Mountain deformation identification method based on satellite remote sensing

Abstract

The invention relates to a mountain deformation identification method based on satellite remote sensing. The method comprises the following steps: acquiring satellite remote sensing images of a research area in different periods, performing image enhancement processing on the satellite remote sensing images, synthesizing the enhanced satellite remote sensing images into point cloud data, and performing target extraction to acquire mountain point cloud data; and performing difference processing on the mountain point cloud data to generate three-dimensional target point cloud models in different periods, matching the three-dimensional target point cloud models in different periods to obtain a point cloud change region, and taking the point cloud change region as a mountain deformation region. The invention aims to improve the mountain deformation monitoring efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of mountain deformation recognition, and in particular to a mountain deformation recognition method based on satellite remote sensing. Background Art

[0002] With the development of society and the increase in human activities, mountain deformation has received increasing attention. Mountain deformation can lead to natural disasters such as landslides and debris flows, posing a serious threat to human life and property. Traditional mountain deformation monitoring methods rely primarily on ground measurements and field surveys, which have the following limitations:

[0003] High labor costs: A large amount of manpower is required for field measurement and data collection, which consumes time and energy.

[0004] Limited monitoring range: It is difficult to cover large mountain areas and some potential deformation areas are easily missed.

[0005] Data acquisition is not timely: It is impossible to obtain mountain deformation data in real time, making it difficult to detect deformation signs and take measures in time.

[0006] Limited by natural conditions: Surveying work is difficult to carry out under bad weather or complex terrain conditions.

[0007] In recent years, satellite remote sensing technology has been widely used in the field of earth science. Satellite remote sensing can provide large-scale, high-resolution images of the earth's surface, providing a new means for monitoring mountain deformation. However, existing mountain deformation monitoring methods based on satellite remote sensing have some shortcomings:

[0008] Limited image quality: Satellite remote sensing images may have problems such as noise and blur, which affect the accuracy of deformation recognition.

[0009] Complex data processing: Extracting mountain deformation information from satellite remote sensing images requires complex image processing and analysis techniques, which places high demands on professionals.

[0010] Lack of effective feature extraction methods: It is difficult to accurately extract the key features of mountain deformation, resulting in low accuracy of deformation recognition. Summary of the Invention

[0011] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to propose a mountain deformation identification method based on satellite remote sensing, aiming to improve the efficiency and accuracy of mountain deformation monitoring and overcome the shortcomings of the prior art.

[0012] To achieve the above object, the present invention provides the following solutions:

[0013] A mountain deformation identification method based on satellite remote sensing, comprising:

[0014] Obtain satellite remote sensing images of the study area at different times, perform image enhancement processing on the satellite remote sensing images, synthesize the enhanced satellite remote sensing images into point cloud data, and perform target extraction to obtain mountain point cloud data;

[0015] The mountain point cloud data is subjected to difference processing to generate three-dimensional target point cloud models of different periods, the three-dimensional target point cloud models of different periods are matched, and the point cloud change area is obtained, and the point cloud change area is used as the mountain deformation area.

[0016] Optionally, performing image enhancement processing on the satellite remote sensing image includes:

[0017] Inputting the satellite remote sensing image into an image enhancement model to obtain an enhanced satellite remote sensing image; the image enhancement model is trained using a training set; the training set includes: the original satellite remote sensing image preprocessed by histogram equalization;

[0018] The image enhancement model uses convolutional neural networks of different scales to extract the remote sensing image feature matrix of the satellite remote sensing image, and fuses them to generate the enhanced satellite remote sensing image.

[0019] Optionally, the image enhancement model includes:

[0020] A feature extraction module is used to input the satellite remote sensing image into a convolutional neural network with two-dimensional convolution kernels of different scales to obtain remote sensing image feature matrices of different scales;

[0021] A feature fusion module is used to fuse the remote sensing image feature matrices of different scales to obtain a multi-scale remote sensing image feature matrix;

[0022] The image enhancement module is used to input the multi-scale remote sensing image feature matrix into an image generator based on a generative adversarial network to obtain the enhanced satellite remote sensing image.

