Synthetic aperture radar image correction method, device, equipment and program product

Abstract

The present disclosure provides a kind of synthetic aperture radar image correction method, device, computer equipment and program product, wherein the method comprises: the echo data of the wide swath scanning mode acquisition of target satellite on synthetic aperture radar (SAR) is radiometrically calibrated, and initial calibration image is obtained;Using the deep learning network trained in advance, the initial backscattering coefficient and the incident angle carried by the pixel are used to determine the target calibration image;For each group of images in the target calibration image, the overlapping area of each group of images is divided into blocks to obtain each sub-block corresponding to each group of images;Wherein, a group of images includes two adjacent sub-aperture images;According to the target backscattering coefficient of each pixel in each sub-block, each sub-block is spliced and corrected to obtain each corrected block;According to the non-overlapping area in each group of images and each corrected block, the corrected SAR image is generated.

Description

Technical Field

[0001] This disclosure relates to the field of remote sensing technology, and more specifically, to a method, apparatus, computer equipment, and program product for correcting synthetic aperture radar images. Background Technology

[0002] Target satellite (such as High Resolution 3 satellite (abbreviated as...) The Wide Scan Coverage Mode (WSC) of Synthetic Aperture Radar (SAR) employs a multi-aperture scanning mechanism to achieve continuous imaging of a large area in a single scan. In Wide Scan Coverage Mode, image corrections such as radiometric and geometric corrections are crucial for improving image quality.

[0003] However, some existing image correction methods often suffer from low incident angle normalization accuracy and poor consistency of multi-aperture stitching during radiometric correction, which reduces the image quality of SAR images after correction. Summary of the Invention

[0004] This disclosure provides at least one method, apparatus, computer equipment, and program product for correcting synthetic aperture radar (SAR) images, in order to improve the image quality of SAR images after correction.

[0005] In a first aspect, embodiments of this disclosure provide a method for correcting synthetic aperture radar images, including:

[0006] Radiometric calibration is performed on the echo data acquired by the synthetic aperture radar (SAR) on the target satellite in wide-swath scanning mode to obtain an initial calibration image; each pixel in the initial calibration image carries the initial backscattering coefficient and incident angle.

[0007] Using a pre-trained deep learning network, a target calibration image is determined based on the initial backscattering coefficient and incident angle carried by each pixel; each pixel in the target calibration image carries a normalized target backscattering coefficient.

[0008] For each group of images in the target calibration image, the overlapping area of ​​each group of images is divided into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images;

[0009] Based on the target backscattering coefficient of each pixel in each sub-block, the sub-blocks are stitched together and corrected to obtain corrected blocks; the stitching and correction process is used to eliminate radiation abrupt changes between the sub-blocks.

[0010] Based on the non-overlapping areas in each group of images and each corrected block, a corrected SAR image is generated.

[0011] Secondly, embodiments of this disclosure also provide a correction device for synthetic aperture radar images, comprising:

[0012] The calibration module is used to perform radiometric calibration on the echo data acquired by the synthetic aperture radar (SAR) on the target satellite in wide-swath scanning mode to obtain an initial calibration image; each pixel in the initial calibration image carries an initial backscattering coefficient and an incident angle.

[0013] The normalization module is used to determine the target calibration image using a pre-trained deep learning network based on the initial backscattering coefficient and incident angle carried by the pixel; each pixel in the target calibration image carries a normalized target backscattering coefficient.

[0014] The segmentation module is used to divide the overlapping areas of each group of images in the target calibration image into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images.

[0015] The stitching correction module is used to perform stitching correction processing on each sub-block according to the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block; the stitching correction processing is used to eliminate radiation abrupt changes between the sub-blocks.

[0016] The generation module is used to generate corrected SAR images based on the non-overlapping areas in each group of images and each corrected block.

[0017] Thirdly, embodiments of this disclosure also provide a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps of the first aspect described above are performed.

[0018] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps described in the first aspect.

[0019] The synthetic aperture radar (SAR) image correction method, apparatus, computer equipment, and program product provided in this disclosure, by radiometrically calibrating the echo data in the wide-swath scanning mode of the target satellite, can obtain the initial backscattering coefficient and incident angle corresponding to each pixel, thus providing standardized input data for subsequent incident angle normalization and image stitching correction. Normalization using a pre-trained deep learning network can realize the complex relationship between the incident angle and the backscattering system based on the adaptive learning of the deep learning network. Based on the initial backscattering coefficient and incident angle of each pixel, the incident angle is accurately normalized, thereby eliminating radiometric inconsistencies caused by differences in incident angle and obtaining a radiometrically uniform target calibration image. By dividing the overlapping areas of adjacent sub-aperture images in the target calibration image into blocks and performing stitching correction processing based on the target backscattering coefficients of pixels within each sub-block, it is possible to eliminate radiometric abrupt changes between sub-apertures block by block, improving the naturalness of the radiometric transition between sub-blocks and the continuity between sub-blocks. Finally, by combining the non-overlapping regions of each sub-aperture with the corrected blocks to generate SAR images, a radiometrically consistent and seamlessly stitched wide-swath SAR image product can be obtained while preserving the details of the original image. In summary, the synthetic aperture radar image correction method provided in this application, compared with traditional correction methods that use empirical functions for backscattering coefficients and suffer from insufficient accuracy in normalization and poor strip continuity when stitching sub-aperture images, improves the backscattering coefficient normalization accuracy and the stitching continuity between overlapping regions by using a deep learning network to adaptively learn the nonlinear normalization relationship between the backscattering coefficient and the incident angle. Combined with block-level fine correction of overlapping regions of adjacent sub-aperture images, this method achieves high-quality SAR image correction with radiometric normalization and stitching consistency, significantly improving the correction effect of SAR images in the wide-swath scanning mode of target satellites.

[0020] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0022] Figure 1A flowchart illustrating a method for correcting synthetic aperture radar images provided in an embodiment of this disclosure is shown;

[0023] Figure 2 A schematic diagram of an overlapping region block division provided by an embodiment of this disclosure is shown;

[0024] Figure 3 A schematic diagram of a synthetic aperture radar image correction device provided in an embodiment of this disclosure is shown;

[0025] Figure 4 A schematic diagram of the structure of a computer device provided in an embodiment of this disclosure is shown. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0027] Furthermore, the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein.

[0028] In this article, "multiple or several" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0029] Research has found that target satellites (such as...) Synthetic Aperture Radar (SAR) in Wide Swath (WSC) mode can acquire large-scale observation data through multi-aperture imaging technology. In this mode, radar waves illuminating the same ground feature at different incident angles cause significant changes in the backscattering coefficient of the feature. Therefore, to achieve effective comparison of observations at different incident angles, the observations corresponding to each incident angle need to be corrected to the same reference angle. Traditional incident angle normalization methods often use fixed functions based on empirical formulas or lookup tables. However, since different ground features have significantly different response patterns to incident angles, traditional incident angle normalization methods cannot be well applied to all types of ground features. Moreover, a single fixed function cannot adapt to changes in external environmental conditions such as wind fields and seasons, making it difficult to uniformly characterize the incident angle response patterns of various ground features, which greatly affects the correction accuracy of traditional incident angle normalization methods. Furthermore, after the incident angle normalization is completed, since the wide-swath SAR image in WSC mode of the target satellite is formed by stitching together multiple sub-aperture images, radiation discontinuities are prone to occur in the stitched area. Therefore, consistency correction is also required for the stitched areas between sub-aperture images. However, some current methods for improving stitching consistency mostly adopt global gain / offset matching or histogram matching. These improvement methods ignore the local differences within the image, which can easily leave obvious strip-like brightness steps at the stitching seams, resulting in the overall coherence of the stitched wide-swath SAR image not being effectively improved.

[0030] Based on the above research, this disclosure provides a method, apparatus, computer equipment, and program product for correcting synthetic aperture radar (SAR) images. By adaptively learning the nonlinear normalization relationship between the backscattering coefficient and the incident angle through a deep learning network, and combining it with block-level fine correction of the overlapping areas of adjacent sub-aperture images, the normalization accuracy of the backscattering coefficient and the stitching continuity between the overlapping areas of adjacent sub-aperture images can be improved. This results in high-quality SAR image correction with radiation normalization and stitching consistency effects, significantly improving the correction effect of SAR images in the wide-swath scanning mode of target satellites.

[0031] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0032] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0033] To facilitate understanding of this embodiment, a detailed description of the synthetic aperture radar (SAR) image correction method disclosed in this disclosure embodiment will be provided first. The execution subject of the SAR image correction method provided in this disclosure embodiment is generally a terminal device or other processing device with certain computing capabilities. The terminal device can be a user equipment (UE), mobile device, terminal, personal digital assistant (PDA), handheld device, computer device, remote sensing device, etc. In some possible implementations, the SAR image correction method can be implemented by a processor calling computer-readable instructions stored in memory.

[0034] The following describes the method for correcting synthetic aperture radar images provided in this embodiment, using a remote sensing imaging device as the execution subject.

[0035] like Figure 1 The flowchart shown is a method for correcting synthetic aperture radar images according to an embodiment of this disclosure, which may include the following steps:

[0036] S101: Radiometric calibration is performed on the echo data acquired by the synthetic aperture radar (SAR) on the target satellite in wide-swath scanning mode to obtain an initial calibration image; each pixel in the initial calibration image carries the initial backscattering coefficient and incident angle.

[0037] Here, echo data refers to the unstandardized raw measurement signals received by the synthetic aperture radar on the target satellite. It records the energy information returned after the radar waves interact with ground features, and is the initial, true observation data during the imaging process. The target satellite is a synthetic aperture radar satellite that supports wide-swath scanning operation; specifically, the target satellite could be the High Resolution 3 satellite. .

