Landslide identification method and device based on generative adversarial network data augmentation strategy

Abstract

The application discloses a landslide identification method and device based on a generative adversarial network data expansion strategy, and comprises the following steps: acquiring a multi-source data set and performing pretreatment, extracting landslide disaster factors, and constructing a landslide sample library after pretreatment and landslide label production are performed on the landslide disaster factors; after the pretreated landslide disaster factors and landslide labels are superimposed in layers, one-dimensional array data are produced through an image cube, and the one-dimensional array data are divided into a training set and a test set; landslide data in the training set are input into a network in batches for training, minority class sample expansion is completed, and the expanded samples are added to the original training set to form a balanced data set; the balanced data set is input into a back-end machine learning classification model for training, an optimal model generated by training is used to test the test set, and a landslide prediction result is obtained. The application solves the problem of imbalance between classes in the data set, and can make the landslide identification boundary more clear and complete, and improve the precision of landslide target identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent geological disaster identification technology, and in particular to a landslide identification method and device based on a generative adversarial network data augmentation strategy. Background Technology

[0002] Landslides, due to their wide distribution and high frequency of occurrence, are considered one of the most serious geological hazards in the world today, causing enormous economic losses and casualties every year. For disaster prevention and mitigation efforts, compiling a landslide inventory is of paramount importance for landslide early warning, geomorphic erosion research, and subsequent landslide risk assessment and management. A landslide inventory includes various information such as the location, date, and movement type of the landslide. Currently, landslide inventory compilation is mainly based on remote sensing imagery, employing three main methods: visual interpretation, machine learning-based computer recognition, and deep learning-based intelligent recognition. In recent years, with the rapid development of computer technology, the use of real-time imagery acquired through remote sensing after disasters for automatic or semi-automatic landslide information extraction has become a new research hotspot.

[0003] The problem of landslide information extraction based on remote sensing images is essentially a binary classification problem, which is to classify the original pixels of the image or the segmented objects composed of similarity in color, texture, etc., into landslide or non-landslide categories. Therefore, landslide monitoring technology based on remote sensing images can be divided into two types: pixel-based and object-oriented. Object-oriented high-resolution images consider the geometric and texture features of the segmented objects when classifying landslides, and have achieved good research results to date. Research on pixel-based landslide identification methods has long focused on the following two aspects: (1) Landslide extraction method based on change detection of multiple images. Landslide information is extracted by comparing remote sensing images before and after a landslide event. This method is often used to monitor landslide events caused by triggering events such as earthquakes, rapid snow melting, and heavy rainfall, but it is less used for monitoring landslides that occur at different times. (2) Landslide extraction method based on single-scene images. This method extracts feature factors from remote sensing images and other basic data, and then combines machine learning and deep learning algorithms to classify pixels to extract landslide information.

[0004] Currently, the main machine learning algorithms used include: logistic regression, information content model, Bayesian model, support vector machine (SVM), random forest (RF), decision tree, and artificial neural network (ANN). In recent years, with the development of deep learning theory, convolutional neural networks (CNN) and their improved algorithms can further extract high-dimensional features from data through internal convolutional and pooling layers, and can efficiently train a large number of parameters through features such as weight sharing, avoiding overfitting problems. Therefore, this algorithm is widely used in landslide information extraction.

[0005] While widely adopted machine learning and deep learning algorithms have shown good performance in classifying, recognizing, and segmenting natural images, they are still affected by imbalanced samples when dealing with landslide problems. This is because, in landslide identification, researchers often encounter datasets where landslide samples represent only a small proportion of the total data in the study area, while the majority of samples are non-landslide areas. In this situation, the model tends to learn features from the majority class while neglecting the minority class, which limits its classification ability and reduces the accuracy of target recognition. Therefore, addressing the class imbalance problem in landslide datasets and further exploring the recognition potential of machine learning and deep learning models to improve the accuracy of landslide target recognition is an urgent problem to be solved. Summary of the Invention

[0006] One of the main problems addressed by this invention is the potential for imbalance between classes in the sample database and the complexity of the backend classification model to limit the model's learning ability, reduce classification efficiency, and decrease target recognition accuracy in current landslide identification and mapping applications. To solve this technical problem, this invention proposes a landslide identification method and apparatus based on a generative adversarial network data augmentation strategy.

[0007] According to one aspect of the present invention, a landslide identification method based on a generative adversarial network data augmentation strategy is provided, comprising the following steps:

[0008] After acquiring and preprocessing multi-source datasets, landslide-causing factors are extracted. After preprocessing the landslide-causing factors and creating landslide labels, a landslide sample library is constructed.

[0009] After overlaying the pre-processed landslide disaster-causing factors and landslide labels into layers, one-dimensional array data is created using image cubes, and the one-dimensional array data is divided into training set and test set.

[0010] The landslide data in the training set are input into the optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural network in batches for training, to complete the minority class sample augmentation, and the augmented samples are added to the original training set to form a balanced dataset.

[0011] The balanced dataset is input into the backend machine learning classification model for training. The optimal model generated by the training is then used to test the test set to obtain landslide prediction results.

[0012] Furthermore, the multi-source dataset includes: remote sensing images of the post-earthquake growing season in the study area, regional geological maps, global digital elevation model (DEM), and various geographic data; among which, the various geographic data include: regional climate and precipitation data, vegetation cover, and human engineering activities.

[0013] Furthermore, after acquiring and preprocessing the multi-source dataset, landslide-causing factors are extracted. After preprocessing the landslide-causing factors and creating landslide labels, a landslide sample library is constructed, including:

[0014] The multi-source data is unified in the remote sensing software ENVI5.3 according to the reference image resolution, projection coordinate system and geographic coordinate system. Radiometric calibration, atmospheric correction, image stitching and cropping operations are performed on the remote sensing image data to obtain the preprocessed multi-source data.

