A landslide identification method, system, device, and medium

Abstract

This invention discloses a landslide identification method, system, device, and medium, relating to the field of landslide identification. The method includes: acquiring joint data of a study area at the current moment; the joint data includes: landslide influencing factors and optical remote sensing data; inputting the joint data at the current moment into a landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network. This invention solves the problems of insufficient landslide samples and imbalanced samples, improving the accuracy of landslide identification.

Description

Technical Field

[0001] This invention relates to the field of landslide identification, and in particular to a landslide identification method, system, device and medium. Background Technology

[0002] China has a vast territory and diverse geological environment, with disasters distributed over a wide area. Landslides are a frequent threat to the lives and property of residents in disaster-prone areas, making landslide prevention and control crucial.

[0003] A landslide is a natural phenomenon in which soil or rock masses on a slope slide downhill, either as a whole or in parts, under the influence of gravity, due to factors such as river erosion, groundwater activity, rainwater soaking, earthquakes, and artificial slope cutting. Currently, most traditional deep learning-based landslide identification methods are implemented using supervised learning, which suffers from problems such as insufficient landslide samples and imbalanced sample distribution. Summary of the Invention

[0004] Based on this, embodiments of the present invention provide a landslide identification method, system, device, and medium to solve the problems of insufficient landslide samples and sample imbalance, and improve the accuracy of landslide identification.

[0005] To achieve the above objectives, embodiments of the present invention provide the following solutions:

[0006] A landslide identification method includes:

[0007] Acquire joint data of the study area at the current moment; the joint data includes: landslide impact factors and optical remote sensing data;

[0008] The joint data at the current moment is input into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network.

[0009] Optionally, the method for determining the landslide identification model is as follows:

[0010] Obtain a training sample set; the training sample set includes: labeled samples and unlabeled samples; the labeled samples include: joint data of the study area with labeled labels at historical times; the unlabeled samples include: joint data of the study area without labeled labels at historical times; the labels are determined based on landslide images at historical times; the labels include: whether it is a landslide area;

[0011] Construct a semi-supervised generative adversarial network; the semi-supervised generative adversarial network includes: a generator and a discriminator;

[0012] Random noise is used as the input to the generator, and the unlabeled samples and the generated samples output by the generator are used as the input to the discriminator. The semi-supervised generative adversarial network is trained unsupervised with the goal of minimizing the true / false loss, and the trained generator is determined as the sample augmentation model. The true / false loss is determined based on the unlabeled samples and the generated samples output by the generator. The generated samples include joint data that is different from the unlabeled samples.

[0013] Using random noise as input to the sample augmentation model, the generated samples output by the sample augmentation model are determined as augmented samples, and new labels are assigned to the augmented samples to obtain labeled new samples;

[0014] The labeled samples and the newly labeled samples are used as inputs to the discriminator. The discriminator is trained in a supervised manner with the goal of minimizing the classification loss, and the trained discriminator is determined as the landslide identification model. The classification loss is determined based on the output of the discriminator and the labels input to the discriminator.

[0015] Optionally, joint data for the study area at the current moment can be obtained, specifically including:

[0016] Acquire synthetic aperture radar data for the study area at the current moment;

[0017] The synthetic aperture radar data is processed using small baseline set radar interferometry to determine the surface deformation information of the study area at the current moment;

[0018] The landslide area within the study region is determined based on the surface deformation information.

[0019] The landslide impact factor and optical remote sensing data within the landslide area are acquired and used as joint data for the study area at the current moment.

[0020] Optionally, the discriminator includes: a first convolutional layer, a first activation function, a first dropout layer, a second convolutional layer, a first batch normalization layer, a second activation function, a second dropout layer, a third convolutional layer, a third activation function, a third dropout layer, and a fully connected layer connected in sequence.

[0021] Optionally, the generator includes: a reshaping layer, a second batch normalization layer, a fourth activation function, a first deconvolution layer, a third batch normalization layer, a fifth activation function, a second deconvolution layer, a fourth batch normalization layer, a sixth activation function, and a third deconvolution layer connected in sequence.

[0022] Optionally, the landslide influencing factors include: elevation, slope, aspect, land use data, average annual rainfall, vegetation normalization index, lithology, distance from fault, distance from water system, and distance from road.

