Same-platform area array image auxiliary linear array image geometric accuracy improvement method

Abstract

The invention provides a same-platform area array image assisted linear array image geometric accuracy improving method, and belongs to the technical field of electrical digital data processing.The method includes the steps that firstly, a linear array and area array image rigorous imaging geometric model is constructed according to attitude orbit data and calibration parameters; introducing external elevation data and substituting the area array image pixel ground coordinates into the linear array image rigorous imaging geometric model for back calculation to obtain a simulation area array image; matching the real and simulated area array images to obtain dense homonymy points, and determining corresponding positions on the linear array image; correcting systematic distortion by adopting an affine transformation model, wherein a residual error is high-order distortion caused by high-frequency attitude jitter; sorting according to the coordinates of the homonymy points to form a two-dimensional grid, and voting based on the parallax difference to calculate scores; optimizing attitude parameters by adopting an iterative detection strategy to eliminate the influence of the attitude error of the area array image; and finally, reconstructing a linear array image geometric model by utilizing the optimized attitude data, generating a geometric seamless sensor correction image, and realizing uncontrolled elevation precision improvement of a stereoscopic image.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric digital data processing, and more particularly, relates to a method for improving the geometric accuracy of a line array image assisted by a same-platform planar array image. Background Art

[0002] Optical satellite remote sensing technology has been widely used in the fields of geographic information systems, resource exploration, urban planning, disaster monitoring, etc. Traditional satellite stereo mapping mainly constructs a stereo model by acquiring forward-looking, backward-looking or side-looking images, and then realizes elevation extraction and three-dimensional reconstruction, which has become an important means to obtain large-scale topographic information. At present, high-resolution optical satellites mostly use the line array push-broom method to acquire panchromatic band images, and at the same time acquire multi-spectral band images to provide different color information.

[0003] However, during the on-orbit operation of the satellite, it is inevitably affected by external factors such as solar radiation pressure, earth's gravitational gradient, thermo-elastic deformation, and mechanical vibration, resulting in high-frequency attitude jitter. This high-frequency attitude jitter has the characteristics of short period, strong randomness, and a frequency much higher than the sampling frequency of the satellite attitude measurement device, and cannot be accurately captured by a conventional attitude measurement device. Since the line array camera uses a line-by-line push-broom imaging method, the high-frequency attitude jitter will cause discontinuity between adjacent lines of the image, seriously affecting the geometric accuracy of the stereo model.

[0004] Traditional satellite stereo image processing methods mainly rely on attitude measurement data or ground control points for geometric correction, and it is difficult to effectively compensate for high-frequency attitude jitter under uncontrolled conditions, resulting in reduced elevation accuracy in stereo mapping, poor terrain reconstruction quality, and increased errors in three-dimensional information extraction, seriously restricting the in-depth application of optical satellite images in the field of high-precision stereo mapping. That is to say, there is a technical problem in the prior art that high-frequency attitude jitter leads to low uncontrolled elevation accuracy of optical satellite stereo images. Summary of the Invention

[0005] In view of this, the present invention provides a method for improving the geometric accuracy of a line array image assisted by a same-platform planar array image, which can solve the technical problem in the prior art that high-frequency attitude jitter leads to low uncontrolled elevation accuracy of optical satellite stereo images.

[0006] The present invention is implemented as follows: The present invention provides a method for improving the geometric accuracy of linear array images assisted by same-platform area array images, including: constructing a rigorous imaging geometric model of linear array images and area array images; generating a simulated area array image through external elevation data; obtaining the same-name points of the real area array image and the simulated area array image, and establishing the corresponding relationship between the linear array and the area array image; calculating the upper and lower parallax and eliminating systematic interference; separating high-frequency posture jitter based on an affine transformation model; sorting the same-name points to form a two-dimensional grid, and voting based on parallax mutations; calculating the posture jitter score based on the voting results; optimizing the posture parameters using an iterative strategy; and reconstructing a rigorous geometric model using the optimized posture parameters to generate a geometrically seamless sensor-corrected image.

[0007] Among them, the rigorous imaging geometry model is a satellite-ground homonymous ray mapping model based on collinear equations. The rigorous imaging geometry models of line array images and area array images are similar in form, and both realize the conversion from image point coordinates to ground three-dimensional coordinates based on collinear equations.

[0008] The image point coordinates of the linear array image are represented by a one-dimensional polynomial, while the image point coordinates of the area array image are represented by a two-dimensional polynomial. The linear array image is based on the line scanning principle, while the area array image is based on the area array instantaneous imaging principle.

[0009] The step of generating a simulated area array image through external elevation data is to substitute the three-dimensional coordinates of each pixel position of the area array image on the ground into the strict geometric model of the line array image for inverse calculation, obtain the corresponding image point coordinates, and finally perform grayscale interpolation to obtain the simulated area array image.

[0010] Among them, the step of high-precision matching of real area array images and simulated area array images adopts a strategy that combines feature point matching with region growing. Feature points are extracted through the scale-invariant feature transformation algorithm, feature points are matched using the nearest neighbor ratio method, and mismatched points are eliminated using the random sampling consistency algorithm.

[0011] Among them, the step of calculating the upper and lower parallax and eliminating systematic interference includes calculating the upper and lower parallax of the same-name image points of the line array image and the area array image, and eliminating the interference of projection differences caused by terrain undulations, imaging viewing angle differences and different imaging scales on high-frequency attitude jitter detection.

[0012] Among them, the affine transformation model is used to correct the systematic linear distortion between the linear array image and the planar array image. The affine transformation model error is the high-order distortion caused by high-frequency posture jitter, and the affine transformation model parameters are estimated by the weighted least squares method.

[0013] Among them, the step of voting based on the sudden change of parallax is to sort the homologous points in ascending order according to the row and column coordinates of the linear array image to form a two-dimensional grid, determine the voting parameter space according to the row coordinate range of the two-dimensional grid, and vote based on the difference between the row and column parallaxes of adjacent homologous points; the step of calculating the attitude jitter score by counting the voting results includes counting the voting ratio of each interval and the maximum value of the parallax difference, calculating the voting score, and the higher the voting score, the greater the high-frequency attitude jitter in the interval, and the smaller the corresponding attitude model fitting weight.

[0014] Among them, after calculating the vertical and horizontal parallaxes, it further includes the step of predicting the attitude jitter using the satellite attitude jitter physical model. The satellite attitude jitter physical model is a satellite attitude prediction equation established based on the satellite dynamics principle, which is used to evaluate the high-frequency attitude jitter generated by the satellite during on-orbit operation due to external factor interference, and provides a reference basis for the correction of the affine transformation model.

[0015] Among them, after separating the high-frequency attitude jitter based on the affine transformation model, it further includes the step of predicting and compensating the high-frequency attitude jitter using the self-attention residual network model. The self-attention residual network model uses the vertical and horizontal parallax information as input features and outputs optimized attitude parameters; the structure of the self-attention residual network model is a deep neural network combining multi-layer self-attention mechanisms and residual connections, including a feature extraction layer with the number of attention heads matching the satellite orbital period, a time series encoding layer corresponding to the attitude measurement frequency, a multi-channel output layer corresponding to the satellite three-axis attitude components, and a regularization layer based on attitude physical constraints.

[0016] The present invention realizes the accurate detection and compensation of high-frequency attitude jitter through steps such as constructing a rigorous imaging geometry model, generating simulated area array images, obtaining dense homologous points through high-precision matching, calculating vertical and horizontal parallaxes and excluding interference factors, correcting systematic distortion using an affine transformation model, voting based on the difference between the parallaxes of homologous points, and optimizing attitude parameters using an iterative detection strategy.

[0017] The present invention effectively solves the problem that traditional methods cannot accurately capture high-frequency attitude jitter through the collaborative processing of linear array and area array images. Since the area array image adopts an instantaneous imaging method and is not affected by high-frequency attitude jitter, while the linear array image is sensitive to high-frequency attitude jitter, a quantitative relationship between jitter and image distortion is established through the complementary characteristics of the two imaging methods. The present invention innovatively excludes the interference of projection differences caused by terrain undulation, imaging perspective differences, and different imaging scales, ensures that the parallax change truly reflects the characteristics of high-frequency attitude jitter, and improves the accuracy and reliability of attitude parameters through a voting strategy and iterative optimization, solving the technical problem of low uncontrolled elevation accuracy of optical satellite stereo images caused by high-frequency attitude jitter. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1It is a flowchart of the method of the present invention.

[0019] Figure 2 It is the schematic diagram of the optimization of the linear array row attitude assisted by the area array image in Embodiment 1.

[0020] Figure 3 It is the technical flowchart of the optimization of the linear array row attitude assisted by the area array image in Embodiment 1.

[0021] Figure 4 It is the schematic diagram of the detection of the high-frequency attitude jitter of homologous image points in Embodiment 1.

