Multi-satellite radar image-based landform deformation detection method and system

Abstract

The invention discloses a landform deformation detection method and system based on a multi-satellite radar image, and the method comprises the steps: obtaining a radar image of a satellite, employing an interpretation neural network for the radar image, obtaining an interpretation image, inputting the interpretation image and the radar image to a preset image processing model, and obtaining a deformation video, and an image analysis model is adopted for the deformation video to obtain a detection result. According to the invention, the depth radar image and the fusion interpretation image are generated according to the radar images acquired by the satellite at different positions, so that the comprehensiveness or accuracy of landform detection space information is improved; and the deformation video is synthesized according to the depth radar images and the fusion interpretation images collected by the satellite at different times, so that the timeliness and the precision of deformation detection are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, and in particular to a method and system for detecting geomorphological deformation based on multi-satellite radar images. BACKGROUND

[0002] Satellite radar-based geomorphological deformation detection plays a crucial role in many fields, such as geological disaster warning, urban planning and construction monitoring, and ecological environment assessment. However, current methods for detecting geomorphological deformation from satellite radar images, although they have made significant progress in technology, still face a series of challenges. First, because the shooting position of the satellite radar is relatively single, it can only observe the ground surface from a specific angle and height, resulting in incomplete or inaccurate capture of deformation information in some complex terrain or specific conditions. In addition, due to the satellite's orbit and shooting plan, continuous and high-frequency observation of the same area cannot be achieved, resulting in discontinuity of shooting time, which limits the timeliness and accuracy of deformation detection. SUMMARY

[0003] In order to solve the above problems existing in the prior art, the present application provides a method and system for detecting geomorphological deformation based on multi-satellite radar images. The technical problem to be solved by the present application is solved by the following technical scheme: A method for detecting geomorphological deformation based on multi-satellite radar images, comprising: obtaining a radar image of a satellite; using an interpretation neural network to obtain an interpreted image from the radar image; inputting the interpreted image and the radar image into a pre-set image processing model to obtain a deformation video; using an image analysis model to obtain a detection result from the deformation video; wherein the radar image includes radar images at multiple pre-set times and multiple pre-set locations; the deformation video includes a radar deformation video and an interpreted deformation video; the image analysis model includes a detection neural network and a decision neural network.

[0004] In one specific embodiment, the interpretation neural network includes an identification sub-neural network, a classification standard sub-neural network, and a feature fusion sub-neural network connected in sequence; wherein the identification sub-neural network includes a plurality of identification convolution units and an identification connection unit connected in sequence; the identification convolution unit includes a plurality of convolution kernels, a pooling kernel, and a RELU activation function connected in sequence; the identification connection unit includes a fully connected kernel, a RELU activation function, and a fully connected kernel connected in sequence; The classification standard sub-neural network comprises a plurality of classification convolution units and a classification connection unit connected in sequence; the classification convolution unit comprises a plurality of convolution kernels, a batch normalization, a RELU activation function and a pooling kernel connected in sequence; and the classification connection unit comprises a Flatten, a softmax activation function, a full connection kernel, a RELU activation function and a full connection kernel connected in sequence. The feature fusion sub-neural network comprises a plurality of fusion convolution units and a fusion connection unit connected in sequence; the fusion convolution unit comprises a plurality of convolution kernels, an up-sampling function, a RELU activation function and a pooling kernel connected in sequence; and the fusion connection unit comprises a MISH activation function, a full connection kernel, a SIGMOID activation function and a full connection kernel connected in sequence.

[0005] In one specific embodiment, the method comprises: obtaining a depth radar image according to the radar images at a plurality of preset positions under the same preset time; obtaining a fusion interpreted image according to the interpreted images at a plurality of preset positions under the same preset time; obtaining a radar deformation video by using difference filtering on the depth radar images at a plurality of preset times; obtaining an interpreted deformation video by using difference filtering on the fusion interpreted images at a plurality of preset times.

