Big data-based surveying and mapping platform and method

Abstract

The invention relates to the technical field of space surveying and mapping, and discloses a surveying and mapping platform and method based on big data, and the method comprises the steps: generating an initial pixel feature for a pixel of each SAR image; inputting the initial pixel feature into a first neural network, and outputting a parameter vector; generating Gabor cores through the parameter vectors to perform Gabor filtering processing on the SAR image, and obtaining filtered images with the same number as the Gabor cores; intercepting a filtering image of the target area on the auxiliary image and the main image to generate an area filtering image, inputting the generated area filtering image into a convolutional neural network, and outputting a deformation value of the target area by the convolutional neural network; a first neural network is trained to enable the first neural network to have the performance of adaptively generating a Gabor kernel for registering a current set of SAR images, then spatial differences of different SAR images are corrected through Gabor filtering, and assistance of high-precision DEM data does not need to be relied on.

Description

Technical Field

[0001] This invention relates to the field of space mapping technology, and more specifically, to a mapping platform and method based on big data. Background Technology

[0002] Because the satellite's orbit, attitude, and sensor operating conditions are not exactly the same during multiple imaging sessions, the coverage, incident angle, slant range, and other geometric relationships of each SAR image vary slightly within a certain range. Although pixels map to the same target, they may represent different parts of the target. However, we still treat them as the same ground feature observation value to calculate its multidimensional deformation, which obviously increases the error in deformation calculation.

[0003] Chinese patent application CN109035312A, entitled "A DEM-Assisted High-Precision Registration Method for SAR Images," discloses a method for extracting the initial registration offset based on feature point information and a technical solution for registration based on the influence of the offset on the primary and secondary images. Although it combines globally published SRTM-C DEM data for auxiliary registration, the update time of global DEM data often has a delay of several years, which may result in a time interval of several years with the SAR image to be processed. Furthermore, the accuracy of global DEM data is generally not high; for example, the highest accuracy of SRTM is 30 meters, which is difficult to meet the needs of high-precision SAR image registration. Moreover, it is difficult to obtain local near-time high-precision DEM data for the target area, and most target areas do not have near-time high-precision DEM data. Therefore, the aforementioned method has significant limitations. Summary of the Invention

[0004] This invention provides a mapping platform and method based on big data. By training a first neural network to adaptively generate Gabor kernels for registering a set of SAR images, and then using Gabor filtering to correct spatial differences between different SAR images, it has excellent registration performance, especially for viewing angle differences. Moreover, it does not require the assistance of high-precision DEM data. In particular, it can be trained specifically for the type of target to further improve the registration effect and reduce the interference of different scattering characteristics of different objects on the registration.

[0005] At least one embodiment of the present invention provides a mapping method based on big data, comprising the following steps:

[0006] Step 101: Extract a set of SAR images and sort them from earliest to latest according to the time they were captured;

[0007] Step 102: Generate an initial pixel feature for each pixel in the SAR image;

[0008] Step 103: Input the initial pixel features into the first neural network. The first neural network includes a spatial pattern detection layer, a convolutional layer, and a fully connected layer. The calculation formula for the spatial pattern detection layer is as follows:

[0009]

[0010]

[0011] H 0 =X (1)

[0012] u (t) =σ(W u G[X (t) ]+W u G[H (t-1) ]+b u )

[0013] r (t) =σ(W r G[X (t) ]+W r G[H (t-1) ]+b r )

[0014] C (t) =tanh(W c G[X (t) ]+W c G[r (t) ⊙H (t-1) ]+b c )

[0015]

[0016]

[0017] H (t) =(1-u (t) )⊙C (t) +u (t) ⊙H (t-1)

[0018] Where X (t) Let represent the feature matrix of the t-th initial pixel. This represents the sum of the full adjacency matrix and the identity matrix. express The degree matrix, H (t) H (t-1) Let W and Wt represent the encoded feature matrices at time steps t and (t-1), respectively; ⊙ represents the dot product, and Wt ... represent the product. u W r W c b represents the weight parameter.u b r b c Indicates the bias parameter;

[0019] The row vectors of the encoded feature matrix of the last time step of the spatial pattern discovery layer are mapped to the pixels of the SAR image to generate a composite image. The composite image is convolved by a convolutional layer to obtain an output feature map. The output feature map is input into a fully connected layer, and the fully connected layer outputs a parameter vector.

