A method for improving the geometric accuracy of linear array images assisted by area array images on the same platform

Abstract

The present invention provides a method for improving the geometric accuracy of linear array images assisted by same-platform area array images, belonging to the technical field of electronic digital data processing. The present invention first constructs a rigorous imaging geometric model of linear array and area array images based on attitude trajectory data and calibration parameters; introduces external elevation data and substitutes the ground coordinates of area array image pixels into the rigorous imaging geometric model of the linear array image to inversely calculate and obtain a simulated area array image; matches the real and simulated area array images to obtain dense same-name points and determines the corresponding positions on the linear array image; uses an affine transformation model to correct systematic distortion, and the residual error is the high-order distortion caused by high-frequency attitude jitter; forms a two-dimensional grid by sorting the coordinates of the same-name points, and calculates scores by voting based on the difference in parallax; uses an iterative detection strategy to optimize attitude parameters to eliminate the influence of the attitude error of the area array image; and finally uses the optimized attitude data to reconstruct the geometric model of the linear array image to generate a geometrically seamless sensor-corrected image, thereby achieving uncontrolled elevation accuracy improvement of the stereo image.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing, and in particular relates to a method for improving the geometric accuracy of a linear array image assisted by a planar array image on the same platform. Background Art

[0002] Optical satellite remote sensing technology has widespread applications in geographic information systems, resource exploration, urban planning, disaster monitoring, and other fields. Traditional satellite stereo mapping primarily constructs stereo models by acquiring forward, rearward, or side-view images, thereby enabling elevation extraction and three-dimensional reconstruction. This has become an important means of acquiring large-scale terrain information. Currently, high-resolution optical satellites mostly use linear array push-broom technology to acquire panchromatic imagery, while also acquiring multispectral imagery to provide information on different colors.

[0003] However, during in-orbit satellite operations, they are inevitably subject to interference from external factors such as solar radiation pressure, Earth's gravitational gradient, thermoelastic deformation, and mechanical vibration, resulting in high-frequency attitude jitter. This high-frequency attitude jitter is characterized by short periods, strong randomness, and a frequency far exceeding the sampling frequency of satellite attitude measurement equipment, making it difficult to accurately capture using conventional attitude measurement equipment. Because line array cameras use a line-by-line push-scan imaging method, high-frequency attitude jitter can cause discontinuities between adjacent lines in the image, seriously affecting the geometric accuracy of the 3D model.

[0004] Traditional satellite stereo image processing methods rely primarily on attitude measurement data or ground control points for geometric correction. These methods struggle to effectively compensate for high-frequency attitude jitter in uncontrolled conditions, resulting in reduced elevation accuracy in stereo mapping, poor terrain reconstruction quality, and increased errors in 3D information extraction. This severely restricts the further application of optical satellite imagery in high-precision stereo mapping. In other words, existing technologies suffer from the technical problem of low elevation accuracy in uncontrolled optical satellite stereo images due to high-frequency attitude jitter. Summary of the Invention

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

[0006] The present invention is implemented as follows: The present invention provides a method for improving the geometric accuracy of line array images assisted by same-platform area array images, including: constructing a rigorous imaging geometric model of line 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 a corresponding relationship between the line 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 the rigorous geometric model using the optimized posture parameters to generate a geometrically seamless sensor correction image.

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

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

[0009] Among them, 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 rigorous geometric model of the line array image for inverse calculation to 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 between real area array images and simulated area array images adopts a strategy that combines feature point matching and region growing. Feature points are extracted through the scale-invariant feature transformation algorithm, feature point matching is performed 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 in the line array image and the area array image, and eliminating the interference of projection differences caused by terrain undulations, imaging perspective 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 area 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 weighted least squares method.

[0013] Among them, the step of voting based on disparity mutation is to sort the homonymous 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 in disparity between adjacent homonymous points in the rows and columns; the step of calculating the posture jitter score by counting the voting results includes counting the maximum value of the difference between the voting ratio and the disparity in each interval, and calculating the voting score. The higher the voting score, the greater the high-frequency posture jitter in the interval, and the smaller the corresponding posture model fitting weight.

[0014] Among them, after calculating the upper and lower parallax, it also includes a step of using the satellite attitude jitter physical model to predict attitude jitter. The satellite attitude jitter physical model is a satellite attitude prediction equation established based on the principle of satellite dynamics, which is used to evaluate the high-frequency attitude jitter caused by external factors while the satellite is in orbit, and provide a reference basis for the correction of the affine transformation model.

[0015] Among them, after separating the high-frequency posture jitter based on the affine transformation model, it also includes the step of using the self-attention residual network model to predict and compensate the high-frequency posture jitter. The self-attention residual network model uses the upper and lower disparity information as input features and outputs the optimized posture parameters; the structure of the self-attention residual network model is a deep neural network that combines a multi-layer self-attention mechanism with a residual connection, including a feature extraction layer whose number of attention heads matches the satellite orbit period, a time series encoding layer corresponding to the posture measurement frequency, a multi-channel output layer corresponding to the satellite's three-axis posture components, and a regularization layer based on posture physical constraints.

[0016] The present invention achieves accurate detection and compensation of high-frequency posture jitter by constructing a rigorous imaging geometry model, generating simulated array images, obtaining dense homonymous points through high-precision matching, calculating upper and lower parallaxes and eliminating interference factors, correcting systematic distortion using an affine transformation model, voting based on the difference in homonymous point parallaxes, and optimizing posture parameters through 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 line array and area array images. Since the area array image adopts instantaneous imaging, it is not affected by high-frequency attitude jitter, while the line array image is sensitive to high-frequency attitude jitter. The complementary characteristics of the two imaging methods establish a quantitative relationship between jitter and image distortion. The present invention innovatively eliminates the interference of projection differences caused by terrain undulations, imaging viewing angle differences, and different imaging scales, ensuring that parallax changes truly reflect the characteristics of high-frequency attitude jitter, and improves the accuracy and reliability of attitude parameters through voting strategies 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 1is a flow chart of the method of the present invention.

[0019] Figure 2 This is a schematic diagram of the principle of optimizing the posture of a linear array using area array imaging assistance in Example 1.

[0020] Figure 3 This is a flow chart of the area array image-assisted linear array posture optimization technology in Example 1.

[0021] Figure 4 Schematic diagram of high-frequency attitude jitter detection by parallax of the same-name image points in Example 1.

[0022] Figure 5 This is a schematic diagram of voting for high-frequency posture jitter in Example 1. DETAILED DESCRIPTION

[0023] In order to make the purpose, 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] like Figure 1 FIG. 1 is a flow chart of a method for improving the geometric accuracy of a linear array image assisted by a planar array image on the same platform provided by the present invention. The method includes the following steps:

[0025] S01. Constructing a rigorous imaging geometric model of a linear array image and a planar array image, wherein both the rigorous imaging geometric model of the linear array image and the rigorous imaging geometric model of the planar array image realize satellite-ground homonymous ray mapping based on collinear equations;

[0026] S02, introducing external elevation data, substituting the three-dimensional coordinates of each pixel position of the area array image on the ground into the rigorous geometric model of the line array image to perform inverse calculation to obtain the corresponding image point coordinates, and finally performing grayscale interpolation to obtain a simulated area array image;

[0027] S03, performing high-precision matching between the real area array image and the simulated area array image to obtain dense homonymous points, and obtaining corresponding positions of the homonymous points on the real area array image on the linear array image based on the conversion relationship between the simulated area array image and the linear array image;

[0028] S04, calculating the vertical parallax of the same-name image points in the linear array image and the planar array image, and eliminating interference of projection differences caused by terrain undulations, imaging viewing angle differences, and imaging scale differences on high-frequency attitude jitter detection;

[0029] S05. Correcting the systematic linear distortion between the linear array image and the area array image using an affine transformation model, wherein the affine transformation model error is the high-order distortion caused by the high-frequency posture jitter;

[0030] S06, sorting the row and column coordinates of the homonymous points in the linear array image in ascending order to form a two-dimensional grid, then determining a voting parameter space based on the row coordinate range of the two-dimensional grid, and performing voting based on the difference in parallax between adjacent rows and columns of the homonymous points;

[0031] S07, counting the maximum value of the difference between the voting ratio and the disparity in each interval, and calculating the voting score. The higher the voting score, the greater the high-frequency posture jitter in the interval, and the smaller the corresponding posture model fitting weight;

[0032] S08, taking the parallax error below a set threshold as the iteration termination condition, optimizing the posture parameters using an iterative detection strategy to eliminate the influence of the posture error during the imaging of the array image on the overall detection result;

[0033] S09. Reconstruct a rigorous imaging geometry model of the linear array image using the optimized attitude data in combination with satellite orbit determination and other auxiliary data, and generate a geometrically seamless sensor correction image and auxiliary files.

