Method for establishing effective data sample group based on SAR-optical image matching

Abstract

The application discloses a kind of based on SAR-Optical image matching's establishment method of effective data sample group, obtains SAR image and optical remote sensing image pair;SAR image and optical remote sensing image pair are carried out rectangular template area matching, in matching, it is not necessary to select measured data manually using naked eye, reduce the manual screening classification error, improve the data set production efficiency, reduce time and labor cost, and more can adapt to current based on deep learning matching algorithm;Finally, all effective SAR image and optical remote sensing image pair screened out are calibrated alignment, generate based on SAR-Optical image matching's effective data sample group, and remove invalid data by threshold method, retain effective data, improve the effectiveness of data, solve how to automatically generate effective data set, reduce the technical problem of artificial data processing cost.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a method for establishing an effective data sample group based on SAR-Optical image matching. BACKGROUND

[0002] An image is the most important form of natural scene information in a computer, and therefore image processing technology has always been a focus and hotspot of theoretical research and practical application. Generally, a single sensor can only obtain some aspect of information of a scene, while multiple sensors can obtain multiple aspects of information of a scene. For example, a visible light image can obtain spectral reflection information of a scene in a visible light band, a remote sensing image can obtain a ground feature of a photographed area, and a SAR image can obtain an all-weather radar image of a photographed area. Heterogeneous image matching and image fusion are technical foundations for comprehensively utilizing complementary image information collected by multiple sensors. Heterogeneous image matching refers to a process of aligning multiple images of a same scene collected by different sensors in a pixel space, and heterogeneous image fusion refers to a process of comprehensively combining complementary information in heterogeneous images. Due to the foundational position of heterogeneous image matching and fusion, they are widely applied in various military and civilian fields, such as navigation guidance, remote sensing scene analysis, target recognition, and the like.

[0003] In a heterogeneous image matching task, matching of SAR (synthetic aperture radar) and visible light (Optical) images is the most representative and has important military application prospects. Due to fundamental differences in imaging principles between a synthetic aperture radar and an optical image, there are complex nonlinear radiation differences between images of the two. Therefore, traditional matching methods based on feature points and based on templates often cannot achieve ideal results.

[0004] Moreover, the following technical problems exist in the prior art:

[0005] 1) In the basic flow of SAR-Optical dataset production, a large amount of manual intervention is still required, especially in the part of effective area screening. The manual checking method is not only low in efficiency but also high in cost;

[0006] 2) In addition, the existing SAR-Optical dataset is manually screened by people, and people cannot completely understand the features and areas that need to be learned by the network, so the data manually screened by people often cannot adapt to the existing neural network;

[0007] 3) Considering that matching algorithms in recent years have gradually changed from feature dependence to data dependence, the size of the data scale sometimes determines the performance of the matching algorithm. The existing data cleaning algorithms are slow in speed, occupy a lot of resources, and can only process single data. These problems greatly limit the development of the size of the dataset. SUMMARY

[0008] The application aims to provide a method for establishing an effective data sample group based on SAR-Optical image matching, so as to solve the technical problem of how to automatically generate an effective data set and reduce the cost of manual data processing.

[0009] The application adopts the following technical solutions:

[0010] The application provides a method for establishing an effective data sample group based on SAR-Optical image matching, comprising:

[0011] Obtaining a SAR image and an optical remote sensing image pair;

[0012] Performing template matching on the SAR image and the optical remote sensing image pair to obtain an effective SAR image and optical remote sensing image pair;

[0013] Calibrating and aligning all the effective SAR image and optical remote sensing image pairs screened out to generate an effective data sample group based on SAR-Optical image matching.

[0014] Optionally, the template matching on the SAR image and the optical remote sensing image pair comprises:

[0015] Randomly selecting a first rectangular template region and a second rectangular template region on the SAR image according to a template matching algorithm;

[0016] Calculating a first length offset and a first width offset of the first rectangular template region and the second rectangular template region in a first preset offset direction;

[0017] Obtaining an effective SAR image and optical remote sensing image pair according to the first length offset and the first width offset.

[0018] Optionally, the template matching on the SAR image and the optical remote sensing image pair further comprises:

[0019] Randomly selecting a third rectangular template region and a fourth rectangular template region on the optical remote sensing image according to a template matching algorithm;

[0020] Calculating a second length offset and a second width offset of the third rectangular template region and the fourth rectangular template region in a second preset offset direction;

[0021] Determining a current effective SAR image and optical remote sensing image pair according to the second length offset and the second length offset.

