Method and device for early intelligent identification of potential landslides in large areas

Abstract

A method and device for early intelligent identification of large-area potential landslides according to the present invention belong to the technical fields of image target recognition and geological disaster monitoring. The method includes the steps of: collecting historical data of the landslide area, obtaining surface deformation rate and slope data, and forming a sample data set; constructing a potential landslide identification model composed of a data input layer, a feature encoding network, and an identification result output layer, wherein the feature encoding network is composed of 9 cascaded feature learning modules, and each feature learning module is composed of a network hidden layer and a data normalization layer; training and validating the potential landslide identification model to obtain a final potential landslide identification model; obtaining the measured data of the monitoring area; inputting the measured data into the final potential landslide identification model, and outputting the potential landslide identification result of the monitoring area. The present invention realizes the distinction between potential landslide candidate points and non-hazard points within a large area, and provides reliable technical support for landslide disaster prevention and control.

Description

Technical Field

[0001] The present invention relates to a method and device for early intelligent identification of large-area potential landslides, belonging to the technical fields of image target recognition and geological disaster monitoring. Background Technique

[0002] As a serious geological disaster, landslides have had a huge impact on the ecological environment and social economy. Therefore, early identification of landslides is crucial for the prevention, control, and treatment of landslide disasters. Monitoring the deformation of mountain bodies is the most direct and effective way in landslide disaster identification, early warning, and decision-making management. By monitoring the deformation of the slope body, its stability can be evaluated and its movement state can be reflected. Traditional landslide monitoring means include levels, GPS observation stations, crack meters, etc. However, traditional means can only meet the monitoring of proven landslide points within a local area, and there are problems such as low efficiency, being easily restricted by external conditions, and being unable to accurately reflect the overall change characteristics and trends of the landslide morphology at the regional scale, especially being powerless for areas with few human traces and high-position long-distance landslides.

[0003] Satellite synthetic aperture radar interferometry (InSAR) technology can overcome the above difficulties. The surface phase change observed by InSAR data can be used to invert the sliding deformation characteristics of the slope body, which is an effective means for early identification of large-area potential landslides. Using techniques such as persistent scatterer InSAR (PS-InSAR), small baseline subset InSAR (SBAS-InSAR), multi-master image coherent target small baseline InSAR (MCTSB-InSAR), or full scatterer InSAR (FS-InSAR) to invert the surface deformation information of differential interferometric phase, the results can intuitively reflect the deformation rate of ground objects and terrain in the observed scene. However, the satellite SAR image has a wide coverage range. The InSAR monitoring results not only contain the surface deformation information of the mountain body, but also include interference information such as ground settlement on flat ground and infrastructure deformation. It is difficult to distinguish them using the threshold method. In order to exclude this interference information, people usually use GIS software to overlay the InSAR monitoring results and optical images for manual interpretation and delineation of the landslide area, or use the vector of the delineated landslide area combined with optical images to further make a training sample set, and then train a certain deep learning image segmentation network to achieve the purpose of intelligent identification of the landslide area. This method is only applicable to the identification of post-disaster landslides and is difficult to be applied to the accurate identification of pre-disaster potential landslides because there are no obvious landslide sign characteristics on the pre-disaster optical images.

[0004] In summary, although InSAR technology is an effective means for observing large-area surface deformation, manual interpretation is still used to identify potential landslides from InSAR monitoring results, which is a time-consuming and laborious process. To reduce the cost of manual interpretation, the present invention provides a method and device for early intelligent identification of potential landslides in a large area. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a method and device for early intelligent identification of potential landslides in a large area, which can automatically distinguish potential landslide candidate points and non-hazard points within a large area, and efficiently and accurately identify potential landslides in a large area.

[0006] The technical solutions adopted by the present invention to solve its technical problems are as follows:

[0007] In a first aspect, a method for early intelligent identification of potential landslides in a large area provided by an embodiment of the present invention includes the following steps:

[0008] Step S1: Collect historical data of the landslide area. The historical data includes SAR satellite time series data and digital elevation model (DEM). Based on the FS-InSAR technology, obtain the surface deformation rate, use the digital elevation model (DEM) to invert the slope data, and combine the location information of the registered landslides for interpretation, so as to form a sample data set including a training sample set and a validation sample set;

[0009] Step S2: Construct a potential landslide identification model. The potential landslide identification model consists of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive the data combination (x1, x2) including the surface deformation rate x1 and the slope x2 as input data; the feature encoding network consists of 9 cascaded feature learning modules, and each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer is composed of m neurons, and each neuron adaptively learns the fusion features of the input data; the data normalization layer adopts a one-dimensional batch normalization operation with learnable parameters;

[0010] Step S3: Train and validate the potential landslide identification model based on the training sample set and the validation sample set to obtain the final potential landslide identification model;

[0011] Step S4: Collect SAR satellite time series data and digital elevation model (DEM) of the monitoring area, obtain the surface deformation rate based on the FS-InSAR technology, and calculate the slope data using the digital elevation model (DEM) to form measured data;

[0012] Step S5: Input the measured data into the final potential landslide identification model and output the potential landslide identification result of the monitoring area.

