A cross-regional landslide susceptibility assessment method and device based on transfer learning
Through the transfer learning method, the landslide susceptibility evaluation model in typical areas is used to evaluate the landslide susceptibility evaluation in the target study area, which solves the problem of poor generalization ability of the existing technology mid-span-region landslide susceptibility evaluation model, and realizes the Efficient and accurate landslide proneness evaluation, reduce resource waste and provide excellent disaster prevention and mitigation information.
Patent Information
- Application Number
- CN202311814489.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-27
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2043-12-27
AI Technical Summary
The existing landslide susceptibility evaluation model has poor generalization ability when applied across regions, and needs to be remodeled for each region, resulting in waste of human and material resources, and it is difficult to effectively evaluate landslide characteristics across scenarios and categories.
Using a transfer learning method, a landslide proneness evaluation model is established for typical areas. Through direct push transfer learning, the second evaluation factor is input to the target research area, and a logistic regression, support vector machine, deep neural network and Transformer model are constructed for slope slip. Efficacy evaluation, reduces the need for reconstruction, and improves efficiency and accuracy.
It effectively reduces the waste of human and material resources, improves the efficiency and accuracy of landslide proneness evaluation, and can achieve extrapolation from typical areas to sample-free areas under similar terrain conditions, providing excellent disaster prevention and mitigation reference information.
Smart Images

Figure CN117648870B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of geological disaster prediction, and in particular to a method and device for evaluating cross-regional landslide susceptibility based on transfer learning. Background Art
[0002] The relevant research on landslide susceptibility assessment began in the mid-20th century and has been developed in foreign countries. At that time, the research mainly focused on the conditions that promote the formation of landslides and the development process of landslide disasters. In the late 1970s, a series of major landslide events occurred in my country. Domestic institutions carried out research on landslide susceptibility zoning, which promoted the development of landslide prevention and disaster reduction at home and abroad. Landslide is a common geological disaster, which is characterized by wide distribution, high frequency, fast movement speed and great destructiveness. Landslide is the result of changes in environmental parameters. Under the influence of extreme climate events, the number of landslides occurring worldwide is increasing, causing extensive losses in economy, natural resources and human life on a global scale. The essence of landslide susceptibility assessment is to complete the spatial probability assessment of the area where landslide disasters may occur in the future based on the existing landslide hazard data in the region.
[0003] At present, there are various types of landslide susceptibility assessment models, including machine learning and deep learning methods. However, when the above methods are applied to regional landslide susceptibility assessment, the diversity of spatial distribution remains the main challenge. At the same time, these methods usually model specific landslide areas. Once cross-scenario and cross-category are taken into account, the generalization ability of the model will be significantly reduced, or a large amount of training data will need to be calibrated for each field to re-model according to the characteristics of the new landslide area, which will consume a lot of manpower and material resources. Summary of the invention
[0004] In order to overcome the technical problems existing in the related art at least to a certain extent, the present invention provides a cross-regional landslide susceptibility assessment method and device based on transfer learning.
[0005] In a first aspect, an embodiment of the present invention provides a cross-regional landslide susceptibility assessment method based on transfer learning, comprising the following steps:
[0006] S100: establishing a sample database for landslide susceptibility evaluation in a typical region, wherein the sample database includes a training set and a test set;
[0007] S200: selecting a first evaluation factor of landslide susceptibility in the target study area, quantifying the first evaluation factor, and obtaining a quantified result;
[0008] S300: extracting and verifying a second evaluation factor of the typical area from the quantification result, wherein an absolute value of a correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold;
[0009] S400: constructing a landslide susceptibility assessment model for the typical area and performing accuracy verification, wherein the landslide susceptibility assessment model includes a logistic regression model, a support vector machine model, a deep neural network model, a recurrent neural network model, and a Transformer model;
[0010] S500: evaluating the landslide susceptibility of the target study area based on the transfer learning method to obtain a susceptibility zoning result;
[0011] S600: verifying and analyzing the susceptibility zoning result.
[0012] Furthermore, the step of establishing a sample database for landslide susceptibility assessment in typical regions specifically includes:
[0013] Based on the field point data of the typical area, the landslide location distribution and boundary information are obtained through visual interpretation;
[0014] Dividing the typical area into rectangular grids, superimposing the landslide location distribution information on the rectangular grids, and counting the area proportion of the interpreted landslide in each rectangular grid;
[0015] The positive and negative properties of the samples in the rectangular grid are determined according to the area ratio, and the training set and the test set are divided according to a certain ratio.
[0016] Furthermore, the first evaluation factor includes topography, meteorology and hydrology, geological structure and human activity dimensions. The topography includes elevation, slope, aspect, curvature and terrain undulation. The meteorology and hydrology include multi-year average rainfall, terrain humidity index and distance to water system. The geological structure includes engineering rock group, slope structure and distance to fault. The human activities include distance to road and land use type.
[0017] Furthermore, the step S300 specifically includes:
[0018] Extracting the first evaluation factors in sequence, and using a factor independence test tool to calculate the correlation coefficients between the first evaluation factors;
[0019] The first evaluation factor whose absolute value of the correlation coefficient is less than the preset threshold is eliminated and used as the second evaluation factor.
[0020] Furthermore, the step S400 specifically includes:
[0021] Landslide susceptibility evaluation models were constructed based on logistic regression model, support vector machine model, deep neural network model, recurrent neural network model and Transformer model respectively. The test accuracy of landslide susceptibility assessment was compared using confusion matrix, ROC curve and susceptibility partition test accuracy.