[0023] Optionally, obtaining the remote sensing image feature matrices of different scales includes:

[0024] Each layer of the convolutional neural network using the two-dimensional convolution kernels of different scales performs feature processing on the input satellite remote sensing image in the forward pass:

[0025] The satellite remote sensing image is convolved based on a two-dimensional convolution kernel to generate a convolution feature map, the convolution feature map is mean pooled along the channel dimension to generate a pooled feature map, the eigenvalues ​​of each position in the pooled feature map are nonlinearly activated to generate an activated feature map, and the last layer of the convolutional neural network outputs the remote sensing image feature matrices of different scales.

[0026] Optionally, obtaining the mountain point cloud data includes:

[0027] Synthesizing the enhanced satellite remote sensing image into point cloud data; wherein the point cloud data is in TXT format and includes coordinate information and color information;

[0028] Perform deep semantic segmentation-clustering point cloud extraction on the point cloud data to obtain mountain point cloud data.

[0029] Optionally, obtaining the point cloud change area includes:

[0030] Performing a rigid body transformation operation on the three-dimensional target point cloud models at different times to achieve rough registration of the two three-dimensional point cloud models, wherein the rigid body transformation operation includes: translation and rotation;

[0031] The three-dimensional target point cloud model after coarse registration is precisely registered, and the precisely registered three-dimensional target point cloud model is mapped to a three-dimensional reference coordinate system. A resampling grid is established based on the reference point cloud, and the two three-dimensional point cloud models are resampled under the resampling grid to obtain resampled point clouds.

[0032] The resampled point cloud is subjected to difference processing to calculate the Z value change of the same grid at different periods, obtain a difference image, and process the difference image to obtain the change area of ​​the point cloud model.

[0033] Optionally, performing precise registration using the feature point cloud includes:

[0034] From the three-dimensional target point cloud model after the rough registration, overlapping three-dimensional point clouds are respectively selected as the target point cloud and the reference point cloud, and the curvature of each point in the target point cloud and the reference point cloud is calculated. According to the curvature similarity, feature matching point pairs are obtained. The feature matching point pairs are used to achieve accurate registration of the target point cloud and the reference point cloud using an iterative closest point matching algorithm.

[0035] Optionally, obtaining feature matching point pairs according to the curvature similarity includes:

[0036] Using a three-dimensional grid method to spatially divide the target point cloud and the reference point cloud, determine the neighborhood of each point, perform local parabola fitting on the neighborhood, and obtain the principal curvature;

[0037] According to the principal curvature, curvature matching is used to perform matching selection to obtain the feature matching point pairs.

[0038] The beneficial effects of the present invention are:

[0039] The present invention obtains a wide range of mountain image data through satellite remote sensing technology, which can quickly cover a large area of ​​monitoring area, reduce the workload of manual field measurement, and improve monitoring efficiency.

[0040] The present invention uses an image enhancement model to process satellite remote sensing images, which can effectively improve image quality, enhance image contrast and clarity, and provide a higher quality data basis for subsequent deformation recognition.

[0041] The present invention uses a multi-scale convolutional neural network to extract remote sensing image features and generates enhanced images through a generative adversarial network, which can more accurately extract the key features of mountain deformation and improve the accuracy of deformation recognition.

[0042] The present invention can realize automatic monitoring of mountain deformation, reduce dependence on professionals, lower monitoring costs, and at the same time can obtain mountain deformation data in real time and promptly discover potential deformation signs. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 This is a flow chart of a mountain deformation identification method based on satellite remote sensing according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0047] like Figure 1As shown, this embodiment discloses a mountain deformation identification method based on satellite remote sensing, including: obtaining satellite remote sensing images of the study area at different periods, performing image enhancement processing on the satellite remote sensing images, synthesizing the enhanced satellite remote sensing images into point cloud data, and performing target extraction to obtain mountain point cloud data; performing difference processing on the mountain point cloud data to generate three-dimensional target point cloud models at different periods, matching the three-dimensional target point cloud models at different periods, obtaining point cloud change areas, and using the point cloud change areas as mountain deformation areas.