[0038] Understandably, echo data can be obtained using target satellites (such as...) L1A level data collected in WSC mode, that is WSC L1A data.

[0039] Radiometric calibration refers to the process of converting the original echo signal intensity in the echo data into standardized initial backscattering coefficients with clear physical meaning, and simultaneously calculating the corresponding incident angle.

[0040] Understandably, radiometric calibration of echo data aims to eliminate the influence of non-ground object factors such as the characteristics of the SAR sensor itself and signal propagation path attenuation on the observation values, so that the initial backscattering coefficient of each pixel can truly reflect the scattering characteristics of the ground object itself, thereby providing a reliable and consistent input data foundation for subsequent deep learning normalization processing that relies on accurate radiometric characteristics.

[0041] Initial calibration image refers to a standardized data set formed after radiometric calibration. Each pixel in the initial calibration image is associated with two core physical quantities: one is the initial backscattering coefficient, which characterizes the echo intensity of ground objects, and the other is the incident angle, which describes the radar beam illumination geometry.

[0042] The initial backscattering coefficient, obtained through radiometric calibration, is a standardized coefficient and a core radiometric indicator for distinguishing different land cover types. Each pixel corresponds to one initial backscattering coefficient. The initial backscattering coefficient directly quantifies the ability of a land cover (e.g., sea ice or seawater) at the corresponding geographic location to reflect radar signals; its value directly determines the brightness or darkness of the pixel in the final SAR image. The angle of incidence is the angle between the radar beam illuminating the land cover at the corresponding geographic location and the local vertical direction (normal). It describes the geometric relationship of radar observation, and changes in the angle of incidence significantly affect the observed backscattering intensity. The backscattering coefficient and the angle of incidence together constitute the fundamental physical quantities describing the scattering characteristics of land covers.

[0043] Specifically, echo data consists of the raw measurement signals received by the satellite. Through radiometric calibration, these signals are converted into initial backscattering coefficients and corresponding incident angles with clear physical meaning. The initial backscattering coefficients and corresponding incident angles of each pixel together constitute the core content of the initial calibrated image, accurately describing the ground object scattering characteristics and observational geometry of each pixel.

[0044] In practice, for the echo data acquired by the SAR on the target satellite under WSC, based on a series of calibration parameters (such as antenna pattern, transmit power, receive gain, etc.) pre-determined by the satellite platform and SAR sensor, the original echo digital value (DN) received by each pixel is converted into an initial backscattering coefficient with a clear physical meaning (defined as...). For example, the formula Calculated Where K is a calibration constant. Simultaneously, based on the geometric model at the imaging moment, the angle between the radar beam center and the normal to the ground position of the pixel is precisely calculated, i.e., the incident angle (defined as...). Ultimately, each cell generates a corresponding... and The set of such data pairs for all pixels in the entire image constitutes the initial calibration image.

[0045] S102: Using a pre-trained deep learning network, the target calibration image is determined based on the initial backscattering coefficient and incident angle carried by the pixel; each pixel in the target calibration image carries a normalized target backscattering coefficient.

[0046] Here, a pre-trained deep learning network refers to a neural network model whose parameters have been trained and fixed, which can be used to learn the nonlinear mapping relationship between the initial backscattering coefficient and the incident angle to the normalized backscattering coefficient. In this application, the function of this network is to take the initial backscattering coefficient and the incident angle of each pixel as input and output a target backscattering coefficient that has been normalized by the incident angle.

[0047] The target calibration image is the output image result after processing by the deep learning network. The target backscattering coefficient value carried by each pixel is the result after normalization of the incident angle.

[0048] Specifically, the target backscattering coefficient recorded for each pixel in the target calibration image represents the ideal scattering intensity of the corresponding ground feature at a uniform standard incident angle. The target calibration image has eliminated the systematic radiation differences caused by different incident angles during the original observation, and can purely reflect the radiation characteristics corresponding to the inherent characteristics of the ground feature.

[0049] Furthermore, the standard incident angle is determined based on practical application requirements and the typical observation geometry of the imaging system. In this application, the standard incident angle can typically be a pre-set fixed constant value, selected to be in the middle of the commonly used incident angle range for SAR satellite wide-swath scanning modes. For example, for the wide-swath scanning mode of the Gaofen-3 satellite, the incident angle range is typically between 20 and 50 degrees, and the standard reference incident angle can be specifically set to 30 or 35 degrees. The purpose of this setting is to provide a unified and reasonable comparison benchmark for the radiometric values ​​of all pixels, enabling subsequent processing (especially image stitching) to be performed under completely consistent observation geometry conditions. The standard reference incident angle is fixed during the training phase of the deep learning network and is implicitly included in the mapping relationship learned by the network.

[0050] In practice, the initial calibration image can be input into a pre-trained deep learning network. This deep learning network will then process the data corresponding to each pixel in the initial calibration image. and As input features, through their learned nonlinear mapping relationships, a normalized target backscattering coefficient (defined as...) is calculated and output pixel by pixel. This process operates in parallel on all pixels of the entire image, ultimately generating an image with the same spatial dimensions as the initial calibrated image, but with each pixel having a different size. All have been replaced with The target calibration image. For example, a pixel with an original observation incident angle of 28 degrees and an initial backscattering coefficient of -15.0 dB, after being processed by a deep learning network, outputs a target backscattering coefficient of -16.2 dB. This new value is the backscattering intensity coefficient value that the ground object should present at the reference angle.

[0051] In one embodiment, S102 described above can be implemented according to the following steps:

[0052] S102-1: Using the input layer of a deep learning network, the initial backscattering coefficients and incident angles carried by each pixel are concatenated to obtain the input features.

[0053] Here, the input feature refers to the two-dimensional feature vector formed by concatenating the initial backscattering coefficients and incident angles corresponding to each pixel. Specifically, the input feature integrates the radiance (backscattering coefficient) describing the scattering characteristics of ground objects and the geometric quantity (incident angle) describing the observation conditions into a unified and structured representation, enabling the network to simultaneously obtain and jointly process these two key pieces of information, and then learn the complex correspondence between them.

[0054] In practice, the input layer of a deep learning network can be used to process the initial backscattering coefficients carried by each pixel. (e.g., -12.5dB) and its corresponding angle of incidence Data (e.g., 35.2°) is concatenated to obtain a two-dimensional feature vector, which serves as the input feature. Based on this, the originally dispersed radiometric and geometric information of each pixel can be integrated into a structured data unit, forming a unified format acceptable to deep learning networks. For example, for a pixel representing exposed soil, its... -10.8dB If the angle is 42.1°, then its input features are [-10.8, 42.1].

[0055] S102-2: Using the convolution module of a deep learning network, the input features are processed by multiple consecutive convolutions to obtain convolutional features; the convolution module includes multiple convolutional layers connected in sequence.

[0056] Here, a convolutional module refers to a core processing unit in a deep learning network, consisting of multiple convolutional layers connected sequentially. In this application, the convolutional module includes, but is not limited to, a first convolutional layer, a second convolutional layer, and a third convolutional layer. The first convolutional layer is used to extract and learn the trend of scattering intensity variation of the same ground object under small-range incident angle changes from the input features of the pixel and its neighboring region, i.e., to establish the basic "..." — "Local correlation model. The second convolutional layer, based on the features output by the first convolutional layer, further identifies and enhances the differential scattering responses of different land cover types (such as seawater, annual ice, and multi-year ice) to changes in incident angle, thereby modeling the nonlinear scattering characteristics related to land cover. The third convolutional layer integrates spatial context information over a wider range, paying particular attention to and extracting systematic radiation bias patterns that may be introduced by the wide-swath multi-aperture imaging method itself, providing high-order feature support for achieving accurate normalized output. These three layers work together, sequentially from local to global, from general to specific, and from representation to systematic bias, to learn the complex mapping relationship from raw observation values ​​to ideal values ​​under a standard reference angle. This ensures that the deep learning network effectively represents the incident angle-scattering nonlinear relationship while avoiding the risk of overfitting under limited sample size conditions."

[0057] Convolutional features refer to the data features generated after the input features are processed sequentially by each convolutional layer in the convolutional module.

[0058] In practice, the input features can be processed sequentially through three convolutional layers using the convolutional module of a deep learning network: the first convolutional layer can use 32 convolutional kernels of size 3×3 (i.e., kernel size 3). 3. With 32 channels, a Rectified Linear Unit (ReLU) activation function is used to extract the fundamental response characteristics of the backscattering coefficient as a function of the incident angle within a local spatial range. This allows for the study of the scattering intensity variation trend of the same ground feature under small-range incident angle variations, in order to characterize... and The local correlation between pixels yields low-order convolutional features; these features primarily reflect the fundamental variation of scattering response between local pixels (usually from the same ground feature) with the incident angle. The second convolutional layer can use 64 pixels of size 3... A 3x3 convolutional kernel is used, employing the ReLU activation function, to further model the scattering differences of different land cover types under different incident angles based on the low-order convolutional features, thereby enhancing the nonlinear response features related to land cover types and obtaining intermediate-order convolutional features. These intermediate-order convolutional features can further characterize the scattering response differences corresponding to different land cover types. The third convolutional layer can also use 64 kernels of size 3x3. A convolution kernel of size 3 is used, and the ReLU activation function is applied to further convolve the intermediate-order convolution features to fuse larger-scale feature information. The focus is on capturing systematic scattering bias features caused by multi-aperture imaging in wide-swath scanning modes, such as the overall radiation difference pattern between strips, resulting in higher-order convolution features. These higher-order convolution features capture a larger range of systematic scattering bias features related to the imaging mechanism. These higher-order convolution features are used as the convolution features of the final data in the convolution module. The features described above are essentially modifications of the original "..." — "Information is abstracted step by step from local to global and from basic to complex, thereby forming convolutional features that can support the network to accurately complete the normalization of the incident angle."

