a watershed unit based on multi-model stacking integration according to claim 1, characterized in that: When screening the meta-model, the three-fold cross-validation method is used to screen out the base model with the smallest error between the predicted probability of debris flow disaster and the probability of debris flow disaster as the meta-model from the four base models.

7. The method for assessing debris flow susceptibility of a watershed unit based on multi-model stacking integration according to claim 1, characterized in that: The sub-intervals in S5 are specifically divided as follows: The interval [0,1] is divided into five levels using the equidistant segmentation method: extremely low incidence area [0,0.2), low incidence area [0.2,0.4), medium incidence area [0.4,0.6), high incidence area [0.6,0.8) and extremely high incidence area [0.8,1].

Images