[0022] Furthermore, the step of evaluating the landslide susceptibility of the target study area based on the transfer learning method specifically includes:
[0023] The direct push transfer learning method is adopted to input the second evaluation factors into the landslide susceptibility evaluation model of the typical area in sequence, and the landslide susceptibility evaluation is performed pixel by pixel.
[0024] In a second aspect, an embodiment of the present invention provides a cross-regional landslide susceptibility assessment device based on transfer learning, comprising:
[0025] A sample database establishment module is used to establish a sample database for landslide susceptibility evaluation in a typical area, wherein the sample database includes a training set and a test set;
[0026] A quantification module is selected to select a first evaluation factor of landslide susceptibility in a target study area, and quantify the first evaluation factor to obtain a quantification result;
[0027] An extraction and inspection module, used for extracting and inspecting a second evaluation factor of the typical area from the quantification result, wherein an absolute value of a correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold;
[0028] An evaluation model construction module, used to construct a landslide susceptibility evaluation model for the typical area and perform accuracy verification, wherein the landslide susceptibility evaluation model includes a logistic regression model, a support vector machine model, a deep neural network model, a recurrent neural network model and a Transformer model;
[0029] A landslide susceptibility assessment module is used to evaluate the landslide susceptibility of the target study area based on a transfer learning method to obtain a susceptibility zoning result;
[0030] The verification and analysis module is used to verify and analyze the susceptibility zoning results.
[0031] Furthermore, the sample database establishment module further includes:
[0032] A visual interpretation submodule, for obtaining the location distribution and boundary information of the landslide through visual interpretation based on the field point data of the typical area;
[0033] A grid statistics submodule is used to divide the typical area into rectangular grids, and superimpose the landslide location distribution information on the rectangular grids, and count the area proportion of the interpreted landslide in each rectangular grid;
[0034] The sample division submodule is used to determine the positive and negative nature of the samples in the rectangular grid according to the area ratio, and to divide the training set and the test set according to a certain ratio.
[0035] Furthermore, the extraction and inspection module further comprises:
[0036] A correlation coefficient calculation submodule, used to extract the first evaluation factors in sequence, and calculate the correlation coefficients between the first evaluation factors using a factor independence test tool;
[0037] The evaluation factor elimination submodule is used to eliminate the first evaluation factor whose absolute value of the correlation coefficient is less than the preset threshold and use it as the second evaluation factor.
[0038] The technical solution provided by the embodiments of the present invention has the following beneficial effects:
[0039] The embodiment of the present invention first establishes a sample database for landslide susceptibility evaluation in typical areas, selects and quantifies the first evaluation factor of the target study area, and obtains a quantitative result; then extracts and verifies the second evaluation factor of the typical area from the quantitative result, thereby ensuring that the typical area and the target study area are in the same set of evaluation index systems; combines machine learning and deep neural network methods to build a landslide susceptibility evaluation model for typical areas, and performs accuracy verification, then performs landslide susceptibility evaluation in the target study area based on the transfer learning method, and finally combines field survey landslide point data and detailed interpretation of landslide superposition to complete the verification and analysis of susceptibility zoning results. The embodiment of the present invention effectively reduces the waste of manpower and material resources, and improves the efficiency and accuracy of landslide susceptibility evaluation in the target study area.
[0040] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.
[0042] Figure 1 It is a flow chart of a cross-regional landslide susceptibility assessment method based on transfer learning provided in an embodiment of the present invention.
[0043] Figure 2 It is a precision and loss diagram of the training set and the test set provided by the embodiment of the present invention.
[0044] Figure 3 Schematic diagram of the ROC curve and AUC value of the test set provided by an embodiment of the present invention.
[0045] Figure 4 This is the prediction result of the landslide susceptibility index in the upper reaches of the Minjiang River based on the logistic regression model provided by the embodiment of the present invention.
[0046] Figure 5 This is the prediction result of the landslide susceptibility index in the upper reaches of the Minjiang River based on the support vector machine model provided by an embodiment of the present invention.
[0047] Figure 6 This is the prediction result of the landslide susceptibility index in the upper reaches of the Minjiang River based on a deep neural network provided by an embodiment of the present invention.
[0048] Figure 7 This is the prediction result of the landslide susceptibility index in the upper reaches of the Minjiang River based on a recurrent neural network provided by an embodiment of the present invention.
[0049] Figure 8 This is the prediction result of the landslide susceptibility index in the upper reaches of the Minjiang River based on the Transformer model provided by an embodiment of the present invention.
[0050] Fig. 9 The embodiments of the present invention provide the proportion of landslide field points and interpretation surfaces in the landslide-prone areas in the upper reaches of the Minjiang River based on different models.
[0051] Fig.10 It is the frequency ratio of landslide field points and interpretation surfaces in the landslide-prone area in the upper reaches of the Minjiang River based on different models provided by the embodiment of the present invention.
[0052] Fig.11 It is a block diagram of a cross-regional landslide susceptibility assessment device based on transfer learning provided by an embodiment of the present invention.
[0053] The accompanying drawings in the figure are marked as follows:
[0054] Sample database establishment module 100; selection and quantification module 200; extraction and inspection module 300;
[0055] Evaluation model building module 400; landslide susceptibility evaluation module 500; verification analysis module 600. DETAILED DESCRIPTION
[0056] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of methods and devices consistent with some aspects of the present application as detailed in the appended claims.
[0057] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are further described in detail below with reference to the accompanying drawings.
[0058] Example 1
[0059] like Figure 1 FIG. 1 is a flow chart of a cross-regional landslide susceptibility assessment method based on transfer learning provided by an embodiment of the present invention. The method may include the following steps:
[0060] Step S100: establishing a sample database for landslide susceptibility assessment in a typical region, wherein the sample database includes a training set and a test set.