[0013] As a possible implementation of this embodiment, the output feature of the hidden layer of the network is expressed as:

[0014] ,

[0015] In the formula, is the ReLU activation function; Is the connection l The weight of the i-th neuron in the hidden layer and the j-th neuron in the l+1-th hidden layer, For the l+1 The bias of the jth neuron in the hidden layer; j, i represent the neuron indexes of the current hidden layer and the previous hidden layer, respectively, j = 1, 2…m l+1 , m l+1 For the l+1 The number of neurons in the hidden layer, i=1, 2…m l ;when l =1, 2…9, For the l Hidden layer i The output features of neurons, m l For the l The number of neurons in the hidden layer; l =0, Represents the input data x of the data input layer, namely the surface deformation rate and slope. At this time, m l is the dimension of the input data, i.e. m l =2.

[0016] As a possible implementation of this embodiment, the data normalization layer performs one-dimensional batch normalization processing on the data features output by the hidden layer, and its output value is expressed as:

[0017] ,

[0018] ,

[0019] In the formula, is the output value of batch normalization for the jth feature of the kth sample, γ j and β j is the parameter to be learned; k represents the sample index, k =1,2,…, K ; μ j Indicates j Features about the number of samples K The average value of ; Indicatesj the variance of a feature, i.e., ; ε is a small constant to prevent the denominator from being zero.

[0020] As a possible implementation of this embodiment, the recognition result output layer performs one-dimensional batch normalization on the output features of the last hidden layer, then performs linear weighted summation in the output layer, and uses the Sigmoid activation function to calculate the probability of the summation result. The calculated value y of the output layer is expressed as:

[0021] ,

[0022] In the formula, is the Sigmoid activation function, is the weight matrix of the output layer, M L is the number of neurons in the 9th hidden layer; b (L) is the bias vector of the 9th hidden layer.

[0023] As a possible implementation of this embodiment, the sample data set is generated by using the surface deformation rate generated by the FS-InSAR technology and the slope data of the corresponding survey area, and is discriminated in combination with the existing on-record landslide hidden danger point database. The sample data set consists of a certain number of positive samples (landslide hidden danger points) and negative samples (non-landslide hidden danger points). The label of the positive sample is represented by 1, and the label of the negative sample is represented by 0.

[0024] As a possible implementation of this embodiment, the training steps of the potential landslide recognition model include:

[0025] Set the model hyperparameters and initialize the network parameters of each neuron in each hidden layer;

[0026] Train the potential landslide recognition model using the mini-batch sample supervised learning mechanism and iteratively update the network parameters;

[0027] Among them, the model hyperparameters include the number of hidden layers and the number of neurons in each hidden layer. Set the number of hidden layers to 9, and the number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16, and 8 in sequence; the network parameters refer to weights and biases. The weights are initialized as random numbers that follow a normal distribution with a mean of 0 and a variance of, and the biases are initialized as 0.

[0028] As a possible implementation of this embodiment, the specific process of training using the mini-batch sample supervised learning mechanism includes:

[0029] Set training termination conditions, including the total number of training cycles, the error threshold of the verification recognition accuracy rate, and the tolerance, and set the initial learning rate and the number of samples in a mini-batch input;

[0030] In each cycle, use the training sample set in batches, and use the Adam optimizer and the learning rate adaptive update strategy to iteratively train the model;

[0031] Use the binary cross-entropy function as the target loss function to calculate the error between the potential landslides recognized by the model and the sample labels after each training cycle;

[0032] Use the verification sample set to calculate the verification accuracy rate of the model in recognizing potential landslides after each training cycle;

[0033] Judge whether the training cycle has reached the preset total number of cycles or the error threshold of the verification accuracy rate. If so, save the trained model; otherwise, continue training.

[0034] As a possible implementation manner of this embodiment, the learning rate adaptive update strategy is to judge whether the verification accuracy rate of the model in each training cycle is improved compared with the previous training cycle. If not, increment the count by 1; otherwise, set the count to 0. If the count reaches the preset tolerance, update the learning rate.

[0035] As a possible implementation manner of this embodiment, the measured data is a vector composed of InSAR surface deformation rate and slope data. The potential landslide recognition result is to use the trained model to recognize the measured data, and the output result is 0 or 1 (1 represents a potential landslide, 0 represents a non-potential landslide).