[0022] Figure 5 It is the schematic diagram of the voting for the high-frequency attitude jitter in Embodiment 1. Specific implementation manners

[0023] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0024] As Figure 1 shown, it is a flowchart of a method for improving the geometric accuracy of an area array image assisted by a linear array image on the same platform provided by the present invention. This method includes the following steps: S01. Construct a rigorous imaging geometric model for the linear array image and the area array image. Both the rigorous imaging geometric model of the linear array image and the rigorous imaging geometric model of the area array image are based on the collinearity equation to realize the mapping of the same-name light rays between the satellite and the ground; S02. Introduce external elevation data, substitute the three-dimensional coordinates of each pixel position of the area array image on the ground into the rigorous geometric model of the linear array image for inverse calculation to obtain the corresponding image point coordinates, and finally perform gray interpolation to obtain a simulated area array image; S03. Perform high-precision matching on the real area array image and the simulated area array image to obtain dense homologous points, and obtain the corresponding positions of the homologous points on the real area array image on the linear array image according to the conversion relationship between the simulated area array image and the linear array image; S04. Calculate the vertical parallax of the homologous image points between the linear array image and the area array image, and exclude the interference of the projection difference caused by terrain undulation, imaging perspective difference, and different imaging scales on the detection of high-frequency attitude jitter; S05. Correct the systematic linear distortion between the linear array image and the area array image by using an affine transformation model. The error of the affine transformation model is the high-order distortion caused by the high-frequency attitude jitter; S06. Sort the homologous points in ascending order according to the row and column coordinates of the linear array image to form a two-dimensional grid, then determine the voting parameter space according to the row coordinate range of the two-dimensional grid, and vote based on the difference in the row and column parallax of adjacent homologous points; S07. Statistically calculate the maximum difference between the voting ratios in each interval and the parallax, and compute the voting score. The higher the voting score, the greater the high-frequency attitude jitter in the interval, and the smaller the corresponding attitude model fitting weight value. S08. Using the condition that the medium error of the parallax is lower than a set threshold as the iteration termination condition, adopt an iterative detection strategy to optimize the attitude parameters and eliminate the influence of the attitude error during the imaging of the area array image on the overall detection result. S09. Utilize the optimized attitude data, combine satellite orbit determination and other auxiliary data to reconstruct the rigorous imaging geometric model of the linear array image, and generate a geometrically seamless sensor calibration image and auxiliary files.

[0025] Optionally, at the end of step S01, it further includes respectively determining the attitude matrix of the satellite and the external calibration parameter matrix; among them, the rigorous imaging geometric model specifically refers to a satellite-ground homologous ray mapping model based on the collinearity equation, and the expression is , is the image point coordinate of the image, is the imaging time, is the scale factor, is the three-dimensional coordinate of the image point in the Earth-centered Earth-fixed coordinate system corresponding to the ground, is the position of the satellite body center, is the satellite attitude matrix, is the external calibration parameter matrix, is the coordinate of the image point in the satellite coordinate system.

[0026] Among them, the of the linear array image specifically refers to a one-dimensional polynomial , , and the of the area array image specifically refers to a two-dimensional polynomial .

[0027] Among them, the affine transformation model specifically refers to , is the coordinate of the linear array image, is the coordinate of the area array image, is the coefficient of the affine transformation model.

[0028] Among them, the voting score specifically refers to , represents the voting score, and is the maximum value of the difference between the voting ratio and the parallax, is the corresponding coefficient.

[0029] Among them, the high-frequency attitude jitter specifically refers to the short-period random fluctuations that exceed the normal attitude change range during the on-orbit operation of the satellite due to external interference. The high-frequency attitude jitter will cause mutations in the vertical parallax of the stereo image that are inconsistent with the overall change trend.

[0030] Among them, the root mean square error of parallax specifically refers to the root mean square difference between the actual parallax and the theoretical parallax of the corresponding points obtained by matching. The root mean square error of parallax is used to evaluate the accuracy of attitude parameters and determine whether the iteration terminates.

[0031] Among them, the sensor-corrected image specifically refers to an image product that eliminates image misalignment error, internal distortion error, and attitude orbit error. The sensor-corrected image has the characteristic of geometric seamlessness, which is convenient for subsequent precise measurement and information extraction.

[0032] Optionally, between step S04 and step S05, a physical model of satellite attitude jitter is used to predict attitude jitter. The physical model of satellite attitude jitter is used to evaluate the high-frequency attitude jitter generated by the satellite during on-orbit operation due to external factor interference, providing a reference basis for subsequent correction of the affine transformation model; Optionally, between step S05 and step S06, a self-attention residual network model is used to predict and compensate for the high-frequency attitude jitter. The self-attention residual network model uses the vertical parallax information calculated in step S04 as input features and outputs optimized attitude parameters, providing a more accurate basis for identifying the high-frequency attitude jitter for the subsequent voting strategy.

[0033] Among them, the physical model of satellite attitude jitter specifically refers to the satellite attitude prediction equation established based on satellite dynamics principles , the physical model of satellite attitude jitter is used to predict the high-frequency attitude jitter generated by the satellite during on-orbit operation due to external factor interference. The inputs of the physical model of satellite attitude jitter include the solar radiation pressure factor , the earth's gravity gradient factor , the geomagnetic disturbance factor , the thermoelastic deformation factor , and the mechanical vibration factor . The solar radiation pressure factor is obtained from the change in the current of the satellite's solar panels. The earth's gravity gradient factor is calculated from the satellite's orbital position. The geomagnetic disturbance factor is obtained from the measurement of the on-board magnetometer. The thermoelastic deformation factor is obtained from the measurement of the temperature sensor array. The mechanical vibration factor is obtained from the high-frequency measurement data of the gyroscope. The output of the physical model of satellite attitude jitter is the satellite attitude angle offset , the satellite attitude angle offset is used to compensate the attitude data measured by the attitude measurement device; the combination of the satellite attitude jitter physical model and the adaptive attitude weighted fitting further improves the accuracy and reliability of attitude optimization.

[0034] Among them, the specific structure of the self-attention residual network model is a deep neural network combining multi-layer self-attention mechanism and residual connection, including a feature extraction layer with the number of attention heads matching the satellite orbit period, a time series encoding layer corresponding to the attitude measurement frequency, a multi-channel output layer corresponding to the satellite three-axis attitude components, and a regularization layer based on attitude physical constraints; the steps for establishing the training data set of the self-attention residual network model specifically include collecting attitude measurement data, real attitude change data, camera internal and external parameters, and external disturbance factor data during the imaging processes of the linear array image and the area array image on the historical orbit segment, stratifying and sampling the attitude measurement data, the real attitude change data, the camera internal and external parameters, and the external disturbance factor data according to different orbits and imaging conditions, and constructing training samples covering different working conditions; the steps for training the self-attention residual network model specifically include first pre-training the self-attention layer parameters in an unsupervised learning manner to capture attitude time series features, then finely tuning the entire network parameters in a supervised learning manner to minimize the error between the predicted attitude and the real attitude, and finally introducing a regularization term based on physical constraints to ensure that the model prediction results conform to the satellite dynamics characteristics.

[0035] The specific implementation manners of the above steps are described in detail below.

[0036] The specific implementation manner of step S01 is to first obtain the original orbit parameters, attitude parameters, and camera parameters transmitted by the satellite up and down, and then establish a rigorous imaging geometric model for the linear array image and the area array image respectively based on the collinearity equation principle. For the linear array image, its internal orientation elements are constructed according to the line-by-line scanning principle, including focal length, principal point coordinates, and distortion coefficients, etc., and at the same time, a corresponding relationship is established between the line number and the imaging time, so as to realize a rigorous one-dimensional projection model; for the area array image, a two-dimensional projection model is constructed according to the area array instantaneous imaging principle, and the projection relationship between the image plane coordinates and the ground target points is established. Then, the external orientation elements, including the satellite position vector and attitude angles, are determined according to the satellite orbit determination data and attitude measurement data. Finally, by combining the internal and external orientation elements, a complete rigorous imaging geometric model is established to realize the mapping relationship from the image plane coordinates to the ground three-dimensional coordinates. The purpose of this step is to establish a rigorous geometric basis for subsequent processing and ensure the accuracy of coordinate transformation.

[0037] The specific implementation of step S02 is to first obtain external elevation reference data corresponding to the satellite coverage area. Global digital elevation models such as SRTM or ASTER GDEM can be used, and their accuracy should be better than 30m. Then, the ground points corresponding to each pixel of the area array image are forward projected through a rigorous geometric model to calculate their ground three-dimensional coordinates. When determining the ground three-dimensional coordinates, an iterative solution method needs to be used to project the ground points onto the elevation model. Usually, the iterative convergence condition is set to an elevation change less than 0.1m. Subsequently, the obtained ground three-dimensional coordinates are substituted into the rigorous geometric model of the line array image for back calculation to obtain the corresponding line array image coordinates. Finally, the bicubic convolution interpolation algorithm is used to calculate the gray value on the line array image, and the interpolated gray value is assigned to the corresponding pixel of the area array image to generate a simulated area array image. The purpose of this step is to eliminate the geometric differences caused by different imaging methods and provide a geometrically consistent basis for subsequent image matching.