[0015] The preprocessed multi-source data was input into the ArcMap 10.6 platform to extract landslide hazard factors;

[0016] Input the landslide-causing factors into the R language platform to calculate the Pearson correlation index. If the correlation index between factors exceeds 0.7, it indicates that there is a strong correlation between the factors, and selective elimination is performed.

[0017] The original hazard-causing factors were input into SPSS for multicollinearity analysis to determine whether collinearity existed among the factors.

[0018] By combining the results of Pearson correlation index and collinearity analysis, factor elimination was performed to construct a clear landslide sample library without data redundancy.

[0019] Furthermore, the preprocessed multi-source data was input into the ArcMap 10.6 platform to extract landslide hazard factors, including:

[0020] Spectral factors and remote sensing index factors were extracted from each band of remote sensing images, including Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Tassel Change Greenness and Humidity.

[0021] Topographic and hydrological factors were extracted based on the global digital elevation model (DEM), including slope, aspect, plane curvature, profile curvature, elevation variation index, surface relief, topographic cutting depth, surface roughness, and distance from the river.

[0022] Geomorphic factors, including distance from faults and lithology, are extracted based on regional geological maps.

[0023] Auxiliary factors are extracted based on geographic information data, including average annual rainfall and human engineering activities.

[0024] Furthermore, after overlaying the preprocessed landslide causative factors and landslide labels into layers, a one-dimensional array of data is created using an image cube. This one-dimensional array data is then divided into a training set and a test set, including:

[0025] After preprocessing, the landslide hazard factors and landslide labels are input into ENVI 5.3. The image cube with an h×w×c layer overlay tool is used, where h is the length of each factor, w is the width, and c is the number of channels of the image cube. Then, the pixels at the same position of the image cube are extracted into one-dimensional array data. The one-dimensional array data is divided into training set X1 and test set T1 in a 6:4 ratio.

[0026] Furthermore, the landslide data in the training set is input in batches into the optimized generative adversarial network model Smo-SE-WGAN based on a convolutional neural network for training, to complete the augmentation of minority class samples, and the augmented samples are added to the original training set to form a balanced dataset, including:

[0027] The landslide data X from training set X1 will be processed in fixed batches. L The landslide features are one-dimensional array data generated from image cubes and input into the network. The batch size needs to be set according to the size of the landslide feature dimension, hardware computing power and network complexity.

[0028] Based on the characteristics of landslides, an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks is constructed, including: an imbalanced data sampling module, a Smote module, and a GAN module;

[0029] First, the landslide data is iteratively extracted from landslide and background samples by the unbalanced data sampling module. Then, the Smote module is used to balance the data and generate pseudo data by copying minority class samples. The GAN module includes a data generator G that simulates the distribution of real data, and a discriminator D that determines whether the data is real landslide data or pseudo data generated by the generator G.

[0030] The Wassertein distance, a probability distribution difference index, is used to construct a loss function to calculate the distance between the pseudo-data distribution generated by generator G and the real data distribution. The specific calculation formula is as follows:

[0031]

[0032] In the formula, Z represents the actual landslide data, and Z belongs to the distribution of actual landslide data. G(n) represents pseudo-data generated by the generator G from random noise, and G(n) belongs to the generated data distribution. γ is the joint distribution of the real data and the generated data, and E is the mean.

[0033] Based on the network batch and landslide feature dimensions, a noise matrix Noise of equal size following a normal distribution is constructed and input into the generator G in the GAN module. The generator is constructed with a convolutional neural network as the underlying network, and adopts an alternating structure of one-dimensional convolutional layers, activation functions, and batch normalization layers for the characteristics of landslide array data. For one-dimensional landslide array data, a channel attention mechanism is introduced into the generator. Then, random noise is subjected to convolutional feature extraction operations through the above structure to simulate the distribution pattern of real landslide feature data, thereby generating pseudo-data X with the same landslide features. W ;

[0034] The pseudo data X generated by the generator G above W and actual landslide data X L Both data are input into the discriminator D. The discriminator's underlying architecture uses alternating one-dimensional convolutional layers and max-pooling layers to learn the distribution of real and fake data. A softmax function is added to the bottom layer of the framework to distinguish between real and fake labels. Then, based on the loss value, the parameters of the generator G and the discriminator D are alternately updated using stochastic gradient descent until the maximum number of training iterations is completed, ending the generative adversarial network training and retaining the optimal parameters of the generator.

[0035] The noise matrix Noise, with the same dimension as the landslide feature, is input into the generator G, and the optimal parameters of the generator are loaded. Landslide augmentation data X is generated based on the difference between the number of landslide samples and non-landslide samples. k And add it to the original training set X1 to form a balanced dataset X P Used for model training.

[0036] Furthermore, the backend machine learning classification model is a support vector machine.

[0037] Furthermore, after the step of testing the test set using the optimal model generated during training to obtain landslide prediction results, the following steps are also included:

[0038] The landslide prediction results are compared with the landslide labels that have been validated in the field, and the accuracy index is calculated based on the confusion matrix to quantitatively evaluate the prediction performance of the optimal model.

[0039] Furthermore, the accuracy metrics include at least one of AUC, G-mean, F1 score, Kappa coefficient, and Matthews correlation coefficient.

[0040] According to another aspect of the present invention, a landslide identification device based on a generative adversarial network data augmentation strategy is provided, comprising the following modules:

[0041] The sample library construction module is used to acquire multi-source datasets and preprocess them, extract landslide disaster-causing factors, preprocess the landslide disaster-causing factors, create landslide labels, and then construct a landslide sample library.