[0061] In this embodiment, the sample database includes landslide attribute data and evaluation factor attribute values. The landslide attribute data is landslide or non-landslide; the evaluation factor attribute value represents the factors that affect the occurrence of landslide disasters. During implementation, the training set is used to establish an evaluation model, and the test set is used to test the modeling results.
[0062] Furthermore, the step S100 specifically includes the following steps:
[0063] Based on the field point data of the typical area, the landslide location distribution and boundary information are obtained through visual interpretation.
[0064] Dividing the typical area into rectangular grids, superimposing the landslide location distribution information on the rectangular grids, and counting the area proportion of the interpreted landslide in each rectangular grid;
[0065] The positive and negative properties of the samples in the rectangular grid are determined according to the area ratio, and the training set and the test set are divided according to a certain ratio.
[0066] Among them, visual interpretation is a type of remote sensing image interpretation, also known as visual interpretation, or visual interpretation, which is the reverse process of remote sensing imaging. It refers to the process by which professionals obtain information about specific target objects on remote sensing images through direct observation or with the help of auxiliary interpretation instruments. Visual interpretation refers to extracting useful information from remote sensing images through brain analysis, reasoning, and judgment, relying on the human eye (optical instruments can also be used), relying on the interpreter's knowledge, experience, and relevant information.
[0067] In this embodiment, the typical area is Wenchuan County. First, based on the field data of landslides in Wenchuan County, the location distribution and boundary information of landslides in the typical area Wenchuan County are obtained through visual interpretation. Then, the typical area Wenchuan County is divided according to a rectangular grid of 30×30 meters, and the landslide disaster data obtained by visual interpretation is superimposed on the rectangular grid, and the area proportion of the interpreted landslides in each rectangular grid is counted. Among them, the rectangular grids with a landslide area proportion of more than 50% are identified as positive samples, and the rectangular grids with a landslide area proportion of less than 50% are identified as negative samples.
[0068] After counting the number of grids occupied by landslide disasters, an equal number of non-landslide grids are uniformly selected within Wenchuan County. The landslide sample database includes landslide grid units and the same number of non-landslide grid units selected by random selection. In addition, the sample database is divided into a training set and a test set according to a 4:1 ratio. The training set samples are preferably 257,424, and the test set samples are preferably 64,356.
[0069] Step S200: selecting a first evaluation factor of landslide susceptibility in a target study area, quantifying the first evaluation factor, and obtaining a quantified result.
[0070] In order to improve the scientific nature of the first evaluation factor, before selecting the first evaluation factor, it is necessary to consider the significant differences in geographical and geological conditions in different regions, fully grasp the relevant materials of the target research area, and make comprehensive considerations based on local topography, geological structure and other conditions. In addition, the correlation between the selected evaluation factors should be as low as possible, so as to quantitatively analyze the impact of the evaluation factors on the landslide and their weights. Therefore, after selecting the evaluation factors, correlation analysis is required to eliminate high correlation factors.
[0071] According to the long-term geological environment background data of the upper reaches of the Minjiang River, and based on a deep understanding of the landslide breeding conditions, the embodiment of the present invention selects 13 factors from numerous factors as indicators for evaluating the landslide susceptibility of the research target area, and constructs an evaluation index system.
[0072] Specifically, the first evaluation factor includes topography, meteorology and hydrology, geological structure and human activity dimensions. The topography includes elevation, slope, aspect, curvature and terrain undulation. The meteorology and hydrology include multi-year average rainfall, terrain humidity index and distance to water system. The geological structure includes engineering rock group, slope structure and distance to fault. The human activities include distance to road and land use type.
[0073] In one implementation, the selected first evaluation factor of the upper reaches of the Minjiang River is normalized and quantified to obtain a series of layers.
[0074] Step S300: extracting and verifying a second evaluation factor of the typical region from the quantification result, wherein an absolute value of a correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold.
[0075] Furthermore, the step S300 specifically includes:
[0076] Extracting the first evaluation factors in sequence, and using a factor independence test tool to calculate the correlation coefficients between the first evaluation factors;
[0077] The first evaluation factor whose absolute value of the correlation coefficient is less than the preset threshold is eliminated and used as the second evaluation factor.
[0078] In order to achieve the premise of promoting the susceptibility evaluation of typical areas to the target research area, it is necessary to keep the evaluation indicators input in the overall area within the same value range. Therefore, this embodiment chooses to quantify the indicators of the upper reaches of the Minjiang River without the intervention of landslide disasters. The evaluation factors of Wenchuan County in the typical area are cut out from the various factor indicators of the upper reaches of the Minjiang River in turn according to the vector boundary, and the second evaluation factor of Wenchuan County is normalized and quantified to obtain a series of layers.
[0079] This embodiment uses ArcGIS to implement the factor independence test. Since the absolute values of the correlation coefficients between slope structure and slope and aspect factors, engineering rock group and land use factors, and elevation and engineering rock group and land use factors are too large, after eliminating engineering rock group, slope structure and land use factors, the elevation data from which several evaluation factors are derived are retained, and the remaining 10 evaluation factors are used as the second evaluation factors.
[0080] Step S400: constructing a landslide susceptibility assessment model for the typical area and performing accuracy verification, wherein the landslide susceptibility assessment model includes a logistic regression model, a support vector machine model, a deep neural network model, a recurrent neural network model and a Transformer model.
[0081] In this embodiment, landslide susceptibility evaluation models are constructed based on the logistic regression model, support vector machine model, deep neural network model, recurrent neural network model and Transformer model, respectively, and the test accuracy of landslide susceptibility assessment is compared using confusion matrix, ROC curve and susceptibility partition test accuracy.