[0038] The specific implementation of step S03 is to adopt a strategy combining feature point matching and dense matching to obtain the corresponding points of the real area array image and the simulated area array image. First, the scale-invariant feature transform algorithm is used to extract the feature points of the two images, and then the nearest neighbor ratio method is used for the initial matching of the feature points. The matching threshold is set to 0.6 to balance the matching accuracy rate and the number of matches. Then, the random sample consensus algorithm is used to eliminate the mismatched points, and the inlier threshold is set to 0.5 pixels. Then, taking the filtered feature points as seed points, the gradient correlation coefficient algorithm is used for region growing, and the matching correlation coefficient threshold is set to 0.8 to generate dense matching point pairs. Finally, through the conversion relationship between the simulated area array image and the line array image, the coordinates of the corresponding points on the real area array image are converted to the line array image, thereby establishing an accurate correspondence between the real area array image and the line array image. To ensure the uniform distribution of the matching points, a grid constraint strategy is adopted. The image is divided into a 10×10 uniform grid, and no less than 10 matching points are extracted from each grid. The purpose of this step is to establish an accurate correspondence between the area array image and the line array image and provide basic data for subsequent attitude error detection.

[0039] The specific implementation of step S04 is as follows: First, using the corresponding points obtained in step S03, calculate the vertical parallax between the line array image and the area array image. The vertical parallax is calculated by subtracting the coordinates of the area array image on the line array image from the coordinates of the line array image. Then, construct a terrain influence elimination model. This model is based on the principle of common vision geometry and calculates the parallax component caused by terrain undulation through elevation data and imaging geometry. Next, calculate the parallax component caused by the difference in imaging perspectives. This component is related to the angular difference between the imaging centers of the two images. After that, calculate the parallax component caused by different imaging scales. This component is related to the ratio of the ground resolutions of the two images. Finally, subtract the above three systematic components from the total parallax to obtain the remaining parallax, which mainly reflects the geometric disturbance caused by high-frequency attitude jitter. To improve the accuracy, a robust estimation method is used to process outliers. When the residual is greater than twice the mean square error, it is marked as an outlier and its weight is reduced. The purpose of this step is to eliminate the systematic factors affecting high-frequency attitude jitter detection and extract the geometric disturbance information purely caused by attitude jitter.

[0040] The specific implementation of step S05 is to correct the systematic linear distortion between the line array image and the area array image using an affine transformation model. First, based on the remaining parallax obtained in step S04, use the weighted least squares method to estimate the parameters of the affine transformation model. The weight is proportional to the reliability of the parallax. Then, calculate the predicted coordinates of each corresponding point under the affine transformation model, compare them with the actual coordinates, and calculate the parallax residual. Next, analyze the spatial distribution characteristics and frequency characteristics of the parallax residual. Use the fast Fourier transform to convert the parallax residual to the frequency domain and identify the high-frequency components. Finally, determine the high-order distortion part in the parallax residual, that is, the non-linear distortion component caused by high-frequency attitude jitter. For the estimation of the affine transformation model parameters, when the number of points is insufficient, the singular value decomposition method is used to enhance the solution stability; when the redundancy is high, a random sampling strategy is used to reduce the influence of outliers. The purpose of this step is to separate the linear systematic error and the non-linear high-frequency jitter error, providing a basis for the accurate identification of high-frequency jitter in the subsequent process.

[0041] The specific implementation of step S06 is to first sort the corresponding points obtained in step S05 in ascending order according to the row and column coordinates in the line array image. The row coordinate corresponds to the time dimension, and the column coordinate corresponds to the space dimension. Then, a two-dimensional grid is constructed, and the grid size is set to 10 times the resolution in the row direction of the line array image to balance accuracy and computational efficiency. To handle the pointless area caused by matching failure, the grid tolerance is set to 20% of the original grid size. Next, the voting parameter space is determined according to the row coordinate range of the two-dimensional grid. The row coordinate range is evenly divided into several intervals, and the number of intervals is usually set to 1% of the total number of rows, but not less than 100 intervals. Then, in each column direction, the absolute value of the difference in row and column disparities between adjacent corresponding points is calculated. When the disparity difference is greater than the 0.5-pixel threshold, it is accumulated in the voting parameter space corresponding to the row coordinate, and the maximum disparity difference value of the column where it is located is recorded. Finally, all columns are traversed to complete the voting statistics work, forming a complete distribution map of the voting parameter space. The purpose of this step is to identify the time intervals where high-frequency attitude jitter occurs through the voting mechanism, improving the reliability and robustness of jitter detection.

[0042] The specific implementation of step S07 is to first count the voting numbers in each interval in step S06 and calculate the voting ratio, that is, the ratio of the voting number in each interval to the total voting number. At the same time, the maximum disparity difference value in each interval is counted, which reflects the severity of attitude jitter. Then, the comprehensive voting score is calculated according to the voting ratio and the maximum disparity difference value. The weight coefficients in the calculation formula are usually set as follows: the voting ratio weight is 0.7, and the maximum disparity difference value weight is 0.3 to balance the two factors of frequency and intensity. Next, the voting score is normalized to make it distributed between 0 and 1 for subsequent processing. Finally, based on the normalized voting score, the attitude model fitting weight value is determined, and an exponential decay function is used to convert the score into a weight value. High-score intervals correspond to low weight values, and low-score intervals correspond to high weight values. The weight value range is usually set between 0.1 and 1.0 to ensure that the attitude data in the high-frequency jitter area is appropriately down-weighted in subsequent processing. The purpose of this step is to quantify the influence degree of high-frequency attitude jitter and provide a reasonable weight allocation scheme for attitude model fitting.

[0043] The specific implementation of step S08 is to optimize the attitude parameters using an iterative detection strategy. First, based on the attitude model fitting weights determined in step S07, the original attitude data is filtered using the weighted spline smoothing method. The smoothing parameter is adaptively determined according to the sampling frequency of the attitude data and is usually set to 5 to 10 times the sampling interval. Then, the processed data is used to repeat steps S02 to S07 to obtain new attitude model fitting weights. Next, the mean square error of the parallax after two adjacent iterative processes is calculated, which is the root mean square difference between the actual parallax and the theoretical parallax of the matching points. When the mean square error of the parallax is less than 0.3 pixels or the change in the mean square error of the parallax between two adjacent iterations is less than 5%, the iteration termination condition is reached. Finally, the optimal attitude parameters at the end of the iteration are output. If the iteration does not converge after 10 times, it is forced to terminate and the result of the last iteration is output. The purpose of this step is to gradually improve the accuracy of the attitude parameters through iteration and eliminate the interference of the attitude error during the imaging of the area array image on the detection result.

[0044] The specific implementation of step S09 is to first combine the attitude data optimized in step S08 with the original orbit determination data to reconstruct the rigorous imaging geometric model of the linear array image. Then, according to the optimized geometric model, the precise ground coordinates of each pixel on the original segmented linear array image are calculated. After that, the output image space is constructed, usually using an equally spaced projection method to establish the mapping relationship between the ground coordinates and the output image coordinates. Then, the original image is resampled into the output image space by back-projection, and bilinear interpolation is used for gray-scale resampling. To eliminate the seam effect, a weighted fusion method is used in the overlapping area, and the weight is proportional to the distance of the pixel to the seam. Finally, a geometrically seamless sensor-corrected image is generated, and the rational function model coefficients are output as an attached file. The order of the coefficients is usually third-order, and the control point residual is required to be better than 0.5 pixels. The purpose of this step is to generate a corrected image product with higher geometric accuracy using the optimized attitude parameters and provide a high-quality data basis for subsequent applications.