[0036] In a second aspect, an apparatus for early intelligent recognition of large-area potential landslides provided by an embodiment of the present invention includes:

[0037] A data set construction module for collecting historical data of the landslide area. The historical data includes SAR satellite time series data and digital elevation model DEM, obtaining the surface deformation rate based on the FS-InSAR technology, inversely calculating the slope data using the digital elevation model DEM, and combining the position information of the registered landslides for interpretation to form a sample data set including a training sample set and a verification sample set;

[0038] The model construction module is used to construct a potential landslide identification model. The potential landslide identification model consists of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive the data combination (x1, x2) containing the surface deformation rate x1 and the slope x2 as input data. The feature encoding network consists of 9 cascaded feature learning modules. Each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer consists of m neurons, and each neuron adaptively learns the fusion features of the input data. The data normalization layer uses a one-dimensional batch normalization operation with learnable parameters.

[0039] The model training module is used to train and validate the potential landslide identification model based on the training sample set and the validation sample set to obtain the final potential landslide identification model.

[0040] The measured data acquisition module is used to collect the SAR satellite time series data and the digital elevation model DEM of the monitoring area, obtain the surface deformation rate based on the FS-InSAR technology, and calculate the slope data using the digital elevation model DEM to form the measured data.

[0041] The model identification module is used to input the measured data into the final potential landslide identification model and output the potential landslide identification result of the monitoring area.

[0042] The beneficial effects of the technical solution of the embodiment of the present invention are as follows:

[0043] The present invention constructs a lightweight potential landslide identification model LSNet, comprehensively utilizes the InSAR surface deformation rate and slope information, autonomously learns the multi-dimensional difference features between potential landslide candidate points and other interference targets (such as ground settlement, buildings, infrastructure deformation, etc.), realizes the automatic distinction between potential landslide candidate points and non-hazard points in a large area, efficiently and accurately identifies potential landslides in a large area, reduces the manual interpretation cost, and improves the prevention and control efficiency of landslide disasters. The present invention is applicable to the early identification of large-area potential landslides and provides a reliable technical support for landslide disaster prevention and control. Description of the Drawings

[0044] Figure 1 is a flowchart of a method for early intelligent identification of large-area potential landslides shown according to an exemplary embodiment;

[0045] Figure 2 is a schematic structural diagram of a device for early intelligent identification of large-area potential landslides shown according to an exemplary embodiment;

[0046] Figure 3 is a network structure diagram of a potential landslide identification model shown according to an exemplary embodiment;

[0047] Figure 4 This is the specific implementation flowchart for the early intelligent identification of large - area potential landslides in the present invention. Specific implementation manner

[0048] To more clearly illustrate the technical features of the solution of the present invention, the present invention will be elaborated in detail below through specific implementation manners and in combination with its attached drawings.

[0049] As Figure 1 shown, a method for early intelligent identification of large - area potential landslides provided by an embodiment of the present invention includes the following steps:

[0050] Step S1: Collect historical data of the landslide area. The historical data includes SAR satellite time - series data and digital elevation model (DEM). Based on the FS - InSAR technology, obtain the surface deformation rate, use the digital elevation model DEM to invert the slope data, and combine with the location information of the registered landslides for interpretation, so as to form a sample data set including a training sample set and a verification sample set;

[0051] Step S2: Build a potential landslide identification model. The potential landslide identification model consists of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive the data combination (x1, x2) including the surface deformation rate x1 and the slope x2 as input data; the feature encoding network consists of 9 cascaded feature learning modules. Each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer consists of m neurons, and each neuron adaptively learns the fusion features of the input data; the data normalization layer adopts a one - dimensional batch normalization operation with learnable parameters;

[0052] Step S3: Train and verify the potential landslide identification model based on the training sample set and the verification sample set to obtain the final potential landslide identification model;

[0053] Step S4: Collect SAR satellite time - series data and digital elevation model DEM of the monitoring area, obtain the surface deformation rate based on the FS - InSAR technology, and calculate the slope data using the digital elevation model DEM to form measured data;

[0054] Step S5: Input the measured data into the final potential landslide identification model and output the potential landslide identification result of the monitoring area.

[0055] The present invention comprehensively utilizes InSAR surface deformation rate and slope information to improve the accuracy of potential landslide identification; by constructing a lightweight potential landslide identification model, it realizes the automatic distinction between potential landslide candidate points and non-hazard points in a large area, improves the accuracy and efficiency of potential landslide identification, and reduces the manual interpretation cost; adding a one-dimensional Batch normalization operation between each hidden layer can adaptively learn the fusion features of the input data, improving the stability and generalization ability of the potential landslide identification model; during the training process, a small-batch sample supervised learning mechanism and an adaptive learning rate adjustment strategy are adopted to improve the convergence speed and recognition accuracy of the model.