[0048] Specifically, this embodiment discloses a mountain deformation identification method based on satellite remote sensing, comprising: first, collecting satellite remote sensing images of the study area in different time periods. Then, performing image enhancement operations on these satellite remote sensing images to improve the image quality. Thereafter, the enhanced satellite remote sensing images are converted into point cloud data, and the target part is extracted therefrom to obtain mountain point cloud data. Then, difference analysis is performed on the mountain point cloud data of different periods to construct three-dimensional target point cloud models of different time stages. Finally, these three-dimensional target point cloud models at different times are matched and compared to determine the area where the point cloud has changed, and this changed area is identified as the mountain deformation area.

[0049] Furthermore, image enhancement processing of satellite remote sensing images includes: inputting the satellite remote sensing images into an image enhancement model to obtain enhanced satellite remote sensing images; the image enhancement model is trained using a training set; the training set includes: original satellite remote sensing images preprocessed by histogram equalization; wherein the image enhancement model uses convolutional neural networks of different scales to extract remote sensing image feature matrices of satellite remote sensing images, and fuses them to generate enhanced satellite remote sensing images.

[0050] Specifically, satellite remote sensing images are input into an image enhancement model to produce enhanced satellite remote sensing images. This image enhancement model is trained using a training set consisting of raw satellite remote sensing images preprocessed with histogram equalization. The image enhancement model uses convolutional neural networks of varying scales to extract feature matrices from the satellite remote sensing images and fuses these feature matrices to ultimately generate the enhanced satellite remote sensing images.

[0051] Furthermore, the image enhancement model includes: a feature extraction module, which is used to input the satellite remote sensing image into a convolutional neural network with two-dimensional convolution kernels of different scales to obtain remote sensing image feature matrices of different scales; a feature fusion module, which is used to fuse the remote sensing image feature matrices of different scales to obtain multi-scale remote sensing image feature matrices; and an image enhancement module, which is used to input the multi-scale remote sensing image feature matrix into an image generator based on a generative adversarial network to obtain enhanced satellite remote sensing images.

[0052] Furthermore, obtaining remote sensing image feature matrices of different scales includes: using each layer of the convolutional neural network with two-dimensional convolution kernels of different scales to perform feature processing on the input satellite remote sensing image in the forward pass: performing convolution processing on the satellite remote sensing image based on the two-dimensional convolution kernel to generate a convolution feature map, performing mean pooling processing on the convolution feature map along the channel dimension to generate a pooled feature map, performing nonlinear activation on the eigenvalues ​​of each position in the pooled feature map to generate an activation feature map, and outputting remote sensing image feature matrices of different scales by the last layer of the convolutional neural network.

[0053] Specifically, the image enhancement model consists of the following key parts: first, the feature extraction module, which is responsible for inputting satellite remote sensing images into convolutional neural networks with two-dimensional convolution kernels of different scales, thereby obtaining remote sensing image feature matrices of different scales; followed by a feature fusion module, which is used to integrate these remote sensing image feature matrices of different scales to obtain multi-scale remote sensing image feature matrices; finally, the image enhancement module, which inputs the multi-scale remote sensing image feature matrices into an image generator based on a generative adversarial network, and finally generates enhanced satellite remote sensing images.

[0054] More specifically, the satellite remote sensing image is passed through a first convolutional neural network model with a two-dimensional convolution kernel of a first scale to obtain a first-scale remote sensing image feature matrix, including: using each layer of the first convolutional neural network model with a two-dimensional convolution kernel of the first scale to perform the following on the input data in the forward pass of the layer: convolution processing on the input data based on the two-dimensional convolution kernel to generate a convolution feature map; mean pooling processing along the channel dimension of the convolution feature map to generate a pooled feature map; nonlinear activation of the eigenvalues ​​at each position in the pooled feature map to generate an activation feature map; wherein the output of the last layer of the first convolutional neural network model is the first-scale remote sensing image feature matrix, the input from the second layer to the last layer of the first convolutional neural network model is the output of the previous layer, and the input of the first layer of the convolutional neural network model is the satellite remote sensing image.

[0055] The satellite remote sensing image is passed through a second convolutional neural network model with a two-dimensional convolution kernel of a second scale to obtain a second-scale remote sensing image feature matrix, including: using each layer of the second convolutional neural network model with a two-dimensional convolution kernel of the second scale to perform two-dimensional convolution processing, mean pooling processing along the channel dimension and nonlinear activation processing on the input data in the forward pass of the layer to output the second-scale remote sensing image feature matrix from the last layer of the second convolutional neural network model.