[0059] S102-3: Utilize the fully connected layers of a deep learning network to perform fully connected processing on the convolutional features to obtain fully connected features.

[0060] Here, the fully connected layer takes the convolutional features output by the convolutional module as input, and uses them to globally integrate and weight the convolutional features that characterize local angular response, differences in land cover types, and system biases, thereby establishing a system from complex features to a single physical quantity. The direct mapping relationship. Fully connected features are high-level, global feature vectors that can be directly used for the final regression calculation. These feature vectors carry the precise calculation of their mapping relationship for each pixel. All the necessary abstract information connects the front-end feature extraction with the back-end. A key bridge for output. For example, when identifying urban areas, the fully connected layer can fuse geometric features characterizing building outlines, scattering features characterizing materials, and system features characterizing imaging strip deviations to jointly calculate the backscattering intensity that a pixel should have at a standard incident angle.

[0061] In practice, the convolutional features output by the convolutional module can be input into a fully connected layer for fully connected processing to obtain a normalized backscattering coefficient representing the value of each pixel under the standard reference incident angle. Fully connected features. To ensure that fully connected features are quantities with clear physical meaning and continuous variation, fully connected layers use linear functions as their activation methods. For example, when processing a pixel in an ocean scene that mixes wave textures with scattered ice floes, the fully connected layer can comprehensively consider the local scattering features of water surface ripples extracted by the convolutional layer, the differential features of ice and water classification, and possible strip brightness shift features, ultimately accurately calculating the pixel's pure surface characteristics after removing the influence of incident angle and systematic bias. .

[0062] S102-4: Using the output layer of a deep learning network, linear activation is applied to the fully connected features to obtain the target calibration image.

[0063] Here, the output layer is used to directly generate and output each cell. The target calibration image is the output of all pixels processed by a deep learning network. The resulting image. Understandably, a target calibration image is a radar data image with consistent radiation, whose effects have been eliminated due to differences in incident angle.

[0064] In practice, fully connected features can be input into the output layer, and the output layer processes these features by applying a linear activation function. This linear activation function can be an identity mapping function, used to ensure the final output... It is a physically continuous real number with an unrestricted range, thus strictly preserving the authenticity and interpretability of the scattering coefficient as a physical quantity. After processing by this output layer, all pixels corresponding to the normalized incident angles... The values ​​together form a target calibration image.

[0065] In one embodiment, the pre-trained deep learning network in S102 above can be trained according to the following steps A1 to A4:

[0066] A1: Acquire a set of sample calibration images acquired by SAR in wide-swath scanning mode; the set of sample calibration images includes sample calibration images acquired at different times for different land cover types; each sample pixel in the sample calibration image carries the sample backscattering coefficient and sample incident angle.

[0067] Here, the sample calibration image set is a collection of radiometrically calibrated sample SAR images used to provide learning samples for deep learning networks, enabling them to learn the complex patterns of backscattering coefficient variations with incident angle and different land cover.

[0068] Specifically, the sample calibration image set not only covers various typical land cover types (such as water bodies, vegetation, bare soil, and man-made features), but also includes sample calibration images collected under different seasons (such as summer and winter) and time conditions. Furthermore, for any land cover type, it can include different subcategories of land features within that type. For example, for water bodies, various categories of land features could include open water, one-year ice, and multi-year ice in sea ice scenarios. Thus, by setting different land cover types and subdividing them into different types, and collecting sample calibration images at different times, sufficient diversity and representativeness in terms of land cover categories and observation conditions can be ensured. Each sample pixel in the sample calibration image contains two key data points: the sample backscattering coefficient and its corresponding sample incident angle.

[0069] In practice, raw echo data acquired by the target satellite in wide-swath scanning mode is obtained and radiometrically calibrated to form multiple sample calibration images. To ensure the trained deep network model has broad adaptability and robustness, the construction of the sample calibration image set must simultaneously consider the diversity of temporal distribution and land cover types: in the temporal dimension, the sample calibration image set can cover sample calibration images acquired during different observation periods (e.g., different seasons, different years); in terms of land cover types, the sample calibration image set must cover various typical land cover types that may appear in the SAR image to be calibrated. To facilitate understanding of the technical solution of this application, the following uses a polar scene as an example. Typical land cover types in this scene can include open ice, one-year sea ice, and multi-year sea ice. Specific sea ice types can be distinguished by the sea ice land cover category index c, such as c=1, 2, 3, which can correspond to open ice, one-year sea ice, and multi-year sea ice, respectively. Each sample pixel in each sample calibration image contains the sample backscattering coefficient and sample observation incident angle after radiometric calibration.

[0070] Optionally, to mitigate the impact of seasonal variations, the amount of data in the summer and winter portions of the sample calibration image set can be configured in approximately a 1:1 ratio to enhance the model's adaptability to different environmental conditions throughout the year.

[0071] A2: Using the deep learning network to be trained, the predicted calibration image and the prediction probability under various land cover types are determined based on the sample backscattering coefficient and sample incident angle carried by the sample pixels in the sample calibration image; each sample pixel in the predicted calibration image carries a normalized predicted backscattering coefficient.

[0072] Here, the predicted calibration image is a predicted image generated by the deep learning network to be trained during the training phase, based on the input sample backscattering coefficients and sample incident angles. Each sample pixel in the predicted calibration image carries a normalized predicted backscattering coefficient, which characterizes the prediction result of the deep learning network to be trained for the backscattering intensity that the corresponding sample pixel should have at the standard reference incident angle under the current parameter configuration. For each sample pixel in the predicted calibration image, there is a one-to-one corresponding sample pixel in the sample calibration image.

[0073] Prediction probability (defined as P) c P c The predicted probability (representing the probability that a sample pixel belongs to class c) is an estimate output by the deep learning network to be trained during the training phase, representing the likelihood that each sample pixel it processes belongs to each preset land cover type (e.g., vegetation, bare soil, man-made land cover, open ice, one-year sea ice, and annual sea ice). Each sample pixel corresponds to a predicted probability for each preset land cover type.

[0074] In this application, the predicted probability is not used for final image classification, but rather serves as an internal training supervision mechanism. This ensures that the deep learning network being trained can simultaneously learn and retain the essential differences between various land features in the feature space when performing the primary task of incident angle normalization, preventing the normalization operation from obscuring the inherent scattering characteristics of different land features. Furthermore, the predicted probability participates in the subsequent calculation of physical consistency constraints, helping the network determine whether its output normalized backscattering coefficient conforms to the physical laws of the corresponding land feature. The detailed process of the predicted probability participating in physical consistency constraints will be described later. For example, when the deep learning network processes a polar image containing open ice, one-year ice, and multi-year ice, the predicted probability it outputs for each pixel (e.g., a probability of 0.6 for "multi-year ice" and 0.1 for "open ice") guides the network to ensure that the normalized backscattering coefficient of "multi-year ice" pixels conforms to the range of strong scattering characteristics expected from its high-roughness surface, while the result for "open ice" pixels remains within the lower scattering value range dominated by specular reflection.

[0075] In practice, the deep learning network to be trained can adopt a multi-task output structure. After processing the input features using the deep learning network with the multi-task output structure, the data flow can be divided into two parallel branches: the main branch is responsible for performing the core incident angle normalization task and outputs a predicted calibration image of the same size as the input sample calibration image. Each sample pixel in this image carries a normalized predicted backscattering coefficient, which is the sample value of that sample pixel under the standard reference incident angle inferred by the deep learning network. Meanwhile, the auxiliary branch performs the land cover type identification task, outputting a set of P values ​​for each sample cell. c This set of probability values ​​represents the likelihood that the pixel belongs to each of the preset land cover types (e.g., open water, annual ice, multi-year ice). Both types of outputs are used in the calculation of the subsequent loss function during the training phase to jointly optimize the network parameters.

[0076] A3: Construct the prediction loss of the deep learning network to be trained based on each predicted backscattering coefficient and the prediction probability of the predicted calibration image under various land cover types.

[0077] Here, predicting the probability of the calibrated image under various land cover types specifically means predicting the probability of each sample pixel in the calibrated image under various land cover types. Prediction loss (defined as...) Loss function (FS) is a loss function used by deep learning networks during the training phase, which combines multiple constraints. It serves as a unified guiding objective for iterative optimization of network parameters, systematically guiding the learning direction of the network.

[0078] Specifically, the prediction loss ensures that, while mastering the core capability of incident angle normalization, the output of the deep learning network simultaneously satisfies the physical laws of SAR imaging, preserves the characteristics of ground object categories, and possesses good visual continuity. Furthermore, by continuously minimizing this prediction loss during network training, the deep learning network can systematically learn and optimize its incident angle normalization processing capabilities, thereby outputting more professional and accurate normalized backscattering coefficients.

[0079] In practice, the samples corresponding to each pixel in the predicted calibration image output by the deep learning network to be trained can be used as a reference. And the P of each sample pixel in the predicted calibration image under various land cover types. c Calculate the prediction loss of the network. For example, this can be done based on the label corresponding to each sample pixel. and samples The loss of the deep learning network during backscattering coefficient correction is determined based on the labeled land cover type corresponding to each sample pixel and the predicted probability P for each land cover type. c The loss of the deep learning network in land cover type prediction is determined, and the prediction loss is determined based on two parts of the loss. Among them, the label... The sample pixels are pre-labeled with the true normalized backscattering coefficients; Let be the normalized backscattering coefficients predicted by the deep learning network to be trained for the sample pixels.