[0022] Optionally, the obtaining of the effective SAR image and optical remote sensing image pair comprises:

[0023] determining whether a first difference between the first length offset and the second length offset is greater than a first threshold value;

[0024] determining whether a second difference between the second length offset and the second length offset is greater than a second threshold value;

[0025] if the first difference is less than the first threshold value and the second difference is less than the second threshold value, retaining the current SAR image and optical remote sensing image pair as a valid SAR image and optical remote sensing image pair;

[0026] Optionally, if the first difference is not less than the first threshold value and the second difference is not less than the second threshold value, adjusting the size of the template region or the size of the different threshold values.

[0027] Optionally, the calibration alignment of all the screened valid SAR image and optical remote sensing image pairs comprises:

[0028] extracting the ground feature in the valid SAR image and the valid optical remote sensing image, respectively;

[0029] calculating the alignment degree between the ground features in the valid SAR image and the valid optical remote sensing image;

[0030] screening a plurality of valid SAR image and optical remote sensing image pairs according to the alignment degree, and generating an effective data sample group based on SAR-Optical image matching.

[0031] Optionally, the calculation method of the alignment degree comprises:

[0032]

[0033] wherein, I1 represents the ground feature of the valid SAR image, I2 represents the ground feature of the valid optical remote sensing image, represents the pixel variance of the I1 image, represents the pixel variance of the I2 image, represents the pixel expected variance of I2 to I1, represents the pixel expected variance of I1 to I2, and CI(I1, I2) represents the alignment degree between the ground features in the valid SAR image and the valid optical remote sensing image.

[0034] Optionally, the SAR image and optical remote sensing image pair comprises:

[0035] performing data format conversion on the SAR image and the optical remote sensing image;

[0036] randomly sampling the SAR image and the optical remote sensing image after data format conversion;

[0037] The SAR image and the optical remote sensing image collected by random sampling are subjected to de-duplication processing to generate a SAR image and an optical remote sensing image pair.

[0038] Embodiment two of the present application provides a device for establishing an effective data sample group based on SAR-Optical image matching, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements a method for establishing an effective data sample group based on SAR-Optical image matching as any one of the above method embodiments when executing the computer program.

[0039] The present application has the following beneficial effects: 1) SAR images and optical remote sensing image pairs are obtained using the mature Google Earth Engine on the market, which is convenient to use, easy to master, and has wide social recognition;

[0040] 2) The SAR image and the optical remote sensing image pair are subjected to template matching, and the selection of the measured data by the naked eye is not required in the matching, which reduces the manual screening and classification errors, improves the data set production efficiency, reduces the time and labor cost, and is more suitable for the current matching algorithm based on deep learning.

[0041] 3) All effective SAR images and optical remote sensing image pairs selected are calibrated and aligned to generate an effective data sample group based on SAR-Optical image matching, and invalid data is removed by threshold method to retain effective data and improve the effectiveness of the data. BRIEF DESCRIPTION OF DRAWINGS

[0042] Figure 1 A method for establishing an effective data sample group based on SAR-Optical image matching is provided for embodiment one of the present application, and a step flowchart is shown in the figure;

[0043] Figure 2 A method for establishing an effective data sample group based on SAR-Optical image matching is provided for embodiment one of the present application, and a step flowchart is shown in the figure;

[0044] Figure 3 A rectangular template region setting diagram is provided for embodiment one of the present application;

[0045] Figure 4 An image calibration method based on a ground control point (GCP) library is provided for embodiment one of the present application, and a step diagram is shown in the figure;

[0046] Figure 5 A template matching algorithm network structure diagram is provided for embodiment one of the present application;

[0047] Figure 6A SAR-optical image matching-based effective data sample group establishment transpose schematic diagram provided for the second embodiment of the present application. DETAILED DESCRIPTION

[0048] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] 1. The first embodiment of the present application provides a method for establishing an effective data sample group based on SAR-optical image matching, which combines Figure 1 and Figure 2 The method comprises the following steps:

[0050] Step 101: Obtain a pair of SAR images and optical remote sensing images.

[0051] Optionally, obtaining a pair of SAR images and optical remote sensing images comprises:

[0052] Converting the data format of the SAR images and the optical remote sensing images.

[0053] Randomly sampling the SAR images and the optical remote sensing images after data format conversion.

[0054] De-duplicating the SAR images and the optical remote sensing images collected by random sampling to generate a pair of SAR images and optical remote sensing images.