[0023] Optionally, the optical remote sensing data includes optical data in 12 bands.

[0024] The present invention also provides a landslide identification system, comprising:

[0025] The data acquisition module is used to acquire joint data of the study area at the current moment; the joint data includes: landslide impact factors and optical remote sensing data;

[0026] The landslide identification module is used to input the joint data at the current moment into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network.

[0027] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described landslide identification method.

[0028] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described landslide identification method.

[0029] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0030] This invention constructs a landslide identification model based on a semi-supervised generative adversarial network (SSGAN). Compared with traditional deep learning landslide identification methods, it solves the problems of insufficient landslide samples and imbalanced samples, and improves the accuracy of landslide identification. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a landslide identification method provided in an embodiment of the present invention;

[0033] Figure 2 A flowchart of the SBAS-InSAR processing provided in an embodiment of the present invention;

[0034] Figure 3 This is a schematic diagram of the sliding cutting sample fabrication strategy provided in an embodiment of the present invention;

[0035] Figure 4 This is a schematic diagram of the semi-supervised generative adversarial network training process provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] Example 1

[0039] See Figure 1 The landslide identification method in this embodiment includes:

[0040] Step 101: Obtain joint data of the study area at the current moment; the joint data includes: landslide impact factors and optical remote sensing data.

[0041] Step 101 specifically includes:

[0042] Acquire synthetic aperture radar (SAR) data of the study area at the current moment; process the SAR data using small baseline set radar interferometry to determine the surface deformation information of the study area at the current moment; determine the landslide range area within the study area based on the surface deformation information; acquire the landslide influence factor and optical remote sensing data within the landslide range area as joint data of the study area at the current moment.

[0043] The landslide influencing factors include: elevation, slope, aspect, land use data, average annual rainfall, vegetation normalization index, lithology, distance from fault, distance from water system, and distance from road. The optical remote sensing data includes optical data in 12 bands.

[0044] Step 102: Input the joint data at the current moment into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network.

[0045] The method for determining the landslide identification model is as follows:

[0046] 1) Obtain a training sample set; the training sample set includes: labeled samples and unlabeled samples; the labeled samples include: joint data of the study area with labels at historical times; the unlabeled samples include: joint data of the study area without labels at historical times; the labels are determined based on landslide images at historical times; the labels include: whether it is a landslide area.

[0047] 2) Construct a semi-supervised generative adversarial network; the semi-supervised generative adversarial network includes: a generator and a discriminator.

[0048] 3) Using random noise as input to the generator, and the unlabeled samples and the generated samples output by the generator as input to the discriminator, the semi-supervised generative adversarial network is trained unsupervised with the goal of minimizing the true / false loss, and the trained generator is determined as the sample augmentation model; the true / false loss is determined based on the unlabeled samples and the generated samples output by the generator; the generated samples include joint data that are different from the unlabeled samples.

[0049] 4) Using random noise as input to the sample augmentation model, the generated samples output by the sample augmentation model are determined as augmented samples, and new labels are assigned to the augmented samples to obtain labeled new samples.

[0050] 5) The labeled samples and the newly labeled samples are used as inputs to the discriminator. The discriminator is trained in a supervised manner with the goal of minimizing the classification loss. The trained discriminator is then identified as the landslide identification model. The classification loss is determined based on the output of the discriminator and the labels input to the discriminator.

[0051] The discriminator comprises: a first convolutional layer, a first activation function, a first dropout layer, a second convolutional layer, a first batch normalization layer, a second activation function, a second dropout layer, a third convolutional layer, a third activation function, a third dropout layer, and a fully connected layer connected in sequence.

[0052] The generator includes: a reshaping layer, a second batch normalization layer, a fourth activation function, a first deconvolution layer, a third batch normalization layer, a fifth activation function, a second deconvolution layer, a fourth batch normalization layer, a sixth activation function, and a third deconvolution layer connected in sequence.

[0053] The following is a more specific example to illustrate the implementation process and effectiveness of the landslide identification method described above.