[0006] In one specific embodiment, the method comprises: obtaining a plurality of binocular depth radar images by using binocular depth estimation on the radar images at a plurality of preset positions under the same preset time; obtaining a corresponding standard binocular depth radar image by using distortion correction and image alignment on each binocular depth radar image; obtaining a depth radar image by using image fusion filtering on a plurality of standard binocular depth radar images, wherein the image fusion filtering formula is: , pix_out is the output pixel of the fusion filtering, pix_in is the input pixel of the fusion filtering, coef is the spatial filtering parameter, wgt is the weight parameter of the input image, i and j are the row and column coordinates of the image, k is the input image index, and r is the filtering radius.

[0007] In one specific embodiment, the method comprises: The distortion correction and image alignment are performed on the interpreted images of each preset position at the same preset time to obtain corresponding standard interpreted images; The image fusion filtering is performed on the standard interpreted images to obtain a fused interpreted image.

[0008] In one embodiment, the difference filtering is performed on the depth radar images of the plurality of preset times to obtain a radar deformation video, including: The difference filtering is performed on the depth radar images or the depth interpreted images of two preset times to obtain a deformation image; The image interpolation is performed on two deformation images to obtain a stage deformation video; The space-time filtering is performed on the time-coincidence videos in the plurality of stage deformation videos to obtain a stage filtered deformation video; The depth radar video or the depth interpreted video is obtained according to the stage deformation video and the stage filtered deformation video; wherein, , pix_dst is the output pixel of the stage filtered deformation video, pix_ori is the input pixel of the coincidence video, para is the spatial filtering parameter, i and j are the row and column coordinates of the image, k is the frame number of the coincidence video, t1, t2 and t3 are the filtering radii.

[0009] In one embodiment, the deformation video is subjected to image analysis modeling to obtain a detection result, including: The detection neural network is used on the radar deformation video to obtain deformation data; The detection neural network is used on the interpreted deformation video to obtain object deformation data; The decision neural network is used on the deformation data and the object deformation data to obtain a detection result.

[0010] In one embodiment, the detection neural network includes a plurality of detection convolution units and a fully connected layer connected in sequence; wherein the detection convolution unit includes a convolution kernel, a maximum pooling kernel, a RELU activation function and a Min-Max normalization function connected in sequence.

[0011] In one embodiment, the decision neural network includes a plurality of decision convolution units and a fully connected layer connected in sequence; wherein the decision convolution unit includes a plurality of convolution kernels, a Leaky ReLU activation function, a plurality of convolution kernels and an Absolute activation function connected in sequence.

[0012] In one embodiment, a topographic deformation detection system based on multi-satellite radar images, including: An acquisition unit is configured to acquire a radar image of a satellite; An interpretation unit is configured to obtain an interpreted image by using an interpretation neural network on the radar image; A processing unit is configured to input the interpreted image and the radar image into a preset image processing model to obtain a deformation video; A detection unit is configured to obtain a detection result by using an image analysis model on the deformation video; The radar image includes radar images of multiple preset times and multiple preset positions; the deformation video includes a radar deformation video and an interpreted deformation video; and the image analysis model includes a detection neural network and a decision neural network.

[0013] The present application has the following advantages: The present application provides a method and system for detecting landform deformation based on multiple satellite radar images. First, depth radar images and fused interpreted images are generated based on radar images collected by a satellite at different positions, which improves the comprehensiveness or accuracy of spatial information for landform detection. Then, deformation videos are synthesized based on depth radar images and fused interpreted images collected by the satellite at different times, which enhances the timeliness and precision of deformation detection.