[0056] As a possible implementation manner of this embodiment, the output feature of the network hidden layer in the feature learning module is expressed as:

[0057] ,

[0058] In the formula, is the ReLU activation function; is the weight connecting the i-th neuron in the l-th l hidden layer and the j-th neuron in the l + 1 hidden layer, is the bias of the j-th neuron in the l+1 hidden layer; j and i respectively represent the neuron indices of the current hidden layer and the previous hidden layer, j = 1, 2... m l+1 , m l+1 is the number of neurons in the l+1 hidden layer, i = 1, 2... m l ; when l = 1, 2... 9, is the output feature of the l -th neuron in the i hidden layer, m l is the number of neurons in the l hidden layer; when l = 0, represents the input data x of the data input layer, that is, the surface deformation rate and slope. At this time, m l is the dimension of the input data, that is, m l = 2.

[0059] As a possible implementation manner of this embodiment, the data normalization layer performs one-dimensional batch normalization processing on the data features output by the hidden layer, making the statistical distribution of the input data of the subsequent hidden layer more stable, and its output value is expressed as:

[0060] ,

[0061] ,

[0062] In the formula, is the output value after batch normalization of the j-th feature of the k-th sample, γ j and β j are parameters to be learned; the subscript k represents the sample index, k = 1, 2, …, K ; μ j represents the average value of the j -th feature with respect to the number of samples K , that is, ; represents the variance of the j -th feature, that is, ; ε is a small constant to prevent the denominator from being zero.

[0063] As a possible implementation of this embodiment, after the recognition result output layer performs one-dimensional batch normalization (1D-Batch normalization) on the output features of the last hidden layer, linear weighted summation is performed in the output layer, and the Sigmoid activation function is used to calculate the probability of the summation result, that is, the probability that each group of input data ( x 1, x 2) belongs to the potential landslide candidate point. The calculated value y of the output layer is expressed as:

[0064] ,

[0065] In the formula, is the Sigmoid activation function, is the weight matrix of the output layer, M L is the number of neurons in the 9th hidden layer; b (L) is the bias vector of the 9th hidden layer.

[0066] As a possible implementation of this embodiment, the sample data set is generated by discriminating the surface deformation rate generated by the FS-InSAR technology and the slope data of the corresponding survey area, in combination with the existing database of landslide hidden danger points. The training sample set and the validation sample set are divided according to a 7:3 ratio. The training sample set is used to supervise the training of the potential landslide recognition model, and the validation sample set is used to evaluate the potential landslide recognition accuracy of the potential landslide recognition model after each training cycle; the sample data set consists of a certain number of positive samples (landslide hidden danger points) and negative samples (non-landslide hidden danger points). The label of the positive sample is represented by 1, and the label of the negative sample is represented by 0.

[0067] As a possible implementation of this embodiment, the training steps of the potential landslide recognition model include:

[0068] Set the model hyperparameters and initialize the network parameters of each neuron in each hidden layer;

[0069] Train the potential landslide recognition model using a mini-batch sample supervised learning mechanism, and iteratively update the network parameters;

[0070] Among them, the model hyperparameters include the number of hidden layers and the number of neurons in each hidden layer. Set the number of hidden layers to 9, and the number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16, and 8 in sequence; the network parameters refer to weights and biases. The weights are initialized as random numbers obeying a normal distribution with a mean of 0 and a variance of, and the biases are initialized to 0.

[0071] As a possible implementation of this embodiment, the specific process of training using the mini-batch sample supervised learning mechanism includes:

[0072] Set the training termination conditions, including the total number of training epochs (such as 200), the verification recognition accuracy error threshold (such as 10 -6 ), and the tolerance (such as 20), and set the initial learning rate lr (default value is 0.001) and the number of samples in the mini-batch input k ;

[0073] In each epoch, use the training sample set in batches, and perform iterative training on the model using the Adam optimizer and the learning rate adaptive update strategy; the number of batches in each epoch is calculated according to the integer relationship of the ratio between the number of training samples and the number of samples in the mini-batch input, that is, T = [Q1 / k], where Q1 is the number of samples in the training sample set, and [ ] represents taking the value.

[0074] Use the binary cross-entropy function as the target loss function, and calculate the error between the potential landslides recognized by the model and the sample labels after each training epoch;

[0075] Use the validation sample set to calculate the validation accuracy of the potential landslides recognized by the model after each training epoch;

[0076] Judge whether the training epoch reaches the preset total number of epochs or the error threshold of the validation accuracy. If it reaches, save the trained model; otherwise, continue training.

[0077] As a possible implementation of this embodiment, the learning rate adaptive update strategy is to judge whether the validation accuracy of the model in each training epoch is improved compared with the previous training epoch. If it is not improved, the count is incremented by 1; otherwise, the count is set to 0. If the count reaches the preset tolerance, the learning rate is updated , continue to train the LSNet network model.

[0078] As a possible implementation manner of this embodiment, the measured data is a batch or a group of vectors composed of InSAR surface deformation rate and slope data. The potential landslide identification result is to identify the measured data using the trained model, and output 0 or 1 (1 represents a potential landslide, and 0 represents a non-potential landslide).