[0056] Fusing the first-scale remote sensing image feature matrix and the second-scale remote sensing image feature matrix to obtain a multi-scale remote sensing image feature matrix, including: fusing the first-scale remote sensing image feature matrix and the second-scale remote sensing image feature matrix using the following cascade formula to obtain a multi-scale remote sensing image feature matrix;

[0057] A multi-scale remote sensing image feature matrix is ​​geometrically ordered based on parameterized features to obtain an optimized multi-scale remote sensing image feature matrix, including: constructing a parameterized coding vector for each position in the multi-scale remote sensing image feature matrix to obtain a plurality of pixel position parameterized coding vectors, wherein the parameterized coding vector includes coordinates, gradient values ​​along the X-axis direction, gradient values ​​along the Y-axis direction, and eigenvalues; passing the plurality of pixel position parameterized coding vectors through a parameterized feature extractor based on a fully connected layer to obtain a plurality of pixel position parameterized coding feature vectors; calculating the cosine similarity between any two pixel position parameterized coding feature vectors in the plurality of pixel position parameterized coding feature vectors to obtain an ordered geometric topology matrix; passing the ordered geometric topology matrix through a geometric topology feature extractor based on a convolutional layer to obtain an ordered geometric topology feature matrix; and fusing the multi-scale remote sensing image feature matrix and the ordered geometric topology feature matrix to obtain an optimized multi-scale remote sensing image feature matrix.

[0058] The image generator based on the generative adversarial network includes a discriminator and a generator, wherein the generator is used to generate images, and the discriminator is used to calculate the difference between the generated image and the reference image, and the network parameters of the generator are updated through the directional propagation algorithm of gradient descent to obtain a generator with the function of generating enhanced images. Then, the optimized multi-scale remote sensing image feature matrix is ​​input into the image generator based on the generative adversarial network so that the generator outputs the remote sensing image enhanced image through deconvolution coding.

[0059] Furthermore, obtaining mountain point cloud data includes: synthesizing enhanced satellite remote sensing images into point cloud data; wherein the point cloud data format is TXT format, including coordinate information and color information; performing deep semantic segmentation-clustering point cloud extraction on the point cloud data to obtain mountain point cloud data.

[0060] Furthermore, obtaining the point cloud change area includes: performing rigid body transformation operations on the three-dimensional target point cloud models at different periods to complete the rough alignment of the two three-dimensional point cloud models, and the rigid body transformation operations include: translation and rotation; performing precise alignment on the three-dimensional target point cloud model after the rough alignment, mapping the three-dimensional target point cloud model after the precise alignment to a three-dimensional reference coordinate system, and establishing a resampling grid based on the reference point cloud, resampling the two three-dimensional point cloud models under the resampling grid, and obtaining resampled point clouds; performing difference processing on the resampled point clouds, calculating the Z value change of the same grid at different periods, obtaining a difference image, processing the difference image, and obtaining the point cloud model change area.

[0061] Specifically, the 3D target point cloud models at different time periods are first subjected to rigid body transformation operations, including translation and rotation, to achieve coarse registration of the two 3D point cloud models. The coarsely registered 3D target point cloud models are then precisely registered and mapped to a 3D reference coordinate system. A resampling grid is established based on the reference point cloud, and the two 3D point cloud models are resampled within this grid to obtain resampled point clouds. The resampled point clouds are then interpolated to calculate the Z value change of the same grid at different time periods, generating a difference image. Finally, the difference image is processed to obtain the changed area of ​​the point cloud model.

[0062] Furthermore, using feature point clouds for precise registration includes: selecting three-dimensional point clouds that roughly overlap with each other from the three-dimensional target point cloud model after rough registration as the target point cloud and the reference point cloud, and calculating the curvature of each point in the target point cloud and the reference point cloud, obtaining feature matching point pairs based on the curvature similarity, and using the feature matching point pairs and an iterative nearest point matching algorithm to achieve precise registration of the target point cloud and the reference point cloud.

