Hyperspectral remote sensing image correction method and system based on superpixel uniformity constraint

Abstract

This application provides a hyperspectral remote sensing image correction method and system based on superpixel uniformity constraints, relating to the field of image processing technology. The method includes: segmenting and aggregating spatially adjacent pixels with similar spectral characteristics in the hyperspectral remote sensing image to be processed, generating multiple superpixel regions; determining the reference brightness value corresponding to each superpixel region based on the pixel brightness distribution; constructing and solving a target optimization function based on the pixel values ​​of the superpixel regions and their corresponding reference brightness values ​​to obtain a target column correction coefficient vector; and performing column-dimensional gain correction on the hyperspectral remote sensing image to be processed, minimizing the overall error between each pixel value and its corresponding reference brightness value after correction by the target column correction coefficient vector. Implementing this method can remove longitudinal stripe noise, improve image radiometric uniformity, maintain the spatial structure and texture details of the image, and alleviate edge blurring.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method and system for correcting hyperspectral remote sensing images based on superpixel uniformity constraints. Background Technology

[0002] Hyperspectral remote sensing imaging technology can simultaneously acquire two-dimensional spatial images and one-dimensional spectral curves of observed targets, and has wide applications in precision agriculture, mineral exploration, and environmental monitoring. Currently, most spaceborne or airborne hyperspectral imagers employ pushbroom imaging, with their focal plane arrays composed of numerous photosensitive elements. During imaging, due to differences in detector pixel manufacturing processes, dark current noise, and spectral bending of the optical system (such as differences in inter-column radiative response caused by the Smile effect), the response of detector pixels in different columns to the same radiative energy often exhibits non-uniformity. This non-uniformity manifests not only as bright and dark stripe noise (bad lines) along the flight direction in the generated image data but also causes distortion of spectral features, severely interfering with the accuracy of subsequent ground feature classification and quantitative inversion.

[0003] In related technologies, variational optimization correction methods based on local gradients are commonly used. This method is primarily based on the prior assumption that natural images possess smoothness in local space, meaning that the grayscale changes of ground objects should be gradual within a small pixel geometric neighborhood. In practice, this method constructs an energy functional model incorporating total variation or unidirectional gradients, solving for the correction coefficients by minimizing the gradient difference between adjacent columns of pixels or minimizing the image gradient along the row direction. During the solution process, the algorithm smooths out abrupt stripe noise by constraining physically adjacent pixels to have numerically similar values ​​after correction.

[0004] However, when dealing with scenes featuring complex textures or rich edge information, existing technologies struggle to distinguish between high-frequency stripe noise and abrupt edge changes in real features. When genuine textured edges exist in the image (such as the boundary between farmland and roads) and the edge direction is roughly parallel to the flight direction, algorithms based on geometric neighborhood smoothing may incorrectly treat the real radiometric differences on both sides of the feature edge as noise and forcefully smooth them. This approach can lead to blurring of high-frequency details in the corrected image, disrupting the original texture structure of the features, and even introducing false transition artifacts at the boundaries between different feature categories, resulting in reduced radiometric fidelity of the image. Summary of the Invention

[0005] This application provides a hyperspectral remote sensing image correction method and system based on superpixel uniformity constraints, which addresses the problem of blurred image details and low edge fidelity caused by the difficulty in distinguishing between the true edges of ground objects and stripe noise when processing complex texture scenes.

[0006] In a first aspect, this application provides a hyperspectral remote sensing image correction method based on superpixel uniformity constraints, applied to a hyperspectral remote sensing image correction system, the method comprising: Pixels that are spatially adjacent and have similar spectral features in the hyperspectral remote sensing image to be processed are segmented and aggregated to generate multiple superpixel regions, each of which corresponds to a superpixel label. Statistical analysis is performed on the pixel brightness distribution of multiple superpixel regions to determine the reference brightness value corresponding to each superpixel region; A target optimization function is constructed and solved based on the pixel values ​​of the superpixel region and the reference brightness values ​​corresponding to the pixel values ​​to obtain the target column correction coefficient vector. After the hyperspectral remote sensing image is corrected by the target column correction coefficient vector, the overall error between each pixel value and the corresponding reference brightness value is minimized. Based on the target column correction coefficient vector, the hyperspectral remote sensing image to be processed is subjected to column-dimensional gain correction to obtain the corrected hyperspectral remote sensing image.