[0054] 1. Dataset Construction

[0055] This example uses a landslide database built based on Small Baselines Subset Synthetic Aperture Radar (SBAS-InSAR) surface deformation information to construct a model learning sample set. To represent the environmental characteristics of landslide development and the optical texture features of landslide occurrence, this example proposes constructing three sample sets representing landslide characteristics: a landslide influencing factor sample set, a Sentinel-2A optical remote sensing sample set, and a joint sample set of influencing factors and Sentinel-2A. These three sample sets are compared and analyzed to identify landslides.

[0056] Traditional methods of visually interpreting landslides using optical remote sensing are time-consuming and labor-intensive, and optical images are easily affected by weather and lighting conditions. InSAR technology, through interferometry, enables all-weather, 24 / 7 surface observation, avoiding the drawbacks of optical imagery. Furthermore, SBAS-InSAR technology can acquire surface deformation information, allowing for rapid location of landslides and clarifying the selection of sample sets. The other three types of sample sets, based on the acquired landslide locations, are used to summarize landslide characteristics.

[0057] 1) Landslide Influencing Factors

[0058] Based on existing research and the actual situation of the study area, this example selects 10 environmental factors, as shown in Table 1. This example covers a variety of categories, including topography, geology, hydrology, vegetation cover, and human engineering activities.

[0059] Table 1. Data Sources of Landslide Influencing Factors

[0060]

[0061]

[0062] 2) Sentinel-2A remote sensing data

[0063] The Sentinel-2A remote sensing data used in this example comes from the United States Geological Survey (USGS). Sentinel-2A is a high-resolution multispectral imaging satellite, as shown in Table 2.

[0064] Table 2 Sentinel-2A Sensor Parameters

[0065]

[0066]

[0067] 3) SBAS-InSAR deformation data

[0068] SBAS-InSAR utilizes the phase information of SAR imagery, combined with precise imaging geometric parameters and orbital ephemeris parameters, to invert surface elevation information and achieve geocoding. The SBAS-InSAR processing flow is as follows: Figure 2 As shown. The SAR data used in this example comes from Sentinel-1A data from the European Space Agency's (https: / / scihub.copernicus.eu / ) open-source data platform. Sentinel-1A is a SAR image that uses InSAR technology to obtain surface deformation; Sentinel-2A is an optical image used to represent the optical characteristics of landslides.

[0069] The specific steps for constructing the dataset are as follows: First, all 10 environmental image factor features and 12 band optical feature data are converted to the same resolution. In this example, Sentinel-2A is used as the benchmark to resample all image factors to 10m resolution. Then, the selected feature data and historical landslide distribution map are output as a 4314-row, 7210-column pixel image as the original dataset. A 13×13 pixel window is used to slide across the image, moving one pixel grid at a time to ensure that every pixel in the image is traversed. Simultaneous sliding and cutting are performed on the image factors, optical images, and historical landslide images to ensure that each feature corresponds to the original landslide area, forming a single image sample. The labeled data sample set uses the pixel value of the original landslide center point as the label for that area, i.e., landslide points are labeled as 1, and non-landslide points are labeled as 0. Each sample is 13 pixels × 13 pixels. Finally, landslide samples, non-landslide samples, and unlabeled samples are used to construct the dataset. The overall dataset is divided into a model training set and a test set in a 7:3 ratio. The sliding and cutting sample creation strategy is as follows: Figure 3 As shown.

[0070] 2. Semi-supervised generative adversarial process

[0071] See Figure 4 This example transmits three datasets—the landslide impact factor sample set, the Sentinel-2A optical remote sensing sample set, and the joint sample set of impact factors and Sentinel-2A—to SSGAN. The study analyzes the model training effect from the perspectives of both the SSGAN discriminator and the generator, and evaluates the performance of different datasets in the landslide identification task based on the model training results.

[0072] 1) The discriminator network structure is a basic feedforward deep CNN structure used to extract features from images and reduce spatial dimensionality. For example... Figure 4As shown, the discriminator consists of 11 layers from left to right: a first convolutional layer, a first activation function, a first dropout layer, a second convolutional layer, a first batch normalization layer, a second activation function, a second dropout layer, a third convolutional layer, a third activation function, a third dropout layer, and a fully connected layer, connected sequentially. The final output is the landslide category, where 1 represents a landslide and 0 represents a non-landslide. The size of the discriminator's input image and the result of one convolutional layer is calculated using the following formula:

[0073]

[0074] n1 is the side length of the input image to the discriminator, n c To determine the number of channels, p0 zero elements are padded to the image boundaries using a size of f. k ×f k ×n c The number is n′ c The convolution kernel is implemented, where [l] represents the l-th neural layer; [l+1] represents the (l+1)-th neural layer; This represents the image size output by the (l+1)th neural layer in the discriminator; f represents the image size output by the l-th neural layer in the discriminator; k This represents the size of the convolution kernel in the discriminator; This indicates rounding x down.