[0080] A4: Use prediction loss to iteratively train the deep learning network to be trained until the preset training cutoff condition is met, and obtain the trained deep learning network.

[0081] Here, the preset training cutoff condition is used to specify when to terminate network training. It can be reasonably set according to actual training needs to adapt to the training adaptation requirements of deep learning networks in different application scenarios. For example, the preset training cutoff condition can be set as follows: the prediction loss remains below a preset threshold (i.e., the prediction loss is minimized, such as 0.001) for N consecutive rounds, and / or the number of network iterations reaches a preset maximum value (such as 600 rounds).

[0082] In practice, the prediction loss constructed in step A3 can be used to iteratively train the deep learning network to be trained. During training, the parameters of each layer within the network (including convolutional kernel parameters, activation layer parameters, etc.) can be adjusted based on the prediction loss until the training cutoff condition is met. The resulting network is then used as the trained deep learning network, which can be practically applied to subsequent incident angle normalization processing of SAR images. It should be noted that when using the trained deep learning network model to determine the target calibration image, only the normalized backscattering coefficients corresponding to the main branch are output, and the prediction probabilities corresponding to the auxiliary branches are not output.

[0083] In one embodiment, A3 above can be implemented according to the following steps:

[0084] A3-1: Determine the first loss of the deep learning network to be trained based on the predicted backscattering coefficient of each sample pixel in the overlapping area of ​​each group of sample images in the predicted calibration image.

[0085] Here, a set of sample images includes two adjacent sample sub-aperture images in the prediction calibration image. The overlapping area in a set of sample images refers to the overlapping part formed by the acquisition coverage of two adjacent sub-aperture images. For example, for a polar sea ice scene, when acquiring multiple consecutive sample sub-aperture images covering the area, these sample sub-aperture images form a sample calibration image. Every two adjacent sample sub-aperture images will have a part that is simultaneously acquired and covered. These areas are the overlapping areas in each set of sample images.

[0086] First loss (defined as) This loss reflects the loss of the deep learning network being trained when normalizing the incident angle. This loss constrains the radiometric consistency of the network, ensuring that the predicted backscattering coefficients of the network output remain stable in overlapping image regions, thus avoiding artifacts such as banding during subsequent image stitching. Specifically, This can be called the normalized principal loss based on the incident angle.

[0087] Optionally, the application of the first loss is not limited to the overlapping areas of the images. It can also be set for the entire sample image according to the actual training needs, so as to further enhance the overall radiometric consistency of the predicted calibration images output by the network.

[0088] For example, the first loss can be specifically achieved through the following formula (1):

[0089] (1)

[0090] in, This represents the first loss (i.e., the principal loss normalized to the incident angle); E[ ] is the expectation operator, used to calculate the expected value of the expression within the parentheses; (i, j) ∈ Ωoverlap represents the pair of sample pixels (i, j) in the overlapping area of ​​each group of sample images, where Ωoverlap specifically refers to the overlapping area of ​​each group of sample images, and i and j represent two different sample pixels (which can be understood as two pixels in the same physical location or adjacent pixels) and / or the pixels corresponding to a sample pixel in the overlapping area in the two adjacent sample calibration images to which it belongs. This represents the predicted backscattering coefficient of pixel i output by the deep learning network to be trained. This represents the predicted backscattering coefficient of pixel j output by the network; The square of the L2 norm is used to calculate and The square of the difference between the two measures the degree of difference between the two predicted backscattering coefficients.

[0091] In practice, sample images can be divided into groups based on the sub-aperture images of each sample in each predicted calibration image. For each sample pixel in the overlapping region of each group of sample images, the predicted backscattering coefficient corresponding to that sample element can be substituted into the above formula (1) to obtain the first loss of the deep learning network to be trained. .

[0092] A3-2: Determine the second loss of the deep learning network to be trained based on the predicted probability and the backscattering coefficient thresholds corresponding to various land cover types.

[0093] Here, the backscattering coefficient threshold is a pre-defined boundary value for the backscattering coefficient range for various land cover types, including the minimum and maximum backscattering coefficient thresholds for each land cover type. Specifically, it can be expressed as... and ,in This represents the threshold value for the maximum backscattering coefficient corresponding to land cover type c. This represents the minimum backscattering coefficient threshold corresponding to the c-th type of land cover. The backscattering coefficient threshold is used to constrain the predicted backscattering coefficients output by the deep learning network to be trained to conform to the physical scattering characteristics of various land covers, thereby ensuring the physical rationality of the network output results.

[0094] Second loss (defined as) The backscattering coefficient (RSC) is a loss term in the prediction loss used to constrain the physical rationality of the output of the deep learning network being trained. This loss can be used to determine whether the predicted backscattering coefficients of the network output conform to the inherent physical scattering characteristics of the corresponding ground object, thus providing a physical constraint for iterative optimization of network parameters and ensuring that the network output conforms to the actual scattering patterns of the ground object. Specifically, This can be referred to as the loss due to physical consistency constraints.

[0095] For example, the second loss can be specifically achieved through the following formula (2):

[0096] (2)

[0097] in, Indicates the second loss; E[ ] represents the expected operator; This represents a summation operation, where c represents the land cover category (i.e., land cover type) number, and C represents the total number of land cover categories; The predicted probability that a representative sample pixel belongs to the c-th type of land cover; The predicted backscattering coefficients of the representative sample pixels are output by the deep learning network to be trained. Represents taking "0" and " "The larger of the two values ​​is used to determine whether the predicted backscattering coefficient exceeds the maximum threshold of the c-th type of land cover. If it exceeds, the difference is calculated; otherwise, it is set to 0." Represents taking "0" and " "The smaller of the two values ​​is used to determine whether the predicted backscattering coefficient is lower than the minimum threshold of the c-th type of land cover. If it is lower, the difference is calculated; otherwise, it is set to 0." This indicates that the result within the parentheses is squared, which is used to amplify the degree to which the predicted backscattering coefficient deviates from the threshold range.

[0098] In practice, for each sample pixel in each predicted calibration image, the predicted backscattering coefficient of that sample pixel can be substituted into the above formula (2) to obtain the second loss.

[0099] A3-3: Determine the predicted loss based on the first loss and the second loss.

[0100] In practice, the predicted loss can be obtained by directly weighting the first loss and the second loss. This predicted loss can comprehensively cover the radiation consistency and physical consistency constraints of network training.

[0101] In one embodiment, for A3-3 above, it can be implemented according to the following steps B1~B2:

[0102] B1: Determine the third loss of the deep learning network to be trained based on the prediction probability of the calibrated image under each type of land cover and the label probability of the sample calibrated image under each type of land cover.

[0103] Here, the label probability of the calibrated image under various land cover types specifically includes the label probability of each sample pixel in the calibrated image under various land cover types. Label probability (defined as...) The probability that a sample pixel in the calibrated image belongs to the c-th type of land cover is used as a benchmark to determine the accuracy of the land cover prediction probability output by the deep learning network being trained. For example, in a sea ice scene containing open ice, one-year ice, and multi-year ice, if a sample pixel is manually labeled as belonging to the one-year ice category, then the label probability of that pixel belonging to the one-year ice category is... (Corresponding to the annual ice category number c) is 1, and the label probability for both open ice and multi-year ice categories is 0. The label probability can be obtained from actual external ice maps or through manual labeling.

[0104] The third loss (defined as) ) is the loss term in the prediction loss used to constrain the accuracy of the ground feature classification prediction of the deep learning network to be trained. This provides classification constraints for iterative optimization of network parameters, thereby ensuring that the network can accurately output the predicted land cover category of sample pixels. Specifically, It can be classified as cross-entropy loss.

[0105] For example, the calculation of the third loss can be achieved using the following formula (3):

[0106] (3)

[0107] in, This represents the third loss; This represents the probability that a sample pixel belongs to the c-th type of land cover. This represents the predicted probability that a sample pixel belongs to land cover type c. Perform logarithmic operations. That is, the predicted probability that the sample pixel output by the deep learning network to be trained belongs to the c-th type of land cover.

[0108] In practice, for each sample pixel in each predicted calibration image, the predicted probability of that sample pixel under various land cover types, and the label probability of the corresponding sample pixel under various land cover types in each sample calibration image corresponding to the predicted calibration image, can be substituted into the above formula (3) to obtain the third loss. For example, for a scene containing open ice, one-year ice and multi-year ice sea ice, the label probability of a certain sample pixel shows that its true category is one-year ice (corresponding to category c). =1), if the prediction probability of the pixel belonging to one-year ice is low and the prediction probability of belonging to multi-year ice is high, then the value obtained by formula (3) is calculated. The prediction bias will be quantified and fed back to the deep learning network to be trained, guiding the network to adjust its parameters to improve the accuracy of classification prediction.

[0109] B2: Determine the fourth loss of the deep learning network to be trained based on the spatial gradient operator corresponding to the predicted backscattering coefficient carried by each sample pixel in the predicted calibration image.

[0110] Here, the spatial gradient operator is used to capture the spatial variation characteristics of the predicted backscattering coefficients of sample pixels in the predicted calibration image, quantify the difference in predicted backscattering coefficients between adjacent pixels, and thus constrain the spatial continuity of the output results of the deep learning network to be trained, avoiding problems such as spatial abrupt changes and texture chaos in the predicted image.