[0007] By employing the aforementioned technical solution, the system divides the image into physically meaningful superpixel regions using spectral feature similarity and spatial adjacency, ensuring that pixels within the same region tend to belong to the same land cover category. Next, the system constructs constraints within each superpixel region, solving for these constraints to make the corrected pixel values ​​approximate the reference brightness value of that region. This process utilizes the assumption of land cover uniformity within the superpixel to distinguish between column-dimensional non-uniform responses (striped noise) and actual land cover texture variations. By solving the objective optimization function to obtain correction coefficients and applying them, the system can remove vertical striped noise, improve image radiometric uniformity, and simultaneously maintain the image's spatial structure and texture details, mitigating edge blurring.

[0008] In some embodiments, the step of constructing and solving a target optimization function based on the pixel value of each superpixel region and the reference brightness value corresponding to the pixel value to obtain a target column correction coefficient vector specifically includes: Construct a target optimization function that includes a data fidelity term and a regularization term. The data fidelity term is used to characterize the minimization of the difference between the pixel value after correction by the column correction coefficient vector and the corresponding reference brightness value. The regularization term is used to constrain the minimization of the deviation between the column correction coefficient vector and the unit vector. The objective optimization function is transformed into a linear least squares model for solution, yielding the objective column correction coefficient vector. The linear least squares model is as follows: , where A is the coefficient matrix, b is the observation vector, and x is the target column correction coefficient vector to be solved.

[0009] By adopting the above technical solution, the system constructs an optimization model that includes data fidelity terms and regularization terms. The data fidelity term ensures that the corrected data conforms as closely as possible to the uniform distribution pattern within the superpixel, while the regularization term imposes constraints on the correction coefficients to prevent excessive fluctuations or overfitting during the solution process. The system transforms this complex optimization problem into a linear least squares model for solution, enabling the system to quickly process massive amounts of hyperspectral data. While ensuring the robustness and accuracy of the correction coefficient calculation, the system improves the computational efficiency and convergence speed of the algorithm, meeting the processing requirements of large amounts of hyperspectral image data.

[0010] In some embodiments, the objective optimization function is: Where x is the target column correction coefficient vector to be solved, and p is the number of pixels. For the k-th superpixel region, The column index where pixel p is located. This is the original brightness value of pixel p. This is the reference brightness value for the k-th superpixel region. Here, W is the regularization parameter, W is the image width, and K is the total number of superpixel regions. Let x be the correction coefficient of the j-th column in the target column correction coefficient vector. This is the correction coefficient corresponding to the column where pixel p is located.

[0011] By adopting the above technical solution, the system clarifies the specific mathematical expression of the objective optimization function and introduces a regularization parameter. This function refines the weight balance between the degree of data fitting and the smoothness of coefficients. Specifically, it considers the deviation between the original brightness of each pixel and the reference brightness of its superpixel region, and quantifies the global error through summation. This ensures that the solution process for the correction coefficient vector x takes into account the statistical characteristics of the entire image, guaranteeing that the correction coefficients converge to the optimal solution in different bands and brightness regions. This improves the fidelity of the correction results in the spectral dimension and reduces the risk of spectral pixel distortion caused by parameter estimation bias.

[0012] In some embodiments, before the step of transforming the objective optimization function into a linear least squares model for solution to obtain the objective column correction coefficient vector, the method further includes: The original brightness value of each pixel in the hyperspectral remote sensing image to be processed is assigned to the specified column element of the corresponding row in the coefficient matrix. The position of the specified column element is determined by the column index of the pixel. Assign the reference brightness value of the superpixel region corresponding to the pixel to the row element in the observation vector that matches the corresponding row; After the assignment is completed, a diagonal matrix is ​​appended to the bottom of the coefficient matrix. The values ​​of the elements on the main diagonal of the diagonal matrix are all the square roots of the regularization parameter. The row and column dimensions of the diagonal matrix are consistent with the width of the hyperspectral remote sensing image. An extended vector is appended to the bottom of the observation vector after the assignment is completed. All elements of the extended vector are the square root of the regularization parameter, and the length of the extended vector is the same as the width of the hyperspectral remote sensing image.