[0075] 2) such as Figure 4 As shown, the generator consists of 10 layers from left to right: a reshaping layer, a second batch normalization layer, a fourth activation function, a first deconvolution layer, a third batch normalization layer, a fifth activation function, a second deconvolution layer, a fourth batch normalization layer, a sixth activation function, and a third deconvolution layer, all connected in sequence. The generator's input is random noise following a normal distribution. After shape transformation, the random noise is upsampled using deconvolution. The feature map obtained after deconvolution is gradually remapped into the space of the original input image, thus being passed to the discriminator for semantic interpretation at the same pixel level as the real image. The size calculation of the output feature map after deconvolution is shown in the following formula:

[0076]

[0077] In the formula, This represents the image size output by the (l+1)th neural layer in the generator; represents the image size output by the l-th neural layer in the generator; k is the kernel size in the generator; s is the stride; and p is the number of zero elements used to pad the image boundaries of the input image to the discriminator.

[0078] By comparing unlabeled samples with generated samples, the discriminator can learn more deeply about the environmental characteristics of the current study area. Based on a full understanding of the environmental characteristics, labeled samples are used to enable the discriminator to learn about landslides and non-landslides in the real environment, thereby making a more accurate classification of whether or not a landslide occurs in the area.

[0079] 3. Landslide identification results

[0080] Based on the above comparison of SSGAN's performance in the discriminator and generator for the three types of datasets, it can be concluded that the model is difficult to complete the semi-supervised landslide identification task by relying solely on optical images. However, it performs better when using a combination of optical images and image factor features than when using only influence factor features. Therefore, this example will use the training results of the combined sample set to complete the subsequent landslide identification task.

[0081] SSGAN identified a total of 160 landslides in the study area, covering a total area of ​​10.328 km². 2 Compared to historical landslide identification results, the area of ​​the landslides identified was 3.139 km² larger. The SSGAN identification results were generally consistent with the landslide distribution in the sample, with relatively fine landslide edge delineation, and some individual landslides had an area exceeding the original landslide area. Therefore, considering the overall results, the SSGAN model is helpful for landslide representation learning and thus obtaining accurate results.

[0082] This embodiment uses deep learning to actively learn landslide features and automatically identify landslides in the study area. To overcome the problems of insufficient and imbalanced landslide samples in traditional deep learning landslide identification, this embodiment proposes to construct an SSGAN based on traditional CNN and combining semi-supervised learning and GAN concepts to complete the landslide identification task. This embodiment feeds three types of sample sets—landslide impact factor sample set, Sentinel-2A optical remote sensing sample set, and a joint sample set of impact factors and optical features—to the SSGAN for training. The results show that the joint sample set yields superior results in both the discriminator and generator. The method in this embodiment performs well in the evaluation metrics, therefore the generated results are accurate and reliable.

[0083] Example 2

[0084] In order to implement the method corresponding to Embodiment 1 above and achieve the corresponding functions and technical effects, a landslide identification system is provided below.

[0085] The system includes:

[0086] The data acquisition module is used to acquire joint data of the study area at the current moment; the joint data includes: landslide impact factors and optical remote sensing data.

[0087] The landslide identification module is used to input the joint data at the current moment into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network.

[0088] Example 3

[0089] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the landslide identification method of Embodiment 1.

[0090] Alternatively, the aforementioned electronic device may be a server.