[0042] The training set creation module is used to overlay the pre-processed landslide disaster factors and landslide labels into layers, and then create a one-dimensional array data through image cubes, dividing the one-dimensional array data into training set and test set.

[0043] The data augmentation module is used to input the landslide data in the training set into the optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural network in batches for training, complete the augmentation of minority class samples, and add the augmented samples to the original training set to form a balanced dataset.

[0044] The landslide prediction module is used to input the balanced dataset into the backend machine learning classification model for training, and then use the optimal model generated by training to test the test set to obtain the landslide prediction results.

[0045] The technical solution provided by this invention has the following beneficial effects:

[0046] First, this invention couples multi-source remote sensing data to create a multi-environment disaster factor sample library. Considering the imbalance between classes in the landslide sample library and the problems that the complex classification model may bring, such as limiting the model's learning ability, low classification efficiency, and low target recognition accuracy, this invention utilizes the adversarial idea of ​​generator and discriminator in the GAN algorithm to expand the few-class landslide samples. This can effectively solve the data imbalance problem in the original sample library and further explore the learning potential of the classification model.

[0047] Secondly, this invention constructs a relatively simple landslide back-end classification model, which further improves the accuracy and efficiency of landslide identification, enabling the trained model to better distinguish landslides from other land features, and greatly improves the accuracy of landslide identification results.

[0048] Third, in the network construction process, the present invention incorporates the Smote (Synthetic Minority Over-sampling Technique) module, which first copies the boundary data, thus solving the problem of unclear landslide boundary identification caused by ignoring boundary data during the data expansion process.

[0049] Fourth, this invention uses Wasstertain distance to effectively calculate the difference between true and false data, and adds a gradient penalty function to solve the gradient vanishing problem.

[0050] Finally, this invention introduces a channel attention mechanism into the generator for one-dimensional landslide array data. This module can effectively extract relevant information between each dimension in the landslide data during training. While the network learns the overall distribution of the data, it can also learn the distribution characteristics within each array data. Attached Figure Description

[0051] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings:

[0052] Figure 1 This is a flowchart illustrating the overall process of a landslide identification method based on a generative adversarial network data augmentation strategy according to the present invention.

[0053] Figure 2 This is a diagram illustrating the preparation of a landslide sample for this invention.

[0054] Figure 3 This is a schematic diagram of a landslide sample from the present invention.

[0055] Figure 4 This is a schematic diagram of the structure of a landslide identification device based on a generative adversarial network data augmentation strategy according to the present invention.

[0056] Figure 5 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0057] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0058] Generative Adversarial Networks (GANs) are machine learning architectures proposed by Ian Goodfellow in 2014. Since their inception, their novel network structure has garnered widespread attention, and the superior performance of their generative models has been widely applied in image, video, and audio processing. Essentially, this network analyzes the spatial distribution of real samples from a probabilistic perspective, constructing pseudo-data with a similar spatial distribution after inputting noise. A GAN consists of a generator and a discriminator, which model the input data through a clever relationship of competition and cooperation. The adversarial approach involves training one generative model and one discriminator model, allowing them to train against each other and continuously improve their capabilities. GANs can support various underlying models; in addressing the problem of generating one-dimensional array data for landslide identification, this invention uses a one-dimensional convolutional neural network as the underlying framework. Compared to other oversampling algorithms that are prone to marginalization, generative adversarial networks can effectively learn the overall distribution of feature samples and generate clearer and more realistic pseudo-samples. They can more deeply and effectively solve the problem of class imbalance in the dataset. Moreover, because they support a variety of simple underlying structures with high inclusiveness, they greatly save computational costs and are highly efficient in training.

[0059] Furthermore, the creation of a landslide sample database needs to take into account various disaster-prone environmental factors such as geology, geomorphology, human engineering activities, land classification, and hydrological conditions. Therefore, this invention couples multi-source remote sensing data during the data collection stage and incorporates landslide-causing factors such as topographic data, basic geological data, and hydrological and meteorological data during the sample creation stage to enrich the deeper characteristics of landslides for subsequent sample expansion and model training. This operation can, to some extent, solve the problem of scarce landslide samples and make landslide identification boundaries clearer and more complete, further improving the accuracy of landslide target identification.

[0060] Currently, in addressing landslide identification, the imbalance of landslide samples often limits the learning ability of classification models, leading to a decline in the final classification accuracy. Previous researchers have attempted to improve accuracy by constructing more complex backend classification models, such as increasing the depth and width of the learning network, but this often results in a significant decrease in landslide identification efficiency. To address this issue, this invention proposes a landslide identification method based on a generative adversarial network (GAN) data augmentation strategy, focusing on minority class sample augmentation. Figure 1 As shown, the specific implementation steps are as follows:

[0061] S1: After acquiring and preprocessing the multi-source dataset, extract the landslide-causing factors, preprocess the landslide-causing factors, create landslide labels, and then construct a landslide sample library.

[0062] This invention takes the post-earthquake landslide in the Jiuzhaigou Reservoir area as an example. Due to the significant differences in landslide area within the study area, and the predominance of small, shallow earthquake landslides, this invention focuses on data extraction by coupling multi-level, multi-scale remote sensing information. Specifically, remote sensing images of the study area during the 2018 post-earthquake growing season, regional geological maps, global digital elevation model (DEM), and other geographic data are collected as basic data for subsequent deep factor extraction. The geographic data includes information such as regional climate and precipitation data, vegetation cover, and human engineering activities. The multi-source data are then unified in the remote sensing software ENVI 5.3 according to the reference image resolution, projected coordinate system, and geographic coordinate system. Additional preprocessing work, such as radiometric calibration, atmospheric correction, image stitching, and cropping, is required for the remote sensing image data. Subsequently, the collected multi-source data is input into the ArcMap 10.6 platform for further extraction of landslide-causing factors.