[0013] By adopting the above technical solution, the system expands and constructs the coefficient matrix and observation vector before solving the least squares model. By mapping the original pixel brightness to the coefficient matrix and appending the regularization parameters as a diagonal matrix to the bottom of the matrix, the optimization problem with regularization constraints is transformed into a standard problem of solving an overdetermined system of equations. This matrix construction method not only allows the mathematical model to be directly adapted to the computer's matrix operation library for parallel acceleration processing, but also improves the stability of numerical calculation by increasing the number of constraint equations, suppresses calculation divergence caused by local bad points or extrema in the observation data, and ensures that the final column correction coefficient vector has high reliability.

[0014] In some embodiments, before the step of segmenting and aggregating spatially adjacent pixels with similar spectral features in the hyperspectral remote sensing image to be processed to generate multiple superpixel regions, the method further includes: Traverse every pixel of the hyperspectral remote sensing image to be processed and identify dead pixels whose brightness values ​​are below the preset response threshold or are zero. The original brightness value of the dead pixel is replaced with the arithmetic mean of the brightness values ​​of the non-dead pixels that are spatially adjacent to the dead pixel.

[0015] By adopting the above technical solution, the system pre-processes the image to be processed with dead pixel detection and repair before performing superpixel segmentation and aggregation. By identifying pixels with abnormal brightness and replacing them with the average value of spatially adjacent normal pixels, the interference of dead pixels on subsequent processing steps can be eliminated. This preprocessing mechanism prevents dead pixels from causing erroneous guidance in region segmentation during the superpixel generation stage, and also avoids skewing the statistical results due to abnormal values ​​of dead pixels during reference brightness statistical analysis, thereby improving the overall quality and visual effect of the final image correction.

[0016] In some embodiments, before the step of segmenting and aggregating spatially adjacent pixels with similar spectral features in the hyperspectral remote sensing image to be processed to generate multiple superpixel regions, the method further includes: Calculate the global gradient mean of the hyperspectral remote sensing image, which is used to characterize the image texture complexity of the hyperspectral remote sensing image to be processed; When the global gradient mean is higher than the first preset threshold, the total number of superpixel regions is increased so that the average number of pixels contained in a single superpixel region falls within the first numerical range. When the global gradient mean is lower than the second preset threshold, the total number of superpixel regions is reduced so that the average number of pixels contained in a single superpixel region falls within the second numerical range, where the value of the first numerical range is less than the value of the second numerical range.

[0017] By adopting the above technical solution, the system introduces an adaptive adjustment mechanism based on the global gradient mean, which can dynamically adjust the size of the superpixel region according to the texture complexity of the image. When the image texture is complex (high gradient), the system automatically increases the number of superpixels to reduce the region area, thereby capturing the edges of ground features more precisely and preventing texture details from being smoothed out. When the image texture is flat (low gradient), the system reduces the number of superpixels to increase the region area, thereby obtaining more sample points for statistical analysis and improving the accuracy of reference brightness estimation. This adaptive adjustment strategy makes the correction method more adaptable to different scenes, and can automatically balance the effects of detail preservation and noise removal in different ground feature scenes such as urban areas, farmland, or water bodies.

[0018] In some embodiments, the step of statistically analyzing the pixel brightness distribution of the superpixel region to determine the reference brightness value corresponding to each superpixel region specifically includes: Calculate the spatial gradient magnitude of each pixel within the current superpixel region, and sort the pixels in ascending order based on the spatial gradient magnitude. Select target pixels that rank in the top preset percentage to form a stable pixel subset; Calculate the median of the brightness values ​​of all pixels in the stable pixel subset, and set the median as the reference brightness value of the corresponding band of the current superpixel region.

[0019] By employing the above technical solution, the system prioritizes calculating and sorting the spatial gradient magnitude of pixels, selecting a stable subset of pixels with gradual changes, and taking the median as the reference value. This process effectively eliminates the influence of high-frequency pixels at texture edges and abnormal pixels affected by strong noise on the statistical results, ensuring that the selected reference brightness value can truly represent the radiometric response level of the main features within the superpixel area. This improves the robustness of the reference value, making the correction coefficients calculated based on it more accurate, thereby reducing the possibility of artifacts in the corrected image.