[0034] Optionally, at the end of step S01, the process further includes determining the satellite's attitude matrix and external calibration parameter matrix respectively; wherein the rigorous imaging geometry model specifically refers to a satellite-ground homonymous ray mapping model based on collinear equations, which is expressed as , is the image point coordinate, is the imaging time, is the proportionality coefficient, is the three-dimensional coordinate of the image point corresponding to the ground Earth-centered Earth-fixed coordinate system, is the position of the satellite's center, is the satellite attitude matrix, is the external calibration parameter matrix, is the coordinate of the image point in the satellite coordinate system.

[0035] Wherein, the linear array image Specifically refers to the one-dimensional polynomial , , and the area array image Specifically refers to the two-dimensional polynomial .

[0036] The affine transformation model specifically refers to , is the linear array image coordinate, is the coordinate of the array image, are the affine transformation model coefficients.

[0037] The voting score specifically refers to , represents the voting score, and is the maximum value of the difference between the voting ratio and the disparity, is the corresponding coefficient.

[0038] Among them, the high-frequency attitude jitter specifically refers to the short-period random fluctuations beyond the normal attitude change range caused by external interference during the satellite's in-orbit operation. The high-frequency attitude jitter will cause the upper and lower parallax of the stereoscopic image to have a sudden change that is inconsistent with the overall change trend.

[0039] The parallax error specifically refers to the root mean square difference between the actual parallax of the same-name points obtained by matching and the theoretical parallax. The parallax error is used to evaluate the accuracy of the posture parameters and determine whether the iteration is terminated.

[0040] The sensor-corrected image specifically refers to an image product that eliminates image misalignment error, internal distortion error, and attitude and trajectory error. The sensor-corrected image has geometric seamless characteristics, which facilitates subsequent precise measurement and information extraction.

[0041] Optionally, between step S04 and step S05, a satellite attitude jitter physical model is used to predict attitude jitter. The satellite attitude jitter physical model is used to evaluate high-frequency attitude jitter caused by external interference while the satellite is in orbit, providing a reference for subsequent correction of the affine transformation model.

[0042] Optionally, a self-attention residual network model is used between step S05 and step S06 to predict and compensate for the high-frequency posture jitter. The self-attention residual network model uses the upper and lower disparity information calculated in step S04 as input features and outputs optimized posture parameters to provide a more accurate basis for identifying the high-frequency posture jitter for subsequent voting strategies.

[0043] The satellite attitude jitter physical model specifically refers to the satellite attitude prediction equation established based on the principle of satellite dynamics. The satellite attitude jitter physical model is used to predict the high-frequency attitude jitter caused by external factors during the satellite's orbit. The input of the satellite attitude jitter physical model includes the solar radiation pressure factor , Earth's gravitational gradient factor , geomagnetic field disturbance factor , thermoelastic deformation factor and mechanical vibration factor The solar radiation pressure factor is obtained from the current change of the satellite solar panel, the earth gravity gradient factor is calculated from the satellite orbit position, the geomagnetic field disturbance factor is obtained from the onboard magnetometer measurement, the thermoelastic deformation factor is obtained from the temperature sensor array measurement, the mechanical vibration factor is obtained from the gyroscope high-frequency measurement data, and the output of the satellite attitude jitter physical model 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 satellite attitude jitter physical model is combined with the adaptive attitude weighted fitting to further improve the accuracy and reliability of attitude optimization.

[0044] The specific structure of the self-attention residual network model is a deep neural network that combines a multi-layer self-attention mechanism with residual connections, 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's three-axis attitude components, and a regularization layer based on attitude physical constraints. The step of establishing a training data set for the self-attention residual network model specifically includes collecting attitude measurement data, true attitude change data, camera internal and external parameters, and external disturbance factor data during the imaging process of the linear array image and the planar array image on the historical orbit segment, and stratifying and sampling the attitude measurement data, the true attitude change data, the camera internal and external parameters, and the external disturbance factor data according to different orbits and imaging conditions to construct training samples covering different working conditions. The step of training the self-attention residual network model specifically includes first pre-training the self-attention layer parameters using an unsupervised learning method to capture the attitude temporal characteristics, then fine-tuning the entire network parameters using a supervised learning method to minimize the error between the predicted attitude and the true attitude, and finally introducing a regularization term based on physical constraints to ensure that the model prediction results conform to the satellite dynamic characteristics.

[0045] The specific implementation of the above steps is described in detail below.

[0046] The specific implementation of step S01 is to first obtain the original orbital parameters, attitude parameters, and camera parameters transmitted from the satellite. Then, based on the principle of collinearity equations, rigorous imaging geometry models are established for both linear and planar array images. For linear array images, internal orientation elements, including focal length, principal point coordinates, and distortion coefficients, are constructed according to the line-by-line scanning principle. A correspondence is established between the line number and the imaging time, thereby achieving a rigorous one-dimensional projection model. For planar array images, a two-dimensional projection model is constructed according to the principle of instantaneous planar array imaging, establishing the projection relationship between image plane coordinates and ground target points. Exterior orientation elements, including the satellite position vector and attitude angle, are then determined based on satellite orbit determination data and attitude measurement data. Finally, the internal and external orientation elements are combined to establish a complete and rigorous imaging geometry model, achieving a mapping relationship from image plane coordinates to three-dimensional ground coordinates. The purpose of this step is to establish a rigorous geometric foundation for subsequent processing and ensure the accuracy of coordinate transformation.

[0047] The specific implementation of step S02 is to first obtain external elevation reference data corresponding to the satellite coverage area. A global digital elevation model such as SRTM or ASTER GDEM can be used, and its accuracy should be better than 30m. Then, the ground point corresponding to each pixel of the array image is forward projected through a rigorous geometric model to calculate its ground three-dimensional coordinates. When determining the ground three-dimensional coordinates, an iterative solution method is required to project the ground point onto the elevation model. Usually, the iterative convergence condition is set to an elevation change of less than 0.1m. The obtained ground three-dimensional coordinates are then substituted into the rigorous geometric model of the linear array image for inverse calculation to obtain the corresponding linear array image coordinates. Finally, the bicubic convolution interpolation algorithm is used to calculate the grayscale value on the linear array image, and the interpolated grayscale value is assigned to the corresponding pixel of the array image to generate a simulated array image. The purpose of this step is to eliminate the geometric differences caused by different imaging methods and provide a geometric consistency basis for subsequent image matching.

[0048] The specific implementation of step S03 employs a strategy combining feature point matching and dense matching to obtain the same-name points between the real and simulated area array images. First, a scale-invariant feature transformation algorithm is used to extract feature points from the two images. Initial feature point matching is then performed using the nearest neighbor ratio method, with a matching threshold set to 0.6 to balance matching accuracy and the number of matches. A random sampling consistency algorithm is then used to remove mismatched points, with an inlier threshold set to 0.5 pixels. Region growing is then performed using the filtered feature points as seeds, with a gradient correlation coefficient algorithm set to a matching correlation coefficient threshold of 0.8 to generate densely matched point pairs. Finally, using the transformation relationship between the simulated area array image and the line array image, the coordinates of the same-name points on the real area array image are converted to those on the line array image, thereby establishing a precise correspondence between the real and line array images. To ensure a uniform distribution of matching points, a grid constraint strategy is employed, dividing the images into a uniform 10×10 grid, with no fewer than 10 matching points extracted within each grid. The purpose of this step is to establish a precise correspondence between the area array image and the line array image, providing basic data for subsequent posture error detection.