[0063] Specifically, spectral factors and remote sensing index factors are extracted from each band of remote sensing imagery, such as Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and factors like tassel change greenness and humidity. Topographic and hydrological factors are extracted from the Global Digital Elevation Model (DEM), such as slope, aspect, plane curvature, profile curvature, elevation variability index, surface relief, topographic cutting depth, surface roughness, and distance from rivers. Geomorphic factors such as distance from faults and lithology are extracted from regional geological maps, and auxiliary factors such as annual average rainfall and human engineering activities are extracted from geographic information data. The extracted factors are then normalized and noise-removed.

[0064] Finally, the landslide-causing factors were input into the R language platform to calculate the Pearson Correlation Coefficient (PCC). In statistics, the Pearson Correlation Coefficient measures the correlation between variables X and Y, with a value range of [-1, 1]. The formula is as follows: a positive value indicates a positive correlation, and a negative value indicates a negative correlation. If there is no linear relationship between the two, the value is 0. When eliminating factors, the positive range is often considered; the closer the value is to 1, the stronger the linear relationship between the variables. In this invention, if the correlation coefficient between factors exceeds 0.7, it indicates a strong correlation between the factors, requiring selective elimination.

[0065]

[0066] In addition, the original disaster-causing factors are input into SPSS for multicollinearity analysis to determine whether there is collinearity between the factors, that is, the variance inflation factor (VIF) and tolerance (TOL) between the factors are calculated for the dataset. When performing factor elimination, it is usually judged according to VIF with 10 as the boundary. When VIF < 10, it is considered that the factor has no collinearity; when 10 < VIF < 100, the factor has strong collinearity; when VIF > 100, it indicates strong correlation. Subsequently, the Pearson correlation index and the results of collinearity analysis are combined to jointly perform the factor elimination work, that is, the union of the factors that meet the strong correlation of the two is eliminated to construct a clear landslide sample library without data redundancy.

[0067] S2: After overlaying the preprocessed landslide disaster-causing factors and landslide labels, one-dimensional array data is made through an image cube, and the one-dimensional array data is divided into a training set and a test set;

[0068] The preprocessed landslide disaster-causing factors and landslide labels are input into ENVI 5.3 and the layer stacking tool is used to form an image cube of h×w×c. h is the length of each factor, w is the width, and c is the number of channels of the image cube, which specifically refers to the number of landslide disaster-causing factors adopted in the present invention, as specifically shown Figure 2 as follows. Then, the pixels at the same position of the image cube are extracted as one-dimensional array data, and the specific operation is as follows Figure 3 as shown. According to the requirements of the minority sample amplification framework, the samples in the present invention are divided into a training set X1 and a test data set T1 in a ratio of 6:4. The data in the training set is mainly input into the network for expansion, and the data in the test set is used to evaluate the landslide recognition accuracy of this framework.

[0069] S3: The landslide data in the training set is input into the optimized generative adversarial network model Smo-SE-WGAN based on a convolutional neural network in batches for training to complete the amplification of minority class samples, and the amplified samples are added to the original training set to form a balanced data set;

[0070] 60% of the landslide data in the training set is input into the network in batches for training to complete the amplification of minority class samples. The specific training process is as follows:

[0071] (1) The landslide data X in the training set X1 is input into the network in fixed batches: The landslide features are one-dimensional array data generated by the image cube, and the batch quantity setting needs to be set according to the size of the landslide feature dimension, the hardware computing power, and the network complexity; L Input into the network: The landslide features are one-dimensional array data generated by the image cube, and the batch quantity setting needs to be set according to the size of the landslide feature dimension, the hardware computing power, and the network complexity;

[0072] (2) Constructing an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks for landslide characteristics: Based on landslide characteristics, an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks is constructed, including: an imbalanced data sampling module, a Shote (Synthetic Minority Over-sampling Technique, Shote) module, and a GAN (Generative Adversarial Network, GAN) module. First, landslide data is iteratively extracted from landslide and background samples through the imbalanced data sampling module. Then, the Shote module is used for data balancing, and pseudo-data is generated by copying minority class samples. During data augmentation, data at the distribution boundary is often ignored, leading to unclear landslide boundary identification. Therefore, this invention adds the Shote module to the network construction to solve this problem by first copying the boundary data. The GAN module includes a data generator G that simulates the real data distribution, and a discriminator D that determines whether the data is real landslide data or pseudo-data generated by the generator G.

[0073] (3) Constructing a loss function Loss to distinguish between real and fake data: The Wassertein distance, a probability distribution difference index, is used to construct a loss function to calculate the distance between the distribution of fake data generated by generator G and the distribution of real data. The specific calculation formula is as follows:

[0074]

[0075] In the formula, Z represents the actual landslide data, and Z belongs to the distribution of actual landslide data. G(n) represents pseudo-data generated by the generator G from random noise, and G(n) belongs to the generated data distribution. γ represents the joint distribution of real and generated data, and E is the average value. Traditional GANs often suffer from vanishing gradients when trained on landslide array data, and the original loss function cannot reasonably estimate the difference between pseudo-data and real data. This invention employs Wasstertain distance, which can effectively calculate the difference between real and pseudo data, and incorporates a gradient penalty function to solve the vanishing gradient problem.