[0111] Fourth loss (defined as) The prediction loss term constrains the spatial continuity of the output of the deep learning network being trained. Its core purpose is to guide the network's output of predicted and calibrated images to have smooth textures and spatial distributions that closely match actual ground features, avoiding problems such as abrupt spatial changes and chaotic textures. Specifically, This can be called spatial smoothing constraint loss.

[0112] For example, the fourth loss can be specifically implemented using the following formula (4):

[0113] (4)

[0114] in, This indicates the fourth loss; That is, the spatial gradient operator; This represents the predicted backscattering coefficient of a sample pixel; Right now The spatial gradient reflects the degree of change in the predicted backscattering coefficient between adjacent pixels; Represents the L1 norm, used to calculate the spatial gradient. The absolute value of the coefficient is used to quantify the overall intensity of spatial variation.

[0115] In practice, the predicted backscattering coefficient carried by each sample pixel in each predicted calibration image can be substituted into the above formula (4), thus forming the fourth loss of the deep learning network to be trained.

[0116] B3: Determine the predicted loss based on the first loss, second loss, third loss, and fourth loss.

[0117] In specific implementation, the first loss determined in step A3-1 can be used. The second loss determined in step A3-2 The third loss determined in step B1 and the fourth loss determined in step B2 By combining these values, we obtain the prediction loss of the deep learning network to be trained.

[0118] For example, the first loss, the second loss, the third loss, and the fourth loss can be achieved according to the following formula (5):

[0119] (5)

[0120] in, Indicates the predicted loss; The weighting coefficient representing the second loss is used to adjust the degree of influence of the second loss in the total predicted loss; The weighting coefficient representing the third loss is used to adjust the degree of influence of the third loss in the total predicted loss; The weighting coefficient representing the fourth loss is used to adjust the degree of influence of the fourth loss in the predicted loss.

[0121] S103: For each group of images in the target calibration image, the overlapping area of ​​each group of images is divided into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images.

[0122] Here, each image group refers to an image combination formed by dividing the target calibration images according to their adjacency relationship. Each image group always contains two adjacent sub-aperture images. The overlapping area of ​​each image group refers to the intersecting and overlapping area formed by two adjacent sub-aperture images in the same group in terms of spatial acquisition coverage, which is the intersection of the coverage areas of the two sub-aperture images.

[0123] The reason for the overlapping area is that when the imaging equipment on the target satellite acquires sub-aperture images, in order to avoid blind spots between adjacent sub-aperture images and to ensure the continuity and integrity of subsequent image stitching processing, the acquisition range of adjacent sub-aperture images is designed to have some overlap, thus forming an overlapping area between two adjacent sub-aperture images.

[0124] For each group of images, the sub-block is an independent small region formed after dividing the overlapping area of ​​the group of images. It is used to break down a large overlapping area into multiple small independent processing units, so as to realize fine-grained and regional processing of the overlapping area of ​​the group of images, thereby improving the efficiency and accuracy of subsequent processing of the overlapping area.

[0125] Specifically, sub-blocks can be divided according to preset partitioning rules, or they can be adaptively partitioned based on the regional characteristics of each overlapping area. The preset partitioning rules can be flexibly set according to the actual calibration processing requirements. Common preset rules include, but are not limited to: fixed size rules, ground feature rules, texture complexity rules, and overlapping boundary alignment rules. Among these rules, the fixed-size rule divides the overlapping region into a grid pattern using a uniform pixel size (e.g., 256×256 pixels) and a consistent number of sub-blocks (6×6), ensuring that each sub-block is the same size and facilitating subsequent batch standardization. The feature rule divides the overlapping region based on the type of land cover (e.g., broken ice areas, flat ice areas, and interglacial waterways in polar sea ice scenes), making the features within each sub-block as uniform as possible and reducing the impact of land cover mixing on subsequent calibration accuracy. The texture complexity rule divides areas with complex textures (e.g., dense ice ridges) into finer sub-blocks and areas with simple textures (e.g., continuous thin ice areas) into coarser sub-blocks to balance processing efficiency and accuracy. The overlap boundary alignment rule divides the region along the overlap boundary of the two sub-aperture images, ensuring that the edges of the sub-blocks are aligned with the overlap boundary and avoiding interference from cross-boundary processing. Regional features can include the normalized duration of the initial calibration image corresponding to the overlapping region and / or the distribution of the target backscattering coefficients for each pixel within the overlapping region. Adaptive partitioning can be achieved for example, by normalizing the duration so that the more sub-blocks are generated and each sub-block is smaller; or by normalizing the duration so that the more uniform the distribution, the fewer sub-blocks are generated and each sub-block is larger.

[0126] In this application, a target calibration image contains multiple sets of images, each set of images consisting of two adjacent sub-aperture images. Since the overlapping range and ground feature features of adjacent sub-aperture images in different sets of images differ during spatial acquisition, the backscattering coefficient distribution of the overlapping area corresponding to each set of images is different. Therefore, when dividing the overlapping area of ​​each set of images into blocks according to preset division rules (such as fixed size rules, ground feature rules, etc.), the number and size of the sub-blocks obtained by dividing different overlapping areas may differ.

[0127] In practice, for the target calibration image, the imaging geometry parameters of the target satellite's wide-swath scanning mode can be used to determine each group of adjacent sub-aperture images (each group of images) and the spatial overlap area between each group of adjacent sub-aperture images. Then, each overlapping area can be divided according to a preset division rule or an adaptive division rule to obtain each sub-block corresponding to each group of images.

[0128] like Figure 2 The diagram shown is a schematic representation of an overlapping region block division provided in this application. Wherein, Figure 2A set of images contains adjacent sub-aperture images A and B, whose spatial acquisition coverage areas overlap, as well as non-overlapping areas. For the overlapping area between sub-aperture images A and B, the sub-blocks are divided into sub-block 1, sub-block 2, ..., sub-block n.

[0129] S104: Based on the target backscattering coefficient of each pixel in each sub-block, perform stitching correction processing on each sub-block to obtain each corrected block; the stitching correction processing is used to eliminate radiation abrupt changes between sub-blocks.

[0130] Here, a radiation abrupt change refers to a sudden jump in the target backscattering coefficient of pixels at the boundary of adjacent sub-blocks due to fluctuations in imaging equipment parameters, differences in the acquisition environment, etc., resulting in significant inconsistencies in brightness and grayscale at the block junctions. The core of the stitching correction process is to eliminate this radiation abrupt change. It involves calibrating and adjusting the radiation consistency of adjacent sub-blocks based on the target backscattering coefficients of pixels in each sub-block. The corrected block is a block where, after stitching correction, radiation abrupt changes have been eliminated, the transition with adjacent blocks is smooth, and the radiation characteristics are uniform.

[0131] In practice, for each sub-block of the overlapping area in each group of images, the target backscattering coefficient of all pixels in each sub-block is extracted. Based on this coefficient, the difference in radiation characteristics at the junction of adjacent sub-blocks is analyzed to determine whether there is a radiation abrupt change. Then, targeted stitching correction processing is carried out. By performing consistency calibration and adjustment on the target backscattering coefficient of the pixels at the junction, the radiation abrupt change between sub-blocks is eliminated, so that each adjacent sub-block is smoothly connected in terms of radiation characteristics and the overall characteristics are unified. Each independent block obtained after this stitching correction processing is the corrected block.

[0132] In one embodiment, S104 described above can be implemented according to the following steps:

[0133] S104-1: Determine the correction order between two sub-aperture images in each group of images; the correction order is used to indicate the reference image and the image to be corrected in two adjacent sub-aperture images.

[0134] Here, the calibration sequence is used to clarify the calibration relationship between two adjacent sub-aperture images within each image group, such as using the previous sub-aperture image to calibrate the next. Therefore, the calibration sequence can indicate the reference image and the image to be calibrated in two sub-aperture images. The reference image is the sub-aperture image used as a benchmark during the calibration process; its radiometric characteristics are stable and its accuracy meets the requirements, providing a calibration basis for the image to be calibrated. The image to be calibrated is the sub-aperture image that needs to be radiometrically calibrated and adjusted based on the reference image; after calibration, it can achieve radiometric consistency matching with the reference image. Figure 2 Taking the scenario shown as an example, Figure 2 A set of images includes adjacent sub-aperture images A and B. If the determined correction order indicates that sub-aperture image A is the reference image and sub-aperture image B is the image to be corrected, then the radiation characteristics of sub-aperture A are used as the reference to calibrate and adjust sub-aperture B to ensure that the radiation characteristics of the two are consistent.

[0135] Specifically, the determination of the correction order can be based on factors such as the radiometric stability, imaging accuracy, and signal-to-noise ratio of the sub-aperture images. For example, the sub-aperture image with more stable radiometric characteristics and higher accuracy can be prioritized as the reference image, and the corresponding other image can be set as the image to be corrected. Alternatively, it can be flexibly set according to the actual correction requirements.

[0136] Optionally, the correction order for different adjacent sub-aperture images can be different. For example, some groups of images can be corrected in the order of "sub-aperture image A as reference image and sub-aperture image B as image to be corrected", while other groups of images can be corrected in the order of "sub-aperture image B as reference image and sub-aperture image A as image to be corrected".

[0137] In practice, for each set of images in the target calibration image, the core feature information of the two sets of sub-aperture images is extracted first, including but not limited to radiometric stability parameters, imaging accuracy data and signal-to-noise ratio values. Combined with the preset correction order determination rules, the sub-aperture image with more stable radiometric features, higher imaging accuracy and better signal-to-noise ratio is selected as the reference image, and the other sub-aperture image is selected as the image to be corrected. This clarifies the correction order of the two sub-aperture images in the set of images, providing a clear priority benchmark for subsequent stitching and correction processing.