[0091] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the landslide identification method of Embodiment 1.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0093] This example uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A landslide identification method, characterized in that, include: Obtain joint data for the study area at the current moment; The combined data includes: landslide impact factors and optical remote sensing data; The joint data at the current moment is input into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network; The method for determining the landslide identification model is as follows: Obtain a training sample set; the training sample set includes: labeled samples and unlabeled samples; the labeled samples include: joint data of the study area with labeled labels at historical times; the unlabeled samples include: joint data of the study area without labeled labels at historical times; the labels are determined based on landslide images at historical times; the labels include: whether it is a landslide area; Construct a semi-supervised generative adversarial network; the semi-supervised generative adversarial network includes: a generator and a discriminator; Random noise is used as the input to the generator, and the unlabeled samples and the generated samples output by the generator are used as the input to the discriminator. The semi-supervised generative adversarial network is trained unsupervised with the goal of minimizing the true / false loss, and the trained generator is determined as the sample augmentation model. The true / false loss is determined based on the unlabeled samples and the generated samples output by the generator. The generated samples include joint data that is different from the unlabeled samples. Using random noise as input to the sample augmentation model, the generated samples output by the sample augmentation model are determined as augmented samples, and new labels are assigned to the augmented samples to obtain labeled new samples; The labeled samples and the newly labeled samples are used as inputs to the discriminator. The discriminator is trained in a supervised manner with the goal of minimizing the classification loss, and the trained discriminator is determined as the landslide identification model. The classification loss is determined based on the output of the discriminator and the labels input to the discriminator.

2. The landslide identification method according to claim 1, characterized in that, Obtain joint data for the study area at the current moment, specifically including: Acquire synthetic aperture radar data for the study area at the current moment; The synthetic aperture radar data is processed using small baseline set radar interferometry to determine the surface deformation information of the study area at the current moment; The landslide area within the study region is determined based on the surface deformation information. The landslide impact factor and optical remote sensing data within the landslide area are acquired and used as joint data for the study area at the current moment.

3. The landslide identification method according to claim 1, characterized in that, The discriminator comprises: a first convolutional layer, a first activation function, a first dropout layer, a second convolutional layer, a first batch normalization layer, a second activation function, a second dropout layer, a third convolutional layer, a third activation function, a third dropout layer, and a fully connected layer connected in sequence.

4. The landslide identification method according to claim 1, characterized in that, The generator includes: a reshaping layer, a second batch normalization layer, a fourth activation function, a first deconvolution layer, a third batch normalization layer, a fifth activation function, a second deconvolution layer, a fourth batch normalization layer, a sixth activation function, and a third deconvolution layer connected in sequence.

5. The landslide identification method according to claim 1, characterized in that, The landslide influencing factors include: elevation, slope, aspect, land use data, average annual rainfall, vegetation normalization index, lithology, distance from fault, distance from water system, and distance from road.

6. The landslide identification method according to claim 1, characterized in that, The optical remote sensing data includes optical data in 12 bands.

7. A landslide identification system, characterized in that, include: The data acquisition module is used to acquire joint data of the study area at the current moment; The combined data includes: landslide impact factors and optical remote sensing data; The landslide identification module is used to input the joint data at the current moment into the landslide identification model to obtain the landslide area of ​​the study area at the current moment; wherein, the landslide identification model is constructed based on a semi-supervised generative adversarial network; The method for determining the landslide identification model is as follows: Obtain a training sample set; the training sample set includes: labeled samples and unlabeled samples; the labeled samples include: joint data of the study area with labeled labels at historical times; the unlabeled samples include: joint data of the study area without labeled labels at historical times; the labels are determined based on landslide images at historical times; the labels include: whether it is a landslide area; Construct a semi-supervised generative adversarial network; the semi-supervised generative adversarial network includes: a generator and a discriminator; Random noise is used as the input to the generator, and the unlabeled samples and the generated samples output by the generator are used as the input to the discriminator. The semi-supervised generative adversarial network is trained unsupervised with the goal of minimizing the true / false loss, and the trained generator is determined as the sample augmentation model. The true / false loss is determined based on the unlabeled samples and the generated samples output by the generator. The generated samples include joint data that is different from the unlabeled samples. Using random noise as input to the sample augmentation model, the generated samples output by the sample augmentation model are determined as augmented samples, and new labels are assigned to the augmented samples to obtain labeled new samples; The labeled samples and the newly labeled samples are used as inputs to the discriminator. The discriminator is trained in a supervised manner with the goal of minimizing the classification loss, and the trained discriminator is determined as the landslide identification model. The classification loss is determined based on the output of the discriminator and the labels input to the discriminator.

8. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the landslide identification method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the landslide identification method as described in any one of claims 1 to 6.

Images