[0076] (4) Generator G Training: A noise matrix Noise of equal size, following a normal distribution, is constructed based on the network batch and landslide feature dimensions. This matrix is ​​then input into the generator G in the GAN module. The generator is built using a convolutional neural network as the underlying network. For the characteristics of landslide array data, a structure alternating between one-dimensional convolutional layers (Convolution 1D), activation functions (ReLU), and batch normalization (BN) layers is employed. Furthermore, a channel attention mechanism is introduced into the generator for one-dimensional landslide array data. During training, this module can effectively extract relevant information between each dimension of the landslide data, allowing the network to learn the overall data distribution while simultaneously learning the distribution characteristics within each array. Then, random noise is subjected to convolutional feature extraction through the above structure to simulate the distribution pattern of real landslide feature data, thereby generating pseudo-data X with the same landslide features. W ;

[0077] (5) Discriminator D determines: The pseudo data X generated by the generator G is determined. W and actual landslide data X L Both data are input into the discriminator D. The discriminator's underlying architecture uses alternating one-dimensional convolutional layers and max-pooling layers to learn the distribution of real and fake data. A softmax function is added to the bottom layer of the framework to distinguish between real and fake labels. Then, based on the loss value, the parameters of the generator G and the discriminator D are alternately updated using stochastic gradient descent until the maximum number of training iterations is completed, ending the generative adversarial network training and retaining the optimal parameters of the generator.

[0078] The noise matrix Noise, with the same dimension as the landslide feature, is input into the generator G, and the optimal parameters of the generator are loaded. Landslide augmentation data X is generated based on the difference between the number of landslide samples and non-landslide samples. k And add it to the original training set X1 to form a balanced dataset X P Used for model training.

[0079] S4: Input the balanced dataset into the backend machine learning classification model for training, use the best model generated by training to test the test set, obtain landslide prediction results, compare the landslide prediction results with the landslide labels validated in the field, calculate the accuracy index based on the confusion matrix, and thus quantitatively evaluate the prediction performance of the best model.

[0080] The balanced dataset X constructed in step S3 PThe data is input into a backend machine learning classification model, such as a Support Vector Machine (SVM), and the optimal model generated during training is tested on a test set to obtain prediction results. Finally, the prediction results are compared with the labels of landslide samples validated in the field, and the accuracy index is calculated based on the confusion matrix (as shown in Table 1) to quantitatively evaluate the performance of the framework proposed in this invention.

[0081] Table 1. Confusion Matrix of Landslide Identification Results

[0082]

[0083] The following five evaluation indicators are commonly used:

[0084] 1) AUC (Area Under Curve): The area under the ROC curve, with a value ranging from 0 to 1. The closer the value is to 1, the better the recognition effect.

[0085] 2) G-mean represents the set square of the recall rates of positive and negative samples.

[0086]

[0087] 3) F1 score (F1-score) is the harmonic mean of precision and recall.

[0088]

[0089] 4) The Kappa coefficient is a metric used to measure the effectiveness of classification. The higher the value, the better the model performance.

[0090]

[0091]

[0092] Where p0 represents the overall accuracy.

[0093] 5) Matthews correlation coefficient (MCC): This is a correlation coefficient that describes the correlation between the actual classification and the predicted classification. The value ranges from -1 to 1, and a value of 1 indicates that the prediction effect is perfect.

[0094]

[0095] The following describes a landslide identification device based on a generative adversarial network data augmentation strategy provided by the present invention. The landslide identification device described below can be referred to in correspondence with the landslide identification method described above.

[0096] like Figure 4As shown, a landslide identification device based on a generative adversarial network (GAN) data augmentation strategy includes the following modules:

[0097] The sample library construction module 001 is used to obtain multi-source datasets and preprocess them, extract landslide disaster-causing factors, preprocess the landslide disaster-causing factors and create landslide labels to construct a landslide sample library.

[0098] The training set creation module 002 is used to overlay the pre-processed landslide disaster factors and landslide labels into layers, and then create a one-dimensional array data through an image cube, dividing the one-dimensional array data into a training set and a test set.

[0099] The data augmentation module 003 is used to input the landslide data in the training set into the optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural network in batches for training, complete the augmentation of minority class samples, and add the augmented samples to the original training set to form a balanced dataset.

[0100] The landslide prediction module 004 is used to input the balanced dataset into the backend machine learning classification model for training, and to use the optimal model generated by training to test the test set to obtain the landslide prediction results.

[0101] Based on, but not limited to, the above-mentioned devices, the multi-source dataset includes: remote sensing images of the post-earthquake growth season in the study area, regional geological maps, global digital elevation model (DEM), and various geographic data; among which, the various geographic data include: regional climate and precipitation data, vegetation cover, and human engineering activities.

[0102] Based on, but not limited to, the above-described apparatus, the sample library construction module 001 is specifically configured to perform the following steps:

[0103] The multi-source data is unified in the remote sensing software ENVI5.3 according to the reference image resolution, projection coordinate system and geographic coordinate system. Radiometric calibration, atmospheric correction, image stitching and cropping operations are performed on the remote sensing image data to obtain the preprocessed multi-source data.

[0104] The preprocessed multi-source data was input into the ArcMap 10.6 platform to extract landslide hazard factors;

[0105] Input the landslide-causing factors into the R language platform to calculate the Pearson correlation index. If the correlation index between factors exceeds 0.7, it indicates that there is a strong correlation between the factors, and selective elimination is performed.

[0106] The original hazard-causing factors were input into SPSS for multicollinearity analysis to determine whether collinearity existed among the factors.