[0045] Another specific implementation of the present invention is provided below. In the following specific implementation, the difference from the previous specific implementation is that: the specific implementation of predicting attitude jitter using the satellite attitude jitter physical model between step S04 and step S05 is to first establish an attitude jitter prediction equation based on satellite dynamics. This equation uses the solar radiation pressure factor, the earth's gravity gradient factor, the geomagnetic disturbance factor, the thermoelastic deformation factor, and the mechanical vibration factor as input parameters to predict the attitude angle offset of the satellite at the orbital position. Among them, the solar radiation pressure factor is obtained by monitoring the current change of the satellite's solar panels, and the typical change range is 0.01 - 0.05; the earth's gravity gradient factor is calculated according to the satellite's orbital position, and the magnitude is usually about ; The geomagnetic field disturbance factor is obtained from the measurements of the on-board magnetometer, with a typical variation range of 20 - 100 nT; the thermoelastic deformation factor is obtained from the measurements of the temperature sensor array, and the attitude change caused by a 1°C change in the satellite structure temperature is approximately 0.1 - 0.5 arcseconds; the mechanical vibration factor is obtained from the high-frequency measurement data of the gyroscope, with a typical amplitude of 0.5 - 2 arcseconds. Then, by adjusting the weight coefficients of each factor, the attitude jitter model is calibrated. The calibration process uses historical orbit data, and at least 10 typical orbit segments are selected, including data under different seasons and different sunlight conditions. Next, the factor data obtained in real time is substituted into the calibrated model to calculate the predicted attitude angle offset. Finally, the predicted result is converted into the geometric offset in the pixel domain, which is used as the prior information for the affine transformation model correction in step S05. The prediction result of the physical model is usually used as the initial value and fused with the subsequent image processing result. The fusion adopts the Bayesian framework, where the physical model provides the prior distribution and the image processing result provides the observed value to jointly obtain the optimal estimate. The model calculation is carried out in a recursive manner, and the time step is consistent with the attitude measurement frequency, usually 10 - 100 Hz. The purpose of this optimization step is to introduce physical constraints, reduce the uncertainty of the attitude error estimation, and improve the reliability and continuity of the attitude estimation.

[0046] The specific implementation method of using the self-attention residual network model to predict and compensate high-frequency attitude jitter between step S05 and step S06 is to first construct a deep neural network combining a multi-layer self-attention mechanism with a residual connection. The network contains four key components: a feature extraction layer, a time series encoding layer, a multi-channel output layer, and a regularization layer. The feature extraction layer uses 8 attention heads, corresponding to the satellite orbit period of about 90 minutes; the sampling rate of the time series encoding layer is consistent with the attitude measurement frequency of 10 to 100 Hz; the multi-channel output layer corresponds to the satellite three-axis attitude component; the regularization layer constructs physical constraints based on the attitude dynamics equation. Then, a training data set is constructed to collect attitude measurement data, real attitude change data, camera internal and external parameters, and external disturbance factor data during the imaging process of linear array images and planar array images on the historical orbit segment. The data collection covers no less than 100 orbital cycles and includes a variety of working conditions. Then a three-stage training strategy is adopted: first, the self-attention layer parameters are pre-trained by unsupervised learning, using the contrastive learning loss function, the learning rate is set to 0.001, and the number of training rounds is 100; then the entire network parameters are fine-tuned by supervised learning, the loss function is the root mean square error, the learning rate is set to 0.0001, and the number of training rounds is 200; finally, a regularization term based on physical constraints is introduced, the regularization coefficient is 0.1, and the number of training rounds is 50. The training adopts a batch processing method with a batch size of 64, and the optimizer uses an adaptive moment estimation algorithm. Finally, the upper and lower disparity information calculated in step S04 is used as the input feature, and the high-frequency posture jitter is predicted by the trained network model. The prediction result is used as an auxiliary basis for the voting strategy in step S06 to improve the accuracy of jitter recognition. The model prediction result is fused with the voting result of S06, and the weighted average method is adopted. The weight is determined according to the confidence of the two. Usually, the model prediction weight is between 0.3 and 0.7. When disparity data is missing, the model prediction weight is increased; when disparity data is complete and consistent, the voting result weight is increased. The purpose of this optimization step is to use deep learning technology to mine implicit patterns in posture data and improve the accuracy and robustness of posture jitter prediction.

[0047] The introduction of the physical model of satellite attitude jitter provides prior knowledge based on physical mechanisms for high-frequency attitude jitter detection and compensation, significantly improving the accuracy and reliability of affine transformation model calibration. Compared with the traditional method that solely relies on image matching, this physical model comprehensively considers the effects of various external interference factors such as solar radiation pressure, earth's gravitational gradient, geomagnetic perturbation, thermoelastic deformation, and mechanical vibration on the satellite attitude, and establishes a quantitative relationship between these factors and high-frequency attitude jitter. By real-time acquiring relevant parameters from various satellite sensors, it predicts the possible attitude fluctuation trends and amplitudes, providing a reference basis for subsequent affine transformation model calibration, and effectively avoiding the problems of parameter divergence or local optimality caused by the lack of physical constraints in the traditional method. Especially for the orbital segments with complex attitude changes and diverse interference factors, this physical model can accurately identify the differences between real attitude jitter and noise, improving the robustness of jitter feature extraction and laying a foundation for the construction of high-precision stereo models.

[0048] The application of the self-attention residual network model realizes the intelligent prediction and compensation of high-frequency attitude jitter, solving the problems of nonlinear and time-varying attitude jitter that are difficult to handle by traditional methods. This network model breaks through the limitations of the expression ability of traditional mathematical models, captures the long-term and short-term dependence relationships in attitude time-series data through multiple layers of self-attention mechanisms, and the residual connection effectively avoids the problem of gradient disappearance in the training of deep networks. Compared with the existing technologies, this model can adaptively learn the attitude jitter characteristics under different orbital conditions and different interference patterns, and achieve accurate modeling of complex attitude changes. Especially the design of the feature extraction layer in the model where the number of attention heads matches the satellite orbital period can accurately capture the periodic attitude changes related to the orbital period; while the regularization layer based on attitude physical constraints ensures that the model prediction results conform to the satellite dynamics characteristics, avoiding the risk of overfitting. Through the in-depth excavation of the up-down parallax information by this model, it can identify the tiny attitude fluctuations that are difficult to discover by conventional methods, providing key technical support for improving the uncontrolled elevation accuracy.

[0049] Specifically, the principle of the present invention is as follows: The technical principle of the present invention is based on the idea of collaborative processing of linear array and area array images, and detects attitude anomalies by using the differences in the sensitivity of the two imaging methods to high-frequency attitude jitter. The linear array image adopts the push-broom imaging mode, and each row is acquired at different times, resulting in the geometric discontinuity between adjacent rows on the image due to high-frequency attitude jitter; while the area array image adopts the instantaneous imaging mode, and the entire image is acquired at the same time, not affected by high-frequency attitude jitter. By establishing a strict geometric relationship between the two images, the satellite attitude changes can be deduced inversely, and then the accurate detection and compensation of high-frequency attitude jitter can be realized.

[0050] Specifically, the present invention first constructs a rigorous imaging geometric model of linear array and area array images based on collinearity equations, accurately describing the mapping relationship of homologous rays between the satellite and the ground. Then, external elevation data is introduced, and the three-dimensional coordinates of each pixel of the area array image on the ground are substituted into the geometric model of the linear array image for back-calculation to obtain a simulated area array image under the ideal condition of no high-frequency attitude jitter. Through the high-precision matching of the real area array image and the simulated area array image, dense homologous points are obtained, and the corresponding positions of the homologous points on the linear array image are determined by using the conversion relationship between the simulated area array image and the linear array image.

[0051] The key innovation of the present invention lies in that during the calculation of the vertical parallax of homologous image points, the interference of projection errors caused by terrain undulation, imaging perspective differences, and different imaging scales is creatively excluded, enabling the parallax change to accurately reflect high-frequency attitude jitter. After correcting the systematic linear distortion using an affine transformation model, the residual error is the high-order distortion caused by high-frequency attitude jitter. Based on the increasing order of the row and column coordinates of the homologous points on the linear array image to form a two-dimensional grid, and using the differences in the row and column parallaxes of adjacent homologous points for voting, the abnormal attitude change area can be accurately identified.

[0052] The present invention adopts an iterative detection strategy to optimize the attitude parameters, excluding the influence of attitude errors during the imaging of the area array image on the overall detection result. Taking the root mean square error of the parallax being lower than the set threshold as the iteration termination condition, the reliability and accuracy of the detection result are ensured. The optimized attitude data is used to reconstruct the rigorous imaging geometric model of the linear array image, generating a geometrically seamless sensor correction image, significantly improving the geometric accuracy and uncontrolled elevation accuracy of the stereo image, and solving the technical problem of low accuracy of the stereo model caused by high-frequency attitude jitter.

[0053] To better understand and implement the present invention, an embodiment 2 of a simplified version of the present invention is provided below. This embodiment 2 can better reflect the principle of the present invention. By utilizing the different imaging principles of linear and area array images, the row attitude of the linear array is inverted with the instantaneous imaging parameters of the area array, the high-frequency jitter in the attitude is detected through the change of the image parallax, and its random influence on the attitude is adaptively weighted and weakened to improve the pointing of the photographic rays, enabling the homologous rays to intersect pairwise and improving the relative accuracy of the stereo model. As Figure 2 shown.