[0138] S104-2: For each sub-block in each image group, based on the correction order and the target backscattering coefficients corresponding to each pixel in the sub-block in the two sub-aperture images, fit the first radiometric mapping relationship of the image group under the sub-block.

[0139] Here, a sub-region corresponds to a first radiometric mapping relationship, which is used to indicate the radiometric correction relationship between the reference image and the image to be corrected under the sub-block.

[0140] It should be noted that for each pixel in each sub-block of each image group, there is a corresponding aligning pixel in two adjacent sub-aperture images (such as sub-aperture image A and sub-aperture image B) of the same image group. One of these two aligning pixels can be a pixel within the sub-block. These two aligning pixels each correspond to one... That is, each pixel corresponds to two , such as in Figure 2 In the middle, one cell in sub-block 1 corresponds to two This can be the target backscattering coefficient of the pixel in sub-aperture image A. and the target backscattering coefficient in sub-aperture image B .

[0141] Specifically, the first radiometric mapping relationship can be represented as a brightness mapping function, used to mathematically characterize the target backscattering coefficient (denoted as ) of pixels in the image to be corrected within the same sub-block. The target backscattering coefficient (denoted as ) and the target backscattering coefficient of the corresponding pixel in the reference image The transformation relationship between them. The first radiation mapping relationship can be represented by a linear model in the form of fundamental functions: Where a and b are mapping parameters determined by the least squares method, corresponding to the gain factor and offset, respectively. This linear model can effectively compensate for the systematic radiation bias of the two sub-aperture images in this local sub-block.

[0142] In practice, for each sub-block, according to the correction order (e.g., determining sub-aperture image A as the reference image and sub-aperture image B as the image to be corrected), the target backscattering coefficients (e.g., ...) of all pixels within that sub-block in the reference image are calculated. ) and the target backscattering coefficient in the image to be corrected (e.g. The data are extracted in pairs to form fitted data pairs. If the sub-block contains 16 pixels, then a total of 32 data pairs are obtained. Data, 16 fitted data pairs ( Subsequently, the least squares method was used to fit the data pairs to establish the radiance mapping relationship between the sub-block and the image to be corrected to the reference image.

[0143] Optionally, to improve the robustness and accuracy of the fitting, the difference distribution of the target backscattering coefficients of all pixel pairs within the sub-block can be statistically analyzed before fitting. Then, based on the distribution parameters corresponding to the difference distribution, the least squares method is used to fit the data pairs. Alternatively, based on the difference distribution, a fitting weight can be assigned to each pixel pair difference, with pixel pairs with larger differences potentially receiving lower fitting weights to reduce the impact of outliers. Then, a weighted least squares method is used, combined with the determined fitting weights, to fit the data pairs, ultimately obtaining a more robust brightness mapping function corresponding to the sub-block.

[0144] S104-3: Based on the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of the pixel in each sub-block in the reference image, the target backscattering coefficient of the pixel in each sub-block in the image to be corrected is corrected to obtain the corrected block.

[0145] In practice, after obtaining the first radiometric mapping relationship for each sub-block, for each sub-block, every pixel within it is traversed, and the target backscattering coefficient of that pixel in the reference image is read. Then, the first radiometric mapping relationship previously fitted for that sub-block is mathematically transformed to determine the appropriate adjustment value for the target backscattering coefficient of the current pixel in the image to be corrected. This calculated correction value is then assigned to the corresponding pixel position in the image to be corrected, thus completing the radiometric correction for that pixel. After processing all pixels within a sub-block, a corrected block corresponding to the spatial range of that sub-block is generated. Each pixel within this corrected block is now locally aligned with the reference image in terms of radiometric brightness.

[0146] In one embodiment, S104-3 described above can be implemented according to the following steps C1 to C5:

[0147] C1: Determine the adjacent cells between each sub-block and its neighboring sub-blocks.

[0148] Here, adjacent cells refer to cells at the boundary of two adjacent sub-blocks. Specifically, they are cells located at the edge of a sub-block and directly connected to the edges of adjacent sub-blocks. For example... Figure 2 In the partitioning of the code, adjacent sub-blocks 1 and 2 are defined. The pixels on the right edge of sub-block 1 and the pixels on the left edge of sub-block 2 are considered their adjacent pixels. For any given sub-block, there can be one or more adjacent sub-blocks. For example, Figure 2 The adjacent sub-blocks of sub-block 1 are sub-blocks 2 and 3, and the adjacent sub-blocks of sub-block 3 are sub-blocks 1, 4 and 5.

[0149] In practice, for each sub-block corresponding to each group of images obtained from the block division, the spatial location and adjacency relationship of each sub-block within the overlapping area are first determined. The directly adjacent sub-blocks are identified. Then, the pixels at the edge of the sub-block are located, and directly adjacent pixels are selected from the edges of adjacent sub-blocks. These corresponding edge pixels are the adjacent pixels between the sub-block and its neighboring sub-blocks. To ensure accurate identification of all adjacent pixels in each pair of adjacent sub-blocks, the following methods can be used: spatial coordinate matching, edge contour comparison, and pixel position indexing. Among them, the spatial coordinate matching method can extract the spatial coordinates (such as pixel row and column numbers) of the edge pixels of each sub-block and match the pixels with corresponding coordinate positions at the edges of adjacent sub-blocks, which are the adjacent pixels; the edge contour comparison method can delineate the edge contours of each sub-block, compare the edge contours of adjacent sub-blocks, and determine the pixels corresponding to each other at the contour fit as adjacent pixels; the pixel position index method can establish a position index for the pixels in each sub-block, and filter out adjacent pixels by associating the corresponding positions in adjacent sub-blocks through the index.

[0150] C2: For each adjacent cell, determine the second radiative mapping relationship corresponding to the adjacent cell based on the first radiative mapping relationship between the two adjacent sub-blocks and the smoothing weights of the two adjacent sub-blocks.

[0151] Here, the smoothing weight is a weight parameter used to adjust the transition effect of radiation characteristics between two adjacent sub-blocks, so as to make the radiation characteristics at the junction of adjacent sub-blocks smoother and avoid radiation abrupt changes.

[0152] Specifically, the determination of smoothing weights needs to be based on a comprehensive evaluation of multiple factors, including but not limited to the uniformity of the distribution of the backscattering coefficients of the target pixels within adjacent sub-blocks, the degree of radiation abrupt changes at the sub-block boundaries, the radiation stability of the sub-blocks, and the imaging accuracy. Among these, sub-blocks with more uniform distribution of the target backscattering coefficients, less abrupt radiation abrupt changes, and better radiation stability and imaging accuracy can be assigned relatively higher smoothing weights, while those with less uniform distribution and less accurate smoothing can be assigned relatively lower weights, thereby ensuring the rationality and accuracy of the radiation transition.

[0153] The second radiative mapping relationship is used to accurately characterize the radiative transformation relationship between adjacent pixels in adjacent sub-blocks, ensuring the consistency of radiative connection at adjacent pixels. Specifically, it is a mapping relationship determined for each adjacent pixel by combining the first radiative mapping relationship between the two adjacent sub-blocks corresponding to that adjacent pixel, and the smoothing weights of the two adjacent sub-blocks.

[0154] In specific implementation, for each adjacent pixel determined in step C1, the first radiation mapping relationship of each of the two adjacent sub-blocks corresponding to the adjacent pixel is first obtained, and the smoothing weights of the two adjacent sub-blocks are obtained at the same time. The two first radiation mapping relationships and the corresponding smoothing weights are fused and calculated. Based on the calculation results, the second radiation mapping relationship exclusive to the adjacent pixel is determined to ensure that the radiation transformation relationship of each adjacent pixel can adapt to the radiation characteristics of the adjacent sub-blocks and ensure radiation smoothness at the junction of the two sub-blocks.

[0155] When a sub-block has only one adjacent sub-block, the second radiative mapping relationship for each adjacent pixel can be determined by directly fusing the established first radiative mapping relationship and smoothing weights of the two sub-blocks, thus achieving a smooth transition of radiative characteristics at the junction of the two sub-blocks. When a sub-block has multiple adjacent sub-blocks, the second radiative mapping relationship for each adjacent pixel of the two sub-blocks can be determined independently.

[0156] For example, for sub-block n and each of its adjacent sub-blocks n near Determine sub-block n near The smoothing weight is 0.5 for the first sub-block and 0.5 for the second sub-block. Then, for each adjacent pixel between the two sub-blocks, the second radiation mapping relationship corresponding to each adjacent pixel is determined by fusing the first radiation mapping relationship determined by the two sub-blocks with the smoothing weight, thus achieving a smooth transition of radiation characteristics at the junction of the two sub-blocks.

[0157] C3: For each sub-block, using the first radiometric mapping relationship of the sub-block and the target backscattering coefficients of the non-adjacent pixels of the sub-block in the reference image, the target backscattering coefficients of the non-adjacent pixels of the sub-block in the image to be corrected are corrected.

[0158] Here, non-adjacent cells refer to cells located inside a sub-block (not at the edge), or cells located at the edge of a sub-block but without adjacent sub-regions or without direct adjacency to adjacent sub-block edge cells. For ease of understanding, for each sub-block, the non-adjacent cells in that sub-block are all the cells in that sub-block excluding adjacent cells.

[0159] In practice, for each sub-block of the overlapping area in each group of images, the non-adjacent pixels within the sub-block are identified, and the first radiometric mapping relationship of the sub-block is obtained. At the same time, the target backscattering coefficient corresponding to each non-adjacent pixel in the reference image is extracted, and the coefficient is substituted into the first radiometric mapping relationship. Through mapping transformation, the target backscattering coefficient corresponding to each non-adjacent pixel in the sub-block in the image to be corrected is calibrated and adjusted to complete the correction processing of non-adjacent pixels, ensuring that the radiometric characteristics of the corrected non-adjacent pixels match the reference image.