[0014] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a flowchart of a method for detecting landform deformation based on multiple satellite radar images according to an embodiment of the present application; Figure 2 is a schematic diagram of an interpreted image in a method for detecting landform deformation based on multiple satellite radar images according to an embodiment of the present application; Figure 3 is a schematic diagram of a time-coincidence video in a method for detecting landform deformation based on multiple satellite radar images according to an embodiment of the present application; Figure 4 is a block diagram of a method for detecting landform deformation based on multiple satellite radar images according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] The present application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0017] Embodiment One In one embodiment, please refer to Figure 1 , Figure 1 is a flowchart of a method and system for detecting landform deformation based on multiple satellite radar images, and the specific steps are as follows: S1: Landform deformation detection plays a crucial role in various fields such as earth science, geological engineering, and environmental protection. It is not only an important means of assessing natural disaster risks and monitoring geological activities, but also a key link in protecting the ecological environment and ensuring human safety. However, traditional ground monitoring methods often struggle to meet the needs of large-scale deformation detection. Traditional monitoring methods have significant limitations, specifically, they are limited by manpower, material resources, and terrain conditions, making it difficult to achieve continuous and high-precision monitoring of large areas. This is not only because of limited manpower and material resources, but also because of complex terrain conditions that make monitoring difficult. Therefore, it is particularly important to find a method that can overcome these limitations and achieve large-scale, efficient, and non-contact monitoring. In this context, landform deformation detection based on multi-satellite radar images has emerged. This method uses radar images obtained by multiple satellites at different times and locations, using advanced techniques such as phase difference technology, time series technology, and orbit change technology to form a complete set of image data in time and space. These image data sets not only contain rich landform deformation information, but also enable continuous and high-precision monitoring of large areas.

[0018] Therefore, radar images of satellites are obtained, specifically, radar images obtained by multiple satellites using phase difference technology, time series technology, and orbit change technology at different times and locations, forming a complete set of image data in time and space.

[0019] S2: Although radar images provide a wealth of information, their form is often complex and not easy to understand and analyze directly. Therefore, we need to use interpretation neural networks to interpret radar images. Interpretation neural networks, after extensive training and learning, can identify features in radar images and convert them into interpretation images that are easier for humans to understand and analyze. In this way, we can more intuitively understand the deformation of the ground or target area. Therefore, an interpretation neural network is used to obtain an interpretation image from the radar image.

[0020] The interpretation neural network includes multiple identification sub-neural networks, multiple classification standard sub-neural networks, and multiple feature fusion sub-neural networks connected in sequence, which work together to extract and analyze rich information from radar images; the modular design of the interpretation neural network allows it to adapt flexibly to different interpretation tasks, and each sub-network is an independent module that can be optimized for specific ground features such as buildings, traffic, and vegetation.

[0021] S21: The identification sub-neural network is responsible for extracting preliminary feature maps from the original radar image. In one specific embodiment, the identification sub-neural network for building recognition includes, in sequence, 2 identification convolution units and an identification connection unit; the identification convolution unit includes, in sequence, 3 5*5 convolution kernels, an average pooling kernel, and a RELU activation function; the identification connection unit includes, in sequence, a full connection kernel, a RELU activation function, and a full connection kernel.

[0022] S22: The classification standard sub-neural network performs higher-level analysis based on the preliminary feature maps to identify different types of land cover. In one specific embodiment, the classification standard sub-neural network for urban areas includes, in sequence, 5 classification convolution units and a classification connection unit; the classification convolution unit includes, in sequence, 3-5 9*9 convolution kernels, a batch normalization, a RELU activation function, and a pooling kernel; the classification connection unit includes, in sequence, Flatten, a softmax activation function, a full connection kernel, a RELU activation function, and a full connection kernel.

[0023] S23: The feature fusion sub-neural network is responsible for fusing information from different identification and classification sub-networks to produce a more comprehensive land feature map. In one specific embodiment, the geothermal feature fusion sub-neural network includes, in sequence, 3 fusion convolution units and a fusion connection unit; the fusion convolution unit includes, in sequence, 3-5 9*9 convolution kernels, an up-sampling function, a RELU activation function, and a pooling kernel; the fusion connection unit includes, in sequence, a MISH activation function, a full connection kernel, a SIGMOID activation function, and a full connection kernel.