[0054] The main technical process for improving the geometric accuracy of the linear array image assisted by the area array image is as Figure 3 shown, where Figure 3There are multiple sub - figures used as icons, which are only used as icons and do not represent specific meanings. The main research plan is as follows: First, construct a rigorous imaging geometric model for line - array images and area - array images, and combine external elevation data to realize the simulation of area - array images based on line - array images. On this basis, introduce a high - precision image matching algorithm to obtain dense corresponding image points between line - and area - images. Second, according to the rigorous imaging geometric model of line - and area - images, determine the vertical parallax of corresponding image points, and preliminarily detect high - frequency jitter of the attitude according to the change of parallax. Different error factors affecting image distortion should be considered during the detection. Then, based on the voting rule, with the line - array scanning behavior as the voting interval, determine the high - frequency jitter in the row attitude according to the sudden change of the vertical parallax of corresponding image points, and adaptively determine the fitting weight value of high - frequency jitter to achieve high - frequency attitude filtering and improvement of the accuracy of the attitude model. Finally, according to the revised satellite attitude data, combined with orbit - determination data and other auxiliary data, reconstruct the rigorous imaging model of the satellite line - array image, splice the original images, generate a geometrically seamless sensor - corrected image product, and generate accessory files at the same time. This embodiment is described in detail as follows.

[0055] The area - array image simulation module will generate a reconstructed image with the same resolution as the real area - array image and a similar imaging perspective. In this module, first, based on the orbit, attitude, and imaging time parameters transmitted from the satellite, construct rigorous imaging geometric models for both area - array images and line - array images respectively. Then, combine external elevation data to determine the three - dimensional coordinates of each pixel position of the area - array image

[0056] in the ground , and then substitute the above three - dimensional coordinates into the rigorous geometric model of the line - array image for inverse calculation to obtain the corresponding image point coordinates of the line - array image . Finally, perform gray - level interpolation on the coordinates on the line - array image to obtain the gray - level value , and assign this gray - level value to the area - array image . Loop through the entire range of the area - array image in this way to obtain the resampled simulated area - array image.

[0057] The rigorous imaging geometric model of the area - array image is similar in form to that of the line - array image. Both are based on the mapping of the same - name rays between the satellite and the ground in the collinearity equation, and their common form is as follows: ; Among them, are the image point coordinates of the image; t is the imaging time, and for the line - array image, it has a strict corresponding relationship with the image column - direction coordinate y; m is the proportionality coefficient; are the three - dimensional coordinates of the image point in the Earth - centered Earth - fixed coordinate system (ECEF) corresponding to the ground, is the position of the center of the satellite body determined by the on - satellite positioning device, is the satellite attitude matrix determined by the attitude measurement equipment, is the external calibration parameter matrix, mainly compensating for attitude and orbit measurement errors and equipment installation errors; is the coordinate of the image point in the satellite coordinate system. For the line array image, its form adopts the following formula: ; The area array image is obtained by instantaneous shooting of the area array imaging unit, and its is composed of two-dimensional general polynomials, as shown in the following formula: ; Introduce a high-precision image matching algorithm to conduct dense matching on real area array images and simulated area array images, obtain evenly covered corresponding points, and obtain the corresponding positions of the corresponding points on the real area array image on the line array image according to the conversion relationship between the simulated area array image and the line array image. In this way, the projection errors caused by terrain undulation, imaging perspective differences, and different imaging scales can be automatically compensated during image reconstruction and image correlation processes, excluding the interference of relevant errors on high-frequency attitude jitter detection.

[0058] When the satellite attitude changes suddenly and greatly, it will cause a sudden change in the vertical parallax of the stereo image near that moment, resulting in a jump that is inconsistent with the overall parallax change trend. However, when the attitude is relatively stable as a whole and only contains systematic errors and small random errors, the above phenomenon will not occur. For multi-payloads on the same platform, their attitude errors are the same, but the imaging times are different. Assume that the attitude model of the satellite is , and the attitude model error is , then the main factor causing the inconsistency of the geometric models of the area array camera and the line array camera is the change of the attitude error over time, that is , which is also the non-synchronous observation error. Because the non-synchronous observation error is mainly reflected as a sudden change in the vertical parallax of the image, and the area array is instantaneously imaged, its attitude error is a fixed value, so the line array row attitude error can be obtained. The overall detection schematic diagram is as shown in Figure 4 .

[0059] The above detection process does not determine the exact attitude error, but determines whether the attitude is high-frequency jitter within the change range according to the change of the vertical parallax of the image. However, the satellite image distortion is also affected by factors such as internal camera error, orbit measurement error, dynamic change of payload installation, terrain undulation, and projection angle difference. For the latter two, their influence is mainly in the along-track direction, and can be compensated through image simulation and correlation. For the former three, since the in-orbit geometric calibration accuracy can reach 0.3 pixels, the orbit measurement accuracy has reached the centimeter level, and the change of payload installation is mainly a systematic error in a short time, the image distortion generated by them is mainly linear distortion. At this time, a linear deviation compensation model is proposed to correct these distortions and further eliminate the interference of related errors on the detection of high-frequency attitude jitter, such as the affine transformation model: ; Among them, is the line array image coordinate, is the area array image coordinate, are the coefficients of the affine transformation model, and its model error is the high-order distortion caused by the high-frequency attitude jitter.

[0060] The high-frequency attitude jitter detection module determines the position of the high-frequency jitter. In order to quantify its influence on the attitude model, a method for determining the weight of high-frequency attitude jitter based on the voting rule is proposed. The main process is as follows: First, sort the row and column coordinates of the homologous points in the line array image in ascending order to form a two-dimensional grid. To avoid the point-free area caused by image matching failure, a certain grid tolerance is allowed; Second, determine the abscissa of the voting parameter space according to the range of the row coordinates (corresponding to the attitude time) of the two-dimensional grid, and the ordinate is the number of votes; Then, calculate the difference (taking the absolute value) of the row and column parallaxes of adjacent homologous points in each row of the two-dimensional grid. If the difference in parallax is greater than a certain threshold, vote at the corresponding position in the voting parameter space, and record the maximum value of the difference in parallax at the same time. Vote in this way and update the maximum value of the difference in parallax; Finally, count the voting ratio and the maximum value of the difference in parallax in each interval. The overall voting schematic diagram is as Figure 5 shown. Calculate the voting score using the voting ratio and the maximum value of the difference in parallax: ; Among them, represents the voting score, E and D are the voting ratio and the maximum value of the difference in parallax, are the corresponding coefficients. The higher the voting score, the greater the high-frequency jitter in this interval, and the smaller the corresponding attitude model fitting weight.

[0061] For each perspective image in the stereo model, the above process should be carried out to ensure that high-frequency attitude detection covers the entire orbit attitude measurement time. In addition, in order to eliminate the influence of attitude errors during the imaging of area array images on the overall detection results, an iterative detection strategy is proposed, that is, after adaptive attitude weighted fitting, new attitude data is used for image simulation and high-frequency attitude jitter detection, and the iterative termination condition is that the mean square error of parallax is lower than a certain threshold. At this time, the optimal attitude data will be obtained.

[0062] Using the aforementioned refined attitude data, combined with satellite orbit determination and other auxiliary data, a rigorous imaging geometric model of the satellite linear array image is generated. Using the ground auxiliary elevation reference data, the original segmented satellite images are mosaicked, synchronously eliminating image misalignment errors, internal distortion errors, and attitude orbit errors, producing geometrically seamless sensor-corrected images. At the same time, a beneficial function model of the sensor-corrected image product is constructed using a terrain-independent method, and other auxiliary files are generated synchronously.

[0063] The following provides a specific embodiment 2 of the present invention, and the specific implementation manners of each step in this embodiment 2 are described in detail as follows.

[0064] The specific implementation manner of step S01 is to construct a rigorous imaging geometric model of the linear array image and the area array image respectively based on the collinearity equation principle. For the linear array image and the area array image, a rigorous imaging geometric model is established in the form of the collinearity equation, which is specifically expressed as follows: ; In the formula, is the image point coordinate of the image, with the unit of pixel; is the imaging time, with the unit of second. In the linear array image, has a strict corresponding relationship with the row number ; is the scale coefficient, dimensionless; is the three-dimensional coordinate of the image point in the Earth-centered Earth-fixed coordinate system on the ground, with the unit of meter; is the position coordinate of the satellite body center, with the unit of meter; is the satellite attitude matrix, which is an orthogonal matrix; is the external calibration parameter matrix, which is a matrix; is the coordinate of the image point in the satellite coordinate system, with the unit of meter. For the linear array image, is represented in the form of a one-dimensional polynomial, which is specifically as follows: ; ; In the formula, represents the highest degree of the polynomial; ( ) are polynomial coefficients, which are obtained through laboratory or on-orbit calibration. For area array images, They are represented in the form of a two-dimensional polynomial as follows: ; ; In the formula, ( ) are two-dimensional polynomial coefficients, which are obtained through laboratory or on-orbit calibration. In the construction process, the original orbit parameters, attitude parameters and camera parameters transmitted by the satellite are first obtained, and then the interior orientation elements and exterior orientation elements are calculated. Finally, a complete geometric model is formed. The purpose of this step is to establish a rigorous geometric basis for subsequent processing, ensure the accuracy of coordinate transformation, and guarantee the theoretical support for subsequent attitude detection and optimization.

