Debris flow prediction method based on hybrid machine learning

Abstract

The invention provides a debris flow prediction method based on hybrid machine learning, and relates to the technical field of debris flow early warning, and the method comprises the steps: constructing an index system, determining indexes for debris flow prediction, and standardizing the index values in the index system, and the standardization comprises the normalization and outlier deletion, and constructing a support vector classification model. The method comprises the following steps: optimizing hyper-parameters by using an AOVA algorithm, obtaining a hybrid machine learning model, constructing a training data set, training the hybrid machine learning model, obtaining a debris flow prediction model, and predicting the debris flow occurrence probability by using the debris flow prediction model. The method is suitable for basin debris flow prediction.

Description

Technical Field

[0001] The present invention relates to the technical field of debris flow early warning, and in particular to a debris flow prediction method based on hybrid machine learning. Background Art

[0002] Debris flow is a rapid and turbulent flow of water-bearing clastic sediments along steep channels and is one of the most dangerous mountain hazards. There are many existing methods for predicting the susceptibility of debris flows, including expert methods, data-driven statistical methods, and deterministic methods. Expert methods are used to assess the possibility of debris flow at an early stage. The relationship between debris flow occurrence and causal factors is directly established through expert experience and background knowledge. This method may be controversial because it is difficult to objectively quantify or evaluate the results. Data-driven statistical methods include principal component analysis, logistic regression, and evidence weighting methods. They predict the susceptibility of debris flow by mathematically modeling the relationship between debris flow occurrence and disaster causal factors. Compared with expert methods, data-driven statistical methods are more objective. In addition, deterministic methods are used to study the physical mechanism of debris flow and develop models to simulate debris flow susceptibility, but due to the complexity of input data and the difficulty of parameter calibration, these physical methods are usually limited to local scales and difficult to use for regional scale studies. Summary of the invention

[0003] The technical problem solved by the present invention is to provide a debris flow prediction method based on hybrid machine learning to solve the problem that the existing debris flow prediction is difficult.

[0004] The present invention solves the above technical problems by adopting a technical solution: a debris flow prediction method based on hybrid machine learning, comprising the following steps:

[0005] S1. Constructing an indicator system, wherein the indicator system includes ecological indicators, hydrological indicators, geotechnical indicators and topographic indicators;

[0006] S2. Standardizing the indicator values ​​in the indicator system; the standardization includes normalization and deletion of outliers;

[0007] S3, build a support vector classification model, use the AOVA algorithm to optimize the hyperparameters, and obtain a hybrid machine learning model;

[0008] S4, constructing a training data set, and training the hybrid machine learning model, and using a cross-validation algorithm to construct a validation data set to validate the trained hybrid machine learning model, to obtain a debris flow prediction model; the training data set includes standardized index values ​​in the index system when historical debris flows occurred and standardized index values ​​in the index system when debris flows did not occur;

[0009] S5. Predicting the probability of debris flow occurrence using the debris flow prediction model.

[0010] Furthermore, the ecological indicators include vegetation weight load, the hydrological indicators include runoff velocity and flow depth, the geotechnical indicators include shear strength of soil, and the terrain indicators include elevation difference, channel slope, connectivity index and propagation probability index.

[0011] Furthermore, the calculation formula for runoff velocity is: Where V is the runoff velocity, L ( h represents the length of the contour line with height h, B ( h ) Represents the horizontal displacement of the contour line.

[0012] Furthermore, the calculation formula for flow depth is: Among them, H represents the flow depth, P represents the rainfall intensity, A represents the basin area, and b is the average width of the wet section.

[0013] Furthermore, the calculation formula for the shear strength of soil is: Where c represents cohesion, represents the friction angle of soil, σ represents the normal stress of soil, σ=γz, γ represents the density of soil, and z represents the elevation difference from the soil surface to the bedrock surface.

[0014] Furthermore, the calculation formula of the connectivity index is: Among them, IC k represents the connectivity index, represents the average weight of upslope catchment areas determined by land use type, represents the average slope of the upslope catchment, represents the square root of the upslope catchment area, d i represents the length of the flow path from the sediment source area to the i-th watershed unit, W i represents the weight of the i-th watershed unit, S i Represents the slope of the i-th watershed unit.

[0015] Furthermore, deleting outliers includes: for indicator values ​​that obey normal distribution, deleting indicator values ​​outside the range of 3δ; for indicators that do not obey normal distribution, deleting indicator values ​​outside the range of Xδ, where δ represents the standard deviation and X is a set coefficient.

[0016] Furthermore, the method also includes generating a visualized debris flow prediction result map on a GIS platform.