[0049] The specific implementation of step S04 is to first use the same-name points obtained in step S03 to calculate the vertical disparity between the line array image and the area array image. The vertical disparity is calculated by subtracting the corresponding coordinates of the area array image on the line array image from the coordinates of the line array image. A terrain effect elimination model is then constructed. This model, based on the principle of common view geometry, uses elevation data and imaging geometry to calculate the disparity components caused by terrain undulation. Next, the disparity component caused by the difference in imaging perspective is calculated, which is related to the difference in imaging center angles between the two images. The disparity component caused by the difference in imaging scale is then calculated, which is related to the ratio of the ground resolutions of the two images. Finally, the three systematic components mentioned above are subtracted from the total disparity to obtain the residual disparity, which primarily reflects the geometric perturbations caused by high-frequency attitude jitter. To improve accuracy, a robust estimation method is used to handle 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 systematic factors that affect the detection of high-frequency attitude jitter and extract geometric perturbation information purely caused by attitude jitter.

[0050] The specific implementation of step S05 involves using an affine transformation model to correct the systematic linear distortion between the linear and planar array images. First, based on the residual disparity obtained in step S04, the affine transformation model parameters are estimated using weighted least squares, with the weights proportional to the reliability of the disparity. The predicted coordinates of each homonymous point under the affine transformation model are then calculated and compared with the actual coordinates to calculate the disparity residual. The spatial distribution and frequency characteristics of the disparity residual are then analyzed, and the disparity residual is converted to the frequency domain using a fast Fourier transform to identify its high-frequency components. Finally, the high-order distortion components in the disparity residual, namely the nonlinear distortion components caused by high-frequency posture jitter, are determined. For affine transformation model parameter estimation, singular value decomposition is used to enhance solution stability when the number of points is insufficient. When redundancy is high, a random sampling strategy is used to reduce the influence of outliers. The purpose of this step is to separate linear systematic errors from nonlinear high-frequency jitter errors, providing a foundation for subsequent accurate high-frequency jitter identification.

[0051] The specific implementation of step S06 is to first sort the homonymous points obtained in step S05 in ascending order according to their row and column coordinates in the line array image. The row coordinates correspond to the time dimension, and the column coordinates correspond to the spatial dimension. A two-dimensional grid is then constructed, with the grid size set to 10 times the row resolution of the line array image to balance accuracy and computational efficiency. To address pointless areas caused by matching failures, the grid tolerance is set to 20% of the original grid size. Next, a voting parameter space is determined based on the row coordinate range of the two-dimensional grid. The row coordinate range is evenly divided into several intervals, typically set to 1% of the total number of rows but no less than 100 intervals. The absolute value of the difference in row and column disparity between adjacent homonymous points is then calculated in each column direction. When the disparity difference exceeds a threshold of 0.5 pixels, it is accumulated in the voting parameter space for the corresponding row coordinate, and the maximum disparity difference for the corresponding column is recorded. Finally, voting statistics are performed across all columns to form a complete voting parameter space distribution map. The purpose of this step is to identify the time intervals during which high-frequency posture jitter occurs through the voting mechanism, thereby improving the reliability and robustness of jitter detection.

[0052] The specific implementation of step S07 is to first count the number of votes in each interval in step S06 and calculate the voting ratio, which is the ratio of the number of votes in each interval to the total number of votes. The maximum parallax difference in each interval is also counted, reflecting the severity of posture jitter. A comprehensive voting score is then calculated based on the voting ratio and the maximum parallax difference. The weighting coefficients in the calculation formula are typically set to 0.7 for the voting ratio and 0.3 for the maximum parallax difference to balance the frequency and intensity factors. The voting scores are then normalized to a range between 0 and 1 for ease of subsequent processing. Finally, the posture model fitting weights are determined based on the normalized voting scores. The scores are converted into weights using an exponential decay function, with high-scoring intervals corresponding to low weights and low-scoring intervals corresponding to high weights. The weight range is typically set between 0.1 and 1.0 to ensure that posture data in areas with high-frequency jitter are appropriately downweighted in subsequent processing. The purpose of this step is to quantify the impact of high-frequency posture jitter and provide a reasonable weight distribution scheme for posture model fitting.

[0053] The specific implementation method of step S08 is to optimize the posture parameters using an iterative detection strategy. First, based on the posture model fitting weights determined in step S07, the original posture data is filtered using a weighted spline smoothing method. The smoothing parameters are adaptively determined according to the sampling frequency of the posture data and are usually set to 5 to 10 times the sampling interval. Then, the optimized posture data is used to re-process steps S02 to S07 to obtain new posture model fitting weights. Then, the parallax error after two adjacent iterative processes is calculated, that is, the root mean square difference between the actual parallax of the matching point and the theoretical parallax. When the parallax error is less than 0.3 pixels or the change in the parallax error between two adjacent iterations is less than 5%, the iteration termination condition is met. Finally, the optimal posture 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 last iteration result is output. The purpose of this step is to gradually improve the accuracy of the posture parameters through iteration, while eliminating the interference of the posture error during the imaging of the array image itself on the detection results.

[0054] 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 a rigorous imaging geometry model for the linear array image. Then, based on the optimized geometric model, the precise ground coordinates of each pixel in the original segmented linear array image are calculated. The output image space is then constructed, typically using equal-interval projection to establish a mapping relationship between ground coordinates and output image coordinates. The original image is then resampled to the output image space via backprojection, with bilinear interpolation used for grayscale resampling. To eliminate seam artifacts, a weighted fusion method is employed in overlapping areas, with weights proportional to the distance from the pixel to the seam. Finally, a geometrically seamless sensor-calibrated image is generated, and the rational function model coefficients are output as an accompanying file. The coefficients are typically third-order, and the control point residuals are required to be better than 0.5 pixels. The purpose of this step is to use the optimized attitude parameters to generate geometrically accurate calibrated image products, providing a high-quality data foundation for subsequent applications.

[0055] Another specific embodiment of the present invention is provided below. In the specific embodiment below, the difference from the above specific embodiment is that: the specific embodiment of using the satellite attitude jitter physical model to predict attitude jitter 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 field disturbance factor, the thermoelastic deformation factor and the mechanical vibration factor as input parameters to predict the attitude angle offset of the satellite in the orbital position. The solar radiation pressure factor is obtained by monitoring the change in the current of the satellite solar panel, and the typical range of change is 0.01 to 0.05; the earth's gravity gradient factor is calculated based on the satellite's orbital position, and is usually on the order of about 0.01 at an orbital altitude of 500km. The geomagnetic field perturbation factor is obtained from onboard magnetometer measurements, with a typical range of 20 to 100 nT. The thermoelastic deformation factor is obtained from temperature sensor array measurements, with each 1°C change in satellite structure temperature causing an attitude change of approximately 0.1 to 0.5 arc seconds. The mechanical vibration factor is obtained from high-frequency gyroscope measurements, with a typical amplitude of 0.5 to 2 arc seconds. The attitude jitter model is then calibrated by adjusting the weight coefficients of each factor. This calibration process uses historical orbit data and selects at least 10 representative orbital segments, including data from different seasons and different sunlight conditions. The real-time factor data is then substituted into the calibrated model to calculate the predicted attitude angle offset. Finally, the predicted result is converted into a pixel-domain geometric offset, which serves as the prior information for affine transformation model correction in step S05. The physical model prediction result is typically used as the initial value and fused with the subsequent image processing results. This fusion adopts a Bayesian framework, with the physical model providing the prior distribution and the image processing results providing the observed values, which are then combined to obtain the optimal estimate. The model calculation uses a recursive approach with a time step that matches the attitude measurement frequency, typically 10 to 100 Hz. The purpose of this optimization step is to introduce physical constraints, reduce the uncertainty of the attitude error estimate, and improve the reliability and continuity of the attitude estimate.