[0065] The specific implementation of step S02 is to generate a simulated area array image by using external elevation data and a rigorous imaging geometric model. First, the external elevation reference data corresponding to the satellite coverage area is obtained, with an accuracy better than 30 meters. Then, for each area array pixel , its corresponding ground three-dimensional coordinates are calculated through forward projection using the collinearity equation. The specific process uses an iterative solution method. The initial elevation value is taken as the regional average elevation, and the iterative convergence condition is that the elevation change is less than 0.1 meter or the number of iterations reaches 10 times. Then, the calculated ground three-dimensional coordinates are substituted into the rigorous geometric model of the linear array image for back calculation to obtain the corresponding image point coordinates of the linear array image. The calculation formula is: ; In the formula, is the scale factor; is the camera focal length, in pixels; the meanings of the other parameters are the same as those in step S01. After obtaining from the above formula, the gray value is calculated on the linear array image using the bicubic convolution interpolation algorithm. The interpolation kernel function is: ; In the formula, is the normalized distance. Finally, the interpolated gray value is assigned to the area array image coordinate , and the generation of the simulated area array image is completed. The purpose of this step is to eliminate the geometric differences caused by different imaging methods, provide a geometrically consistent basis for subsequent image matching, and ensure the accuracy of the corresponding relationship of homologous points.

[0066] The specific implementation of step S03 is to obtain dense corresponding points of the real area array image and the simulated area array image by combining the Scale-Invariant Feature Transform (SIFT) algorithm and the region growing algorithm. First, feature points are extracted. For each image point , a scale space representation is constructed: ; In the formula, is the Gaussian kernel function, is the scale parameter; is the original image; represents the convolution operation. Then, local extreme points are detected in the scale space through the Difference of Gaussian (DoG) function: ; In the formula, is the scale factor, and its value is usually . Next, feature point matching is performed using the nearest neighbor ratio method: ; In the formula, represents the distance between the feature point and its nearest neighbor ; represents the distance between the feature point and its second nearest neighbor ; is the threshold, and its value is 0.6. Then, the Random Sample Consensus (RANSAC) algorithm is used to eliminate the mismatched points. The inlier judgment criterion is that the residual is less than 0.5 pixels. Finally, the filtered feature points are used as seed points, and the region growing is performed using the gradient correlation coefficient algorithm. The size of the correlation window is usually taken as pixels, and the correlation coefficient threshold is taken as 0.8 to generate dense matching point pairs. The corresponding points on the real area array image are transformed to the line array image through the conversion relationship between the simulated area array image and the line array image, and the corresponding relationship between the real area array image and the line array image is established. The purpose of this step is to establish an accurate corresponding relationship between images of different imaging modes and provide basic data for subsequent attitude error detection.

[0067] The specific implementation of step S04 is to calculate the vertical parallax of corresponding image points between the line array image and the area array image and eliminate the interference factors affecting the detection of high-frequency attitude jitter. First, the vertical parallax between the line array image and the area array image is calculated: ; ; In the formula, is the coordinate of the line array image; is the corresponding coordinate of the area array image on the line array image; and They are the row - direction parallax and the column - direction parallax respectively. Then, the influence of various factors on the parallax is separated, and the total parallax can be expressed as the superposition of multiple factors: ; ; In the formula, and are the parallax components caused by terrain undulation; and are the parallax components caused by the difference in imaging perspectives; and are the parallax components caused by different imaging scales; and are the parallax components caused by high - frequency attitude jitter; and are random noises. The parallax component caused by terrain is calculated as: ; ; In the formula, and are the partial derivatives of parallax with respect to elevation; is the elevation error. The parallax component caused by the difference in imaging perspectives is related to the difference in imaging center angles between the two images, and is usually expressed as: ; ; In the formula, and are scale factors. The parallax component caused by different imaging scales is related to the ratio of the ground resolutions of the two images : ; ; In the formula, and are image - point coordinates. By subtracting the above - mentioned systematic components, the residual parallax is obtained: ; ; In the formula, and are the residual parallax, which mainly includes the parallax component caused by high - frequency attitude jitter and random noise. The purpose of this step is to eliminate the systematic factors that affect the detection of high - frequency attitude jitter and extract the geometric perturbation information purely caused by attitude jitter.

[0068] The specific implementation of step S05 is to use an affine transformation model to correct the systematic linear distortion between the linear array image and the area array image. The affine transformation model is expressed as: ; ; In the formula, is the coordinate of the linear array image; is the coordinate of the area array image; ( ) are the coefficients of the affine transformation model. According to the remaining disparity obtained in step S04, the parameters of the affine transformation model are estimated by weighted least squares: ; In the formula, is the number of corresponding points; is the weight of the th point, which is proportional to the reliability of the disparity. After solving the above least squares problem to obtain the parameters of the affine transformation model, calculate the difference between the predicted coordinates and the actual coordinates of each corresponding point under the affine transformation model, that is, the disparity residual: ; ; In the formula, and are the disparity residuals, which mainly reflect the non-linear distortion caused by high-frequency jitter of the attitude. Perform frequency analysis on the disparity residuals, using the fast Fourier transform: ; In the formula, is the disparity residual on the time series; is the frequency; is the sequence length; is the frequency domain representation. Extract the high-frequency components from the frequency domain representation to determine the high-frequency attitude jitter characteristics. The purpose of this step is to separate the linear system error and the non-linear high-frequency jitter error, providing a basis for the accurate identification of high-frequency jitter in the subsequent steps.

[0069] The specific implementation of step S06 is to construct a two-dimensional grid and vote based on the disparity mutation. First, sort the corresponding points in ascending order according to the row and column coordinates to construct a two-dimensional grid: ; In the formula, and are the center coordinates of the row and column grids respectively; and They are the numbers of row and column grids respectively. The grid size is usually set to 10 times the resolution in the row direction of the linear array image, and the allowable grid tolerance is 20% of the grid size. Then, the voting parameter space is determined according to the row coordinate range of the two-dimensional grid: ; In the formula, represents the th voting interval, is the number of intervals, usually taking 1% of the total number of rows but not less than 100. Then, in each column direction, calculate the absolute value of the difference in row and column parallaxes of adjacent homologous points: ; ; In the formula, and are the row and column parallax residuals at the grid point respectively. When the difference in parallaxes is greater than the threshold (usually taking 0.5 pixels), accumulate votes in the voting parameter space corresponding to the row coordinate: , if or and falls within the interval ; At the same time, record the maximum value of the difference in parallaxes in this interval: ; Traverse all grid points to complete the vote statistics. The purpose of this step is to identify the time intervals where high-frequency attitude jitter occurs through the voting mechanism, providing a basis for attitude model refinement.

[0070] The specific implementation of step S07 is to calculate the voting score and determine the fitting weight of the attitude model. First, calculate the voting ratio of each interval: ; In the formula, is the voting ratio of the th interval. Then, calculate the voting score according to the voting ratio and the maximum value of the difference in parallaxes: ; In the formula, represents the voting score; and are the voting ratio and the maximum value of the difference in parallaxes respectively; and are the corresponding coefficients, usually , . Then, normalize the voting score: ; Finally, determine the pose model fitting weight based on the normalized voting score: ; In the formula, is the attenuation coefficient, usually taking a value of 2. The weight range is usually between 0.1 and 1.0. The high-score interval corresponds to a low weight, and the low-score interval corresponds to a high weight. The purpose of this step is to quantify the influence degree of high-frequency pose jitter and provide a reasonable weight allocation scheme for pose model fitting.

[0071] The specific implementation of step S08 is to optimize the pose parameters using an iterative detection strategy. First, based on the pose model fitting weight determined in step S07, perform filtering on the original pose data using the weighted spline smoothing method: ; In the formula, is the filtered pose data; is the original pose data; is the weight; is the spline basis function; is the number of pose data points. The spline smoothing parameter is adaptively determined according to the pose data sampling frequency, usually 5 to 10 times the sampling interval. Then, use the optimized pose data to re-perform the processing from step S02 to step S07 to obtain a new pose model fitting weight. Next, calculate the error of the disparity between two adjacent iterations: ; In the formula, is the error of the disparity; and are respectively the row-column disparity residuals of the -th point after the -th iteration; is the number of corresponding points. When the error of the disparity is less than 0.3 pixels or the change in the error of the disparity between two adjacent iterations is less than 5%, the iteration termination condition is reached. If it does not converge after 10 iterations, force termination and output the result of the last iteration. The purpose of this step is to gradually improve the accuracy of the pose parameters through iteration and eliminate the influence of pose errors during the imaging of the area array image on the detection result.