[0055] The specific implementation of step 101 is: downloading the required SAR images and optical remote sensing images from Google Earth Engine, or obtaining the SAR images and optical remote sensing images of a preset city or region from a public dataset composed of public radar satellite sampling data on the network in order to expand the data source, and performing data format conversion to convert them into easy-to-handle data.

[0056] Specifically, for the downloaded public dataset, all the data need to be adjusted to a unified format for easy processing.

[0057] Taking Sentinel-1 and Sentinel-2 data as an example, converting the data format means converting the data into GeoTiffs data format images, using the Export.image.toDrive function of GEE to export the downloaded SAR image data or optical remote sensing image to generate GeoTiff data; and controlling the gray value of the GeoTiff data within the range of ±2.5σ, normalizing the pixel value to the interval [0, 1] to represent a relatively large range, and if there are multiple bands, the above correction operations are performed on all bands. After the above operations, the data format conversion of the SAR images and the optical remote sensing images is completed. After data format conversion, random sampling and selection of the required ground object area data are performed.

[0058] To remove the random sampling data with overlap, a function to set the size of the trimmable sampling data is set for each random sampling using the ee.ImageCollection.mosaic() function and the ee.Image.clip() function built in GEE. In short, the ee.ImageCollection.mosaic() function is used to process those partially overlapping images, remove invalid SAR images and optical remote sensing images, and improve the heterogeneity of matching.

[0059] For the images obtained from Google Earth Engine, first, sampling is performed, de-duplication is performed, the format is unified with the public data set, and the available data is integrated and screened out.

[0060] After downloading the required images from Google Earth Engine, the region needs to be set for random sampling, the sampling seed is set, 100 points are randomly sampled from the ground in the pre-set city or the required area, and 50 points are selected from the pre-set city area. The shape details of different land or city areas are provided by the public domain Geodata service, and the sampling ratio is 1:50m.

[0061] If the positions of the two points obtained are quite close, de-duplication needs to be performed to remove one of the points, so as to ensure that there is no overlapping part during sampling, and to retain SAR images and optical remote sensing image pairs without overlap. For the images obtained by sampling, image de-duplication is performed, and the ee.ImageCollection.mosaic() function and the ee.Image.clip() function built in GEE described above are used for processing.

[0062] After sampling is completed, the Google Earth Engine data is unified in format, and the public data set data and the Google Earth Engine data are integrated, and the unusable data is screened. For example, the effective data sample group based on SAR-optical image matching needs the cloud coverage of the Sentinel-2 image to be less than 1% and all the VV-IW bands of the Sentinel-1 to be available. The images are filtered according to the above conditions using the tools of GEE, so as to screen and process the sampling data. If there is data that does not meet the conditions, it is removed.

[0063] Step 102, template matching is performed on the SAR image and optical remote sensing image pair to obtain an effective SAR image and optical remote sensing image pair;

[0064] In an embodiment, it is to be noted that the data of the cloud coverage of the granules of the Sentinel-2 is only a global shared parameter, and filtering the SAR image and the optical remote sensing image according to the parameter is not accurate, and there are a large number of image data in the whole granule, including a large number of cloud-shielded image data, only a small part of the cloud coverage local image data, and these are invalid data, and only a small part of the data is valid data without any shielding. The template matching method is used to set a rectangular template area in the SAR image and the optical remote sensing image, it is to be noted that the template area contains the target ground feature, the offset amount of each template in the SAR image and the optical remote sensing image is judged, and the valid SAR image and optical remote sensing image pair is filtered. Thus, the template matching method is used to remove the large invalid area and the image with serious cloud coverage in the SAR image and the optical remote sensing image, the coarse matching of the SAR image and the optical remote sensing image is realized, the valid data is preliminarily filtered, and the data cleaning effect is achieved.

[0065] Optionally, the template matching on the SAR image and the optical remote sensing image pair comprises:

[0066] According to the template matching algorithm, a first rectangular template area and a second rectangular template area are randomly selected on the SAR image;

[0067] The first rectangular template area and the second rectangular template area have a first length offset and a first width offset in a first preset offset direction;

[0068] According to the first length offset and the first width offset, a valid SAR image and optical remote sensing image pair are obtained.

[0069] Optionally, the template matching on the SAR image and the optical remote sensing image pair further comprises:

[0070] According to the template matching algorithm, a third rectangular template area and a fourth rectangular template area are randomly selected on the optical remote sensing image;

[0071] The third rectangular template area and the fourth rectangular template area have a second length offset and a second width offset in a second preset offset direction;

[0072] According to the second length offset and the second length offset, a current valid SAR image and optical remote sensing image pair are determined.