[0107] By combining the results of Pearson correlation index and collinearity analysis, factor elimination was performed to construct a clear landslide sample library without data redundancy.

[0108] Based on, but not limited to, the above-described apparatus, the training set creation module 002 includes a hazard factor extraction unit, specifically used for:

[0109] Spectral factors and remote sensing index factors were extracted from each band of remote sensing images, including Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Tassel Change Greenness and Humidity.

[0110] Topographic and hydrological factors were extracted based on the global digital elevation model (DEM), including slope, aspect, plane curvature, profile curvature, elevation variation index, surface relief, topographic cutting depth, surface roughness, and distance from the river.

[0111] Geomorphic factors, including distance from faults and lithology, are extracted based on regional geological maps.

[0112] Auxiliary factors are extracted based on geographic information data, including average annual rainfall and human engineering activities.

[0113] The training set production module 002 also includes a dataset production unit, specifically used for:

[0114] After preprocessing, the landslide hazard factors and landslide labels are input into ENVI 5.3. The image cube with an h×w×c layer overlay tool is used, where h is the length of each factor, w is the width, and c is the number of channels of the image cube. Then, the pixels at the same position of the image cube are extracted into one-dimensional array data. The one-dimensional array data is divided into training set X1 and test set T1 in a 6:4 ratio.

[0115] Based on, but not limited to, the above-described apparatus, the data expansion module 003 is specifically configured to perform the following operations:

[0116] The landslide data X from training set X1 will be processed in fixed batches. L The landslide features are one-dimensional array data generated from image cubes and input into the network. The batch size needs to be set according to the size of the landslide feature dimension, hardware computing power and network complexity.

[0117] Based on the characteristics of landslides, an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks is constructed, including: an imbalanced data sampling module, a Smote module, and a GAN module;

[0118] First, the landslide data is iteratively extracted from landslide and background samples by the unbalanced data sampling module. Then, the Smote module is used to balance the data and generate pseudo data by copying minority class samples. The GAN module includes a data generator G that simulates the distribution of real data, and a discriminator D that determines whether the data is real landslide data or pseudo data generated by the generator G.

[0119] The Wassertein distance, a probability distribution difference index, is used to construct a loss function to calculate the distance between the pseudo-data distribution generated by generator G and the real data distribution. The specific calculation formula is as follows:

[0120]

[0121] In the formula, Z represents the actual landslide data, and Z belongs to the distribution of actual landslide data. G(n) represents pseudo-data generated by the generator G from random noise, and G(n) belongs to the generated data distribution. γ is the joint distribution of the real data and the generated data, and E is the mean.

[0122] Based on the network batch and landslide feature dimensions, a noise matrix Noise of equal size following a normal distribution is constructed and input into the generator G in the GAN module. The generator is constructed with a convolutional neural network as the underlying network, and adopts an alternating structure of one-dimensional convolutional layers, activation functions, and batch normalization layers for the characteristics of landslide array data. For one-dimensional landslide array data, a channel attention mechanism is introduced into the generator. Then, random noise is subjected to convolutional feature extraction operations through the above structure to simulate the distribution pattern of real landslide feature data, thereby generating pseudo-data X with the same landslide features. W ;

[0123] The pseudo data X generated by the generator G above W and actual landslide data X L Both data are input into the discriminator D. The discriminator's underlying architecture uses alternating one-dimensional convolutional layers and max-pooling layers to learn the distribution of real and fake data. A softmax function is added to the bottom layer of the framework to distinguish between real and fake labels. Then, based on the loss value, the parameters of the generator G and the discriminator D are alternately updated using stochastic gradient descent until the maximum number of training iterations is completed, ending the generative adversarial network training and retaining the optimal parameters of the generator.

[0124] The noise matrix Noise, with the same dimension as the landslide feature, is input into the generator G, and the optimal parameters of the generator are loaded. Landslide augmentation data X is generated based on the difference between the number of landslide samples and non-landslide samples. k And add it to the original training set X1 to form a balanced dataset X P Used for model training.

[0125] Based on, but not limited to, the above-described apparatus, the landslide prediction module 004 is specifically used to perform the following operations:

[0126] The balanced dataset is input into the backend machine learning classification model for training. The optimal model generated by the training is then used to test the test set to obtain landslide prediction results.

[0127] Furthermore, based on, but not limited to, the above-described apparatus, the landslide prediction module 004 also includes a quantitative assessment unit, specifically used to perform the following operations:

[0128] The landslide prediction results are compared with the landslide labels that have been validated in the field, and the accuracy index is calculated based on the confusion matrix to quantitatively evaluate the prediction performance of the optimal model.

[0129] As a preferred implementation, the backend machine learning classification model is a support vector machine, which has a simple structure, reduces the amount of computation, and significantly improves the efficiency of landslide identification.

[0130] As an optional implementation, the accuracy metrics include at least one of AUC, G-mean, F1 score, Kappa coefficient, and Matthews correlation coefficient.

[0131] In addition, to better implement the above methods, such as Figure 5 As shown in the figure, an embodiment of the present invention illustrates a physical structure diagram of an electronic device, which may include: a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other through the communication bus 640. The processor 610 can call the logic instructions in the memory 630 to execute the steps of the landslide identification method described above. Specifically, this includes acquiring and preprocessing a multi-source dataset, extracting landslide-causing factors, preprocessing the landslide-causing factors, creating landslide labels, and constructing a landslide sample library; overlaying the preprocessed landslide-causing factors and landslide labels into layers, creating a one-dimensional array data using image cubes, and dividing the one-dimensional array data into a training set and a test set; inputting the landslide data in the training set into batches into an optimized generative adversarial network model Smo-SE-WGAN based on a convolutional neural network for training, completing minority class sample expansion, and adding the expanded samples to the original training set to form a balanced dataset; inputting the balanced dataset into a backend machine learning classification model for training, using the optimal model generated by training to test the test set, and obtaining landslide prediction results.