[0017] Beneficial effects of the present invention: The present invention provides a debris flow prediction method based on hybrid machine learning, which constructs an index system, determines the index used for debris flow prediction, standardizes the index values ​​in the index system, and the standardization includes normalization and deletion of outliers. A support vector classification model is constructed, and the AOVA algorithm is used to optimize hyperparameters to obtain a hybrid machine learning model. A training data set is constructed, and the hybrid machine learning model is trained to obtain a debris flow prediction model. The debris flow prediction model is used to predict the probability of debris flow occurrence, thereby solving the problem of the difficulty of existing debris flow prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flow chart of a debris flow prediction method based on hybrid machine learning provided by the present invention. DETAILED DESCRIPTION

[0019] In view of the difficulty of existing debris flow prediction, the present invention provides a debris flow prediction method based on hybrid machine learning, comprising the following steps:

[0020] S1. Constructing an indicator system, wherein the indicator system includes ecological indicators, hydrological indicators, geotechnical indicators and topographic indicators;

[0021] Specifically, a watershed unit is an independent hydrological area with the river as the main line and the watershed as the boundary, which is suitable for predicting the occurrence of debris flows. Therefore, debris flow prediction is carried out based on the watershed unit.

[0022] However, the selection of prediction factors is crucial. Therefore, an index system is determined based on the debris flow formation mechanism and the commonly used prediction factors in existing debris flow prediction. The index system includes ecological indicators, hydrological indicators, geotechnical indicators and topographic indicators. The ecological indicators include vegetation weight load and root morphology, the hydrological indicators include runoff velocity and flow depth, the geotechnical indicators include shear strength and soil thickness, and the topographic indicators include elevation difference, channel slope, connectivity index and propagation probability index.

[0023] Since the collinearity of features in machine learning may lead to model instability, the Spearman analysis technique was used to calculate the correlation coefficients between the indicators. The correlation coefficient between root morphology and vegetation weight load was 0.83, and the correlation coefficient between soil thickness and shear strength of soil was 0.82. Since the correlation coefficients of these two pairs of indicators were higher than the threshold of 0.8, the root morphology and soil thickness were eliminated. The final indicators determined were: runoff velocity, flow depth, shear strength of soil, elevation difference, channel slope, connectivity index and propagation probability index.

[0024] The formula for calculating runoff velocity is: Where V is the runoff velocity, L( h ) represents the length of the contour line with height h, B ( h ) Represents the horizontal displacement of the contour line.

[0025] The calculation formula for flow depth is: Among them, H represents the flow depth, P represents the rainfall intensity, A represents the basin area, and b is the average width of the wet section.

[0026] The calculation formula for the shear strength of soil is: Where c represents cohesion, represents the friction angle of soil, σ represents the normal stress of soil, σ=γz, γ represents the density of soil, and z represents the elevation difference from the soil surface to the bedrock surface.

[0027] The calculation formula of connectivity index is: Among them, IC k represents the connectivity index, represents the average weight of upslope catchment areas determined by land use type, represents the average slope of the upslope catchment, represents the square root of the upslope catchment area, d i represents the length of the flow path from the sediment source area to the i-th watershed unit, W i represents the weight of the i-th watershed unit, S i Represents the slope of the i-th watershed unit.

[0028] The calculation of the propagation probability index belongs to the prior art.

[0029] S2. Standardizing the indicator values ​​in the indicator system; the standardization includes normalization and deletion of outliers;

[0030] Specifically, normalizing the values ​​of each indicator is helpful to accelerate the convergence of the model and improve the accuracy of the model. The normalization formula is: Among them, I final represents the index value after normalization, I represents the index value before normalization, and I max Indicates the maximum value of the indicator value, I min Indicates the minimum value among the indicator values.

[0031] Deleting outliers will improve the ability to fit and explore the main relationship between debris flow occurrence and disaster-causing factors. Deleting outliers specifically includes: for indicator values ​​that obey the normal distribution, deleting indicator values ​​outside the range of 3δ; for indicators that do not obey the normal distribution, deleting indicator values ​​outside the range of Xδ, where δ represents the standard deviation and X is the set coefficient.

[0032] S3, build a support vector classification model, use the AOVA algorithm to optimize the hyperparameters, and obtain a hybrid machine learning model;

[0033] Specifically, in the traditional machine learning model training process, hyperparameters need to be set manually. Due to time and labor costs, the traditional hyperparameter debugging process is difficult to find the best hyperparameters from all parameter groups, especially when the hyperparameters can be floating-point parameters. In order to solve this defect, the AOVA algorithm is used to optimize the hyperparameters.

[0034] S4, constructing a training data set, and training the hybrid machine learning model, and using a cross-validation algorithm to construct a validation data set to validate the trained hybrid machine learning model, to obtain a debris flow prediction model; the training data set includes standardized index values ​​in the index system when historical debris flows occurred and standardized index values ​​in the index system when debris flows did not occur;

[0035] Specifically, using a cross-validation algorithm to construct a validation data set to validate the trained hybrid machine learning model can effectively avoid overfitting.

[0036] S5. Predicting the probability of debris flow occurrence using the debris flow prediction model.