[0024] In one specific embodiment, please refer to Figure 2 , Figure 2 A method and system for detecting geomorphic deformation based on multi-satellite radar images generate a schematic diagram of the interpreted image, wherein input the radar image to the first identification sub-neural network to obtain a building feature map, wherein the first identification sub-neural network identifies straight lines, corner points, and structured patterns in the image, which are associated with man-made structures such as buildings and infrastructure; input the radar image to the second identification sub-neural network to obtain a traffic feature map, wherein the second identification sub-neural network identifies traffic features in the image, such as roads, bridges, and vehicles, which involve identifying specific linear features and moving objects; input the radar image to the third identification sub-neural network to obtain a forest belt feature map, wherein the third identification sub-neural network identifies forest belt features, including trees, vegetation-covered areas, and parks, etc. inputting the radar image into a fourth identification sub-neural network to obtain a crop feature map, wherein the fourth identification sub-neural network identifies an agricultural region in the image and identifies the crop feature map, such as the planting area of different crops, the growth stage, and the farmland boundary; inputting the radar image into a fifth identification sub-neural network to obtain a river feature map, wherein the fifth identification sub-neural network identifies water bodies in the image, such as rivers, lakes, and reservoirs, etc.; adopting a first classification standard sub-neural network to obtain a city feature map according to the building feature map and the traffic feature map; adopting a second classification standard sub-neural network to obtain an agro-forest feature map according to the traffic feature map, the forest belt feature map, and the crop feature map; adopting a third classification standard sub-neural network to obtain a water system feature map according to the river feature map; adopting a first feature fusion sub-neural network to obtain a first interpretation image according to the city feature map and the agro-forest feature map, wherein the first interpretation image is a dense coverage image of agro-forestry, and this interpretation image can indirectly provide the probability of soil erosion; adopting a second feature fusion sub-neural network to obtain a second interpretation image according to the city feature map, the agro-forest feature map, and the water system feature map, wherein the second interpretation image is a surface thermal degree image, and this interpretation image can indirectly provide the probability of geographical structure stability.

[0025] S3: Deformation is a dynamic process, and a single radar image or interpretation image can only reflect the deformation state at a certain moment. In order to more comprehensively understand the deformation process, we need to input the interpretation image and the radar image into a preset image processing model. This model can use the temporal correlation and spatial relationship between images to generate a video showing the deformation process. Through the deformation video, we can more intuitively observe and analyze the development trend of deformation, providing strong support for subsequent decision-making and response measures. Therefore, the interpretation image and the radar image are input into a preset image processing model to obtain a deformation video.

[0026] S31: Since radar images at different locations can capture different angles and details of the target area, providing more comprehensive information. By fusing these radar images at the same time, we can obtain a depth radar image. This image not only contains more information, but also reflects the deformation of the target area in three-dimensional space, providing a more accurate basis for subsequent analysis. Therefore, a depth radar image is obtained according to the radar images at multiple preset locations at the same preset time.

[0027] S311: Obtain a plurality of binocular depth radar images by using binocular depth estimation on the radar images of a plurality of preset positions at the same preset time; the technology is based on the difference between images of two different perspectives, calculates the parallax to calculate the depth information, the method is accurate and easy to implement, and can provide a preliminary depth radar image set for us.

[0028] S312: To ensure the accuracy and consistency of the binocular depth radar image, distortion correction and image alignment are used on each binocular depth radar image to obtain the corresponding standard binocular depth radar image.

[0029] S313: Obtain a depth radar image by using image fusion filtering on a plurality of standard binocular depth radar images, wherein the fusion filtering is shown in formula 1, and coef is a Gaussian filtering parameter, and r is 5. By integrating a plurality of standard binocular depth radar images, we remove redundancy and noise, and obtain a more smooth, accurate and comprehensive depth radar image.

[0030] S32: Obtain a fused interpretation image according to the interpretation images of a plurality of preset positions at the same preset time.

[0031] S321: Obtain a corresponding standard interpretation image by using distortion correction and image alignment on the interpretation image of each preset position at the same preset time.

[0032] S322: Obtain a fused interpretation image by using the image fusion filtering on a plurality of standard interpretation images.