[0132] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0133] Furthermore, to better implement the above method, this embodiment of the invention also provides a storage medium storing a computer program. When the computer program is executed by a processor, it implements the steps of the landslide identification method described above, specifically including: acquiring and preprocessing a multi-source dataset, extracting landslide-causing factors, preprocessing the landslide-causing factors, creating landslide labels, and constructing a landslide sample library; overlaying the preprocessed landslide-causing factors and landslide labels into layers, creating a one-dimensional array data using image cubes, and dividing the one-dimensional array data into a training set and a test set; inputting the landslide data in the training set into a batch of an optimized generative adversarial network model Smo-SE-WGAN based on a convolutional neural network for training, completing the minority class sample expansion, and adding the expanded samples to the original training set to form a balanced dataset; inputting the balanced dataset into a backend machine learning classification model for training, using the optimal model generated by the training to test the test set, and obtaining landslide prediction results.

[0134] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0135] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In the unit claims listing several devices, several of these devices may be embodied by the same hardware item. The use of the terms first, second, and third, etc., does not indicate any order and can be interpreted as identifiers.

[0136] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A landslide identification method based on a generative adversarial network (GAN) data augmentation strategy, characterized in that, Includes the following steps: After acquiring and preprocessing multi-source datasets, landslide-causing factors are extracted. After preprocessing the landslide-causing factors and creating landslide labels, a landslide sample library is constructed. After overlaying the pre-processed landslide disaster-causing factors and landslide labels into layers, one-dimensional array data is created using image cubes, and the one-dimensional array data is divided into training set and test set. The landslide data in the training set were input in batches into the optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks for training, completing the minority class sample augmentation. The augmented samples were then added to the original training set to form a balanced dataset, including: The landslide data X from training set X1 will be processed in fixed batches. L The landslide features are one-dimensional array data generated from image cubes and input into the network. The batch size needs to be set according to the size of the landslide feature dimension, hardware computing power and network complexity. Based on the characteristics of landslides, an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks is constructed, including: an imbalanced data sampling module, a Smote module, and a GAN module; First, the landslide data is iteratively extracted from landslide and background samples by the unbalanced data sampling module. Then, the Smote module is used to balance the data and generate pseudo data by copying minority class samples. The GAN module includes a data generator G that simulates the distribution of real data, and a discriminator D that determines whether the data is real landslide data or pseudo data generated by the generator G. The Wassertein distance, a probability distribution difference index, is used to construct a loss function to calculate the distance between the pseudo-data distribution generated by generator G and the real data distribution. The specific calculation formula is as follows: In the formula, Z represents the actual landslide data, and Z belongs to the distribution of actual landslide data. G(n) is pseudo-data generated by a generator G from random noise, and G(n) belongs to the generated data distribution G( ); Let E be the joint distribution of the real data and the generated data, and E be the mean. Based on the network batch and landslide feature dimensions, a noise matrix of equal size following a normal distribution is constructed. This matrix is ​​then input into the generator G in the GAN module. The generator is constructed using a convolutional neural network as the underlying network. For the characteristics of landslide array data, a one-dimensional convolutional layer, activation function, and batch normalization layer are alternately arranged. For one-dimensional landslide array data, a channel attention mechanism is introduced into the generator. Then, random noise is subjected to convolutional feature extraction through the aforementioned alternate structure to simulate the distribution of real landslide feature data, thereby generating pseudo-data X with the same landslide features. W ; The pseudo data X generated by the generator G above W and actual landslide data X L Both data are input into the discriminator D. The discriminator's underlying architecture uses alternating one-dimensional convolutional layers and max-pooling layers to learn the distribution of real and fake data. A softmax function is added to the bottom layer of the framework to distinguish between real and fake labels. Then, based on the loss value, the parameters of the generator G and the discriminator D are alternately updated using stochastic gradient descent until the maximum number of training iterations is completed, ending the generative adversarial network training and retaining the optimal parameters of the generator. ; The noise matrix Noise, with the same dimension as the landslide feature, is input into the generator G, and the optimal parameters of the generator are loaded. Landslide amplification data X is generated based on the difference between the number of landslide samples and non-landslide samples. k And add it to the original training set X1 to form a balanced dataset X P Used for model training; The balanced dataset is input into the backend machine learning classification model for training. The optimal model generated by the training is then used to test the test set to obtain landslide prediction results.

2. The landslide identification method according to claim 1, characterized in that, The multi-source dataset includes: remote sensing images of the study area during the post-earthquake growing season, regional geological maps, global digital elevation model (DEM), and various geographic data; among which, the various geographic data include: regional climate and precipitation data, vegetation cover, and human engineering activities.

3. The landslide identification method according to claim 2, characterized in that, After acquiring and preprocessing multi-source datasets, landslide-causing factors are extracted. Following further preprocessing of these factors and creation of landslide labels, a landslide sample library is constructed, including: The multi-source data is unified in the remote sensing software ENVI 5.3 according to the reference image resolution, projection coordinate system and geographic coordinate system. Radiometric calibration, atmospheric correction, image stitching and cropping operations are performed on the remote sensing image data to obtain the preprocessed multi-source data. The preprocessed multi-source data was input into the ArcMap 10.6 platform to extract landslide hazard factors; Input the landslide-causing factors into the R language platform to calculate the Pearson correlation index. If the correlation index between factors exceeds 0.7, it indicates that there is a strong correlation between the factors, and selective elimination is performed. The original hazard-causing factors were input into SPSS for multicollinearity analysis to determine whether collinearity existed among the factors. By combining the results of Pearson correlation index and collinearity analysis, factor elimination was performed to construct a clear landslide sample library without data redundancy.