[0056] The specific implementation of the self-attention residual network model used between steps S05 and S06 to predict and compensate for high-frequency attitude jitter involves first constructing a deep neural network combining a multi-layer self-attention mechanism with residual connections. This network comprises 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 eight attention heads, corresponding to the satellite's orbital period of approximately 90 minutes; the time series encoding layer has a sampling rate consistent with the attitude measurement frequency of 10-100 Hz; the multi-channel output layer corresponds to the satellite's three-axis attitude components; and the regularization layer establishes physical constraints based on attitude dynamics equations. A training dataset is then constructed, collecting attitude measurement data, actual attitude change data, camera internal and external parameters, and external disturbance factors from the imaging process of linear and planar array images over historical orbital segments. Data collection covers at least 100 orbital periods and encompasses a variety of operating conditions. A three-stage training strategy is then employed: First, the self-attention layer parameters are pre-trained using unsupervised learning, using a contrastive learning loss function, a learning rate of 0.001, and 100 training epochs. Next, the overall network parameters are fine-tuned using supervised learning, using the root mean square error loss function, a learning rate of 0.0001, and 200 training epochs. Finally, a physical constraint-based regularization term is introduced with a regularization coefficient of 0.1 and 50 training epochs. Training is performed in a batch process 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 input features, and the trained network model is used to predict high-frequency pose jitter. This prediction serves as an auxiliary basis for the voting strategy in step S06, improving the accuracy of jitter detection. The model predictions are then combined with the voting results in step S06 using a weighted average, with weights determined based on the confidence level of each. The model prediction weights typically range from 0.3 to 0.7. When disparity data is missing, the model prediction weight is increased; when disparity data is complete and consistent, the voting weight is increased. The purpose of this optimization step is to use deep learning technology to mine implicit patterns in pose data and improve the accuracy and robustness of pose jitter prediction.

[0057] The introduction of a physical model for satellite attitude jitter provides physical-based prior knowledge for high-frequency attitude jitter detection and compensation, significantly improving the accuracy and reliability of affine transformation model correction. Compared with traditional methods that rely solely on image matching, this physical model comprehensively considers the impact of multiple external interference factors on satellite attitude, including solar radiation pressure, Earth's gravitational gradient, geomagnetic field perturbations, thermoelastic deformation, and mechanical vibration, establishing a quantitative relationship between these factors and high-frequency attitude jitter. By acquiring relevant parameters from various satellite sensors in real time, it predicts the potential attitude fluctuation trend and amplitude, providing a reference for subsequent affine transformation model correction. This effectively avoids the parameter divergence or local optimum problems caused by the lack of physical constraints in traditional methods. Especially for orbit segments with complex attitude variations and diverse interference factors, this physical model can accurately distinguish between true attitude jitter and noise, improving the robustness of jitter feature extraction and laying the foundation for the construction of high-precision stereo models.

[0058] The application of a self-attention residual network model enables intelligent prediction and compensation of high-frequency attitude jitter, addressing the nonlinear and time-varying attitude jitter that traditional methods struggle to address. This network model overcomes the expressive limitations of traditional mathematical models. Through a multi-layer self-attention mechanism, it captures long-term and short-term dependencies in attitude time series data. Residual connections effectively avoid the vanishing gradient problem in deep network training. Compared to existing technologies, this model is able to adaptively learn attitude jitter characteristics under different orbital conditions and interference patterns, accurately modeling complex attitude variations. In particular, the design of a feature extraction layer with the number of attention heads matching the satellite's orbital period accurately captures periodic attitude variations associated with the orbital period. The regularization layer, based on attitude physical constraints, ensures that the model's predictions are consistent with the satellite's dynamic characteristics, avoiding the risk of overfitting. By deeply mining vertical disparity information, the model can identify subtle attitude fluctuations that are difficult to detect using conventional methods, providing key technical support for improving uncontrolled elevation accuracy.

[0059] Specifically, the present invention is based on the concept of collaborative processing of linear and area array images, leveraging the difference in sensitivity of the two imaging methods to high-frequency attitude jitter to detect attitude anomalies. Linear array imaging utilizes push-broom imaging mode, with each row acquired at a different time. This causes high-frequency attitude jitter to appear as geometric discontinuities between adjacent rows. Area array imaging, on the other hand, utilizes instantaneous imaging mode, with the entire image acquired at the same time, unaffected by high-frequency attitude jitter. By establishing a strict geometric relationship between the two images, satellite attitude changes can be inferred, enabling precise detection and compensation of high-frequency attitude jitter.

[0060] Specifically, the present invention first constructs a rigorous imaging geometry model for line and area array images based on collinearity equations, accurately describing the mapping relationship between satellite and ground homonymous rays. External elevation data is then introduced, and the three-dimensional coordinates of each pixel in the area array image on the ground are substituted into the linear array image geometry model for inverse calculation, resulting in a simulated area array image ideally free of high-frequency attitude jitter. Through high-precision matching of the real area array image with the simulated area array image, dense homonymous points are obtained. The corresponding positions of the homonymous points on the linear array image are determined using the conversion relationship between the simulated area array image and the linear array image.

[0061] The key innovation of this invention lies in the creative elimination of projection difference interference caused by terrain undulations, imaging perspective differences, and different imaging scales in the process of calculating the upper and lower parallax of homonymous image points, so that the parallax change can accurately reflect high-frequency posture jitter. After correcting the systematic linear distortion using the affine transformation model, the residual error is the high-order distortion caused by high-frequency posture jitter. Based on the ascending order of the row and column coordinates of the homonymous points in the linear array image to form a two-dimensional grid, and using the difference in the parallax of adjacent homonymous points in the rows and columns to vote, it is possible to accurately identify areas of abnormal posture changes.

[0062] This method uses an iterative detection strategy to optimize posture parameters, eliminating the impact of posture errors on the overall detection results during area array imaging. Iterations are terminated when the parallax error falls below a set threshold, ensuring the reliability and accuracy of the detection results. The optimized posture data is used to reconstruct a rigorous imaging geometry model for the linear array images, generating geometrically seamless sensor-calibrated images. This significantly improves the geometric accuracy and uncontrolled elevation precision of stereo images, resolving the technical issue of low stereo model accuracy caused by high-frequency posture jitter.

[0063] To better understand and implement the present invention, a simplified version of Example 2 of the present invention is provided below. This Example 2 can better reflect the principles of the present invention. This example utilizes the different imaging principles of line and plane images, uses the instantaneous imaging parameters of the plane array to invert the line array's posture, detects high-frequency jitter in the posture through changes in image parallax, and adaptively weights it to reduce its random influence on the posture, improves the direction of photographic light, and makes pairs of light rays with the same name intersect, thereby improving the relative accuracy of the stereo model. Figure 2 shown.

[0064] The main technical processes for improving the geometric accuracy of linear array images with the aid of area array images are as follows: Figure 3 As shown, Figure 3There are multiple sub-images that serve as icons, which are only used as icons and do not represent specific meanings. The main research plan is: first, construct a rigorous imaging geometry model of line array images and area array images, combine external elevation data, and realize area array image simulation based on line array images. On this basis, introduce a high-precision image matching algorithm to obtain dense same-name image points of line and surface images; second, according to the rigorous imaging geometry model of line and surface images, determine the upper and lower parallax of the same-name image points, and preliminarily detect the high-frequency jitter of the attitude based on the change of parallax. The impact of different error factors on image distortion should be considered in the detection; then, based on the voting rule, use the line array scanning behavior as the voting interval, determine the high-frequency jitter in the line attitude based on the sudden change of the upper and lower parallax of the same-name image points, and adaptively determine the high-frequency jitter fitting weight to achieve high-frequency attitude filtering and improve the accuracy of the attitude model; finally, based on the revised satellite attitude data, combined with orbit determination data and other auxiliary data, reconstruct the rigorous imaging model of satellite line array images, splice the original images, generate a geometrically seamless sensor-corrected image product, and generate an auxiliary file. This embodiment is described in detail as follows.

[0065] The area array image simulation module will generate reconstructed images with the same resolution and imaging perspective as the real area array image. In this module, firstly, based on the orbit, attitude and imaging time parameters of the satellite uplink and downlink, the rigorous imaging geometric models of the area array image and the linear array image are constructed respectively. Then, combined with the external elevation data, the position of each pixel in the area array image is determined.