[0073] Optionally, obtaining the valid SAR image and optical remote sensing image pair comprises:

[0074] It is judged whether a first difference value of the first length offset and the second length offset is greater than a first threshold value;

[0075] It is judged whether a second difference value of the second length offset and the second length offset is greater than a second threshold value;

[0076] If the first difference is less than the first threshold value and the second difference is less than the second threshold value, the current SAR image and optical remote sensing image pair is reserved as a valid SAR image and optical remote sensing image pair.

[0077] Optionally, if the first difference is not less than the first threshold value and the second difference is not less than the second threshold value, the size of the template region or the size of the different threshold values is adjusted.

[0078] In an embodiment, as shown in FIG. 1, two rectangular template regions are first randomly set in the SAR image to be detected, for example, region A is a first rectangular template region, and region B is a second rectangular template region, wherein region B has a certain offset relative to region A in a preset direction. For example, in a two-dimensional coordinate system, region B has a first length offset Δx in the horizontal direction and a first width offset Δy in the vertical direction relative to region A. Corresponding templates are set in the optical remote sensing image in the same way, for example, region A' is a third rectangular template region, and region B' is a fourth rectangular template region. Region B' has a certain offset relative to region A' in a preset direction. For example, in a two-dimensional coordinate system, region B' has a second length offset Δx' in the horizontal direction and a second width offset Δy' in the vertical direction relative to region A'. Figure 3 To this end, a first difference between the first length offset and the second length offset between the SAR image and the optical remote sensing image is further calculated, i.e., whether (Δx-Δx') is greater than a first threshold value set in the template matching algorithm; at the same time, a second difference between the first width offset and the second width offset between the SAR image and the optical remote sensing image is determined, i.e., whether (Δy-Δy') is greater than a second threshold value set in the template matching algorithm. Only when the first difference is less than the first threshold value and the second difference is less than the second threshold value, it is indicated that the SAR image and the optical remote sensing image currently processed are a valid SAR image and optical remote sensing image pair.

[0079] If the first difference is not less than the first threshold value and the second difference is not less than the second threshold value, the size of the template region or the size of the different threshold values is adjusted.

[0080] It should be noted that, generally, if the region has very high adaptability, Δx, Δy and Δx', Δy' will be very close. If the adaptability of the region is low, Δx and Δx', Δy and Δy' will show a large deviation. By setting a certain threshold value for the deviation of Δx, Δy and Δx', Δy', most of the non-adaptive regions can be screened out.

[0081]

[0082] ​It should be noted that the size and position of the rectangular template region are not fixed during the entire cleaning process, and can be constantly changed to adapt to different invalid region screening requirements and image types. For example, for a small invalid region on the data to be processed, a large rectangular template region may not be able to detect the existence of the invalid region, or the calculated offset is very small and does not meet the threshold. Therefore, an additional coarse-to-fine rectangular template region setting strategy can be added.

[0083] In the coarse-to-fine rectangular template region setting strategy, a pair of rectangular template regions with large sizes are first set as sliding windows, which can be adjusted in size according to the size of the region where the feature to be detected is located, and can be slid on the entire reference image to contain all the features to be detected, so as to more accurately calculate the offset of the rectangular template region in the SAR image and the optical remote sensing image.

[0084] Through the template detection method, all SAR images and optical remote sensing images are detected multiple times to increase the screening accuracy of the data.

[0085] And by setting multiple thresholds in this embodiment, the data set can be constructed and adjusted according to the needs of the network. In the entire data set construction process, data cleaning is used repeatedly for the same batch of data, and all inappropriate images are gradually removed to retain appropriate images, thereby effectively replacing the manual intervention step in the data set preparation process and realizing the automation of the large-scale data set preparation process.

[0086] Step 103, calibrate and align all the effective SAR image and optical remote sensing image pairs screened out, and generate effective data sample groups based on SAR-Optical image matching.