[0033] S33: Deformation is a process that changes over time, so we need to perform difference filtering on a plurality of preset time points of the depth radar image. Difference filtering can capture the small changes between images, thereby generating a video showing the deformation process. Through the radar deformation video, we can intuitively observe the deformation of the target area at different time points and the development trend of the deformation. This is of great significance to us in terms of timely monitoring and early warning of natural disasters, and evaluating the safety of buildings. Therefore, difference filtering is used on the depth radar images of a plurality of preset times to obtain a radar deformation video.

[0034] S331: Obtain a deformation image by using difference filtering on the depth radar images or depth interpretation images of two preset times; The difference filtering method comprises the following steps: 1. Subtract two images and normalize to obtain a difference image; 2.Edge filtering is performed on the difference image to obtain a difference filtered image, taking into account burrs and sawteeth in the difference image. In practical applications, the selection of edge filtering parameters has an important influence on the result. Therefore, we need to adjust the filtering parameters (such as the size of the filter, the standard deviation, etc.) according to the characteristics of the image and the actual needs to obtain the best filtering effect. 3. In order to make the subjective vision of the difference filtered image more prominent, detail enhancement is performed on the difference filtered image to obtain a deformation image.

[0035] S332: Obtain a stage deformation video by using image interpolation on the two deformation images. The image interpolation algorithm calculates the inter-frame motion relationship of the original image, inserts an intermediate frame that meets the motion relationship between two frames, and thus improves the frame rate of the video. In this process, the motion vector between adjacent frames needs to be accurately estimated, and then the intermediate frame is synthesized according to the motion vector.

[0036] The number of inserted frames is determined by the time interval of the two deformation images and the predetermined frame rate.

[0037] In one specific embodiment, the interpolation algorithm first obtains a pixel-level motion vector using the MEMC algorithm and the optical flow algorithm, and obtains the pixel value of each pixel in the frame to be inserted according to the pixel-level motion vector and the time position of the inserted frame.

[0038] S333: Please refer to Figure 3 , Figure 3 A time-coincidence video in a topography deformation detection method and system based on multi-satellite radar images is obtained by using spatio-temporal filtering on the time-coincidence videos in the plurality of stage deformation videos. The para parameter setting method of formula 3 includes: 1. Traverse the coincident pixels of the two time-coincidence videos; 2. Detect the texture complexity and texture directionality of each coincident pixel; 3. If the texture complexity is high, the para parameter corresponding to the coincident pixel is a one-dimensional filter of the symmetry of the texture direction, and the filter radius is set to be large; 4. If the texture complexity is medium, the para parameter corresponding to the coincident pixel is a one-dimensional filter of the symmetry of the texture direction, and the filter radius is set to be small; 5. If the texture complexity is low, the para parameter is a Gaussian filter, and the filter radius is set to be large.

[0039] S334: Obtain the corresponding depth radar video or depth interpretation video according to the stage deformation video and the stage filtered deformation video.

[0040] S34: The same as step S33, difference filtering is adopted on the fusion interpretation image of the plurality of preset times to obtain an interpretation deformation video.

[0041] S4: Although the deformation video provides rich deformation information, it also contains a large amount of information redundancy and information noise. In order to more accurately extract the deformation information and give the detection result, we need to use an image analysis model to further process and analyze the deformation video. This model can automatically identify and extract key information in the video, such as the rate, direction, and range of deformation, and give the final detection result based on these information. This detection result will provide detailed and accurate description of the deformation phenomenon, which helps us make correct decisions and response measures. Therefore, the deformation video is processed by the image analysis model to obtain the detection result.

[0042] S41: The radar deformation video is processed by a detection neural network to obtain deformation data.

[0043] The detection neural network is composed of multiple detection convolution units and fully connected layers.

[0044] In each detection convolution unit, data is processed by a convolution kernel, a max-pooling kernel, a RELU activation function, and a Min-Max normalization function in sequence to extract deformation features in the video; these operations can gradually extract deformation features in the video and convert them into deformation data for subsequent analysis.

[0045] S42: The interpretation deformation video is processed by a detection neural network to obtain material deformation data; similar to S41, but the data source is different, so there are differences between the two detection neural network parameters.