[0020] Secondly, this application provides a hyperspectral remote sensing image correction system, the system comprising: one or more processors and a memory; The memory is coupled to the one or more processors. The memory is used to store computer program code, which includes computer instructions. The one or more processors call the computer instructions so that the system can implement the hyperspectral remote sensing image correction method based on superpixel uniformity constraints provided in the above embodiments, which will not be described in detail here.

[0021] Thirdly, this application provides a computer-readable storage medium including instructions that, when executed on a hyperspectral remote sensing image correction system, enable the system to implement the hyperspectral remote sensing image correction method based on superpixel uniformity constraints provided in the above embodiments, which will not be elaborated here.

[0022] Fourthly, this application provides a computer program product, including a computer program / instruction, which, when run on a hyperspectral remote sensing image correction system, enables the system to implement a hyperspectral remote sensing image correction method based on superpixel uniformity constraints provided in the above embodiments, which will not be elaborated here.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. By segmenting the image into physically meaningful superpixel regions and combining this with a gradient-based statistical strategy for stationary pixel subsets, a high-precision reference brightness constraint was constructed. Utilizing the spectral similarity and spatial adjacency within superpixels, uniformity correction was ensured within the same ground feature area. This effectively removed column-dimensional non-uniform response noise (stripes) while preserving the true texture and edge details of ground features to the greatest extent possible, thus improving the sharpness and radiometric fidelity of the corrected image.

[0024] 2. By automatically refining the segmentation scale in areas with complex textures to capture minute details and expanding the segmentation scale in flat areas to enhance the representativeness of statistical samples, this strategy can address the problem of detail loss or incomplete correction that easily occurs in different terrain scenarios with a single fixed scale parameter. It achieves an intelligent balance and optimal matching of detail preservation and noise removal effects in various complex scenarios such as cities and farmland.

[0025] 3. The system constructs a global optimization function that includes data fidelity and regularization terms, and transforms the complex optimization problem into a standard linear least squares model by constructing an augmented matrix. This transformation enables the system to directly solve for the target column correction coefficient vector using mature matrix operations. This not only avoids the slow convergence or getting trapped in local optima that may be caused by traditional iterative methods, but also uses regularization constraints to suppress numerical instability and overfitting risks during the calculation process. Thus, while ensuring the accuracy and reliability of the correction coefficients, the system improves the computational efficiency and real-time processing capability of massive hyperspectral data. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a hyperspectral remote sensing image correction method based on superpixel uniformity constraints in an embodiment of this application. Figure 2 This is another flowchart illustrating a hyperspectral remote sensing image correction method based on superpixel uniformity constraints in an embodiment of this application; Figure 3 This is a schematic diagram of the physical device structure of a hyperspectral remote sensing image correction system in an embodiment of this application. Detailed Implementation

[0027] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification and appended claims of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0028] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0029] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a hyperspectral remote sensing image correction method based on superpixel uniformity constraints according to an embodiment of this application.

[0030] S101. Segment and aggregate pixels that are spatially adjacent and have similar spectral characteristics in the hyperspectral remote sensing image to be processed to generate multiple superpixel regions.

[0031] The hyperspectral remote sensing image to be processed refers to the original hyperspectral remote sensing data that needs to be corrected for bad lines, stripe noise, and SMILE effect. Spatial proximity means that pixels are located close to each other in the two-dimensional spatial coordinates of the image. Spectral similarity means that pixels have consistent reflectance or radiation characteristics in the spectral dimension. The degree of similarity can be quantified by measurement methods such as Spectral Angle Mapper (SAM), Euclidean distance, or Bhattacharyya Distance. When the spectral similarity measurement value between two pixels is lower than the preset similarity threshold, the two pixels are judged to have similar spectral features. The preset similarity threshold can be adaptively set according to the sensor type and number of bands of the image. For example, for hyperspectral images with more than 100 bands, the SAM threshold can be set to a value in the range of 0.05 to 0.15 radians.

[0032] This step is performed at the initial stage of the hyperspectral remote sensing image correction process, after preprocessing operations such as dead pixel repair and adaptive adjustment of superpixel count. It is applicable to all hyperspectral remote sensing image correction scenarios with column-dimensional non-uniform response noise (such as bad lines and stripes) and SMILE effect.