[0087] In one embodiment, it should be noted that the SAR image and optical remote sensing image pair screened out by the above-mentioned step 102 is formed by the backscattering of the ground object target, and the image presents a curved surface image information of a region. In order to better align the SAR image and the optical remote sensing image, a ground control point (GCP) library-based image calibration method is first used to calibrate the SAR image and the optical remote sensing image, and the specific process is as follows: Figure 4As shown, the image to be calibrated is taken as the input of the calibrator, the optical remote sensing image of the target area obtained through the (GCP) library, and the center point position of the optical remote sensing image can be directly extracted; in addition, since the SAR image presents the curved image information of the ground surface of the target area, the SAR orbit parameters and imaging parameters need to be obtained, the SAR image is unfolded, the position information of the four corner points of the unfolded SAR image is calculated, the position of the center point of the SAR image is calculated according to the position information of the four corner points, the center point position of the SAR image is calibrated with the center point position of the optical remote sensing image in the (GCP) library, the affine transformation and resampling are used to extract the initial matching area, and it is ensured that the same target area image or the same ground feature can be found from the two images; finally, the center point position of the SAR image and the initial matching area of the optical remote sensing image in the (GCP) library are cross-correlated for fine matching.

[0088] Optionally, the calibrated and aligned all effective SAR images and optical remote sensing image pairs are generated to generate an effective data sample group based on SAR-Optical image matching, which includes:

[0089] The ground features in the effective SAR image and the effective optical remote sensing image are extracted respectively;

[0090] The alignment degree between the ground features in the effective SAR image and the effective optical remote sensing image is calculated.

[0091] According to the alignment degree, a plurality of effective SAR image and optical remote sensing image pairs are selected to generate an effective data sample group based on SAR-Optical image matching.

[0092] In an embodiment, as shown in Figure 5 based on the twin network architecture, the selected effective SAR image and optical remote sensing image pair is taken as the input of the rectangular template area matching algorithm network, and the same CNN feature extractor is used to extract the ground features of the effective SAR image and the effective optical remote sensing image respectively; after the ground feature map of the SAR image and the ground feature map of the optical remote sensing image are determined, the alignment degree between the ground features of the SAR image and the effective optical remote sensing image is calculated according to the following method:

[0093]

[0094] wherein, I1 represents the ground feature of the effective SAR image, I2 represents the ground feature of the effective optical remote sensing image, represents the pixel variance of the I1 image, represents the pixel variance of the I2 image, represents the pixel expected variance of I2 to I1, CI(I1, I2) represents the alignment degree between the ground feature in the effective SAR image and the effective optical remote sensing image.

[0095] In order to realize the alignment of the SAR image and the optical image, the interactive variance is calculated to reflect the stability of the corresponding gray scale of the two images. The main idea is that if the two images are aligned, then the gray scale of one image corresponding to the pixel position of the gray scale of the other image is the most stable, that is, the variance is the smallest. Therefore, all effective SAR images and optical remote sensing image pairs that meet the preset alignment threshold requirement can be screened according to the size of the variance or the size of the alignment degree, and an effective data sample group based on SAR-Optical image matching is generated. The accurate matching of the SAR image and the optical image is realized.

[0096] It should be noted that, as Figure 5 After the ground feature of the extracted SAR image and the effective optical remote sensing image is extracted, the correlation of the extracted feature map is calculated by using Fourier convolution. Since there is a certain down-sampling operation in the feature extraction network, the correlation map also needs to be up-sampled to restore the original spatial resolution accuracy.

[0097] The scheme trains a larger model on the server side and arranges a lightweight model on the mobile side. On the server side, the scheme uses the classic ResNet network structure, and enhances the robustness and recognition of the image features through the channel attention module. The lightweight model deployed on the actual mobile side mainly uses the network structure of mobilenetv2, which has smaller parameter quantity and calculation quantity.

[0098] When calculating the correlation of the ground feature map of the SAR image and the ground feature map of the optical remote sensing image, the present application adopts a dense InforNCE loss function based on contrast learning, and a multi-granularity regularization scheme is used, which can effectively avoid the overfitting problem in the training stage.

[0099] And the correlation score is obtained by calculating the matching correct position of the ground feature map of the SAR image and the ground feature map of the optical remote sensing image during the whole training process:

[0100]

[0101] Wherein, s p is the correlation score of the ground feature map of the SAR image and the ground feature map of the optical remote sensing image at the correct matching position p, s jis the total correlation score in the matching process of the ground feature map of the SAR image and the ground feature map of the optical remote sensing image, tau is a temperature coefficient used for moderately nonlinear scaling of the correlation score at different positions, and R represents all positions of the correlation score map.

[0102] In order to speed up the matching process, Fourier (FFT) convolution is used instead of ordinary convolution. In the FFT convolution process, first, the SAR image as the convolution kernel and the optical remote sensing image as the reference are subjected to FFT convolution. When convolving, the SAR image is padded by zero to make it have the same size as the reference image. The total amount of calculation is obviously reduced compared with the operation of ordinary convolution.