[0079] As Figure 2 shown, an apparatus for early intelligent identification of large-area potential landslides provided by an embodiment of the present invention includes:

[0080] A dataset construction module, configured to collect historical data of a landslide area. The historical data includes SAR satellite time series data and digital elevation model DEM. Based on the FS-InSAR technology, obtain the surface deformation rate, use the digital elevation model DEM to invert the slope data, and combine the position information of the registered landslides for interpretation and interpretation to form a sample dataset including a training sample set and a validation sample set;

[0081] A model construction module, configured to construct a potential landslide identification model. The potential landslide identification model is composed of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive a data combination (x1, x2) including the surface deformation rate x1 and the slope x2 as input data; the feature encoding network is composed of 9 cascaded feature learning modules, and each feature learning module is composed of a network hidden layer and a data normalization layer. The network hidden layer is composed of m neurons, and each neuron adaptively learns the fusion features of the input data; the data normalization layer adopts a parameter-learnable one-dimensional batch normalization operation;

[0082] A model training module, configured to train and validate the potential landslide identification model based on the training sample set and the validation sample set to obtain a final potential landslide identification model;

[0083] A measured data acquisition module, configured to collect SAR satellite time series data and digital elevation model DEM of a monitoring area, obtain the surface deformation rate based on the FS-InSAR technology, and calculate the slope data using the digital elevation model DEM to form measured data;

[0084] A model identification module, configured to input the measured data into the final potential landslide identification model and output the potential landslide identification result of the monitoring area.

[0085] The present invention comprehensively utilizes InSAR deformation rate and slope information to construct a neural network model LSNet for identifying potential landslide candidate points, which can automatically distinguish potential landslide deformations from deformations of other targets in the InSAR deformation monitoring results.

[0086] As Figure 3 shown, the LSNet network is a simple classification network model, a fully connected network composed of 11 network layers, including 1 input layer, 9 hidden layers and 1 output layer, as shown in Figure 1. The input layer contains a data combination of a set of InSAR deformation rates x 1 and slopes x 2, that is, ( x 1, x 2). The number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16 and 8 in sequence. For ease of description, assume that each hidden layer is composed of M neurons, and each neuron adaptively learns the fusion features of the input data, that is, the excitation of the weighted combination of the input data. If the weight from the input layer to the first hidden layer is defined as W (1) , its dimension is M × N (here N = 2, the number of input data features), and the bias vector is , then for the output feature j of the j -th neuron ( l = 1, 2,..., m l ) in the first hidden layer is expressed as:

[0087] (1),

[0088] wherein, is the ReLU activation function.

[0089] Multiple neurons can be used to learn various features of the input data. Through the deep and orderly cascading of multiple hidden layers, higher-level abstract semantic features can be generated, which helps the network capture the deep difference features in the input data that can effectively reflect potential landslide candidate points and other targets. Then, for the l -th neuron in the j -th hidden layer ( j = 1, 2,..., m l + 1), its output feature is , that is:

[0090] (2),

[0091] In formula (2), is the ReLU activation function; is the weight connecting the l ith neuron in the th hidden layer to the l+1 jth neuron in the l+1 th hidden layer; j and i represent the neuron indices of the current hidden layer and the previous hidden layer respectively, j = 1, 2, …, m l+1 is the l+1 number of neurons in the l th hidden layer, i = 1, 2, …, m l = 1, 2, …, 9, is the l output feature of the i th neuron in the l th hidden layer, m l is the l number of neurons in the th hidden layer; when l = 0, l represents the input data x of the data input layer, i.e., the surface deformation rate and slope. At this time, m

[0092] Generally, the slope value is always greater than 0, and its value range is [0, 90], while the surface deformation rate values detected by InSAR are positive and negative (negative values indicate surface subsidence; positive values indicate surface uplift). Thus, there are usually obvious differences in the values of InSAR surface deformation rate and slope. Then, when using the sample set composed of these two types of data for network training, there are usually obvious differences in the statistical distributions of the output features of each neuron in the hidden layer. Moreover, the output of the previous hidden layer is also the input of the next hidden layer, that is to say, there will also be obvious differences in the statistical distributions of the input data of the next hidden layer, which easily leads to instability in the update of the model parameters (i.e., weights and biases) during network training and makes it difficult to converge to the global optimal solution. This is because during the network training process, the model is always adapting to different distributions and is prone to falling into a local optimal solution. On the other hand, the network model will overfit these inconsistent distribution features, thus reducing its prediction ability for unseen data, that is, reducing the generalization ability of the network model. To solve the above problems, we add a one-dimensional Batch normalization operation with learnable parameters between each hidden layer. Batch normalization adaptively normalizes the data features output by each hidden layer, making the statistical distributions of the input data of each hidden layer more stable. Assume that when training in batches, the input k samples ( x 1, x 2), the hidden layer will generate dimensions ofk × m If the output feature is adopted, then one-dimensional Batch normalization is used to perform batch normalization on each feature j (that is, each feature in m dimensions), and its output value is expressed as:

[0093] (3),

[0094] (4),

[0095] In the formula, γ j and β j are parameters to be learned; the subscript k represents the sample index, k = 1, 2, …, K ; μ j represents j the average value of the K th feature with respect to the number of samples ; represents j the variance of the th feature, that is

[0096] Finally, after performing one-dimensional Batch normalization on the output features of the last hidden layer (that is, the 9th hidden layer), linear weighted summation is performed in the output layer, and the Sigmoid activation function is used to calculate the probability of the summation result, that is, the probability that each group of input data ( x 1, x 2) belongs to the potential landslide candidate point. Assume that the last hidden layer contains M L neurons, and the weight matrix from it to the output layer (a single neuron) is W (L) , and its dimension is M L × q (here q = 1), and the bias is b (L) , then the calculated value y of the output layer is:

[0097] (5),

[0098] In the formula, is the Sigmoid activation function.

[0099] For example Figure 4As shown in the figure, in the specific implementation process, the specific implementation process of early intelligent identification of large-area potential landslides is as follows.

[0100] 1. Construct a sample data set.

[0101] The sample data set is generated by using the surface deformation rate generated by the FS-InSAR technology and the slope data of the corresponding survey area, combined with the existing database of landslide hidden danger points on the register. The sample data set consists of a certain number of positive samples (landslide hidden danger points) and negative samples (non-landslide hidden danger points). The label of the positive sample is represented by 1, and the label of the negative sample is represented by 0.

[0102] Based on the FS-InSAR technology, the surface deformation rate is obtained from the historical time series data of SAR satellite remote sensing in the landslide area, and the slope data is obtained from the terrain data of the landslide area by using the digital elevation model DEM. The sample data set is composed of the surface deformation rate, slope and their corresponding labels. The sample data set is divided into a training sample set and a verification sample set according to the ratio of 7:3. The training sample set is used to supervise the training of the potential landslide identification model, and the verification sample set is used to evaluate the potential landslide identification accuracy of the potential landslide identification model after each training cycle.

[0103] 2. Construct the LSNet network model.

[0104] The constructed LSNet network model for early intelligent identification of large-area potential landslides is as Figure 3 shown; the potential landslide identification model consists of a data input layer, a feature encoding network and an identification result output layer. The data input layer is used to receive the data combination containing the surface deformation rate and slope as input data; the feature encoding network consists of 9 cascaded feature learning modules, and each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer consists of m neurons, and each neuron adaptively learns the fusion features of the input data; the data normalization layer adopts a one-dimensional batch normalization operation with learnable parameters.

[0105] 3. Train the LSNet network model.

[0106] First, set the hyperparameters of the LSNet neural network model and initialize the network parameters of each neuron in each hidden layer; the hyperparameters of the LSNet neural network model refer to the number of hidden layers and the number of neurons in each hidden layer. Set the number of hidden layers to 9, and the number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16 and 8 in turn. The network parameters refer to the weights and biases in the LSNet neural network. The weight W is initialized as a random number obeying a normal distribution with a mean of 0 and a variance of and the bias b is initialized to 0.

[0107] Then, a small-batch sample supervised learning mechanism is adopted to train the LSNet neural network, and the network parameters are iteratively updated. The specific process includes:

[0108] 31. Set the training termination conditions (i.e., the total number of training cycles (such as 200), the verification recognition accuracy error threshold is 10 -6 ), the tolerance is 20, and the initial learning rate of the network parameters lr is defaulted to 0.001, and the number of samples in the small-batch input k ;

[0109] 32. Using the training sample set, in each cycle, the LSNet neural network is iteratively trained in batches using the Adam optimizer and the learning rate adaptive update strategy, and the network parameters in the LSNet neural network are continuously learned and updated. The number of batches in each cycle is calculated according to the rounding relationship between the number of training samples and the number of samples in the small-batch input, that is, T = [Q1 / k], where Q1 is the number of training samples, and [ ] represents the value taken; the learning rate adaptive update strategy is to judge whether the verification accuracy of the LSNet network model in each training cycle is improved compared with the previous training cycle. If not, the count is incremented by 1; otherwise, the count is set to 0. If the count reaches the preset tolerance, the learning rate is updated and the LSNet network model is continued to be trained;

[0110] 33. Adopt the binary cross-entropy function as the network target loss function to measure the error between the potential landslides identified by the LSNet network model and the sample labels after each training cycle ends;

[0111] 34. Using the verification sample set, calculate the verification accuracy of the LSNet network model in identifying potential landslides after each training cycle ends;

[0112] 35. Judge whether the network training cycle reaches the preset total number of cycles or the error threshold of the verification accuracy. If so, exit and save the trained LSNet network model; otherwise, repeat steps 32 - 34.