[0037] Specifically, the standardized index value corresponding to the watershed unit to be predicted is input into the debris flow prediction model to obtain the probability of debris flow occurring in the watershed unit to be predicted. In order to more intuitively observe the probability of debris flow occurring in a certain area, a visualized debris flow prediction result map of all watershed units in the area is generated on the GIS platform.

[0038] In the present invention, in order to verify the accuracy of outlier deletion, AOVA algorithm optimization of hyperparameters and support vector classification model, that is, the accuracy of RO-AOVA-SVC, extreme gradient boosting models and random forest models are established respectively, such as: AOVA-RF, AOVA-SVC, AOVA-XGB, RO-AOVA-RF and RO-AOVA-XGB, where AOVA represents the use of AOVA algorithm to optimize hyperparameters, RF represents the random forest model, SVC represents the support vector classification model, XGB represents the extreme gradient boosting model, and RO represents the use of outlier deletion.

[0039] The model prediction accuracy was used to evaluate the model performance, and it was obtained that RO-AOVA-SVC can effectively predict the occurrence of debris flows in the extreme terrain transition zone.

[0040] The model performance evaluation analysis ACC was used for analysis, and it was found that the machine learning model with the addition of AOVA and RO algorithms was better than the single machine learning model. According to the ACC analysis, the RO-AOVA optimization algorithm improved the performance of SVC, RF, and XGB by 3.84%, 2.59%, and 5.94%, respectively. When only the AOVA algorithm was used, the ACC values ​​of SVC, RF, and XGB increased by 2.63%, 0.56%, and 1.34%, respectively. In addition, the RO algorithm improved the performance of AOVA-SVC, AOVA-RF, and AOVA-XGB by 1.21%, 2.03%, and 4.60%, respectively.

Claims

1. A debris flow prediction method based on hybrid machine learning, characterized in that: The following steps are involved: S1. Constructing an indicator system, wherein the indicator system includes ecological indicators, hydrological indicators, geotechnical indicators and topographic indicators; S2. Standardizing the indicator values ​​in the indicator system; the standardization includes normalization and deletion of outliers; S3, build a support vector classification model, use the AOVA algorithm to optimize the hyperparameters, and obtain a hybrid machine learning model; S4, constructing a training data set, and training the hybrid machine learning model, and using a cross-validation algorithm to construct a validation data set to validate the trained hybrid machine learning model, to obtain a debris flow prediction model; the training data set includes standardized index values ​​in the index system when historical debris flows occurred and standardized index values ​​in the index system when debris flows did not occur; S5. Predicting the probability of debris flow occurrence using the debris flow prediction model.

2. The debris flow prediction method based on hybrid machine learning according to claim 1 is characterized in that: The ecological indicators include vegetation weight load, the hydrological indicators include runoff velocity and flow depth, the geotechnical indicators include the shear strength of soil, and the terrain indicators include elevation difference, channel slope, connectivity index and propagation probability index.

3. The debris flow prediction method based on hybrid machine learning according to claim 2 is characterized in that: When constructing the indicator system, the Spearman analysis technique is used to calculate the correlation coefficient between the indicators, and any two indicators with a correlation coefficient greater than the threshold are eliminated.

4. The debris flow prediction method based on hybrid machine learning according to claim 2 is characterized in that: The formula for calculating runoff velocity is: Where V is the runoff velocity, L(h) is the length of the contour line with height h, and B(h) is the horizontal displacement of the contour line.

5. The debris flow prediction method based on hybrid machine learning according to claim 4 is characterized in that: The calculation formula for flow depth is: Among them, H represents the flow depth, P represents the rainfall intensity, A represents the basin area, and b is the average width of the wet section.

6. The debris flow prediction method based on hybrid machine learning according to claim 2 is characterized in that: The calculation formula for the shear strength of soil is: Where c represents cohesion, represents the friction angle of soil, σ represents the normal stress of soil, σ=γz, γ represents the density of soil, and z represents the elevation difference from the soil surface to the bedrock surface.

7. The debris flow prediction method based on hybrid machine learning according to claim 2 is characterized in that: The calculation formula of connectivity index is: Among them, IC k represents the connectivity index, represents the average weight of upslope catchment areas determined by land use type, represents the average slope of the upslope catchment, represents the square root of the upslope catchment area, d i represents the length of the flow path from the sediment source area to the i-th watershed unit, W i represents the weight of the i-th watershed unit, S i Represents the slope of the i-th watershed unit.

8. The debris flow prediction method based on hybrid machine learning according to claim 1 is characterized in that: Deleting outliers includes: for indicator values ​​that obey normal distribution, deleting indicator values ​​outside the range of 3δ; for indicators that do not obey normal distribution, deleting indicator values ​​outside the range of Xδ, where δ represents the standard deviation and X is the set coefficient.

9. The debris flow prediction method based on hybrid machine learning according to claim 1, characterized in that: The method also includes generating a visualized debris flow prediction result map on a GIS platform.

Images