[0066] Three-dimensional coordinates on the ground , and then the above three-dimensional coordinates Substitute the rigorous geometric model of the linear array image for inverse calculation to obtain the coordinates of the corresponding image points of the linear array image , and finally align the coordinates on the linear image Perform grayscale interpolation to obtain grayscale value , and assign the grayscale value to the array image The whole area array image range is cycled in this way to obtain the resampled simulated area array image.

[0067] The geometric model of the strict imaging of the area array image is similar to that of the linear array image. Both are based on the satellite-to-ground homonymous ray mapping of the collinear equation. Their common form is as follows:

[0068] ;

[0069] in, is the coordinate of the image point; t is the imaging time, which has a strict corresponding relationship with the image column direction coordinate y for linear array images; m is the scale factor; is the three-dimensional coordinate of the image point in the Earth-centered Earth-fixed coordinate system (ECEF). It is the position of the satellite center determined by the onboard positioning equipment. is the satellite attitude matrix determined by the attitude measurement equipment, It is the external calibration parameter matrix, which mainly compensates for attitude and orbit measurement errors and equipment installation errors; is the coordinate of the image point in the satellite coordinate system. For linear array images, its form is as follows:

[0070] ;

[0071] The area array image is captured by the area array imaging unit instantaneously. It is composed of two-dimensional general polynomials as follows: A high-precision image matching algorithm is introduced to perform dense matching between real and simulated area array images, obtaining evenly covered same-name points. Based on the conversion relationship between the simulated area array image and the line array image, the corresponding positions of the same-name points on the real area array image on the line array image are obtained. In this way, projection errors caused by terrain undulations, imaging perspective differences, and imaging scale differences can be automatically compensated during the image reconstruction and correlation process, eliminating the interference of related errors on high-frequency attitude jitter detection.

[0072] When the satellite attitude changes suddenly and significantly, the vertical parallax of the stereo image near that moment will change suddenly, which is inconsistent with the overall parallax change trend. However, when the attitude is relatively stable, containing only systematic errors and small random errors, the above phenomenon will not occur. For multiple payloads on the same platform, their attitude errors are the same, but the imaging time is different. Assume that the satellite attitude model is , the posture model error is , then the main factor causing the inconsistency between the geometric models of the area array camera and the line array camera is that the attitude error changes with time, that is, , which is the asynchronous observation error. Because the asynchronous observation error is mainly reflected in the sudden change of the upper and lower parallax of the image, and the array is instantaneous imaging, its attitude error is a fixed value, so the linear array attitude error The overall detection diagram is as follows: Figure 4 shown.

[0073] The above detection process does not determine the exact attitude error, but determines whether the attitude within the change range is high-frequency jitter based on the change of the upper and lower parallax of the image. However, satellite image distortion is also affected by factors such as internal camera errors, orbit measurement errors, dynamic changes in payload installation, terrain undulations, and projection viewing angle differences. For the latter two, the impact is mainly in the along-track direction, and can be compensated through image simulation and correlation. For the first three, because the accuracy of on-board geometric calibration can reach 0.3 pixels, the orbit measurement accuracy has reached the centimeter level, and the changes in payload installation in a short period of time are mainly system errors, the image distortion they produce is mainly linear distortion. At this time, a linear deviation compensation model is proposed to correct these distortions to further eliminate the interference of related errors on high-frequency attitude jitter detection, such as the affine transformation model:

[0074] ;

[0075] in, is the line array image coordinate, is the array image coordinate, is the affine transformation model coefficient, and its model error is the high-order distortion caused by high-frequency posture jitter.

[0076] The high-frequency attitude jitter detection module determines the location of high-frequency jitter. In order to quantify its impact on the attitude model, a high-frequency attitude jitter weight determination method based on the voting rule is proposed. The main process is: first, the row and column coordinates of the line array image of the same-name points are sorted in ascending order to form a two-dimensional grid. In order to avoid the pointless area caused by image matching failure, a certain grid tolerance is allowed; secondly, the horizontal coordinate of the voting parameter space is determined according to the row coordinate range of the two-dimensional grid (corresponding to the attitude time), and the vertical coordinate is the number of votes; then, the difference in parallax between the adjacent rows and columns of the same-name points is calculated in each row of the two-dimensional grid (taking the absolute value). If the parallax difference is greater than a certain threshold, vote at the corresponding position in the voting parameter space, and record the maximum parallax difference. The voting is repeated in this way and the maximum parallax difference is updated; finally, the voting ratio and the maximum parallax difference of each interval are counted. The overall voting diagram is shown as follows Figure 5 The voting score is calculated using the maximum value of the difference between the voting ratio and the parallax: ;in, represents the voting score, E and D are the maximum difference between the voting ratio and the parallax, is the corresponding coefficient. The higher the voting score, the greater the high-frequency jitter in the interval, and the smaller the corresponding posture model fitting weight.

[0077] The above process should be repeated for each viewpoint image in the stereo model to ensure that high-frequency attitude detection covers the entire trajectory attitude measurement time. Furthermore, to eliminate the influence of attitude errors during area array imaging on the overall detection results, an iterative detection strategy is proposed. This involves performing image simulation and high-frequency attitude jitter detection using new attitude data after adaptive attitude weighted fitting. The iteration is terminated when the parallax error falls below a certain threshold, at which point the optimal attitude data is obtained.

[0078] The refined attitude data mentioned above is used in combination with satellite orbit determination and other ancillary data to generate a rigorous imaging geometry model of the satellite line array image. The ground-based auxiliary elevation reference data is used to stitch the original fragmented satellite imagery, and the image misalignment error, internal distortion error, and attitude orbit error are simultaneously eliminated to produce geometrically seamless sensor-corrected images. At the same time, a terrain-independent method is used to construct a favorable function model for the sensor-corrected image product, and other ancillary files are generated simultaneously.

[0079] A specific embodiment 2 of the present invention is provided below. The specific implementation of each step in this embodiment 2 is described in detail as follows.

[0080] The specific implementation of step S01 is to construct rigorous imaging geometric models of linear array images and planar array images based on the principle of collinear equations. For both linear array images and planar array images, rigorous imaging geometric models are established in the form of collinear equations, which are specifically expressed as follows:

[0081] ;

[0082] Where, is the coordinate of the image point, in pixels; is the imaging time in seconds. With line number There is a strict correspondence; is the proportionality coefficient, dimensionless; is the three-dimensional coordinate of the image point in the Earth-centered Earth-fixed coordinate system, in meters; is the position coordinate of the center of the satellite body, in meters; is the satellite attitude matrix, Orthogonal matrix; is the external calibration parameter matrix, matrix; is the coordinate of the image point in the satellite coordinate system, in meters. For linear array images, It is expressed in the form of a one-dimensional polynomial, as follows:

[0083] ;

[0084] ;

[0085] Where, Indicates the highest degree of polynomial; ( ) are polynomial coefficients, which are obtained through laboratory or on-orbit calibration. For area array images, It is expressed in two-dimensional polynomial form as follows:

[0086] ;

[0087] ;

[0088] Where, ( ) are the coefficients of a two-dimensional polynomial, obtained through laboratory or on-orbit calibration. The construction process first obtains the original orbital and attitude parameters transmitted from the satellite, as well as the camera parameters. Then, the internal and external orientation elements are calculated, and finally, a complete geometric model is formed. This step aims to establish a rigorous geometric foundation for subsequent processing, ensure the accuracy of coordinate transformations, and provide theoretical support for subsequent attitude detection and optimization.

[0089] The specific implementation of step S02 is to use external elevation data and a rigorous imaging geometry model to generate a simulated array image. First, obtain the external elevation reference data corresponding to the satellite coverage area with an accuracy better than 30 meters, and then calculate the height of each array pixel. , calculate the corresponding three-dimensional coordinates on the ground by forward projection of the collinear equation The specific process adopts an iterative solution method. The initial elevation value is the average elevation of the region. The iterative convergence condition is that the elevation change is less than 0.1 meters or the number of iterations reaches 10. Then the calculated ground three-dimensional coordinates are Substitute the rigorous geometric model of the linear array image for inverse calculation to obtain the image point coordinates corresponding to the linear array image , the calculation formula is:

[0090] ;

[0091] Where, is the scale factor; is the focal length of the camera, in pixels; the other parameters have the same meanings as step S01. Then, the grayscale value is calculated using the bicubic convolution interpolation algorithm on the line array image. , the interpolation kernel function is:

[0092] ;

[0093] Where, is the normalized distance. Finally, the grayscale value obtained by interpolation is assigned to the array image coordinates , complete the generation of simulated array images. 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 correspondence between homonymous points.