[0113] IV. Obtain the measured data and input it into the trained LSNet network model.

[0114] Collect the SAR satellite remote sensing time series data and terrain data of the monitoring area, obtain the surface deformation rate based on the FS-InSAR technology, calculate the slope data using the digital elevation model DEM, and obtain the measured data of the monitoring area; the measured data is a batch or a group of vectors composed of InSAR surface deformation rate and slope data.

[0115] V. Output the potential landslide recognition result.

[0116] The measured data is input into the trained LSNet neural network model. The LSNet neural network model processes the measured data and outputs the potential landslide identification results corresponding to each measured data, where the results are 0 or 1 (1 indicates potential landslide, and 0 indicates non-potential landslide). In specific implementation, the output results can also be presented to the user in a visual form or stored in a database for subsequent analysis and processing.

[0117] The present invention comprehensively utilizes InSAR surface deformation rate and slope information, improving the accuracy of potential landslide identification; by constructing a lightweight potential landslide identification model, it realizes the automatic distinction between potential landslide candidate points and non-hazard points within a large area, improving the accuracy and efficiency of potential landslide identification and reducing the manual interpretation cost; adding one-dimensional Batch normalization operations between each hidden layer can adaptively learn the fusion features of input data, improving the stability and generalization ability of the potential landslide identification model; during the training process, a mini-batch sample supervised learning mechanism and an adaptive learning rate adjustment strategy are adopted, improving the convergence speed and identification accuracy of the model.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that it is still possible to modify the specific implementation manners of the present invention or make equivalent replacements. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention shall be covered by the protection scope of the claims of the present invention.

Claims

1. A method for early intelligent identification of potential large-area landslides, characterized in that It includes the following steps: Step S1, collect historical data of the landslide area. The historical data includes SAR satellite time series data and digital elevation model (DEM). Obtain the surface deformation rate based on the FS-InSAR technology, invert the slope data using the digital elevation model (DEM), and combine with the location information of the registered landslides for interpretation, constituting a sample data set including a training sample set and a validation sample set; Step S2, construct a potential landslide identification model. The potential landslide identification model consists of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive a data combination including the surface deformation rate and slope as input data; The feature encoding network consists of 9 cascaded feature learning modules. Each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer consists of m neurons, and each neuron adaptively learns the fusion features of the input data; The data normalization layer adopts a one-dimensional batch normalization operation with learnable parameters; Step S3, train and validate the potential landslide identification model based on the training sample set and the validation sample set to obtain the final potential landslide identification model; Step S4, collect SAR satellite time series data and digital elevation model (DEM) of the monitoring area, obtain the surface deformation rate based on the FS-InSAR technology, and calculate the slope data using the digital elevation model (DEM) to form measured data; Step S5, input the measured data into the final potential landslide identification model to output the potential landslide identification result of the monitoring area; The output features of the network hidden layer are represented as: ， Wherein, is the ReLU activation function; is the weight connecting the l ith neuron in the lth hidden layer and the jth neuron in the (l + 1)th hidden layer, is the bias of the l+1 jth neuron in the lth hidden layer; j and i respectively represent the neuron indices of the current hidden layer and the previous hidden layer, j = 1, 2, …, m l+1 , m l+1 is the number of neurons in the l+1 lth hidden layer, i = 1, 2, …, m l ; When l = 1, 2…9, is the output feature of the l th neuron in the i hidden layer, and m l is the number of neurons in the l hidden layer; When l = 0, represents the input data x of the data input layer, i.e., the surface deformation rate and slope. At this time, m l is the dimension of the input data, i.e., m l = 2; the number of hidden layers is 9, and the number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16, and 8 in sequence.

2. The method for early intelligent identification of large-area potential landslides according to claim 1, wherein The data normalization layer performs one-dimensional batch normalization processing on the data features output by the hidden layer, and its output value is represented as: ， ， where is the output value of batch normalization for the j-th feature of the k-th sample, γ j and β j are the parameters to be learned; the subscript k represents the sample index, k = 1, 2, …, K ; μ j represents the mean of the j -th feature with respect to the number of samples K , that is ; represents the variance of the j -th feature, that is ; ε is a small constant to prevent the denominator from being zero.

3. The method for early intelligent identification of large-area potential landslides according to claim 1, characterized in that After the identification result output layer performs one-dimensional batch normalization processing on the output features of the last hidden layer, performs linear weighted summation in the output layer, and uses the Sigmoid activation function to calculate the probability of the summation result. The calculated value y of the output layer is represented as: ， In the formula, is the Sigmoid activation function, is the weight matrix of the output layer, M L is the number of neurons in the 9th hidden layer; b (L) Is the bias of the 9th hidden layer.