[0160] C4: Using the second radiometric mapping relationship of each neighboring pixel in the sub-block, and the target backscattering coefficient of each neighboring pixel in the reference image, the target backscattering coefficient of each neighboring pixel in the image to be corrected is corrected.

[0161] In practice, for each sub-block of the overlapping area in each group of images, the already determined neighboring pixels between the sub-block and its adjacent sub-blocks are first extracted. Then, the second radiometric mapping relationship specific to each neighboring pixel is obtained one by one. At the same time, the corresponding image of each neighboring pixel in the reference image is extracted. The coefficient is substituted into the second radiometric mapping relationship of the corresponding adjacent pixel. The target backscattering coefficient of each adjacent pixel in the image to be corrected is calibrated and adjusted through mapping transformation. The correction process of the adjacent pixel is completed, ensuring that the radiometric characteristics of the corrected adjacent pixel match the reference image and are smoothly connected with the adjacent sub-blocks.

[0162] Optionally, for each sub-block, the target backscattering coefficients of each pixel in the sub-block in the image to be corrected can be adjusted using the first radiometric mapping relationship corresponding to the sub-block and the target backscattering coefficients of each pixel in the sub-block in the reference image, thus obtaining an initial correction block. Then, for each adjacent pixel already determined between the initial correction block and the adjacent pixel in the reference image, the final corrected backscattering coefficients of the adjacent pixel can be determined based on the second radiometric mapping relationship, the target backscattering system corresponding to the adjacent pixel in the reference image, and the initially corrected backscattering coefficients of the adjacent pixel in the image to be corrected. The initially corrected backscattering coefficients are obtained by correcting the target backscattering system corresponding to the adjacent pixel using the first radiometric mapping relationship.

[0163] C5: Each corrected block is obtained based on the target backscattering coefficient of each pixel in the reference image and the corrected target backscattering coefficient of each pixel in the image to be corrected.

[0164] Here, the corrected target backscattering coefficient is used to characterize the radiation characteristics of the corrected pixel in the image to be corrected. Specifically, it refers to the original target backscattering coefficient of each pixel in the image to be corrected, which, after corresponding correction processing (including correction of non-adjacent pixels in step C3 and correction of adjacent pixels in step C4), becomes the target backscattering coefficient that conforms to the radiation characteristics of the reference image.

[0165] In practice, for each sub-block of overlapping areas in each group of images, the backscattering coefficients of all pixels (including non-adjacent and adjacent pixels) in the sub-block are verified for consistency between the two types of coefficients in the reference image and the corrected backscattering coefficients of each pixel in the image to be corrected after corresponding correction processing. This ensures that the two types of coefficients of each pixel are accurately matched and the radiation characteristics are consistent. Then, all pixels that have passed the verification are integrated according to their original spatial positions in the sub-block to finally form a complete block with unified radiation characteristics and natural connection. This block is the corrected block corresponding to each sub-block.

[0166] In this way, through the above-mentioned block-level fitting and smoothing at adjacent pixels, the backscattering coefficient in the stitched area can be made to change continuously in space, thereby eliminating the strip-shaped brightness abrupt change.

[0167] S105: Generate corrected SAR images based on the non-overlapping areas in each group of images and each corrected block.

[0168] In practice, the integrity of the non-overlapping areas is first checked based on the non-overlapping regions of the two sub-aperture images in each group of images, as well as the corresponding corrected blocks in each group of images, to ensure that there are no missing or distorted areas. At the same time, the radiometric characteristics of each corrected block and the adjacent non-overlapping areas are checked. The non-overlapping areas and the corresponding corrected blocks are then precisely stitched and integrated according to the original spatial position of the images to ensure that the overall image after stitching is without misalignment and has continuous and consistent radiometric characteristics, and finally, the corrected SAR image is generated.

[0169] Optionally, after stitching together the non-overlapping areas and corrected blocks in each group of images, geometric correction and denoising can be performed on the stitched images to further eliminate geometric bias, thus obtaining the final corrected SAR image. Geometric correction and denoising can be performed, for example, by performing geometric correction and projection on the stitched images, and then using non-local mean or refined Lee filtering to denoise the projection results. This suppresses noise while preserving clear boundaries between different ground features in the image, resulting in a corrected SAR image with higher image quality.

[0170] The synthetic aperture radar (SAR) image correction method provided in this application, through a deep learning network, can adaptively learn the nonlinear relationship between the scattering coefficient and the incident angle. Compared with traditional empirical function methods, it significantly improves the radiometric consistency under different incident angles and can output more reasonable normalized backscattering coefficients for different ground features. Addressing the radiometric discontinuity problem in multi-aperture images under WSC mode, by dividing the overlapping areas between adjacent sub-apertures into blocks and smoothing the transition between adjacent pixels in the blocks, the brightness discontinuity problem in the stitching area can be effectively suppressed, improving the naturalness and coherence of the corrected image.

[0171] Based on the same inventive concept, this disclosure also provides a synthetic aperture radar (SAR) image correction device corresponding to the SAR image correction method. For example... Figure 3 The diagram shown is a schematic of a synthetic aperture radar image correction device provided in an embodiment of this disclosure, comprising:

[0172] The calibration module 301 is used to perform radiometric calibration on the echo data acquired by the synthetic aperture radar SAR on the target satellite in wide-swath scanning mode to obtain an initial calibration image; each pixel in the initial calibration image carries an initial backscattering coefficient and an incident angle.

[0173] The normalization module 302 is used to determine the target calibration image using a pre-trained deep learning network based on the initial backscattering coefficient and incident angle carried by the pixel; each pixel in the target calibration image carries a normalized target backscattering coefficient.

[0174] The segmentation module 303 is used to divide the overlapping area of ​​each group of images in the target calibration image into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images.

[0175] The stitching correction module 304 is used to perform stitching correction processing on each of the sub-blocks according to the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block; the stitching correction processing is used to eliminate radiation abrupt changes between the sub-blocks.

[0176] The generation module 305 is used to generate corrected SAR images based on the non-overlapping areas in each group of images and each corrected block.

[0177] In one possible implementation, the stitching correction module 304, when performing stitching correction processing on each sub-block according to the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block, is used to:

[0178] Determine the correction order between two sub-aperture images in each group of images; the correction order is used to indicate the reference image and the image to be corrected in two adjacent sub-aperture images.

[0179] For each sub-block in each group of images, a first radiometric mapping relationship corresponding to the sub-block is fitted according to the correction order and the target backscattering coefficients corresponding to each pixel in the sub-block in the two sub-aperture images.

[0180] Based on the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of the pixel of each sub-block in the reference image, the target backscattering coefficient corresponding to the pixel of each sub-block in the image to be corrected is corrected to obtain the corrected block.

[0181] In one possible implementation, the stitching correction module 304, when correcting the target backscattering coefficient corresponding to the pixel of each sub-block in the image to be corrected according to the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of the pixel of each sub-block in the reference image, to obtain a corrected block, is used to:

[0182] Determine the adjacent cells between each sub-block and its neighboring sub-blocks;

[0183] For each adjacent pixel, a second radiative mapping relationship is determined based on the first radiative mapping relationship between the two adjacent sub-blocks corresponding to the adjacent pixel and the smoothing weight of the two adjacent sub-blocks.

[0184] For each sub-block, using the first radiometric mapping relationship of the sub-block and the target backscattering coefficient of the non-adjacent pixels of the sub-block in the reference image, the target backscattering coefficient corresponding to the non-adjacent pixels of the sub-block in the image to be corrected is corrected.

[0185] Using the second radiometric mapping relationship of each adjacent pixel in the sub-block, and the target backscattering coefficient of each adjacent pixel in the reference image, the target backscattering coefficient corresponding to each adjacent pixel in the image to be corrected is corrected.

[0186] Each corrected block is obtained based on the target backscattering coefficient of each pixel in each sub-block in the reference image, and the corrected target backscattering coefficient of each pixel in each sub-block in the image to be corrected.

[0187] In one possible implementation, the normalization module 302, when determining the target calibration image using a pre-trained deep learning network based on the initial backscattering coefficients and incident angles carried by the pixels, is used to:

[0188] Using the input layer of the deep learning network, the initial backscattering coefficients and incident angles carried by each pixel are concatenated to obtain the input features;

[0189] The input features are processed by the convolution module of the deep learning network multiple times to obtain convolutional features; the convolution module includes multiple convolutional layers connected in sequence.

[0190] The convolutional features are processed using the fully connected layers of the deep learning network to obtain fully connected features.

[0191] The fully connected features are linearly activated using the output layer of the deep learning network to obtain the target calibration image.

[0192] In one possible implementation, the apparatus further includes a training module 306 for training the deep learning network according to the following steps:

[0193] Acquire a set of sample calibration images acquired by SAR in wide-swath scanning mode; the set of sample calibration images includes sample calibration images acquired at different time periods for different land cover types; each sample pixel in the sample calibration image carries the sample backscattering coefficient and sample incident angle.

[0194] Using the deep learning network to be trained, the predicted calibration image and the prediction probability under various land cover types are determined based on the sample backscattering coefficient and sample incident angle carried by the sample pixels in the sample calibration image; each sample pixel in the predicted calibration image carries a normalized predicted backscattering coefficient.

[0195] Based on each predicted backscattering coefficient and the predicted probability of the predicted calibrated image under various land cover types, the prediction loss of the deep learning network to be trained is constructed.