[0082] In one embodiment, Pycharm is used to construct a logistic regression model for landslide susceptibility evaluation. The randomly divided training set sample data is input into the code. When a landslide occurs, the dependent variable is 1; when a landslide does not occur, the dependent variable is 0. A significance test is performed on each evaluation factor by comparing the Sig value. If the Sig value is less than 0.05, it means that the difference is significant, and the evaluation factor can enter the logistic regression model to complete the modeling task. The regression coefficient β (β1, β2, ..., βn) calculated by Table 1 is brought into the logistic regression formula.
[0083] Table 1 Significance test of logistic regression coefficients
[0084] Evaluation Factor B SE <![CDATA[ Wals ]]> <![CDATA[ df ]]> <![CDATA[ Sig ]]> Exp(B) slope 9.6877 0.093 104.193 1 0.000 0.365 Slope aspect -17.1338 0.083 -206.223 1 0.000 0.326 Elevation 0.2007 0.022 -8.968 1 0.000 0.088 Terrain relief -2.6382 0.477 -5.534 1 0.000 1.896 Curvature -15.2212 0.189 -80.429 1 0.000 0.742 Distance to water system -1.0328 0.063 -16.364 1 0.000 0.247 Topographic Wetness Index 1.3164 0.077 17.061 1 0.000 0.303 Average rainfall over the years -4.1308 0.116 -35.676 1 0.000 0.454 Distance from fault 0.9742 0.082 10.738 1 0.000 0.323 Distance from road -0.2979 0.088 3.400 1 0.001 0.344 constant 14.0203 0.167 89.923 1 0.000 0.655
[0085] Where: B is the logistic regression coefficient; SE is the standard deviation; Wals is the chi-square value; df is the degree of freedom; Sig is the indicator significance; Exp(B) is the evaluation factor odds ratio.
[0086] The logistic regression formula is:
[0087]
[0088] Among them: x1 is the slope; x2 is the slope direction; x3 is the elevation; x4 is the terrain undulation; x5 is the curvature; x6 is the distance to the water system; x7 is the terrain humidity index; x8 is the average rainfall over many years; x9 is the distance to the fault; x10 is the distance to the road.
[0089] P is the probability of landslide in Wenchuan County, and its value range is [0, 1]. After the logistic regression model is established, the classification results are divided with 0.5 as the threshold: when the output pixel value is greater than or equal to 0.5, the pixel is assigned a value of 1 and is identified as a landslide occurrence area; when the output pixel value is less than 0.5, the pixel is assigned a value of 0 and is identified as a landslide non-occurrence area. The classification results of the training set samples are shown in Table 2: The training samples have a total of 257,424 pixels, of which the prediction accuracy of landslide disasters is 89.1%; the prediction accuracy of non-landslide disasters is 82.5%; the total prediction rate of the two is 85.8%. The classification results of the test set samples are shown in Table 3: The test samples have 64,356 pixels, of which the prediction accuracy of landslide disasters is 89.1%; the prediction accuracy of non-landslide disasters is 82.5%; the total prediction rate of the two is 85.8%.
[0090] Table 2 Classification results of the logistic regression model for training set sample pixels
[0091]
[0092] Table 3 Classification results of the logistic regression model for the test set sample pixels
[0093]
[0094] Using the raster calculator of ArcGIS software, the evaluation factor layers were input in sequence according to the above formula to calculate the landslide susceptibility index map of Wenchuan County. According to the natural break point method, the calculated index values were divided into four levels, namely, low landslide susceptibility area, medium landslide susceptibility area, high landslide susceptibility area, and extremely high landslide susceptibility area.
[0095] In one embodiment, a landslide susceptibility assessment model based on a support vector machine is established using a Python platform, and the model is trained using a training set sample and the model parameters are optimized. In this embodiment, a radial basis kernel function is selected as the kernel function, the penalty coefficient C is set to have a value range of (2-5, 215), and the γ value is set to have a value range of (2-15, 25). The grid search method is combined with a 5-fold cross validation to perform parameter optimization, and the optimal parameter C is 1 and the γ value is 2. The grid search method has two advantages: one is that the global optimum can be obtained, and the other is that C and γ are independent of each other, which is convenient for parallel calculation. Feature weights, see Table 4.
[0096] Table 4 Feature weights calculated by support vector machine
[0097]
[0098] After the support vector machine model is established, the output value is normalized and the classification results are divided with a threshold of 0.5: when the output pixel value is greater than or equal to 0.5, the pixel is assigned a value of 1 and is identified as a landslide occurrence area; when the output pixel value is less than 0.5, the pixel is assigned a value of 0 and is identified as a landslide non-occurrence area. From the classification results of the training set samples as shown in Table 5, it can be concluded that in the training samples, the prediction accuracy of landslide disasters is 92.2%; the prediction accuracy of non-landslide disasters is 86.5%; the total prediction rate of the two reaches 89.4%. From the classification results of the test set samples as shown in Table 6, it can be seen that in the test samples, the prediction accuracy of landslide disasters is 95.6%; the prediction accuracy of non-landslide disasters is 81%; the total prediction rate of the two reaches 88.3%. The entire process of support vector machine training and testing took a total of 157 minutes, which is less efficient than other networks.
[0099] Table 5 Classification results of support vector machine model for training sample pixels
[0100]
[0101] Table 6 Classification results of the support vector machine model for test sample pixels
[0102]
[0103] Using the raster calculator of ArcGIS software, the evaluation factor layers were input in sequence according to the weight values, and the landslide susceptibility index of Wenchuan County was calculated and normalized before being plotted. According to the natural break point method, the calculated index values were divided into four levels: low landslide susceptibility area, medium landslide susceptibility area, high landslide susceptibility area, and extremely high landslide susceptibility area.