[0046] S43: The deformation data and the material deformation data are processed by a decision neural network to obtain a detection result.

[0047] The decision neural network includes: a plurality of decision convolution units and fully connected layers connected in sequence; wherein the decision convolution unit includes: five convolution kernels, a Leaky ReLU activation function, a plurality of convolution kernels, and an Absolute activation function connected in sequence; these operations can further extract and fuse features in the data, thereby obtaining an accurate detection result.

[0048] In one specific embodiment, please refer to Figure 4 , Figure 4 is a block diagram of a landform deformation detection system based on multi-satellite radar images. A landform deformation detection system based on multi-satellite radar images includes: An acquisition unit for obtaining radar images of satellites; An interpretation unit is configured to apply an interpretation neural network to the radar image to obtain an interpretation image; A processing unit is configured to input the interpretation image and the radar image into a preset image processing model to obtain a deformation video; A detection unit is configured to apply an image analysis model to the deformation video to obtain a detection result.

[0049] The method and system for detecting landform deformation based on multi-satellite radar images can generate a depth radar image and a fusion interpretation image according to radar images collected by a satellite at different positions, improve the comprehensiveness or accuracy of spatial information of landform detection, synthesize a deformation video according to depth radar images and fusion interpretation images collected by the satellite at different times, and enhance the timeliness and precision of deformation detection.

[0050] In addition, the terms "first", "second", "third", etc. are used only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", etc. can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0051] Although the present application is described herein in conjunction with various embodiments, other variations of the disclosed embodiments can be understood and implemented by those skilled in the art with reference to the drawings, the disclosure, and the appended claims in the process of implementing the claimed present application. In the claims, the word "comprising" does not exclude other components or steps, and "one" or "an" does not exclude a plurality.

[0052] The above is a further detailed description of the present application in conjunction with specific preferred embodiments, and the specific implementation of the present application cannot be limited to these descriptions. For those skilled in the art to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be regarded as falling within the scope of protection of the present application.

Claims

1. A method for detecting landform deformation based on multi-satellite radar images, characterized in that, include: Acquire radar images from satellites; The radar image is processed by a decoding neural network to obtain a decoded image; The interpreted image and radar image are input into a preset image processing model to obtain deformation video; The deformation video was analyzed using an image analysis model to obtain the detection results. The radar images include radar images from multiple preset times and multiple preset locations; the deformation videos include radar deformation videos and interpreted deformation videos; the image analysis model includes a detection neural network and a decision neural network.

2. The method for detecting landform deformation based on multi-satellite radar images according to claim 1, characterized in that, The interpretation neural network includes a recognition sub-neural network, a classification standard sub-neural network, and a feature fusion sub-neural network connected in sequence; The recognition sub-neural network includes multiple recognition convolutional units and one recognition connection unit connected in sequence; the recognition convolutional unit includes multiple convolutional kernels, a pooling kernel, and a ReLU activation function connected in sequence; the recognition connection unit includes a fully connected kernel, a ReLU activation function, and a fully connected kernel connected in sequence. The classification standard sub-neural network includes multiple classification convolutional units and a classification connection unit connected in sequence; the classification convolutional unit includes multiple convolutional kernels, a batch normalization function, a ReLU activation function, and a pooling kernel connected in sequence; the classification connection unit includes a Flatten function, a softmax activation function, a fully connected kernel, a ReLU activation function, and a fully connected kernel connected in sequence. The feature fusion sub-neural network includes multiple fusion convolutional units and a fusion connection unit connected in sequence; the fusion convolutional unit includes multiple convolutional kernels, an upsampling function, a ReLU activation function and a pooling kernel connected in sequence; the fusion connection unit includes a MISH activation function, a fully connected kernel, a SIGMOID activation function and a fully connected kernel connected in sequence.