[0020] Step 104: Generate Gabor kernels using parameter vectors and perform Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels;

[0021] Step 105: Extract the filtered image of the target region from a secondary image and a primary image to generate a region filtered image. Input the generated region filtered image into a convolutional neural network, and the convolutional neural network outputs the deformation value of the target region.

[0022] Furthermore, a set of SAR images includes more than one main image, with the SAR images other than the main image being auxiliary images, and the main image serving as a reference.

[0023] Furthermore, the i-th row vector of the t-th initial pixel feature matrix is ​​the initial pixel feature of the i-th pixel in the t-th SAR image;

[0024] The size of the full adjacency matrix is ​​U*U, where U is equal to the number of pixels in the SAR image. The pixel numbers of the SAR image are incremented row by row, starting from the top left corner of the SAR image.

[0025] Furthermore, the composite image includes M channels, where M is equal to the number of components of the row vectors of the encoded feature matrix.

[0026] Furthermore, when training the first neural network, the image difference between the filtered images generated from the SAR images is calculated pairwise, and then the average of all the calculated differences is used as the loss value for backpropagation.

[0027] The value of the i-th pixel in all filtered images is used to generate the filtered feature of the i-th pixel. The distance between the filtered features of corresponding pixels in two SAR images is taken as the pixel difference. After calculating the pixel differences of all pixels, the mean value is taken as the image difference of the filtered image generated from the two SAR images.

[0028] Furthermore, when calculating pixel differences, only the pixel differences of pixels belonging to the same category as the target to be measured are calculated.

[0029] Furthermore, the number of components in the parameter vector is 5F, where F is a positive integer and the default value is 8;

[0030] The parameter vector is divided into F equal parts, each containing 5 component values, which are the five parameters of a Gabor kernel;

[0031] The five parameters are the kernel size, the standard deviation of the Gaussian function, the direction of the Gabor kernel function, the wavelength of the sine function, and the phase shift.

[0032] At least one embodiment of the present invention provides a big data-based mapping platform, comprising:

[0033] The first module extracts a set of SAR images and sorts them from earliest to latest according to the time they were captured.

[0034] The second module generates an initial pixel feature for each pixel in the SAR image;

[0035] The third module inputs the initial pixel features into the first neural network to obtain the parameter vector;

[0036] The fourth module generates Gabor kernels from parameter vectors and performs Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels.

[0037] The fifth module extracts filtered images of the target region from a secondary image and a primary image to generate a region-filtered image. The generated region-filtered image is then input into a convolutional neural network, which outputs the deformation value of the target region.

[0038] At least one embodiment of the present invention provides a storage medium storing non-transitory computer-readable instructions that, when executed by a computer, can perform the steps of the aforementioned real estate surveying method. Attached Figure Description

[0039] Figure 1 This is a flowchart of a big data-based mapping method according to the present invention;

[0040] Figure 2 The image shown is information about the two images to be registered.

[0041] Figure 3 The image shows the coherence of two images registered using the two-dimensional function method.

[0042] Figure 4 The image shows the coherence of two images registered using the aforementioned method. The white box area in the image represents the target region. Detailed Implementation

[0043] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0044] At least one embodiment of the present invention provides a mapping method based on big data, such as... Figure 1 As shown, it includes the following steps:

[0045] Step 101: Extract a set of SAR images and sort them from earliest to latest according to the time they were captured;

[0046] A set of SAR images includes more than one main image. SAR images other than the main image are called auxiliary images, and the main image serves as a reference.

[0047] Step 102: Generate an initial pixel feature for each pixel in the SAR image;

[0048] The component values ​​of the vector of initial pixel features represent the attribute values ​​of pixels in different dimensions of the SAR image, such as phase, amplitude, distance, azimuth, etc.