[0104] In one embodiment, a deep neural network computing model is written in PyCharm using Python language based on the keras framework. After combining different parameters to achieve the best result, the best parameter combination is selected, and L1 regularization and Dropout are introduced to optimize the model. The parameter settings of the entire model are shown in Table 7. By observing the changes in the accuracy loss of the training set and the test set, the number of training times is determined to avoid overfitting of the model. The accuracy and loss diagrams of the training set and the test set are shown in Figure 2 .
[0105] Table 7 Deep neural network parameter settings
[0106] Parameter name Settings Epochs 500 Dropout 0.5 Learning Rate 0.001 Number of hidden layers 5 Dense connection 512 / 1024 / 1024 / 2048 / 2048 / 2 Activation Function <![CDATA[ ReLU ]]> Optimizer Adam Loss Function Binary cross-entropy
[0107] All samples are input into the established deep neural network model, and the output classification results are divided with 0.5 as the threshold: when the output pixel value is greater than or equal to 0.5, the pixel is assigned a value of 1 and is identified as a landslide occurrence area; when the output pixel value is less than 0.5, the pixel is assigned a value of 0 and is identified as a landslide non-occurrence area. The classification results of the training set samples are shown in Table 8: In the training samples, the prediction accuracy of landslide disasters is 91.4%; the prediction accuracy of non-landslide disasters is 89.3%; the total prediction rate of the two is 90.4%. The classification results of the test set samples are shown in Table 9: In the test samples, the prediction accuracy of landslide disasters is 91.5%; the prediction accuracy of non-landslide disasters is 89.2%; the total prediction rate of the two is 90.4%.
[0108] Table 8 Classification results of deep neural network model for training sample pixels
[0109]
[0110] Table 9 Classification results of deep neural network model for test sample pixels
[0111]
[0112] Each evaluation raster layer is converted into ASCII code and input into the deep neural network model in turn to calculate the probability of landslide occurrence in each pixel, and obtain the landslide susceptibility index map and classification schematic diagram.
[0113] In one embodiment, a recurrent neural network is constructed to evaluate the susceptibility of landslides in Wenchuan County. The overall model structure includes an input layer and two relative LSTM layers. The input layer inputs the evaluation sample data into the recurrent neural network in a digital form; the output layer adopts a softmax regression model. A Dropout strategy is introduced to prevent overfitting. The Dropout value is set to 0.5. By selectively abandoning 50% of the parameters, the generalization ability of the evaluation model can be improved to a certain extent. Such an operation can ensure accuracy and improve training efficiency. The specific hyperparameter settings are shown in Table 10. By optimizing the hyperparameters, finding the optimal weight value between the evaluation factor and whether the landslide occurs or not is the ultimate training goal of the model. With the increase in the number of iterations, its accuracy and loss value will also change accordingly.
[0114] Table 10 Parameter settings of recurrent neural network
[0115] Parameter name Settings Epochs 500 Dropout 0.5 Learning Rate 0.001 Embedding_size 512 <![CDATA[ SimpleRNN ]]> 512 / 1024 / 512 / 256 / 2 Activation Function <![CDATA[ ReLU ]]> Optimizer Adam Loss Function Binary cross-entropy
[0116] All samples were input into the established recurrent neural network model, and the output classification results were divided with 0.5 as the threshold: when the output pixel value was greater than or equal to 0.5, the pixel was assigned a value of 1 and was identified as a landslide occurrence area. When the output pixel value of the training set sample was less than 0.5, the pixel was assigned a value of 0 and was identified as a landslide non-occurrence area. The classification results are shown in Table 11: In the training samples, the prediction accuracy of landslide disasters was 91.4%; the prediction accuracy of non-landslide disasters was 88.3%; the total prediction rate of the two reached 88%. The classification results of the test set samples are shown in Table 12: In the test samples, the prediction accuracy of landslide disasters was 91.3%; the prediction accuracy of non-landslide disasters was 88.2%; the total prediction rate of the two reached 89.8%.
[0117] Table 11 Classification results of the recurrent neural network model for training sample pixels
[0118]
[0119] Table 12 Classification results of the recurrent neural network model for test sample pixels
[0120]
[0121] Each evaluation grid layer is converted into ASCII code and input into the recurrent neural network model in turn to calculate the probability of landslide occurrence in each pixel, and then the landslide susceptibility index map is obtained. According to the natural break point method, the calculated index value is divided into four levels: low landslide susceptibility area, medium landslide susceptibility area, high landslide susceptibility area, and extremely high landslide susceptibility area.
[0122] In one embodiment, a Transform model is constructed to conduct a study on landslide susceptibility assessment. In this embodiment, PyTorch is selected as the deep learning framework for the Transformer model applied to the landslide susceptibility assessment task to complete the experimental process. The number of iterations N of the Transformer model encoder during the experiment is set to 3, and other parameters of the experiment are shown in Table 13. This embodiment uses a binary cross-entropy loss function to calculate the loss value of the model. It can be found that during the training process, as the number of model iterations increases, the loss value gradually decreases. In the first 50 steps, the convergence speed is very fast, and the loss value has a significant change. Between 50 and 100 steps, the convergence speed gradually slows down, and the rate of change of the loss value decreases. After 100 steps, the loss value tends to a stable state.