[0033] Specifically, the system acquires a hyperspectral remote sensing image to be processed, which contains a large number of discrete pixels. Each pixel contains both spatial location information and specific spectral feature data. Due to potential issues such as inconsistent detector responses and optical system distortion in the original image, the system first divides the image into superpixel regions through segmentation and aggregation operations to establish a correction constraint basis based on ground feature semantics. The system traverses all pixels in the image, determining the spatial location of each pixel and analyzing its spectral feature data (such as combinations of brightness values ​​across different bands). Pixels that are spatially adjacent and whose spectral feature differences are within a preset range are grouped together, forming continuous superpixel regions through aggregation. Each superpixel region essentially corresponds to the same type of ground feature or a local portion of the same ground feature in reality, possessing inherent homogeneity. Finally, the system assigns a unique superpixel label to each aggregated superpixel region.

[0034] In some embodiments, the segmentation, aggregation, and label assignment of superpixel regions can be achieved in multiple ways: Optionally, the system can perform band dimensionality reduction processing on hyperspectral remote sensing images to extract the main spectral feature components and reduce data redundancy; set initial superpixel seed points, and evenly distribute several seed points in the image based on the image resolution and a preset initial superpixel size; calculate the weighted value of the spatial distance and spectral feature distance between the pixel and the seed point within a preset radius around each seed point, wherein the spectral feature distance can be calculated using SAM distance or normalized Euclidean distance, and the weighting coefficient of spatial distance and spectral distance can be adjusted according to the ratio of image spatial resolution to spectral resolution; assign each pixel to the cluster containing the seed point with the smallest weighted distance; iteratively optimize the clustering results, adjust the seed point position to the cluster center, recalculate the pixel affiliation, until the clustering results are stable; assign a unique digital label or symbolic label to each stable cluster (superpixel region).

[0035] Optionally, the system can also quantify the differences in spectral features between pixels by calculating the spectral similarity matrix of each pixel in the hyperspectral image; construct a spatial adjacency matrix based on the spatial coordinates of the pixels to determine the spatial proximity relationship between pixels; fuse the spectral similarity matrix and the spatial adjacency matrix to construct a pixel association matrix, highlighting the association strength of spatially adjacent and spectrally similar pixels; adopt a region growing algorithm to select an initial pixel from the image edge or randomly as the growth starting point; according to the strength order of the association matrix, neighboring pixels with an association degree higher than the threshold with the current region pixels are merged into the region until no new pixels can be included; perform morphological optimization on the generated region to remove excessively small regions or merge excessively fragmented regions, and finally form a superpixel region and assign a label.

[0036] It is understandable that other methods can be used to achieve superpixel segmentation aggregation and label assignment, such as methods that combine K-Means clustering with spatial constraints, etc., which are not limited here.

[0037] S102. Perform statistical analysis on the pixel brightness distribution of the superpixel region to determine the reference brightness value corresponding to each superpixel region.

[0038] Among them, pixel brightness distribution refers to the distribution of brightness values ​​of all pixels within a superpixel region over a numerical range, including characteristics such as the range, central tendency, and dispersion of brightness values; statistical analysis refers to the use of mathematical statistical methods to process pixel brightness data to extract feature values ​​that can characterize the overall brightness level of the region; reference brightness value refers to a benchmark value that can represent the true radiant brightness level of the main ground features within the superpixel region, used for subsequent construction of correction constraints; the same superpixel region corresponds to different reference brightness values ​​in different spectral bands, and the calculation of reference brightness values ​​is performed independently for each band.

[0039] Specifically, the system acquires the raw brightness values ​​of all pixels within each superpixel region, forming a brightness dataset. Since pixels within a superpixel region belong to the same type of land cover or a localized area of ​​the same land cover, their brightness values ​​should theoretically have strong consistency. However, due to noise, detector response differences, and other factors, the actual brightness values ​​exhibit some dispersion. Therefore, the system can perform statistical analysis on this brightness dataset, removing outliers or pixels with significant interference, and extracting reference values ​​that accurately reflect the brightness level of land cover areas. During the statistical analysis, the system can also comprehensively consider the central tendency and stability of brightness values ​​to avoid the impact of abnormal brightness values ​​from a single pixel on the accuracy of the reference brightness, ultimately determining a unique reference brightness value for each superpixel region.