[0063] Specifically, the process for precise registration using feature point clouds is as follows: From the roughly registered 3D target point cloud model, roughly overlapping 3D point clouds are selected, one set serving as the target point cloud and the other as the reference point cloud. The curvature of each point in the target and reference point clouds is then calculated. Based on the similarity of these curvatures, matching feature point pairs are identified. Finally, using an iterative closest point matching algorithm, these matching feature point pairs are used to achieve precise registration of the target and reference point clouds.

[0064] More specifically, point cloud selection after coarse registration: roughly overlapping 3D point cloud parts are selected from the 3D target point cloud model after coarse registration. The purpose of coarse registration is to roughly align the two point cloud models in space so that subsequent precise registration operations can be performed more efficiently. One group of roughly overlapping point clouds selected is defined as the target point cloud, and the other group is defined as the reference point cloud. The target point cloud is the point cloud that needs to be adjusted to match the reference point cloud, while the reference point cloud is the point cloud used as a benchmark, and the target point cloud will be aligned with it through the registration process.

[0065] Curvature calculation: Next, the curvature of each point in the target point cloud and the reference point cloud is calculated. Curvature is an important feature that describes the local geometry of a point cloud. It reflects the degree of curvature of the point cloud surface. The process of calculating curvature generally includes the following steps:

[0066] Neighborhood Determination: For each point in the point cloud, we first determine the points within its neighborhood. The size of the neighborhood can be adjusted based on the density of the point cloud and specific application requirements. Common neighborhood determination methods include fixed radius search and fixed number of nearest neighbors search.

[0067] Local fitting: Perform local surface fitting within the neighborhood of each point. Common fitting methods include polynomial fitting and parabola fitting. The surface equation obtained by fitting can be used to further calculate the curvature of the point.

[0068] Curvature Calculation: Based on the fitted surface equation, the principal curvature of each point is calculated. Principal curvature is an important parameter that describes the local curvature of a surface and is typically obtained by calculating the eigenvalues ​​of the second-order derivative matrix (Hessian matrix) of the fitted surface. The magnitude and direction of the principal curvature can reflect the characteristics of the local geometric structure, such as concavity, convexity, or flatness.

[0069] Feature point pair matching: Find matching feature point pairs based on the calculated curvature similarity. The matching process usually includes the following steps:

[0070] Similarity metric: Calculate the similarity between the curvature of each point in the target point cloud and the curvature of all points in the reference point cloud. Common similarity metrics include Euclidean distance and cosine similarity. By calculating the similarity metric, a similarity matrix is ​​obtained, where each element in the matrix represents the similarity between a point in the target point cloud and a point in the reference point cloud.

[0071] Matching point pair generation: Based on the similarity matrix, point pairs with the closest curvature between the target and reference point clouds are found. These point pairs are considered feature matching pairs because they share similarities in local geometry. To improve matching accuracy, a similarity threshold is typically set; only point pairs with a similarity above this threshold are selected as feature matching pairs.

[0072] Screening and Optimization: To further improve the accuracy and reliability of matching, the initially obtained matching point pairs are usually screened and optimized. For example, consistency checks (such as the RANSAC algorithm) and distance constraints (such as limiting the maximum distance between matching point pairs) can be used to remove incorrectly matched point pairs, ensuring that the final feature matching point pairs are of high quality.

[0073] Iterative Closest Point Matching (ICP) algorithm: These features are used to match point pairs, achieving precise registration of the target point cloud and the reference point cloud. The ICP algorithm is a commonly used point cloud registration algorithm. Its basic idea is to gradually reduce the distance between the target point cloud and the reference point cloud through iterative optimization, ultimately achieving precise alignment. The specific steps of the ICP algorithm are as follows:

[0074] Initialization: Set the initial positions of the target point cloud and the reference point cloud to the result of coarse registration.

[0075] Closest point search: For each point in the target point cloud, find the closest point in the reference point cloud. This step is usually implemented using a spatial search algorithm (such as a KD tree) to improve search efficiency.

[0076] Transformation estimation: Based on the closest point pair found, estimate the rigid body transformation (including translation and rotation) of the target point cloud relative to the reference point cloud. Commonly used transformation estimation methods include least squares method, singular value decomposition, etc.

[0077] Point cloud update: Update the target point cloud according to the estimated transformation, that is, translate and rotate each point in the target point cloud according to the estimated transformation.