[0049] Step 103: Input the initial pixel features into the first neural network. The first neural network includes a spatial pattern detection layer, a convolutional layer, and a fully connected layer. The calculation formula for the spatial pattern detection layer is as follows:

[0050]

[0051]

[0052] H 0 =X (1)

[0053] u (t) =σ(W u G[X (t) ]+W u G[H (t-1) ]+b u )

[0054] r (t) =σ(W r G[X (t) ]+W r G[H (t-1) ]+b r )

[0055] C(t) =tanh(W c G[X (t) ]+W c G[r (t) ⊙H (t-1) ]+b c )

[0056]

[0057]

[0058] H (t) =(1-u (t) )⊙C (t) +u (t) ⊙H (t-1)

[0059] Where X (t) Let represent the feature matrix of the t-th initial pixel. This represents the sum of the full adjacency matrix and the identity matrix. express The degree matrix, H (t) H (t-1) Let W and Wt represent the encoded feature matrices at time steps t and (t-1), respectively; ⊙ represents the dot product, and Wt ... represent the product. u W r W c b represents the weight parameter. u b r b c Indicates the bias parameter;

[0060] The i-th row vector of the t-th initial pixel feature matrix is ​​the initial pixel feature of the i-th pixel in the t-th SAR image;

[0061] The size of the full adjacency matrix is ​​U*U, where U is equal to the number of pixels in the SAR image. The pixel numbers of the SAR image are incremented row by row, starting from the top left corner of the SAR image.

[0062] The row vectors of the encoded feature matrix of the last time step of the spatial pattern discovery layer are mapped to the pixels of the SAR image to generate a composite image. The composite image is convolved by a convolutional layer to obtain an output feature map. The output feature map is input into a fully connected layer, and the fully connected layer outputs a parameter vector.

[0063] The composite image includes M channels, where M is equal to the number of components of the row vectors of the encoded feature matrix;

[0064] Step 104: Generate Gabor kernels using parameter vectors and perform Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels;

[0065] In one embodiment of the present invention, training samples are constructed by using SAR images of the same area from different azimuths on the same radar platform.

[0066] In one embodiment of the present invention, training samples are constructed by using SAR images of the same area from different radar platforms at the same time.

[0067] During training, the image difference between the filtered images generated from the SAR images is first calculated pairwise, and then the average of all the calculated differences is used as the loss value for backpropagation.

[0068] The value of the i-th pixel in all filtered images is used to generate the filtered feature of the i-th pixel. The distance between the filtered features of corresponding pixels in two SAR images is taken as the pixel difference. After calculating the pixel differences of all pixels, the mean value is taken as the image difference of the filtered image generated from the two SAR images.

[0069] This invention trains a first neural network to adaptively generate Gabor kernels for registration of a current set of SAR images.

[0070] Furthermore, when calculating pixel differences, only the pixel differences of pixels belonging to the same category as the target to be measured are calculated. For example, if the category of the target to be measured is man-made objects, only the pixels belonging to the category of man-made objects are calculated, thereby specifically enhancing the registration performance for the target to be measured.

[0071] The parameter vector has 5F components, where F is a positive integer and the default value is 8.

[0072] The parameter vector is divided into F equal parts, each containing 5 component values, which are the five parameters of a Gabor kernel;

[0073] The five parameters are the kernel size, the standard deviation of the Gaussian function, the direction of the Gabor kernel function, the wavelength of the sine function, and the phase shift.

[0074] In this embodiment, the aspect ratio is 1, which is γ = 1 in the original definition, so that the width and height of the Gabor core are consistent.

[0075] like Figure 2 The image shown is information about the two images to be registered.

[0076] like Figure 3 The image shows the coherence of two images registered using the two-dimensional function method.

[0077] like Figure 4 The image shows the coherence of two images registered using the aforementioned method. (The white box area in the image represents the target area, and this coherence is represented by the pixel difference between the filtered images generated from the two SAR images.)

[0078] Step 105: Extract the filtered image of the target region from a secondary image and a primary image to generate a region filtered image. Input the generated region filtered image into a convolutional neural network, and the convolutional neural network outputs the deformation value of the target region.

[0079] The architecture and training of convolutional neural networks are conventional techniques and will not be elaborated here. One reference is AlexNet, where one channel of the input feature map corresponds to a region of the filtered image.

[0080] At least one embodiment of the present invention provides a big data-based mapping platform, such as... Figure 1 As shown, it includes:

[0081] The first module extracts a set of SAR images and sorts them from earliest to latest according to the time they were captured.

[0082] The second module generates an initial pixel feature for each pixel in the SAR image;

[0083] The third module inputs the initial pixel features into the first neural network to obtain the parameter vector;

[0084] The fourth module generates Gabor kernels from parameter vectors and performs Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels.

[0085] The fifth module extracts filtered images of the target region from a secondary image and a primary image to generate a region-filtered image. The generated region-filtered image is then input into a convolutional neural network, which outputs the deformation value of the target region.