[0123] Table 13 Transformer model parameter settings
[0124] Parameter name Settings Epochs 500 Dropout 0.5 Learning Rate 0.001 Embedding_size 512 Dense connection 3×(512 / 2048 / 512) / 256 / 2 Activation Function <![CDATA[ ReLU ]]> Optimizer Adam Loss Function Binary cross-entropy
[0125] In one embodiment, after the Transformer model is established, the output value is normalized and the classification result is divided with a threshold of 0.5: when the output pixel value is greater than or equal to 0.5, the pixel is assigned a value of 1 and is identified as a landslide occurrence area; when the output pixel value is less than 0.5, the pixel is assigned a value of 0 and is identified as a landslide non-occurrence area. From the classification results of the training set samples, as shown in Table 14, it can be seen that in the training samples, the prediction accuracy of landslide disasters is 96.2%; the prediction accuracy of non-landslide disasters is 93.3%; the total prediction rate of the two reaches 94.8%. From the classification results of the test set samples, as shown in Table 15, it can be obtained that in the test samples, the prediction accuracy of landslide disasters is 95.6%; the prediction accuracy of non-landslide disasters is 92.3%; the total prediction rate of the two reaches 94%.
[0126] Table 14 Transformer model training sample pixel classification results
[0127]
[0128] Table 15 Transformer model test sample pixel classification results
[0129]
[0130] Convert each evaluation raster layer into ASCII code, input them into the recurrent neural network model in turn, calculate the probability of landslide occurrence in each pixel, and obtain the landslide susceptibility index map of the study area. According to the natural break point method, the calculated index value is divided into four levels: low landslide prone area, medium landslide prone area, high landslide prone area, and extremely high landslide prone area. The accuracy of the landslide susceptibility evaluation of Wenchuan County obtained by the above model was verified. The present invention divides the training data and the test data in a ratio of 4:1, and uses the test data to evaluate the accuracy of the model results. It is known that the landslide susceptibility evaluation problem is a binary classification problem, and 0.5 is used as the threshold for dividing TP, TN, FP, and FN. The evaluation results of the training samples in different models are shown in Table 16, and the ranking results of the evaluation indicators of the model training samples can be shown in Table 18. The ROC curve results obtained from the test set are shown in Figure 3 shown.
[0131] Table 16 Test sample pixel classification accuracy indicators based on different models
[0132]
[0133] Table 17 Ranking of classification accuracy evaluation indicators of test samples based on different models
[0134]
[0135] The above table shows that: the correct classification ratio of the model for landslide samples is higher than that for non-landslide samples. Among them, the test results of the Transformer model (accuracy = 96.1%, precision = 94.6%, sensitivity = 97.8%, specificity = 94.4%, F1 score = 96.2%) are the best in all indicators; the recognition effect of the deep neural network on positive and negative samples is second only to the Transformer model, and the performance is relatively excellent. The number of correct recognition of positive samples is 1681 less than that of the Transformer model ranked first, and the number of correct recognition of negative samples is 1720 less than that of the Transformer model. The various result parameters of the deep neural network test are: accuracy = 90.7%, precision = 89.5%, sensitivity = 92.5%, specificity = 89.1%, F1 score = 91.0%. The values of various precision evaluation indicators of the recurrent neural network are slightly lower than those of the deep neural network, but it also shows its ability to correctly classify samples. Although the logistic regression model also reflects its own characteristics and achieves good classification results, it still has a large gap compared with the results of other models.
[0136] By comparing the ROC curves of the above five models, we can see that the AUC values of the five models are all greater than 0.93, which shows that the model evaluation effect is very good. The AUC values of the five models are: 0.93, 0.95, 0.96, 0.96 and 0.97 respectively. Arranged in order from small to large: AUC LR <AUC SVM <AUC DNN =AUC RNN <AUC TR .
[0137] There are certain differences between traditional machine learning methods and deep learning methods in the process of dividing the susceptibility of the whole area. Although the proportion of the designated susceptibility zones gradually decreases as the level increases, the deep learning method is more accurate and detailed than traditional machine learning. For example, the classification area accounts for a small proportion, but the extremely high susceptibility area with large disaster capacity and a larger area account for a larger proportion, and the low susceptibility area with smaller disaster capacity can be divided better than traditional machine learning, and the two susceptibility areas at the middle level have an overall small proportion.
[0138] It is certain that the zoning statistics of the Transformer model are the best. In this model, the high-prone area only accounts for 9.2% of the total area, but 91.3% of the landslide disasters are located here; the low-prone area accounts for 86.3%, occupying the vast majority of the study area, but only 4.6% of the landslide disasters are distributed in this area. The zoning and disaster statistics results can fully prove that the Transformer model can refine the extremely high-prone landslide zoning, which can provide a theoretical basis for the subsequent key landslide investigation work and reduce the workload of field investigation. All of the above test accuracy results can prove that the accuracy of the Transformer model is better than the other four models, and can achieve good landslide susceptibility evaluation results.
[0139] Step S500: evaluating the landslide susceptibility of the target study area based on the transfer learning method to obtain a susceptibility zoning result.
[0140] In this embodiment, a direct transfer learning method is adopted to sequentially input the second evaluation factors into the landslide susceptibility evaluation model of the typical area to perform a pixel-by-pixel landslide susceptibility assessment.
[0141] This embodiment adopts the direct push transfer learning method in transfer learning, and uses the five benchmark models (logistic regression model, support vector machine model, deep neural network, recurrent neural network and Transformer model) trained above to calculate the landslide susceptibility index pixel by pixel for Heishui County, Songpan County, Mao County and Li County, and splices the susceptibility evaluation results with the typical area Wenchuan County to form a complete zoning map of the landslide susceptibility index in the upper reaches of the Minjiang River. Different models will have a high degree of similarity in the size of the susceptibility probability prediction values of the same pixel, but they will not be completely consistent. When the logistic regression model and the support vector machine model are migrated, the direct influencing factor raster layer is converted into ASCII code as a test set input model to obtain the result; the deep learning model (DNN model, RNN model, Transformer model) directly predicts by loading the trained weights to the evaluation layer. Using ArcGIS software, the prediction results output by the model are converted into raster data, and combined with the field point data of the upper reaches of the Minjiang River landslide and the detailed interpreted landslide data for superposition and display, as shown respectively. Figure 4-Figure 8 shown.