[0072] The specific implementation of step S09 is to reconstruct the rigorous imaging geometric model of the linear array image using the optimized pose data and generate a geometrically seamless sensor correction image. First, combine the optimized pose data with the original orbit determination data to update the pose matrix in the collinearity equation and reconstruct the rigorous imaging geometric model of the linear array image. Then, construct the output image space and construct a rational function model, whose form is: ; ; In the formula, is the normalized image coordinate; is the ground three-dimensional coordinate; , , and are rational function coefficients; , and are the highest degrees of the polynomials, usually taking the value of 3. Then, for each pixel point on the output image, its corresponding position on the original image is determined by back-projection, and bilinear interpolation is used for gray-level resampling: ; In the formula, is the gray-level value of the output image; is the gray-level value of the original image; is the coordinate in the original image, is its integer part. Finally, a sensor-corrected image and its attached files are generated. The purpose of this step is to generate a sensor-corrected image product with higher geometric accuracy by using the optimized attitude parameters, providing a high-quality data basis for subsequent applications.

[0073] Through the above detailed steps, this Embodiment 2 realizes a complete technical solution for improving the geometric accuracy of the area array image assisting the linear array image on the same platform. This solution makes full use of the complementarity of the multi-payload data on the same platform, uses the instantaneous imaging characteristics of the area array to assist in detecting the high-frequency attitude jitter in the linear array image, determines the jitter interval by using a voting mechanism, and optimizes the attitude parameters by adaptive fitting, finally improving the geometric positioning accuracy of the linear array image.

[0074] To better optimize the above Embodiment 2, the following provides an Embodiment 3 for optimization based on Embodiment 2: The specific implementation manner of predicting the attitude jitter by using the physical model of satellite attitude jitter between step S04 and step S05 is to first establish an attitude jitter prediction equation based on the satellite dynamics principle, and this equation is specifically expressed as follows: ; In the formula, is the satellite attitude angle offset, with the unit of arcsecond; is the solar radiation pressure factor, dimensionless, with the value range of 0.01 - 0.05, obtained by monitoring the current change of the satellite solar panel, and the calculation formula is , where is the conversion coefficient, is the measured current of the solar panel, is the maximum current value; is the earth gravity gradient factor, with the unit of , at an orbital altitude of 500 km, the magnitude is approximately , calculated from the satellite's orbital position, and the calculation formula is , where is the gravitational constant of the Earth, is the distance from the satellite to the center of the Earth, is the satellite's latitude; is the geomagnetic disturbance factor, with the unit of nT, and the value range is 20 - 100 nT, obtained from the measurements of the on-board magnetometer; is the thermo-elastic deformation factor, with the unit of °C, obtained from the measurements of the temperature sensor array; is the mechanical vibration factor, with the unit of arcseconds, and the value range is 0.5 - 2 arcseconds, obtained from the high-frequency measurement data of the gyroscope; , , , , are the corresponding coefficients, obtained through calibration with historical data. Then, model calibration is carried out. Select the historical data of at least 10 typical orbital segments to construct the calibration equation: ; In the formula, is the measured attitude angle offset of the th historical data point; is the total number of historical data points. Solve this least squares problem to obtain the optimal coefficient values. Then, substitute the factor data obtained in real-time into the calibrated model to calculate the predicted attitude angle offset: ; Then, convert the prediction result into the geometric offset in the pixel domain: ; ; In the formula, and are the geometric offsets in the row and column directions respectively, with the unit of pixels; and are the conversion coefficients, related to the satellite altitude and pixel size. Finally, introduce the prediction result as prior information into the affine transformation model correction in step S05. The specific method is: ; ; In the formula, is the corrected linear array image coordinate; are the affine model parameters, estimated by the least squares method, but the weight design considers the reliability of the physical model. The fusion of the physical model prediction and the image processing result adopts the Bayesian framework, and the fusion formula is: ; In the formula, is the attitude angle offset after fusion; and are the weights of the physical model and image processing respectively, satisfying , and the weight values are adaptively adjusted according to their respective uncertainties. The purpose of this optimization step is to introduce physical constraints, reduce the uncertainty of attitude error estimation, and improve the reliability and continuity of attitude estimation, especially in areas where the matching of homologous points is insufficient.

[0075] The specific implementation of predicting and compensating high-frequency attitude jitter using a self-attention residual network model between step S05 and step S06 is to first construct a deep neural network that combines a multi-layer self-attention mechanism and residual connections. The network structure is specifically expressed as: ; ; ; ; ; ; ; In the formula, is the output feature of the th layer; , , are the query matrix, key matrix, and value matrix respectively; , , are the corresponding weight matrices; is the multi-head attention mechanism; is the number of attention heads, set to 8, which matches the satellite orbit period of about 90 minutes; , , are the parameter matrices of the th attention head; is the feature dimension; is the normalization function. The network mainly includes a feature extraction layer, a time series encoding layer, a multi-channel output layer, and a regularization layer. Then, a training dataset is constructed. First, attitude measurement data, true attitude change data, camera internal and external parameters, and external disturbance factor data during the imaging process of line array images and area array images on historical orbit segments are collected to construct data samples: ; ; In the formula, is the input feature vector, including disparity data , orbital parameters , thermal parameters , magnetic parameters , gyro data and so on; is the output label, representing the three-axis attitude angle offset. The dataset is divided into a training set, a validation set, and a test set, with a ratio of 7:2:1. Then, a three-stage training strategy is adopted. First, unsupervised pre-training is carried out: ; In the formula, is the contrastive learning loss function; is the feature vector and is the similarity; is the temperature parameter, with a value of 0.07; is the indicator function. The pre-training adopts the adaptive moment estimation optimizer, with a learning rate of 0.001 and 100 training epochs. Then, supervised fine-tuning is carried out: ; In the formula, is the root mean square error loss function; is the true label; is the model prediction value; is the number of samples. The fine-tuning adopts the adaptive learning rate decay strategy, with an initial learning rate of 0.0001. When the validation loss does not decrease for 5 consecutive epochs, the learning rate is halved, and the total number of training epochs is 200. Finally, regularization based on physical constraints is introduced: ; ; In the formula, is the total loss function with regularization; is the regularization coefficient, with a value of 0.1; is the physical constraint loss, representing the difference between the second derivative of the predicted value and the predicted value of the dynamic equation; is the predicted value of the dynamic equation at time . After the training is completed, the up and down disparity information calculated in step S04 is used as the input feature, and the high-frequency attitude jitter is predicted through the trained network model: ; In the formula, is the predicted attitude angle offset; represents the self-attention residual network model. The prediction result is weighted and fused with the voting result in step S06: ; In the formula, is the attitude angle offset after fusion; is the weight coefficient, with a range between 0.3 and 0.7, dynamically adjusted according to the confidence levels of the two. When there are missing parallax data, the value increases; when the parallax data is complete and has good consistency, the value decreases. The fused result is used for subsequent attitude model fitting, providing a more accurate basis for high-frequency jitter identification in the voting strategy. The purpose of this optimization step is to use deep learning technology to mine the hidden patterns in the attitude data, handle non-linear relationships, and improve the accuracy and robustness of attitude jitter prediction, especially in the case of incomplete or low-quality parallax data.

[0076] To better understand and implement the present invention, Example 4 of a specific application scenario of the present invention is provided below: Researchers used a method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform to process simulated satellite images. The simulated satellite carried a multi-spectral linear array camera and a panchromatic planar array camera, with an orbital altitude of 778 km, a spatial resolution of 5 m for the multi-spectral linear array camera, and a spatial resolution of 2 m for the panchromatic planar array camera. A set of simulation data from a certain mountainous area in western China was selected for the experiment, and its basic parameters are shown in Table 1: Table 1 Basic Parameter Table of Experimental Data

[0077] First, a rigorous imaging geometric model of the linear array image and the planar array image was constructed. Based on the satellite orbit determination data, the orbital coordinates were obtained, the attitude data sampling interval was 0.1 s, the position accuracy was better than 0.5 m, and the attitude measurement accuracy was 0.001°. Using the orbit position data and attitude data transmitted by the satellite, rigorous imaging geometric models of the linear array image and the planar array image were respectively constructed. The elevation data of the study area was extracted from the SRTM digital elevation model, with a resolution of 30 m and an elevation accuracy of approximately ±16 m, and was input as external elevation data.

[0078] Then, based on the constructed rigorous imaging geometric model and elevation data, the researchers realized the simulation of the linear array image to the planar array image. During the simulation process, for each pixel (x, y) of the planar array image, its ground three-dimensional coordinates (X, Y, Z) were calculated, and then the corresponding image point coordinates (x’, y’) of the linear array image were inversely calculated through the rigorous geometric model of the linear array image, and the gray value was obtained by using the bicubic convolution interpolation method. After the entire simulation process, a simulated planar array image with the same resolution as the real planar array image and a similar imaging perspective was generated, with an actual size of 6000×6000 pixels.