4. The method for early intelligent identification of potential large-area landslides according to claim 1, characterized in that, The sample data set is generated by discriminating the surface deformation rate generated by the FS-InSAR technology and the slope data of the corresponding measurement area, combined with the existing database of registered landslide hazard points. The sample data set consists of a certain number of positive samples and negative samples. The label of the positive sample is represented by 1, and the label of the negative sample is represented by 0.

5. The method for early intelligent identification of large-area potential landslides according to claim 1, wherein The training steps of the potential landslide identification model include: Set the model hyperparameters and initialize the network parameters of each neuron in each hidden layer; Train the potential landslide identification model using a mini-batch sample supervised learning mechanism and iteratively update the network parameters; Among them, the model hyperparameters include the number of hidden layers and the number of neurons in each hidden layer; the network parameters refer to weights and biases, where the weights are initialized as random numbers obeying a normal distribution with a mean of 0 and a variance of , and the biases are initialized as 0.

6. The method for early intelligent identification of large-area potential landslides according to claim 5, characterized in that, The specific process of training using the mini-batch sample supervised learning mechanism includes: Set the training termination conditions, including the total number of training epochs, the validation recognition accuracy error threshold, and the tolerance, set the initial learning rate and the number of samples in the mini-batch input; In each epoch, use the training sample set in batches, and use the Adam optimizer and the learning rate adaptive update strategy to perform iterative training on the model; Adopt the binary cross-entropy function as the target loss function to calculate the error between the potential landslides identified by the model and the sample labels after each training epoch; Using the validation sample set, calculate the validation accuracy rate of the model in identifying potential landslides after each training cycle; Judge whether the training cycle has reached the preset total number of cycles or the error threshold of the validation accuracy rate. If it has reached, save the trained model; otherwise, continue training.

7. The method for early intelligent identification of large-area potential landslides according to claim 6, characterized in that, The learning rate adaptive update strategy is to judge whether the validation accuracy rate of the model in each training cycle has increased compared with the previous training cycle. If it has not increased, increment the count by 1; otherwise, set the count to 0. If the count reaches the preset tolerance, update the learning rate.

8. The method for early intelligent identification of large-area potential landslides according to any one of claims 1-7, characterized in that, The measured data is a vector composed of InSAR surface deformation rate and slope data. The potential landslide identification result is obtained by using the trained model to identify the measured data, and the output result is potential landslide or non-potential landslide.

9. An apparatus for early intelligent identification of potential large-area landslides, characterized in that, It includes: A dataset construction module for collecting historical data of the landslide area. The historical data includes SAR satellite time series data and digital elevation model DEM. Based on the FS-InSAR technology, obtain the surface deformation rate, invert the slope data using the digital elevation model DEM, and combine the location information of the registered landslides for interpretation and interpretation to form a sample dataset containing a training sample set and a validation sample set; A model construction module for constructing a potential landslide identification model. The potential landslide identification model consists of a data input layer, a feature encoding network, and an identification result output layer. The data input layer is used to receive a data combination containing the surface deformation rate and slope as input data; The feature encoding network consists of 9 cascaded feature learning modules. Each feature learning module consists of a network hidden layer and a data normalization layer. The network hidden layer consists of m neurons, and each neuron adaptively learns the fusion features of the input data; The data normalization layer adopts a one-dimensional batch normalization operation with learnable parameters; A model training module for training and validating the potential landslide identification model based on the training sample set and the validation sample set to obtain the final potential landslide identification model; A measured data acquisition module for collecting SAR satellite time series data and digital elevation model DEM of the monitoring area, obtaining the surface deformation rate based on the FS-InSAR technology, and calculating the slope data using the digital elevation model DEM to form the measured data; A model identification module for inputting the measured data into the final potential landslide identification model and outputting the potential landslide identification result of the monitoring area; The output feature of the network hidden layer is expressed as: ， In the formula, is the ReLU activation function; is the weight connecting the l ith neuron in the lth hidden layer and the jth neuron in the (l + 1)th hidden layer, is the bias of the l+1 jth neuron in the lth hidden layer; j and i respectively represent the neuron indices of the current hidden layer and the previous hidden layer, j = 1, 2, …, m l+1 , m l+1 is the number of neurons in the l+1 lth hidden layer, i = 1, 2, …, m l ; When l = 1, 2, …, 9, is the output feature of the l th neuron in the i hidden layer, and m l is the number of neurons in the l hidden layer; When l = 0, represents the input data x of the data input layer, i.e., the surface deformation rate and slope. At this time, m l is the dimension of the input data, i.e., m l = 2; the number of hidden layers is 9, and the number of neurons in each hidden layer is 2048, 1024, 512, 256, 128, 64, 32, 16, and 8 in sequence.

Images