[0086] At least one embodiment of the present invention provides a storage medium that stores non-transitory computer-readable instructions, which, when executed by a computer, can perform the steps of the aforementioned big data-based mapping method.

[0087] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A big data-based surveying and mapping platform, characterized in that, Includes the following steps: Step 101: Extract a set of SAR images and sort them from earliest to latest according to the time they were captured; Step 102: Generate an initial pixel feature for each pixel in the SAR image; Step 103: Input the initial pixel features into the first neural network. The first neural network includes a spatial pattern detection layer, a convolutional layer, and a fully connected layer. The calculation formula for the spatial pattern detection layer is as follows: H 0 =X (1) H (t) =(1-u (t) )⊙C (t) +u (t) ⊙H (t-1) Where X (t) Let represent the feature matrix of the t-th initial pixel. This represents the sum of the full adjacency matrix and the identity matrix. express The degree matrix, H (t) H (t-1) Let represent the encoded feature matrices at the t-th and (t-1)-th time steps, respectively; ⊙ represents the dot product, W, W u W r W c b represents the weight parameter. u b r b c Indicates the bias parameter; The row vectors of the encoded feature matrix of the last time step of the spatial pattern discovery layer are mapped to the pixels of the SAR image to generate a composite image. The composite image is convolved by a convolutional layer to obtain an output feature map. The output feature map is input into a fully connected layer, and the fully connected layer outputs a parameter vector. Step 104: Generate Gabor kernels using parameter vectors and perform Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels; Step 105: Extract the filtered image of the target region from a secondary image and a primary image to generate a region filtered image. Input the generated region filtered image into a convolutional neural network, and the convolutional neural network outputs the deformation value of the target region.

2. The big data-based mapping platform according to claim 1, characterized in that, A set of SAR images includes more than one main image. SAR images other than the main image are called auxiliary images, and the main image serves as a reference.

3. The big data-based mapping platform according to claim 1, characterized in that, The i-th row vector of the t-th initial pixel feature matrix is ​​the initial pixel feature of the i-th pixel in the t-th SAR image; The size of the full adjacency matrix is ​​U*U, where U is equal to the number of pixels in the SAR image. The pixel numbers of the SAR image are incremented row by row, starting from the top left corner of the SAR image.

4. The big data-based mapping platform according to claim 1, characterized in that, The composite image consists of M channels, where M is equal to the number of components of the row vectors of the encoded feature matrix.

5. A big data-based mapping platform according to claim 1, characterized in that, When training the first neural network, the image difference between the filtered images generated from the SAR images is calculated pairwise, and then the average of all the calculated differences is used as the loss value for backpropagation. The value of the i-th pixel in all filtered images is used to generate the filtered feature of the i-th pixel. The distance between the filtered features of corresponding pixels in two SAR images is taken as the pixel difference. After calculating the pixel differences of all pixels, the mean value is taken as the image difference of the filtered image generated from the two SAR images.

6. A big data-based mapping platform according to claim 5, characterized in that, When calculating pixel differences, only the pixel differences of pixels belonging to the same category as the target being measured are calculated.

7. A big data-based mapping platform according to claim 1, characterized in that, The parameter vector has 5F components, where F is a positive integer and the default value is 8. The parameter vector is divided into F equal parts, each containing 5 component values, which are the five parameters of a Gabor kernel.

8. A big data-based mapping platform according to claim 7, characterized in that, The five parameters are the kernel size, the standard deviation of the Gaussian function, the direction of the Gabor kernel function, the wavelength of the sine function, and the phase shift.

9. A big data-based surveying and mapping platform, characterized in that, include: The first module extracts a set of SAR images and sorts them from earliest to latest according to the time they were captured. The second module generates an initial pixel feature for each pixel in the SAR image; The third module inputs the initial pixel features into the first neural network to obtain the parameter vector; The fourth module generates Gabor kernels from parameter vectors and performs Gabor filtering on the SAR image to obtain a filtered image with the same number of Gabor kernels. The fifth module extracts filtered images of the target region from a secondary image and a primary image to generate a region-filtered image. The generated region-filtered image is then input into a convolutional neural network, which outputs the deformation value of the target region.

10. A storage medium, characterized in that, It stores non-transitory computer-readable instructions, which, when executed by a computer, can perform the steps of a real estate surveying method as described in any one of claims 1-8.

Images