[0040] In some embodiments, the reference brightness value of the superpixel region can be determined in several ways: Optionally, the system can traverse all pixels within the target superpixel region, extract the original brightness value of each pixel, and form a brightness data list; use the 3σ criterion to detect outliers in the brightness data list, calculate the mean and standard deviation of the data, and remove outlier brightness values ​​that exceed the mean ± 3 times the standard deviation; sort the brightness data after removing outliers and calculate the median of the middle position; verify whether the median is within the main distribution range of the brightness values ​​of the superpixel region, and if it is, determine it as the reference brightness value of the superpixel region; if it is not, readjust the outlier removal threshold and repeat the above steps until a reasonable median is obtained as the reference brightness value.

[0041] Optionally, the system can also filter out stable pixels whose gradient magnitude is less than a preset threshold by calculating the spatial gradient magnitude of each pixel within the target superpixel region; collect the brightness values ​​of stable pixels and calculate the arithmetic mean of these brightness values; count the frequency of occurrence of the brightness values ​​of stable pixels to determine the brightness range with the highest frequency; check whether the arithmetic mean falls within the high-frequency brightness range, and if so, use it as the reference brightness value; if not, take the midpoint of the high-frequency brightness range as the reference brightness value; compare the reference brightness value with the reference brightness value of adjacent superpixel regions, and if the difference is too large, re-verify the stable pixel filtering threshold and optimize the reference brightness value.

[0042] It is understandable that other methods can be used to determine the reference brightness value, such as combining the weighted average method (calculating the average brightness based on pixel spatial location weight or spectral purity weight), etc., which are not limited here.

[0043] S103. Construct a target optimization function based on the pixel values ​​of the superpixel region and the reference brightness values ​​corresponding to the pixel values.

[0044] Here, pixel value refers to the original brightness data of a single pixel within the superpixel region; the objective optimization function is a mathematical function constructed with the goal of minimizing the difference between the corrected pixel value and the reference brightness value and stabilizing the correction coefficient, used to quantify the optimization objective of the correction process; the data fidelity term is a component of the objective optimization function, used to ensure that the corrected pixel value is as close as possible to the reference brightness value, preserving the true radiation characteristics of ground objects; the regularization term is another component of the objective optimization function, used to constrain the fluctuation range of the correction coefficient, avoiding overfitting or numerical instability.

[0045] The core objective of constructing the target optimization function is to make the corrected pixel value as close as possible to the reference brightness value of its superpixel region, ensuring uniform brightness within the superpixel region and eliminating non-uniform noise in the column dimension. Simultaneously, it constrains the values ​​of the correction coefficients to avoid over- or under-correction caused by excessively large or small coefficients, ensuring the stability and rationality of the correction process. Based on these two objectives, the system constructs a target optimization function that includes a data fidelity term and a regularization term. In the data fidelity term, the system iterates through all pixels in each superpixel region, calculates the difference between the brightness value of each pixel after correction by the correction coefficient and the reference brightness value of that superpixel, and quantifies the global difference using a sum of squares to ensure that the overall difference is minimized. In the regularization term, the system constrains the correction coefficients of all columns, limiting their deviation from a preset benchmark value (e.g., 1), and controls the overall fluctuation of the correction coefficients using a sum of squares to prevent abnormal correction coefficients in individual columns from affecting the overall correction effect. The final target optimization function ensures that the correction effect closely matches the true brightness of ground features while also guaranteeing the robustness of the correction process.

[0046] In some embodiments, the objective optimization function is: Where x is the target column correction coefficient vector to be solved, and p is the number of pixels. For the k-th superpixel region, The column index where pixel p is located. This is the original brightness value of pixel p. This is the reference brightness value for the k-th superpixel region. Here, W is the regularization parameter, W is the image width, and K is the total number of superpixel regions. Let x be the correction coefficient of the j-th column in the target column correction coefficient vector. This is the correction coefficient corresponding to the column where pixel p is located.

[0047] It should be noted that the regularization term in the objective optimization function constrains the column correction coefficient vector x to approach the unit vector 1. Its physical meaning is that, under ideal imaging conditions, the radiation response gain of each column of detectors should remain consistent. In this case, the correction coefficients do not need to be additionally scaled to the original brightness values. This constraint effectively prevents the correction coefficients of individual columns from deviating from the normal response range due to local data anomalies during the solution process, avoiding overcorrection or overfitting.