[0094] The specific implementation of step S03 is to use the scale-invariant feature transformation algorithm combined with the region growing algorithm to obtain the dense same-name points of the real array image and the simulated array image. First, extract the feature points. For each image point , construct a scale space representation: ;

[0095] Where, is the Gaussian kernel function, is the scale parameter; is the original image; Represents the convolution operation. Then detect the local extreme points in the scale space by using the difference Gaussian function:

[0096] ;

[0097] Where, is the scale factor, which is usually Then perform feature point matching using the nearest neighbor ratio method:

[0098] ;

[0099] Where, Representing feature points Its nearest neighbor distance; Representing feature points and its next closest neighbor distance; The threshold is 0.6. Then the random sampling consistency algorithm is used to remove the mismatched points. The internal point judgment standard is that the residual is less than 0.5 pixels. Finally, the filtered feature points are used as seed points, and the gradient correlation coefficient algorithm is used for region growing. The correlation window size is usually Pixels, with a correlation coefficient threshold of 0.8, are used to generate dense matching point pairs. The same-name points on the real area array image are converted to the line array image using the conversion relationship between the simulated area array image and the line array image, and the correspondence between the real area array image and the line array image is established. The purpose of this step is to establish a precise correspondence between images in different imaging modes, providing basic data for subsequent posture error detection.

[0100] The specific implementation of step S04 is to calculate the vertical parallax of the same-name image points of the linear array image and the area array image, and eliminate the interference factors that affect the detection of high-frequency posture jitter. First, calculate the vertical parallax of the linear array image and the area array image:

[0101] ;

[0102] ;

[0103] Where, 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 row disparity and column disparity respectively. Then, the influence of various factors on disparity is separated, and the total disparity can be expressed as the superposition of multiple factors:

[0104] ;

[0105] ;

[0106] Where, and is the parallax component caused by terrain undulation; and It is the parallax component caused by the difference in imaging perspective; and It is the parallax component caused by the difference in imaging scale; and It is the parallax component caused by high-frequency posture jitter; and is random noise. The disparity component caused by terrain is calculated as:

[0107] ;

[0108] ;

[0109] Where, and is the partial derivative of parallax with respect to elevation; is the elevation error. The parallax component caused by the difference in imaging perspective is the difference in imaging center angle between the two images. Correlation can usually be expressed as: ;

[0110] ;

[0111] Where, and is the ratio of the parallax component caused by the different imaging scales to the ground resolution of the two images. Related:

[0112] ;

[0113] ;

[0114] Where, and is the image point coordinate. By subtracting the above systematic components, the residual parallax is obtained:

[0115] ;

[0116] ;

[0117] Where, and The residual disparity mainly includes the disparity 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 disturbance information caused purely by attitude jitter.

[0118] The specific implementation of step S05 is to use an affine transformation model to correct the systematic linear distortion between the line array image and the area array image. The affine transformation model is expressed as:

[0119] ;

[0120] ;

[0121] Where, is the coordinate of the line array image; is the coordinate of the array image; ( ) are affine transformation model coefficients. Based on the residual disparity obtained in step S04, the affine transformation model parameters are estimated using the weighted least squares method:

[0122] ;

[0123] Where, is the number of points with the same name; For the The weight of each point is proportional to the reliability of the parallax. After solving the above least squares problem to obtain the affine transformation model parameters, the difference between the predicted coordinates and the actual coordinates of each point with the same name under the affine transformation model is calculated, that is, the parallax residual:

[0124] ;

[0125] ;

[0126] Where, and is the parallax residual, which mainly reflects the nonlinear distortion caused by high-frequency jitter of the posture. The frequency analysis of the parallax residual is performed using the fast Fourier transform:

[0127] ;

[0128] Where, is the parallax residual on the time series; is the frequency; is the sequence length; This is a frequency domain representation. High-frequency components are extracted from the frequency domain representation to determine the high-frequency attitude jitter characteristics. This step aims to separate linear system errors from nonlinear high-frequency jitter errors, providing a foundation for subsequent accurate high-frequency jitter identification.

[0129] The specific implementation of step S06 is to construct a two-dimensional grid and perform voting based on parallax mutation. First, the points with the same name are sorted in ascending order by row and column coordinates to construct a two-dimensional grid:

[0130] ;

[0131] Where, and are the center coordinates of the row and column grids respectively; and are the number of grid rows and columns respectively. The grid size is usually set to 10 times the row resolution of the line array image, and the allowed grid tolerance is 20% of the grid size. The voting parameter space is then determined based on the row coordinate range of the two-dimensional grid:

[0132] ;

[0133] Where, Indicates the voting sections, is the number of intervals, usually 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 parallax between adjacent rows and columns of the same name:

[0134] ;

[0135] ;

[0136] Where, and Grid points The row and column disparity residuals at . When the disparity difference is greater than the threshold (usually 0.5 pixels), votes are accumulated in the voting parameter space of the corresponding row coordinates:

[0137] ,if or and Falling in the interval Inside;

[0138] At the same time, record the maximum value of the parallax difference in this interval:

[0139] ;

[0140] The voting statistics are completed by traversing all grid points. The purpose of this step is to identify the time intervals where high-frequency posture jitter occurs through the voting mechanism, providing a basis for posture model refinement.

[0141] The specific implementation of step S07 is to calculate the voting score and determine the posture model fitting weight. First, calculate the voting ratio of each interval:

[0142] ;

[0143] Where, For the The voting ratio of each interval. Then the voting score is calculated based on the maximum difference between the voting ratio and the parallax:

[0144] ;

[0145] Where, Indicates the voting score; and are the maximum value of the difference between voting proportion and disparity respectively; and is the corresponding coefficient, usually , . Then normalize the voting scores:

[0146] ;

[0147] Finally, the posture model fitting weights are determined based on the normalized voting scores:

[0148] ;

[0149] Where, is the attenuation coefficient, typically set to 2. Weights typically range from 0.1 to 1.0, with high-scoring intervals corresponding to low weights and low-scoring intervals corresponding to high weights. The purpose of this step is to quantify the impact of high-frequency attitude jitter and provide a reasonable weight distribution scheme for attitude model fitting.

[0150] The specific implementation of step S08 is to optimize the posture parameters using an iterative detection strategy. First, based on the posture model fitting weights determined in step S07, the original posture data is filtered using a weighted spline smoothing method:

[0151] Where, is the filtered posture data; is the original posture data; is the weight; is the spline basis function; is the number of posture data points. The spline smoothing parameter is adaptively determined based on the posture data sampling frequency, usually 5 to 10 times the sampling interval. Then, using the optimized posture data, steps S02 to S07 are repeated to obtain new posture model fitting weights. Next, the parallax error after two adjacent iterative processes is calculated:

[0152] ;

[0153] Where, is the parallax error; and Respectively After the iteration The row and column parallax residuals of points; is the number of points with the same name. The iteration terminates when the parallax error is less than 0.3 pixels or the change in parallax error between two consecutive iterations is less than 5%. If convergence does not occur after 10 iterations, the iteration is terminated and the final iteration result is output. The purpose of this step is to gradually improve the accuracy of the pose parameters through iteration, eliminating the influence of pose errors during area array imaging on the detection results.