[0142] Step S600: verifying and analyzing the susceptibility zoning result.
[0143] The susceptibility zoning method is used to count the proportion of landslide hazard points / surfaces in different zones and verify the accuracy of the final evaluation results. Fig. 9 and Fig.10 As shown in the figure, the field observation points and the landslide surface with detailed interpretation of the intelligent identification results are superimposed with the zoning results and statistically analyzed. According to the above results, it can be seen that: 1. The proportion of landslide field points and interpretation surface zoning basically increases with the severity of the landslide susceptibility zoning level, and the selected deep learning model divides the area of the intermediate level susceptibility interval less; 2. The zoning statistical results of the above models for different test objects can meet the qualified requirements of susceptibility zoning, and the landslide frequency ratio of the landslide field points increases with the increase of the susceptibility level; for different test objects, the performance ranking of each model in the extremely high susceptibility area is R TR >R DNN >R CNN >R SVM >R LR Among them, the Transformer model can divide the landslide high-prone areas more finely, that is, among the five susceptibility evaluation models, the area of the high-prone areas accounts for the lowest proportion, and the proportion of landslide disasters in the zones is relatively high. This shows that the Transformer model has strong stability in the division of susceptibility zones, and is the most suitable evaluation model for direct push transfer learning methods to predict the entire zone among all models.
[0144] This embodiment selects the direct push transfer learning method as the basis of transfer learning in this article. Through the previously trained machine learning model, the landslide susceptibility of other counties in the upper reaches of the Minjiang River without samples is evaluated pixel by pixel, and the migrated susceptibility zoning results of the upper reaches of the Minjiang River are combined with the landslide point data of the upper reaches of the Minjiang River field survey, the original interpreted landslides, and the landslide disaster superposition analysis based on deep learning identification and detailed interpretation. It proves that the Transformer model used in this article has the best transferability. This invention proves that transfer learning can achieve extrapolation from a typical area containing a limited sample set to a large area without samples under similar terrain conditions, and can achieve excellent results, which can provide reference information for future disaster prevention and mitigation work in the study area.
[0145] Example 2
[0146] The embodiment of the present invention also provides a cross-regional landslide susceptibility assessment device based on transfer learning, such as Fig.11 As shown, the device may include a sample database establishment module 100 , a selection and quantification module 200 , an extraction and inspection module 300 , an evaluation model construction module 400 , a landslide susceptibility evaluation module 500 and a verification and analysis module 600 .
[0147] Specifically, the sample database establishing module 100 is used to establish a sample database for landslide susceptibility assessment in a typical region, and the sample database includes a training set and a test set.
[0148] The selection and quantification module 200 is used to select a first evaluation factor of landslide susceptibility in a target research area, quantify the first evaluation factor, and obtain a quantification result.
[0149] The extraction and verification module 300 is used to extract and verify the second evaluation factor of the typical area from the quantification result, and the absolute value of the correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold.
[0150] The evaluation model construction module 400 is used to construct a landslide susceptibility evaluation model for the typical area and perform accuracy verification, wherein the landslide susceptibility evaluation model includes a logistic regression model, a support vector machine model, a deep neural network model, a recurrent neural network model and a Transformer model.
[0151] The landslide susceptibility assessment module 500 is used to perform landslide susceptibility assessment in the target study area based on a transfer learning method to obtain a susceptibility zoning result.
[0152] The verification and analysis module 600 is used to verify and analyze the susceptibility zoning result.
[0153] Furthermore, the sample database establishment module 100 may include a visual interpretation submodule, a grid statistics submodule and a sample division submodule.
[0154] The visual interpretation submodule is used to obtain the landslide location distribution and boundary information through visual interpretation based on the field point data of the typical area.
[0155] The grid statistics submodule is used to divide the typical area into rectangular grids, and superimpose the landslide location distribution information on the rectangular grids, and count the area proportion of the interpreted landslide in each rectangular grid.
[0156] The sample division submodule is used to determine the positive and negative nature of the samples in the rectangular grid according to the area ratio, and to divide the training set and the test set according to a certain ratio.
[0157] Furthermore, the extraction and inspection module 300 may include a correlation coefficient calculation submodule and an evaluation factor elimination submodule.
[0158] The correlation coefficient calculation submodule is used to extract the first evaluation factors in sequence and calculate the correlation coefficients between the first evaluation factors using a factor independence test tool.
[0159] The correlation coefficient calculation submodule is used to extract the first evaluation factors in sequence and calculate the correlation coefficients between the first evaluation factors using a factor independence test tool.
[0160] Regarding the device in the above embodiment, the specific steps for each module to perform the operation have been described in detail in the embodiment of the method, and will not be elaborated in detail here. Each module in the above-mentioned cross-regional landslide susceptibility assessment device based on transfer learning can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
[0161] In summary, the embodiment of the present invention first establishes a sample database for landslide susceptibility evaluation in typical areas, selects and quantifies the first evaluation factor of the target study area, and obtains a quantitative result; then extracts and verifies the second evaluation factor of the typical area from the quantitative result, thereby ensuring that the typical area and the target study area are in the same set of evaluation index systems; combines machine learning and deep neural network methods to build a landslide susceptibility evaluation model for typical areas, and verifies the accuracy, then evaluates the landslide susceptibility of the target study area based on the transfer learning method, and finally combines the field survey landslide point data and detailed interpretation of landslide superposition to complete the verification and analysis of the susceptibility zoning results. The embodiment of the present invention effectively reduces the waste of manpower and material resources, and improves the efficiency and accuracy of landslide susceptibility evaluation in the target study area.