[0196] The deep learning network to be trained is iteratively trained using the prediction loss until a preset training cutoff condition is met, resulting in a trained deep learning network.

[0197] In one possible implementation, the training module 306, when constructing the prediction loss of the deep learning network to be trained based on each predicted backscattering coefficient and the prediction probability corresponding to the predicted calibrated image under various land cover types, is configured to:

[0198] The first loss of the deep learning network to be trained is determined based on the predicted backscattering coefficient of each sample pixel in the overlapping region of each group of sample images in the predicted calibration image.

[0199] The second loss of the deep learning network to be trained is determined based on the predicted probability and the backscattering coefficient thresholds corresponding to various land cover types.

[0200] The predicted loss is determined based on the first loss and the second loss.

[0201] In one possible implementation, the training module 306, when determining the prediction loss based on the first loss and the second loss, is configured to:

[0202] The third loss of the deep learning network to be trained is determined based on the predicted probability of the predicted calibrated image under various land cover types and the label probability of the sample calibrated image under various land cover types.

[0203] The fourth loss of the deep learning network to be trained is determined based on the spatial gradient operator corresponding to the predicted backscattering coefficient carried by each sample pixel in the predicted calibration image.

[0204] The predicted loss is determined based on the first loss, the second loss, the third loss, and the fourth loss.

[0205] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0206] Based on the same technical concept, embodiments of this application also provide a computer device. (Refer to...) Figure 4 The diagram shown is a structural schematic of a computer device provided in an embodiment of this application, comprising:

[0207] The system comprises a processor 401, a memory 402, and a bus 403. The memory 402 stores machine-readable instructions executable by the processor 401. The processor 401 executes these machine-readable instructions, performing steps S101-S105. The memory 402 includes a main memory 4021 and an external memory 4022. The main memory 4021, also called internal memory, is used to temporarily store computational data in the processor 401 and data exchanged with external memory such as a hard disk. The processor 401 exchanges data with the external memory 4022 through the main memory 4021. When the computer is running, the processor 401 communicates with the memory 402 via the bus 403, enabling the processor 401 to execute the instructions mentioned in the above method embodiments.

[0208] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the synthetic aperture radar image correction method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0209] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the synthetic aperture radar image correction method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0210] The computer program product can be implemented specifically through hardware, software, or a combination thereof. In one alternative embodiment, the computer program product is specifically embodied in a computer storage medium; in another alternative embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0211] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for correcting synthetic aperture radar images, characterized in that, include: Radiometric calibration is performed on the echo data acquired by the synthetic aperture radar (SAR) on the target satellite in wide-swath scanning mode to obtain the initial calibration image; Each pixel in the initial calibration image carries an initial backscattering coefficient and an incident angle; Using a pre-trained deep learning network, a target calibration image is determined based on the initial backscattering coefficient and incident angle carried by the pixel; each pixel in the target calibration image carries a normalized target backscattering coefficient; wherein, the deep learning network is used to learn the nonlinear mapping relationship between the initial backscattering coefficient and incident angle and the normalized backscattering coefficient. The overlapping areas of each group of images in the target calibration image are divided into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images; Based on the target backscattering coefficient of each pixel in each sub-block, the sub-blocks are stitched together and corrected to obtain corrected blocks; the stitching and correction process is used to eliminate radiation abrupt changes between the sub-blocks. Based on the non-overlapping areas in each group of images and each corrected block, a corrected SAR image is generated. The step of performing stitching and correction processing on each sub-block based on the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block includes: Determine the correction order between two sub-aperture images in each group of images; the correction order is used to indicate the reference image and the image to be corrected in two adjacent sub-aperture images. For each sub-block in each group of images, a first radiometric mapping relationship corresponding to the sub-block is fitted according to the correction order and the target backscattering coefficients corresponding to each pixel in the sub-block in the two sub-aperture images. Based on the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of the pixel of each sub-block in the reference image, the target backscattering coefficient corresponding to the pixel of each sub-block in the image to be corrected is corrected to obtain the corrected block.

2. The method according to claim 1, characterized in that, The step of correcting the target backscattering coefficient of each sub-block in the image to be corrected based on the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of each sub-block's pixel in the reference image to obtain a corrected block includes: Determine the adjacent cells between each sub-block and its neighboring sub-blocks; For each adjacent pixel, a second radiative mapping relationship is determined based on the first radiative mapping relationship between the two adjacent sub-blocks corresponding to the adjacent pixel and the smoothing weight of the two adjacent sub-blocks. For each sub-block, using the first radiometric mapping relationship of the sub-block and the target backscattering coefficient of the non-adjacent pixels of the sub-block in the reference image, the target backscattering coefficient corresponding to the non-adjacent pixels of the sub-block in the image to be corrected is corrected. Using the second radiometric mapping relationship of each adjacent pixel in the sub-block, and the target backscattering coefficient of each adjacent pixel in the reference image, the target backscattering coefficient corresponding to each adjacent pixel in the image to be corrected is corrected. Each corrected block is obtained based on the target backscattering coefficient of each pixel in each sub-block in the reference image, and the corrected target backscattering coefficient of each pixel in each sub-block in the image to be corrected.

3. The method according to claim 1, characterized in that, The step of determining the target calibration image using a pre-trained deep learning network based on the initial backscattering coefficient and incident angle carried by the pixel includes: Using the input layer of the deep learning network, the initial backscattering coefficients and incident angles carried by each pixel are concatenated to obtain the input features; The input features are processed by the convolution module of the deep learning network multiple times to obtain convolutional features; the convolution module includes multiple convolutional layers connected in sequence. The convolutional features are processed using the fully connected layers of the deep learning network to obtain fully connected features. The fully connected features are linearly activated using the output layer of the deep learning network to obtain the target calibration image.

4. The method according to claim 1, characterized in that, The deep learning network is trained according to the following steps: Acquire a set of sample calibration images acquired by SAR in wide-swath scanning mode; the set of sample calibration images includes sample calibration images acquired at different time periods for different land cover types; each sample pixel in the sample calibration image carries the sample backscattering coefficient and sample incident angle. Using the deep learning network to be trained, the predicted calibration image and the prediction probability under various land cover types are determined based on the sample backscattering coefficient and sample incident angle carried by the sample pixels in the sample calibration image; each sample pixel in the predicted calibration image carries a normalized predicted backscattering coefficient. Based on each predicted backscattering coefficient and the predicted probability of the predicted calibrated image under various land cover types, the prediction loss of the deep learning network to be trained is constructed. The deep learning network to be trained is iteratively trained using the prediction loss until a preset training cutoff condition is met, resulting in a trained deep learning network.

5. The method according to claim 4, characterized in that, The step of constructing the prediction loss of the deep learning network to be trained based on each predicted backscattering coefficient and the predicted probability of the predicted calibrated image under various land cover types includes: The first loss of the deep learning network to be trained is determined based on the predicted backscattering coefficient of each sample pixel in the overlapping region of each group of sample images in the predicted calibration image. The second loss of the deep learning network to be trained is determined based on the predicted probability and the backscattering coefficient thresholds corresponding to various land cover types. The predicted loss is determined based on the first loss and the second loss.

6. The method according to claim 5, characterized in that, Determining the predicted loss based on the first loss and the second loss includes: The third loss of the deep learning network to be trained is determined based on the predicted probability of the predicted calibrated image under various land cover types and the label probability of the sample calibrated image under various land cover types. The fourth loss of the deep learning network to be trained is determined based on the spatial gradient operator corresponding to the predicted backscattering coefficient carried by each sample pixel in the predicted calibration image. The predicted loss is determined based on the first loss, the second loss, the third loss, and the fourth loss.

7. A correction device for synthetic aperture radar images, characterized in that, include: The calibration module is used to perform radiometric calibration on the echo data acquired by the synthetic aperture radar SAR on the target satellite in wide-swath scanning mode to obtain the initial calibration image; Each pixel in the initial calibration image carries an initial backscattering coefficient and an incident angle; A normalization module is used to determine a target calibration image based on the initial backscattering coefficient and incident angle carried by the pixel using a pre-trained deep learning network; each pixel in the target calibration image carries a normalized target backscattering coefficient; wherein the deep learning network is used to learn a nonlinear mapping relationship between the initial backscattering coefficient and incident angle and the normalized backscattering coefficient. The segmentation module is used to divide the overlapping areas of each group of images in the target calibration image into blocks to obtain each sub-block corresponding to each group of images; wherein, a group of images includes two adjacent sub-aperture images. The stitching correction module is used to perform stitching correction processing on each sub-block according to the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block; the stitching correction processing is used to eliminate radiation abrupt changes between the sub-blocks. The generation module is used to generate corrected SAR images based on the non-overlapping areas in each group of images and each corrected block. The stitching correction module, when performing stitching correction processing on each sub-block based on the target backscattering coefficient of each pixel in each sub-block to obtain each corrected block, is used to: Determine the correction order between two sub-aperture images in each group of images; the correction order is used to indicate the reference image and the image to be corrected in two adjacent sub-aperture images. For each sub-block in each group of images, a first radiometric mapping relationship corresponding to the sub-block is fitted according to the correction order and the target backscattering coefficients corresponding to each pixel in the sub-block in the two sub-aperture images. Based on the first radiometric mapping relationship of each sub-block and the target backscattering coefficient of the pixel of each sub-block in the reference image, the target backscattering coefficient corresponding to the pixel of each sub-block in the image to be corrected is corrected to obtain the corrected block.

8. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the synthetic aperture radar image correction method as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is run by a computer device, the computer device performs the steps of the synthetic aperture radar image correction method as described in any one of claims 1 to 6.

Images