[0154] The specific implementation of step S09 is to use the optimized attitude data to reconstruct the strict imaging geometry model of the linear array image and generate a geometrically seamless sensor correction image. First, the optimized attitude data is combined with the original orbit determination data to update the attitude matrix in the collinearity equation. , reconstruct the rigorous imaging geometric model of the linear array image. Then construct the output image space and the rational function model, which is in the form of:

[0155] ;

[0156] ;

[0157] Where, is the normalized image coordinate; is the ground three-dimensional coordinate; 、 、 and are rational function coefficients; 、 and is the highest degree of the polynomial, usually 3. Then, for each pixel on the output image, its corresponding position on the original image is determined by backprojection, and grayscale resampling is performed using bilinear interpolation:

[0158] ;

[0159] Where, is the output image grayscale value; is the gray value of the original image; are the coordinates in the original image, The goal of this step is to use the optimized attitude parameters to generate sensor-corrected image products with higher geometric accuracy, providing a high-quality data foundation for subsequent applications.

[0160] Through the detailed steps above, Example 2 implements a complete technical solution for improving the geometric accuracy of linear array images using area array images on the same platform. This solution fully utilizes the complementary nature of multi-payload data on the same platform, leveraging the instantaneous imaging characteristics of the area array to assist in detecting high-frequency attitude jitter in linear array images. A voting mechanism is used to determine the jitter range, and attitude parameters are optimized through adaptive fitting, ultimately improving the geometric positioning accuracy of linear array images.

[0161] In order to better optimize the above embodiment 2, the following provides an embodiment 3 for optimization based on embodiment 2: A specific implementation method of using a satellite attitude jitter physical model to predict attitude jitter between step S04 and step S05 is to first establish an attitude jitter prediction equation based on the principle of satellite dynamics. The equation is specifically expressed as follows:

[0162] ;

[0163] Where, is the satellite attitude angle offset, in arc seconds; is the solar radiation pressure factor, dimensionless, ranging from 0.01 to 0.05, obtained by monitoring the current changes of satellite solar panels. The calculation formula is: ,in is the conversion factor, is the measured current of the solar panel, is the maximum current value; is the Earth's gravitational gradient factor, in units of , at an orbital altitude of 500km, the magnitude is approximately , calculated from the satellite orbital position, the calculation formula is: ,in is the Earth's gravitational constant, is the distance from the satellite to the center of the Earth, is the satellite latitude; is the geomagnetic field disturbance factor, in nT, ranging from 20 to 100 nT, obtained from the satellite-borne magnetometer measurement; is the thermoelastic deformation factor, in °C, obtained from the temperature sensor array measurement; is the mechanical vibration factor, in arc seconds, ranging from 0.5 to 2 arc seconds, obtained from the high-frequency measurement data of the gyroscope; 、 、 、 、 The corresponding coefficient is obtained through historical data calibration. Then the model is calibrated, and historical data of at least 10 typical track sections are selected to construct the calibration equation:

[0164] ;

[0165] Where, For the The measured attitude angle offset of each historical data point; is the total number of historical data points. Solve the least squares problem to obtain the optimal coefficient value. Then substitute the factor data obtained in real time into the calibrated model to calculate the predicted attitude angle offset:

[0166] ;

[0167] The prediction result is then converted to a geometric offset in the pixel domain:

[0168] ;

[0169] ;

[0170] Where, and are the geometric offsets in row and column directions, in pixels; and is the conversion coefficient, which is related to the satellite altitude and pixel size. Finally, the prediction result is introduced as prior information into the affine transformation model correction in step S05, specifically:

[0171] ;

[0172] ;

[0173] Where, is the corrected linear array image coordinate; is the affine model parameter, estimated by the least squares method, but the weight design takes into account the reliability of the physical model. The fusion of physical model prediction and image processing results adopts the Bayesian framework, and the fusion formula is:

[0174] ;

[0175] Where, is the attitude angle offset after fusion; and are the weights of physical model and image processing respectively, satisfying ,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 pose error estimation, and improve the,reliability and continuity of pose estimation, especially in areas where,the matching of homonymous points is insufficient.

[0176] The specific implementation method of using the self-attention residual network model to predict and compensate for high-frequency posture jitter between step S05 and step S06 is to first construct a deep neural network combining a multi-layer self-attention mechanism with residual connections. The network structure is specifically expressed as follows:

[0177] ;

[0178] ;

[0179] ;

[0180] ;

[0181] ;

[0182] ;

[0183] ;

[0184] Where, For the Output features of the layer; 、 、 They are query matrix, key matrix and value matrix respectively; 、 、 is the corresponding weight matrix; It is a multi-head attention mechanism; is the number of attention heads, which is set to 8, matching the satellite orbit period of approximately 90 minutes; 、 、 For the The parameter matrix of the attention head; is the feature dimension; is a normalization function. The network mainly consists of 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, pose measurement data, real pose change data, camera internal and external parameters, and external disturbance factor data are collected during the imaging process of linear and planar array images on historical track segments to construct data samples:

[0185] ;

[0186] ;

[0187] Where, is the input feature vector, containing disparity data , orbital parameters , thermal parameters , magnetic parameters , gyro data wait; The output label represents the three-axis attitude angular offset. The dataset is divided into training set, validation set and test set with a ratio of 7:2:1. Then a three-stage training strategy is adopted, firstly unsupervised pre-training:

[0188] ;

[0189] Where, is the contrastive learning loss function; is the eigenvector and similarity; is the temperature parameter, and its value is 0.07; is the indicator function. Pre-training uses the adaptive moment estimation optimizer with a learning rate of 0.001 and 100 training rounds. Then supervised fine-tuning is performed:

[0190] Where, is the root mean square error loss function; is the true label; is the model prediction value; is the number of samples. Fine-tuning uses an adaptive learning rate decay strategy with an initial learning rate of 0.0001. When the validation loss does not decrease for 5 consecutive rounds, the learning rate is halved. The total number of training rounds is 200. Finally, physical constraint-based regularization is introduced:

[0191] ;

[0192] ;

[0193] Where, is the total loss function with regularization; is the regularization coefficient, which is set to 0.1; is the physical constraint loss, which represents the difference between the second-order derivative of the predicted value and the value predicted by the dynamic equation; is the dynamic equation at time After training, the upper and lower parallax information calculated in step S04 is used as input features to predict high-frequency posture jitter through the trained network model:

[0194] ;

[0195] Where, is the predicted attitude angle offset; Represents the self-attention residual network model. The prediction result is weightedly fused with the voting result of step S06:

[0196] ;

[0197] Where, is the attitude angle offset after fusion; is the weight coefficient, ranging from 0.3 to 0.7, and is dynamically adjusted according to the confidence of the two. When the disparity data is missing, The value increases; when the disparity data is complete and consistent, The fused results are used for subsequent pose 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 techniques to mine implicit patterns in pose data, handle nonlinear relationships, and improve the accuracy and robustness of pose jitter prediction, especially when disparity data is incomplete or of low quality.

[0198] To better understand and implement the present invention, Example 4 of a specific application scenario is provided below: Researchers used a method for improving the geometric accuracy of linear array images assisted by same-platform area array images to process simulated satellite imagery. The simulated satellite was equipped with a multispectral linear array camera and a panchromatic area array camera, orbiting at an altitude of 778 km. The multispectral linear array camera had a spatial resolution of 5 meters, and the panchromatic area array camera had a spatial resolution of 2 meters. A set of simulated data from a mountainous area in western China was selected for the experiment. Its basic parameters are shown in Table 1:

[0199] Table 1 Basic parameters of experimental data

[0200]

[0201] First, rigorous imaging geometry models for line and area array imagery were constructed. Orbital coordinates were obtained based on satellite orbit determination data. Attitude data sampling intervals were 0.1 seconds, resulting in position accuracy better than 0.5 meters and attitude measurement accuracy of 0.001 degrees. Using the orbital position and attitude data transmitted from the satellite, rigorous imaging geometry models for line and area array imagery were constructed, respectively. Elevation data for the study area was extracted from the SRTM digital elevation model with a resolution of 30 meters and an elevation accuracy of approximately ±16 meters, serving as external elevation data input.

[0202] Then, based on the rigorous imaging geometry model and elevation data, the researchers simulated the conversion of line array images to area array images. During the simulation, the ground three-dimensional coordinates (X, Y, Z) of each area array image pixel (x, y) were calculated. The corresponding image point coordinates (x', y') were then inversely calculated using the rigorous geometric model of the line array image. Bicubic convolution interpolation was then used to obtain grayscale values. This entire simulation process resulted in a simulated area array image with the same resolution and viewing angle as the real area array image, measuring 6000 × 6000 pixels.