[0079] Next, the multi-scale SIFT matching algorithm was used to perform dense matching on the real area array image and the simulated area array image. During the matching process, the number of Gaussian pyramid layers for SIFT feature point extraction was set to 4, the descriptor dimension was 128, the nearest neighbor ratio threshold was set to 0.75, and the NCC similarity threshold was set to 0.85. After removing the mismatched points using the RANSAC algorithm, with the number of iterations set to 2000 and the inlier threshold set to 2.5 pixels, a total of 32,768 uniformly distributed corresponding points were finally obtained, with a matching accuracy better than 0.5 pixels and the point distribution interval being approximately 30 pixels. The matching results are statistically shown in Table 2: Table 2 Statistical Table of Image Matching Results

[0080] Based on the conversion relationship between the simulated area array image and the linear array image, the corresponding positions of the corresponding points on the real area array image on the linear array image were determined. Based on the strict imaging geometric model of the linear array image and the area array image, the vertical and horizontal disparities of 32,768 dense corresponding image points were calculated. To eliminate the influence of systematic errors, an affine transformation model was used for preliminary correction: .

[0081] The residual error after correction mainly reflects the non-linear distortion caused by high-frequency jitter. By calculating the difference in row and column disparities between adjacent corresponding points, it was found that there were obvious disparity mutations in some areas. The disparity mutation threshold was set to 0.5 pixels, and the linear array image scanning line was used as the voting interval (one interval every 25 lines) for high-frequency jitter detection. The detection results showed that there were obvious high-frequency jitters in the intervals of 2560 - 2816 lines, 5886 - 6016 lines, and 8192 - 8320 lines of the linear array image, corresponding to the attitude times of 03:45:20.48 - 03:45:21.05, 03:45:29.63 - 03:45:30.13, and 03:45:35.36 - 03:45:35.84 respectively. The voting statistical results of these intervals are shown in Table 3: Table 3 Statistical Table of Voting in High-Frequency Jitter Intervals

[0082] The attitude fitting weights were calculated based on the voting scores, and the weight coefficients were as follows: . The weights of each high-frequency jitter interval were 0.24, 0.39, and 0.47 respectively. Then, the weighted least squares method was used to fit the attitude data to reduce the weight of the high-frequency jitter area. At the same time, the pre-trained attitude optimization attention neural network model was used to analyze the relationship between disparity changes and high-frequency jitter. This model used 8 attention heads, the attention dimension was 64, and the total number of layers was 12 layers to further optimize the attitude parameters.

[0083] Finally, according to the revised satellite attitude data, combined with precise orbit determination data and camera internal parameters, a rigorous imaging geometric model of the linear array image is reconstructed, and the original images are mosaicked to generate a geometrically seamless sensor correction image product. The comparison of geometric accuracy before and after correction is shown in Table 4: Table 4 Comparison Table of Geometric Accuracy Improvement Effect

[0084] Traditional means to solve the problem of high-frequency satellite attitude jitter mainly include low-pass filtering method, spline fitting method and polynomial fitting method. These methods usually smooth the attitude data as a whole, cannot distinguish the intensity differences of jitter in different time periods, and are prone to over-smoothing or under-smoothing, resulting in unstable geometric accuracy. The method of the present invention, assisted by the same-platform area array image, utilizes the characteristics of instantaneous imaging of the area array to accurately identify the high-frequency attitude jitter interval in the linear array image, and performs targeted processing with adaptive weights, avoiding the blindness of traditional methods. Experimental results show that compared with traditional methods, the geometric accuracy of the method of the present invention is improved by 61.6% - 75.0%. Especially in areas with large high-frequency attitude jitter, the accuracy improvement is more significant, providing a reliable guarantee for subsequent high-precision stereo mapping and three-dimensional reconstruction.

[0085] It should be noted that the detailed explanations of the variables involved in the present invention are shown in Tables 5, 6, and 7 as follows.

[0086] Table 5 Variable Explanation Table (First Part)

[0087] Table 6 Variable Explanation Table (Second Part)

[0088] Table 7 Variable Explanation Table (Third Part)

[0089] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention.

Claims

1. A method for improving the geometric accuracy of a same-platform planar array image and an auxiliary linear array image, characterized in that, Including: Construct a rigorous imaging geometric model for linear array images and area array images; generate a simulated area array image through external elevation data; obtain corresponding points of the real area array image and the simulated area array image, and establish the corresponding relationship between the linear array and area array images; calculate the vertical parallax and eliminate systematic interference; Separate high-frequency attitude jitter based on the affine transformation model; sort the corresponding points to form a two-dimensional grid, and vote based on the sudden change of parallax; count the voting results and calculate the attitude jitter score; Adopt an iterative strategy to optimize the attitude parameters; use the optimized attitude parameters to reconstruct the rigorous geometric model and generate a geometric seamless sensor correction image.

2. The method for improving the geometric accuracy of a same-platform planar array image and an auxiliary linear array image according to claim 1, wherein The rigorous imaging geometric model is a satellite-ground corresponding ray mapping model based on the collinearity equation. The rigorous imaging geometric models of linear array images and area array images are similar in form, and both are based on the collinearity equation to realize the conversion from image point coordinates to ground three-dimensional coordinates.

3. The method for improving the geometric accuracy of the same-platform area array image and the linear array image according to claim 2, wherein The image point coordinates of the linear array image are represented by a one-dimensional polynomial, while the image point coordinates of the area array image are represented by a two-dimensional polynomial. The linear array image is based on the row-by-row scanning principle, and the area array image is based on the area array instantaneous imaging principle.

4. The method for improving the geometric accuracy of a same-platform area array image and a linear array image according to claim 3, wherein The steps to generate a simulated area array image through external elevation data are specifically to substitute the three-dimensional coordinates of each pixel position of the area array image on the ground into the rigorous geometric model of the linear array image for inverse calculation to obtain the corresponding image point coordinates, and finally perform gray interpolation to obtain the simulated area array image.

5. The method for improving the geometric accuracy of a co-platform area array image and a linear array image according to claim 4, wherein The steps to perform high-precision matching on the real area array image and the simulated area array image adopt a strategy combining feature point matching and region growing. Extract feature points through the scale-invariant feature transform algorithm, perform feature point matching using the nearest neighbor ratio method, and eliminate mismatched points using the random sample consensus algorithm.

6. The method for improving the geometric accuracy of the same-platform planar array image assisted by the linear array image according to claim 5, wherein The steps to calculate the vertical parallax and eliminate systematic interference include calculating the vertical parallax of the corresponding image points of the linear array image and the area array image, and eliminating the interference of projection differences caused by terrain undulation, imaging perspective differences, and different imaging scales on the detection of high-frequency attitude jitter.

7. The method for improving the geometric accuracy of a co-platform area array image and an auxiliary linear array image according to claim 6, wherein Use the affine transformation model to correct the systematic linear distortion between the linear array image and the area array image. The error of the affine transformation model is the high-order distortion caused by high-frequency attitude jitter, and the parameters of the affine transformation model are estimated by the weighted least squares method.

8. The method for improving the geometric accuracy of a same-platform area array image and a linear array image according to claim 7, wherein The steps to vote based on the sudden change of parallax are to sort the corresponding points in ascending order according to the row and column coordinates of the linear array image to form a two-dimensional grid, determine the voting parameter space according to the row coordinate range of the two-dimensional grid, and vote based on the difference between the row and column parallaxes of adjacent corresponding points; the steps to count the voting results and calculate the attitude jitter score include counting the voting ratio and the maximum value of the parallax difference in each interval, calculating the voting score. The higher the voting score, the greater the high-frequency attitude jitter in the interval, and the smaller the fitting weight value of the corresponding attitude model.

9. The method for improving the geometric accuracy of a same-platform planar array image assisted by a linear array image according to claim 8, wherein After calculating the vertical parallax, it also includes the step of predicting attitude jitter using the physical model of satellite attitude jitter. The physical model of satellite attitude jitter is a satellite attitude prediction equation established based on the principles of satellite dynamics, which is used to evaluate the high-frequency attitude jitter generated by the satellite during on-orbit operation due to external factor interference, and provides a reference basis for the correction of the affine transformation model.

10. The method for improving the geometric accuracy of a same-platform planar array image and an auxiliary linear array image according to claim 9, wherein After separating the high-frequency attitude jitter based on the affine transformation model, it further includes the step of predicting and compensating the high-frequency attitude jitter by using a self-attention residual network model. The self-attention residual network model uses the up and down parallax information as input features and outputs optimized attitude parameters. The structure of the self-attention residual network model is a deep neural network combining multi-layer self-attention mechanisms and residual connections, including a feature extraction layer with the number of attention heads matching the satellite orbital period, a time series encoding layer corresponding to the attitude measurement frequency, a multi-channel output layer corresponding to the satellite's three-axis attitude components, and a regularization layer based on attitude physical constraints.

Images