[0078] Iterative optimization: Repeat the above steps of nearest point search, transformation estimation, and point cloud update until certain convergence conditions are met, such as the distance between the target point cloud and the reference point cloud no longer decreases significantly or the preset number of iterations is reached.

[0079] Result output: The final transformation is the accurate registration result of the target point cloud relative to the reference point cloud. Through this transformation, the target point cloud can be accurately aligned to the reference point cloud.

[0080] The detailed steps above allow for accurate registration using feature point clouds. This process not only effectively improves the accuracy of point cloud registration but also provides a reliable foundation for subsequent applications such as 3D reconstruction and deformation analysis.

[0081] Furthermore, obtaining feature matching point pairs based on curvature similarity includes: spatially dividing the target point cloud and the reference point cloud using a three-dimensional grid method, determining the neighborhood of each point, performing local parabola fitting on the neighborhood, and obtaining the principal curvature; and performing matching selection using curvature matching based on the principal curvature to obtain feature matching point pairs.

[0082] Specifically, the process for obtaining feature matching point pairs based on curvature similarity is as follows: First, the target point cloud and the reference point cloud are spatially partitioned using a 3D grid method to determine the neighborhood of each point. Next, a local parabola is fitted to the neighborhood of each point to obtain the principal curvature. Finally, based on the similarity of the principal curvatures, matching selection is performed using a curvature matching method to obtain feature matching point pairs.

[0083] More specifically, spatial division and neighborhood determination: First, the target point cloud and the reference point cloud are spatially divided using the three-dimensional grid method. The three-dimensional grid method is a method that divides the three-dimensional space into a regular grid, and each grid cell can be regarded as a small local area. Through this method, the neighborhood range of each point cloud point can be efficiently determined. The determination of the neighborhood range is crucial for the subsequent curvature calculation because it determines the analysis range of the local geometric structure. Specifically, the neighborhood of each point is usually composed of points within a certain distance around it, and this distance can be adjusted according to the density of the point cloud and the specific application requirements.

[0084] Local parabola fitting and principal curvature calculation:

[0085] After determining the neighborhood of each point, we then perform local parabola fitting on the neighborhood of each point. Local parabola fitting is a fitting method based on the least squares method that approximates the local geometric shape of the point cloud by fitting a parabola. The specific steps are as follows:

[0086] Data preparation: Collect the coordinates of points in the neighborhood of each point.

[0087] Fitting process: Use the least squares method to fit a parabola equation.

[0088] Principal curvature calculation: Extracts the principal curvatures from the fitted parabola equation. Principal curvatures are important parameters that describe the local curvature of a surface and are typically obtained by calculating the eigenvalues ​​of the parabola's second-order derivative matrix (the Hessian matrix). The magnitude and direction of the principal curvatures can reflect the characteristics of the local geometric structure, such as concavity, convexity, or flatness.

[0089] Selection of curvature matching and feature matching point pairs:

[0090] After obtaining the principal curvature of each point, the curvature matching method is used for matching selection.

[0091] The specific steps are as follows:

[0092] Similarity metric: Calculate the similarity between the principal curvatures of each point in the target point cloud and the reference point cloud. Usually, the Euclidean distance or other similarity metrics can be used to compare the principal curvatures of two points.

[0093] Matching point pair generation: Based on curvature similarity, find the point pairs with the closest curvature in the target point cloud and the reference point cloud. These point pairs are considered feature matching point pairs because they have similarities in local geometric structures.

[0094] Screening and Optimization: To improve matching accuracy and reliability, the initial matching point pairs are usually screened and optimized. For example, incorrectly matched point pairs can be removed through consistency checks and distance constraints to ensure the final feature matching point pairs are of high quality.

[0095] The feature matching point pairs obtained through the above steps can be used for subsequent point cloud registration, 3D reconstruction, or other related applications. These feature matching point pairs can effectively reflect the geometric similarity between point clouds and provide an important basis for accurate point cloud registration and analysis.