[0162] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.
[0163] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by technicians in the technical field to which the embodiments of the present application belong.
[0164] In addition, each functional unit in each embodiment of the present application may be integrated into a processing module, or each unit may exist physically separately, or two or more units may be integrated into one module. The above-mentioned integrated module may be implemented in the form of hardware or in the form of a software functional module. If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it may also be stored in a computer-readable storage medium.
[0165] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.
[0166] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.
Claims
1. A cross-regional landslide susceptibility assessment method based on transfer learning, characterized in that: The following steps are involved: S100: establishing a sample database for landslide susceptibility evaluation in a typical region, wherein the sample database includes a training set and a test set; S200: selecting a first evaluation factor of landslide susceptibility in the target study area, quantifying the first evaluation factor, and obtaining a quantitative result; wherein the first evaluation factor includes topography, meteorology and hydrology, geological structure, and human activity dimensions, the topography includes elevation, slope, slope aspect, curvature, and terrain undulation, the meteorology and hydrology includes multi-year average rainfall, terrain humidity index, and distance to water system, the geological structure includes engineering rock group, slope structure, and distance to fault, and the human activity includes distance to road and land use type; S300: extracting and testing the second evaluation factor of the typical area from the quantification result, wherein the absolute value of the correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold, and specifically comprising: extracting the first evaluation factors in turn, and using the factor independence test tool to calculate the correlation coefficient between the first evaluation factors; removing the first evaluation factor whose absolute value of the correlation coefficient is not less than the preset threshold as the second evaluation factor; wherein the second evaluation factor is the evaluation factor after removing the engineering rock group, slope structure and land use type from the first evaluation factor; and, According to the vector boundary of the second evaluation factor of the typical area, the factor indicators in the first evaluation factor of the target research area are cut in sequence, and the second evaluation factor of the selected typical area is normalized and quantified to obtain a series of layers; S400: constructing a landslide susceptibility evaluation model for the typical area and verifying its accuracy, wherein the landslide susceptibility evaluation model is a Transformer model, and specifically comprising: constructing a landslide susceptibility evaluation model based on the Transformer model, and verifying its accuracy by using a confusion matrix, a ROC curve, and susceptibility partitioning; S500: Based on the transfer learning method, the landslide susceptibility evaluation of the target study area is performed to obtain a susceptibility zoning result, and specifically includes: using the direct transfer learning method, the second evaluation factor is sequentially input into the landslide susceptibility evaluation model of the typical area, and the landslide susceptibility evaluation is performed pixel by pixel; S600: verifying and analyzing the susceptibility zoning result.
2. The cross-regional landslide susceptibility assessment method based on transfer learning according to claim 1 is characterized in that: The steps of establishing a sample database for landslide susceptibility assessment in typical regions specifically include: Based on the field point data of the typical area, the landslide location distribution and boundary information are obtained through visual interpretation; Dividing the typical area into rectangular grids, superimposing the landslide location distribution information on the rectangular grids, and counting the area proportion of the interpreted landslide in each rectangular grid; The positive and negative properties of the samples in the rectangular grid are determined according to the area ratio, and the training set and the test set are divided according to a certain ratio.
3. A cross-regional landslide susceptibility assessment device based on transfer learning, characterized in that: include: A sample database establishment module is used to establish a sample database for landslide susceptibility evaluation in a typical area, wherein the sample database includes a training set and a test set; A quantification module is selected to select a first evaluation factor of landslide susceptibility in a target study area, and quantify the first evaluation factor to obtain a quantification result; wherein the first evaluation factor includes topography, meteorology and hydrology, geological structure and human activity dimensions, the topography includes elevation, slope, slope aspect, curvature and terrain undulation, the meteorology and hydrology includes multi-year average rainfall, terrain humidity index and distance to water system, the geological structure includes engineering rock group, slope structure and distance to fault, and the human activity includes distance to road and land use type; An extraction and inspection module is used to extract and inspect the second evaluation factor of the typical area from the quantification result, and the absolute value of the correlation coefficient between the second evaluation factor and the first evaluation factor is less than a preset threshold; and the extraction and inspection module further includes: a correlation coefficient calculation submodule, which is used to extract the first evaluation factors in turn and calculate the correlation coefficient between the first evaluation factors using a factor independence test tool; An evaluation factor elimination submodule, used to eliminate the first evaluation factor whose absolute value of the correlation coefficient is not less than the preset threshold value as the second evaluation factor; wherein the second evaluation factor is the evaluation factor after eliminating the engineering rock group, slope structure and land use type from the first evaluation factor; An evaluation model building module is used to build a landslide susceptibility evaluation model for the typical area and perform accuracy verification, wherein the landslide susceptibility evaluation model is a Transformer model; A landslide susceptibility assessment module is used to evaluate the landslide susceptibility of the target study area based on a transfer learning method to obtain a susceptibility zoning result; The verification and analysis module is used to verify and analyze the susceptibility zoning results.
4. The cross-regional landslide susceptibility assessment device based on transfer learning according to claim 3 is characterized in that: The sample database establishment module further includes: A visual interpretation submodule, for obtaining the location distribution and boundary information of the landslide through visual interpretation based on the field point data of the typical area; A grid statistics submodule is used to divide the typical area into rectangular grids, and superimpose the landslide location distribution information on the rectangular grids, and count the area proportion of the interpreted landslide in each rectangular grid; The sample division submodule is used to determine the positive and negative nature of the samples of the rectangular grid according to the area ratio, and to divide the training set and the test set according to a certain ratio.