4. The landslide identification method according to claim 3, characterized in that, The preprocessed multi-source data was input into the ArcMap 10.6 platform to extract landslide hazard factors, including: Spectral factors and remote sensing index factors were extracted from each band of remote sensing images, including Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), and Tassel Change Greenness and Humidity. Topographic and hydrological factors were extracted based on the global digital elevation model (DEM), including slope, aspect, plane curvature, profile curvature, elevation variability index, surface relief, topographic cutting depth, surface roughness, and distance from the river. Geomorphic factors, including distance from faults and lithology, are extracted based on regional geological maps. Auxiliary factors are extracted based on geographic information data, including average annual rainfall and human engineering activities.

5. The landslide identification method according to claim 1, characterized in that, The process involves overlaying preprocessed landslide causative factors and landslide labels into layers, then creating a one-dimensional array of data using an image cube. This one-dimensional array data is then divided into a training set and a test set, including: After preprocessing, the landslide hazard factors and landslide labels are input into ENVI 5.

3. The image cube with an h×w×c layer overlay tool is used, where h is the length of each factor, w is the width, and c is the number of channels of the image cube. Then, the pixels at the same position of the image cube are extracted into one-dimensional array data. The one-dimensional array data is divided into training set X1 and test set T1 in a 6:4 ratio.

6. The landslide identification method according to claim 1, characterized in that, The backend machine learning classification model is a support vector machine.

7. The landslide identification method according to claim 1, characterized in that, After the step of testing the test set using the optimal model generated during training to obtain landslide prediction results, the following steps are also included: The landslide prediction results are compared with the landslide labels that have been validated in the field, and the accuracy index is calculated based on the confusion matrix to quantitatively evaluate the prediction performance of the optimal model.

8. The landslide identification method according to claim 7, characterized in that, The accuracy metrics include at least one of AUC, G-mean, F1 score, Kappa coefficient, and Matthews correlation coefficient.

9. A landslide identification device based on a generative adversarial network data augmentation strategy, characterized in that, Includes the following modules: The sample library construction module is used to acquire multi-source datasets and preprocess them, extract landslide disaster-causing factors, preprocess the landslide disaster-causing factors, create landslide labels, and then construct a landslide sample library. The training set creation module is used to overlay the pre-processed landslide disaster factors and landslide labels into layers, and then create a one-dimensional array data through image cubes, dividing the one-dimensional array data into training set and test set. The data augmentation module is used to input landslide data from the training set into batches and train the optimized generative adversarial network model Smo-SE-WGAN based on a convolutional neural network. This augments the minority class samples and adds the augmented samples to the original training set to form a balanced dataset, including: The landslide data X from training set X1 will be processed in fixed batches. L The landslide features are one-dimensional array data generated from image cubes and input into the network. The batch size needs to be set according to the size of the landslide feature dimension, hardware computing power and network complexity. Based on the characteristics of landslides, an optimized generative adversarial network model Smo-SE-WGAN based on convolutional neural networks is constructed, including: an imbalanced data sampling module, a Smote module, and a GAN module; First, the landslide data is iteratively extracted from landslide and background samples by the unbalanced data sampling module. Then, the Smote module is used to balance the data and generate pseudo data by copying minority class samples. The GAN module includes a data generator G that simulates the distribution of real data, and a discriminator D that determines whether the data is real landslide data or pseudo data generated by the generator G. The Wassertein distance, a probability distribution difference index, is used to construct a loss function to calculate the distance between the pseudo-data distribution generated by generator G and the real data distribution. The specific calculation formula is as follows: In the formula, Z represents the actual landslide data, and Z belongs to the distribution of actual landslide data. G(n) is pseudo-data generated by a generator G from random noise, and G(n) belongs to the generated data distribution G( ); Let E be the joint distribution of the real data and the generated data, and E be the mean. Based on the network batch and landslide feature dimensions, a noise matrix of equal size following a normal distribution is constructed. This matrix is ​​then input into the generator G in the GAN module. The generator is constructed using a convolutional neural network as the underlying network. For the characteristics of landslide array data, a one-dimensional convolutional layer, activation function, and batch normalization layer are alternately arranged. For one-dimensional landslide array data, a channel attention mechanism is introduced into the generator. Then, random noise is subjected to convolutional feature extraction through the aforementioned alternate structure to simulate the distribution of real landslide feature data, thereby generating pseudo-data X with the same landslide features. W ; The pseudo data X generated by the generator G above W and actual landslide data X L Both data are input into the discriminator D. The discriminator's underlying architecture uses alternating one-dimensional convolutional layers and max-pooling layers to learn the distribution of real and fake data. A softmax function is added to the bottom layer of the framework to distinguish between real and fake labels. Then, based on the loss value, the parameters of the generator G and the discriminator D are alternately updated using stochastic gradient descent until the maximum number of training iterations is completed, ending the generative adversarial network training and retaining the optimal parameters of the generator. ; The noise matrix Noise, with the same dimension as the landslide feature, is input into the generator G, and the optimal parameters of the generator are loaded. Landslide amplification data X is generated based on the difference between the number of landslide samples and non-landslide samples. k And add it to the original training set X1 to form a balanced dataset X P Used for model training; The landslide prediction module is used to input the balanced dataset into the backend machine learning classification model for training, and then use the optimal model generated by training to test the test set to obtain the landslide prediction results.

Images