3. The method for detecting landform deformation based on multi-satellite radar images according to claim 1, characterized in that, The step of inputting the interpreted image and radar image into a preset image processing model to obtain the deformation video includes: A depth radar image is obtained from radar images at multiple preset locations under the same preset time. A fused interpretation image is obtained from the interpretation images at multiple preset positions under the same preset time. Differential filtering is applied to the depth radar images at the multiple preset times to obtain radar deformation video; Differential filtering is applied to the fused interpretation images at the multiple preset times to obtain the interpretation deformation video.

4. The method for detecting landform deformation based on multi-satellite radar images according to claim 3, characterized in that, The process of obtaining a depth radar image based on radar images from multiple preset locations at the same preset time includes: Multiple binocular depth radar images are obtained by using binocular depth estimation on the radar images of multiple preset locations at the same preset time. For each of the binocular depth radar images, distortion correction and image alignment are performed to obtain the corresponding standard binocular depth radar image; A depth radar image is obtained by image fusion filtering of multiple standard binocular depth radar images, wherein the fusion filtering formula is: ， pix_out represents the output pixels of the fusion filter, pix_in represents the input pixels of the fusion filter, coef represents the spatial filtering parameters, wgt represents the weight parameters of the input image, i and j represent the row and column coordinates of the image, k represents the index of the input image, and r represents the filtering radius.

5. The method for detecting landform deformation based on multi-satellite radar images according to claim 4, characterized in that, The process of obtaining a fused interpreted image based on the interpreted images at multiple preset positions under the same preset time includes: For each preset position at the same preset time, distortion correction and image alignment are applied to the interpreted image to obtain the corresponding standard interpreted image; The image fusion filter is applied to multiple standard interpreted images to obtain a fused interpreted image.

6. The method for detecting landform deformation based on multi-satellite radar images according to claim 3, characterized in that, Differential filtering is applied to the depth radar images at the multiple preset times to obtain radar deformation video, including: Deformation images are obtained by filtering and subtracting two depth radar images or depth interpretation images at two preset times; Image interpolation is used to obtain a staged deformation video from the two deformation images; Spatiotemporal filtering is applied to time-coincident videos in multiple stage deformation videos to obtain stage-filtered deformation videos; The corresponding depth radar video or depth interpretation video is obtained based on the stage deformation video and the stage filtered deformation video; in, , pix_dst represents the output pixels of the stage-filtered deformed video, pix_ori represents the input pixels of the overlapping video, para represents the spatial filtering parameters, i and j represent the row and column coordinates of the image, k represents the frame number of the overlapping video, and t1, t2, and t3 represent the filtering radii.

7. The method for detecting landform deformation based on multi-satellite radar images according to claim 1, characterized in that, The step of obtaining detection results by applying an image analysis model to the deformed video includes: Deformation data is obtained by using a detection neural network on the radar deformation video; The deformation data is obtained by using a detection neural network on the interpreted deformation video; The detection results are obtained by using a decision neural network on the deformation data and the material change data.

8. The method for detecting landform deformation based on multi-satellite radar images according to claim 7, characterized in that, The detection neural network includes multiple detection convolutional units and a fully connected layer connected in sequence; wherein, the detection convolutional unit includes a convolutional kernel, a max pooling kernel, a ReLU activation function and a Min-Max normalization function connected in sequence.

9. A method for detecting landform deformation based on multi-satellite radar images according to claim 7, characterized in that, The decision neural network includes multiple decision convolutional units and a fully connected layer connected in sequence; wherein, the decision convolutional unit includes multiple convolutional kernels, a Leaky ReLU activation function, multiple convolutional kernels and an Absolute activation function connected in sequence.

10. A terrain deformation detection system based on multi-satellite radar imagery, characterized in that, include: The acquisition unit is used to acquire radar images from the satellite; The interpretation unit is used to apply an interpretation neural network to the radar image to obtain an interpretation image; The processing unit is used to input the interpreted image and radar image into a preset image processing model to obtain the deformation video; The detection unit uses an image analysis model to obtain the detection result from the deformed video; The radar images include radar images at multiple preset times and multiple preset locations; the deformation videos include radar deformation videos and interpreted deformation videos. Image analysis models include detection neural networks and decision neural networks.

Images