[0203] Next, a multi-scale SIFT matching algorithm was used to densely match the real area array image with the simulated area array image. During the matching process, the Gaussian pyramid layer number 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. The RANSAC algorithm was used to eliminate mismatched points with 2000 iterations and an inlier threshold set to 2.5 pixels. Ultimately, 32,768 evenly distributed points of the same name were obtained, with a matching accuracy better than 0.5 pixels and a point distribution spacing of approximately 30 pixels. The matching results are shown in Table 2:

[0204] Table 2 Image matching results statistics

[0205]

[0206] Based on the conversion relationship between the simulated area array image and the line array image, the corresponding positions of the same-name points on the real area array image on the line array image were determined. Based on the strict imaging geometry model of the line array image and the area array image, the vertical parallax of 32,768 densely packed same-name image points was calculated. To eliminate the influence of systematic errors, the affine transformation model was used for preliminary correction:

[0207] .

[0208] The corrected residual mainly reflects the nonlinear distortion caused by high-frequency jitter. By calculating the difference in row and column parallax between adjacent points of the same name, it was found that there were obvious parallax mutations in some areas. The parallax mutation threshold was set to 0.5 pixels, and the line array image scanning line was used as the voting interval (one interval every 25 lines) to detect high-frequency jitter. The detection results showed that there was obvious high-frequency jitter in the intervals of 2560-2816 lines, 5886-6016 lines, and 8192-8320 lines of the line array image, and the corresponding posture times were 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 statistics for these intervals are shown in Table 3:

[0209] Table 3 High-frequency jitter interval voting statistics

[0210]

[0211] The posture fitting weight is calculated based on the voting score, and the weight coefficient is: The weights of each high-frequency jitter interval are 0.24, 0.39, and 0.47, respectively. Then, the weighted least squares method is used to fit the posture data to reduce the weight of the high-frequency jitter area. At the same time, a pre-trained posture-optimized attention neural network model is used to analyze the relationship between parallax changes and high-frequency jitter. This model uses 8 attention heads, 64 attention dimensions, and a total of 12 layers to further optimize the posture parameters.

[0212] Finally, based on the revised satellite attitude data, combined with precise orbit determination data and camera internal parameters, a rigorous imaging geometry model of the linear array image is reconstructed, and the original image is stitched together to generate a geometrically seamless sensor-corrected image product. The comparison of geometric accuracy before and after correction is shown in Table 4:

[0213] Table 4 Comparison of geometric accuracy improvement effects

[0214]

[0215] Traditional methods for solving the problem of high-frequency jitter in satellite attitude mainly include low-pass filtering, spline fitting, and polynomial fitting. These methods usually perform overall smoothing on the attitude data and cannot distinguish the intensity differences of jitter in different time periods. They are prone to over-smoothing or under-smoothing, resulting in unstable geometric accuracy. However, the method of the present invention, assisted by the same-platform array image, utilizes the characteristics of instantaneous imaging of the array array to accurately identify the high-frequency attitude jitter interval in the line array image, and uses adaptive weights for targeted processing, thus 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% to 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.

[0216] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5, 6 and 7 below.

[0217] Table 5 Variable Explanation Table (Part I)

[0218]

[0219] Table 6 Variable Explanation Table (Part II)

[0220]

[0221] Table 7 Variable Explanation Table (Part 3)

[0222]

[0223] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform, characterized in that: include: S01. Constructing a rigorous imaging geometric model of a linear array image and a planar array image, wherein both the rigorous imaging geometric model of the linear array image and the rigorous imaging geometric model of the planar array image realize satellite-ground homonymous ray mapping based on collinear equations; S02, introducing external elevation data, substituting the three-dimensional coordinates of each pixel position of the area array image on the ground into the rigorous imaging geometric model of the line array image to perform inverse calculation to obtain the coordinates of the corresponding image point, and finally performing grayscale interpolation to obtain a simulated image; S03, performing high-precision matching on the area array image and the simulation image to obtain dense homonymous points, and obtaining corresponding positions of the homonymous points on the area array image on the line array image based on a conversion relationship between the simulation image and the line array image; S04, calculating the vertical parallax of the same-name image points in the linear array image and the planar array image, and eliminating interference of projection differences caused by terrain undulations, imaging viewing angle differences, and imaging scale differences on high-frequency attitude jitter detection; S05. Correcting the systematic linear distortion between the linear array image and the area array image using an affine transformation model, wherein the affine transformation model error is the high-order distortion caused by the high-frequency posture jitter; S06, sorting the row and column coordinates of the homonymous points in the linear array image in ascending order to form a two-dimensional grid, then determining a voting parameter space based on the row coordinate range of the two-dimensional grid, and performing voting based on the difference in parallax between adjacent rows and columns of the homonymous points; S07, counting the maximum value of the difference between the voting ratio and the disparity in each interval, and calculating the voting score. The higher the voting score, the greater the high-frequency posture jitter in the interval, and the smaller the corresponding posture model fitting weight; S08, taking the parallax error below a set threshold as the iteration termination condition, optimizing the posture parameters using an iterative detection strategy to eliminate the influence of the posture error during the imaging of the array image on the overall detection result; S09. Reconstruct a rigorous imaging geometry model of the linear array image using the optimized attitude data in combination with satellite orbit determination and other auxiliary data, and generate a geometrically seamless sensor correction image and auxiliary files.

2. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 1, characterized in that: The rigorous imaging geometric model is a satellite-ground homonymous ray mapping model based on collinear equations. The rigorous imaging geometric models of line array images and area array images are similar in form, and both are based on collinear equations to realize the conversion from image point coordinates to ground three-dimensional coordinates.

3. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 2, characterized in that: The image point coordinates of the linear array image are expressed by a one-dimensional polynomial, while the image point coordinates of the area array image are expressed 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.

4. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 3, characterized in that: The specific steps of generating the simulated image are to substitute the three-dimensional coordinates of each pixel position of the area array image on the ground into the rigorous imaging geometric model of the line array image for inverse calculation to obtain the corresponding image point coordinates, and finally perform grayscale interpolation to obtain the simulated image.

5. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 4, characterized in that: The steps of performing high-precision matching between the array image and the simulated image are: adopting a strategy combining feature point matching and region growing, extracting feature points through a scale-invariant feature transformation algorithm, matching feature points using the nearest neighbor ratio method, and eliminating mismatched points using a random sampling consistency algorithm.

6. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 5, characterized in that: The step of calculating the upper and lower parallaxes of the same-named image points of the line array image and the area array image includes calculating the upper and lower parallaxes of the same-named image points of the line array image and the area array image, eliminating interference of projection differences caused by terrain undulations, imaging viewing angle differences, and different imaging scales on high-frequency attitude jitter detection.

7. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 6, characterized in that: The affine transformation model parameters are estimated using weighted least squares.

8. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 7, characterized in that: The step of voting based on the difference in parallax between adjacent rows and columns of the same-name points is to sort the same-name 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 in parallax between adjacent rows and columns of the same-name points; the step of counting the voting results and calculating the posture jitter score includes counting the maximum value of the difference between the voting ratio and the parallax of each interval, and calculating the voting score. The higher the voting score, the greater the high-frequency posture jitter in the interval, and the smaller the corresponding posture model fitting weight.

9. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 8, characterized in that: After calculating the upper and lower parallaxes, the step also includes predicting attitude jitter using a satellite attitude jitter physical model. The satellite attitude jitter physical model is a satellite attitude prediction equation established based on the principle of satellite dynamics, which is used to evaluate the high-frequency attitude jitter caused by external factors while the satellite is in orbit, and provide a reference basis for the correction of the affine transformation model.

10. The method for improving the geometric accuracy of linear array images assisted by planar array images on the same platform according to claim 9, characterized in that: After step S04, the method also includes the step of predicting and compensating the high-frequency attitude jitter by using a self-attention residual network model, wherein the self-attention residual network model uses upper and lower 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 a multi-layer self-attention mechanism with a residual connection, comprising a feature extraction layer whose number of attention heads matches the satellite orbit 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