[0096] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A mountain deformation identification method based on satellite remote sensing, characterized in that: include: Obtain satellite remote sensing images of the study area at different times, perform image enhancement processing on the satellite remote sensing images, synthesize the enhanced satellite remote sensing images into point cloud data, and perform target extraction to obtain mountain point cloud data; The mountain point cloud data is subjected to difference processing to generate three-dimensional target point cloud models of different periods, the three-dimensional target point cloud models of different periods are matched, and the point cloud change area is obtained, and the point cloud change area is used as the mountain deformation area.

2. The mountain deformation identification method based on satellite remote sensing according to claim 1 is characterized in that: Performing image enhancement processing on the satellite remote sensing image includes: Inputting the satellite remote sensing image into an image enhancement model to obtain an enhanced satellite remote sensing image; the image enhancement model is trained using a training set; the training set includes: the original satellite remote sensing image preprocessed by histogram equalization; The image enhancement model uses convolutional neural networks of different scales to extract the remote sensing image feature matrix of the satellite remote sensing image, and fuses them to generate the enhanced satellite remote sensing image.

3. The mountain deformation identification method based on satellite remote sensing according to claim 2 is characterized in that: The image enhancement model includes: A feature extraction module is used to input the satellite remote sensing image into a convolutional neural network with two-dimensional convolution kernels of different scales to obtain remote sensing image feature matrices of different scales; A feature fusion module is used to fuse the remote sensing image feature matrices of different scales to obtain a multi-scale remote sensing image feature matrix; The image enhancement module is used to input the multi-scale remote sensing image feature matrix into an image generator based on a generative adversarial network to obtain the enhanced satellite remote sensing image.

4. The mountain deformation identification method based on satellite remote sensing according to claim 3 is characterized in that: Obtaining the remote sensing image feature matrix of different scales includes: Each layer of the convolutional neural network using the two-dimensional convolution kernels of different scales performs feature processing on the input satellite remote sensing image in the forward pass: The satellite remote sensing image is convolved based on a two-dimensional convolution kernel to generate a convolution feature map, the convolution feature map is mean pooled along the channel dimension to generate a pooled feature map, the eigenvalues ​​of each position in the pooled feature map are nonlinearly activated to generate an activated feature map, and the last layer of the convolutional neural network outputs the remote sensing image feature matrices of different scales.

5. The mountain deformation identification method based on satellite remote sensing according to claim 1, characterized in that: Acquiring the mountain point cloud data includes: Synthesizing the enhanced satellite remote sensing image into point cloud data; wherein the point cloud data is in TXT format and includes coordinate information and color information; Perform deep semantic segmentation-clustering point cloud extraction on the point cloud data to obtain mountain point cloud data.

6. The mountain deformation identification method based on satellite remote sensing according to claim 1, characterized in that: Obtaining the point cloud change area includes: Performing a rigid body transformation operation on the three-dimensional target point cloud models at different times to achieve rough registration of the two three-dimensional point cloud models, wherein the rigid body transformation operation includes: translation and rotation; The three-dimensional target point cloud model after coarse registration is precisely registered, and the precisely registered three-dimensional target point cloud model is mapped to a three-dimensional reference coordinate system. A resampling grid is established based on the reference point cloud, and the two three-dimensional point cloud models are resampled under the resampling grid to obtain resampled point clouds. The resampled point cloud is subjected to difference processing to calculate the Z value change of the same grid at different periods, obtain a difference image, and process the difference image to obtain the change area of ​​the point cloud model.

7. The mountain deformation identification method based on satellite remote sensing according to claim 6, characterized in that: Using the feature point cloud to perform accurate registration includes: From the three-dimensional target point cloud model after the rough registration, overlapping three-dimensional point clouds are respectively selected as the target point cloud and the reference point cloud, and the curvature of each point in the target point cloud and the reference point cloud is calculated. According to the curvature similarity, feature matching point pairs are obtained. The feature matching point pairs are used to achieve accurate registration of the target point cloud and the reference point cloud using an iterative closest point matching algorithm.

8. The mountain deformation identification method based on satellite remote sensing according to claim 7, characterized in that: The feature matching point pairs obtained according to the curvature similarity include: Using a three-dimensional grid method to spatially divide the target point cloud and the reference point cloud, determine the neighborhood of each point, perform local parabola fitting on the neighborhood, and obtain the principal curvature; According to the principal curvature, curvature matching is used to perform matching selection to obtain the feature matching point pairs.

Images