[0103] The rectangular template region matching algorithm provided by the present application realizes the coarse matching of two kinds of heterogeneous images in combination with the SAR image; and further, the effective optical remote sensing image ground feature calibration alignment is realized to achieve the fine matching of the two kinds of heterogeneous images, thereby improving the final matching efficiency and precision of the model.

[0104] Embodiment two of the present application provides a kind of establishment device 500 of effective data sample group based on SAR-Optical image matching, as shown in Figure 6 The processor executes the computer program to implement the method for establishing the effective data sample group based on SAR-Optical image matching as any one of the above method embodiments.

[0105] It should be noted that the above-mentioned establishment device 500 of the effective data sample group based on SAR-Optical image matching can realize the method for establishing the effective data sample group based on SAR-Optical image matching in real time, and the method steps are consistent with those of embodiment one, which will not be repeated here.

Claims

1. A method for establishing an effective data sample group based on SAR-Optical image matching, characterized in that, include: Acquire SAR image and optical remote sensing image pairs; Template matching is performed on the SAR image and optical remote sensing image pair to obtain a valid SAR image and optical remote sensing image pair; The effective SAR image and optical remote sensing image pairs are calibrated and aligned to generate an effective data sample group based on SAR-Optical image matching; Template matching for the SAR image and optical remote sensing image pair includes: A first rectangular template region and a second rectangular template region are randomly selected on the SAR image according to the template matching algorithm; Calculate the first length offset and the first width offset between the first rectangular template area and the second rectangular template area in the first preset offset direction; The effective SAR image and optical remote sensing image pair are obtained based on the first length offset and the first width offset; Template matching for the SAR image and optical remote sensing image pair also includes: The third and fourth rectangular template regions are randomly selected on the optical remote sensing image according to the template matching algorithm. Calculate the second length offset and the second width offset between the third rectangular template region and the fourth rectangular template region in the second preset offset direction; The current valid SAR image and optical remote sensing image pair are determined based on the second length offset and the second length offset. Template matching for the SAR image and optical remote sensing image pair includes: Determine whether the first difference between the first length offset and the second length offset is greater than a first threshold. Determine whether the second difference between the second length offset and the second length offset is greater than the second threshold. If the first difference is less than the first threshold and the second difference is less than the second threshold, then the current SAR image and optical remote sensing image pair are retained as valid SAR image and optical remote sensing image pairs.

2. The method for establishing an effective data sample group based on SAR-Optical image matching as described in claim 1, characterized in that, If the first difference is not less than the first threshold and the second difference is not less than the second threshold, then adjust the size of the template region or the size of different thresholds.

3. The method for establishing an effective data sample group based on SAR-Optical image matching as described in claim 1, characterized in that, The calibration and alignment of all selected valid SAR and optical remote sensing image pairs includes: Ground feature extraction is performed on effective SAR images and effective optical remote sensing images, respectively. Calculate the alignment between ground features in the effective SAR image and the effective optical remote sensing image; Based on the alignment, several pairs of valid SAR images and valid optical remote sensing images are selected to generate a group of valid data samples based on SAR-Optical image matching.

4. The method for establishing an effective data sample group based on SAR-Optical image matching as described in claim 3, characterized in that, The alignment is calculated using the following methods: Wherein, I1 represents the ground feature of the effective SAR image, and I2 represents the ground feature of the effective optical remote sensing image. This represents the pixel variance of the I1 image. This represents the pixel variance of the I2 image. This represents the expected variance of I2 pixels relative to I1. CI(I1,I2) represents the expected variance of pixels between I1 and I2, and CI(I1,I2) represents the alignment between ground features in the effective SAR image and the effective optical remote sensing image.

5. The method for establishing an effective data sample group based on SAR-Optical image matching as described in claim 1, characterized in that, Acquiring SAR image and optical remote sensing image pairs includes: Data format conversion is performed on the SAR image and the optical remote sensing image; Random sampling is performed on SAR images and optical remote sensing images after data format conversion; The randomly sampled SAR images and optical remote sensing images are deduplicated to generate the SAR image and optical remote sensing image pairs.

6. The apparatus for establishing an effective data sample group based on SAR-Optical image matching as described in claim 1, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a method for establishing an effective data sample group based on SAR-Optical image matching as described in any one of claims 1-5.

Images