[0048] The regularization parameter λ is used to balance the weight relationship between the data fidelity term and the regularization term. Its value should be adjusted according to the signal-to-noise ratio and stripe noise intensity of the image to be processed. When the stripe noise of the image is strong and the inter-column response difference is significant, λ can be taken as a small value (e.g., in the range of 0.001 to 0.01) to enhance the weight of the data fidelity term and enable the correction coefficient to fully approximate the true gain deviation. When the image texture is complex and the stripe noise is weak, λ can be taken as a large value (e.g., in the range of 0.1 to 1) to enhance the stability of the coefficient and avoid the ground texture being misjudged as stripe noise. The specific value can be determined by the L-curve method, the generalized cross-validation (GCV) method, or an empirical calibration method based on historical data. In some embodiments, the initial value of λ can be set to 0.05 and iteratively fine-tuned according to the corrected image quality evaluation index (such as stripe intensity index and radiometric uniformity).

[0049] S104. Transform the objective optimization function into a linear least squares model and solve it to obtain the objective column correction coefficient vector.

[0050] Among them, the linear least squares model is a mathematical model that finds the best function match of the data by minimizing the sum of squares of the errors. Its form is to minimize the squared 2 norm of the difference between the product of the coefficient matrix and the correction coefficient vector and the observation vector. The target column correction coefficient vector is a set of column-dimensional correction parameters that can minimize the overall error between the corrected pixel value and the reference brightness value. Each element corresponds to the correction coefficient of a certain column of the image.

[0051] The objective optimization function is in the form of a sum of squares of linear terms, which allows it to be transformed into a linear least squares model. During the transformation, the system constructs the corresponding coefficient matrix A and observation vector b. For the coefficient matrix A, the number of rows is the same as the sum of the total number of image pixels and the regularization-related dimensions, and the number of columns is the same as the image width (i.e., the number of columns). In each row corresponding to a pixel, only the column containing that pixel is filled with its original brightness value, while the remaining columns are filled with 0. Simultaneously, to incorporate the regularization term, a diagonal matrix is ​​appended to the bottom of the coefficient matrix, with the diagonal elements being the square root of the regularization parameter. For the observation vector b, its length is the same as the number of rows in the coefficient matrix. The first half corresponds to the reference brightness value of the superpixel region to which each pixel belongs, and the second half corresponds to the extended vector related to the regularization term, with all elements being the square root of the regularization parameter. After constructing the coefficient matrix and observation vector, the system transforms the objective optimization function into a standard linear least squares model. The model is then solved using mature linear algebraic methods (such as QR decomposition and singular value decomposition) to obtain the target column correction coefficient vector. Each element in this vector corresponds to the correction gain of a certain column of the image. This vector can be used to achieve non-uniformity correction in the column dimension.

[0052] In some embodiments, the linear least squares model is: , where A is the coefficient matrix and b is the observation vector.

[0053] The system initializes the coefficient matrix A and observation vector b by determining the initial number of rows (total number of pixels) and columns (image width) of matrix A, and the initial length of vector b (total number of pixels), based on the total number of image pixels and the image width. It then iterates through each pixel in the image, determining its column index and filling the corresponding column index position in the coefficient matrix A with the original pixel brightness value, and filling the remaining column positions with 0. In the observation vector b, it fills the corresponding pixel row position with the reference brightness value of the superpixel region to which the pixel belongs. A regularized diagonal matrix is ​​constructed with dimensions equal to the image width × image width, where each element on the main diagonal is set to the square root of the regularization parameter, and off-diagonal elements are 0. A regularized augmented vector is constructed with a length equal to the image width, and all elements are set to the square root of the regularization parameter. The regularized diagonal matrix is ​​concatenated to the bottom of the coefficient matrix A, and the regularized augmented vector is concatenated to the bottom of the observation vector b, completing the construction of the augmented matrix A and augmented vector b. The augmented matrix A is decomposed using the QR decomposition method to obtain an orthogonal matrix Q and an upper triangular matrix R. The matrix is ​​then solved using matrix operations. This yields the target column correction coefficient vector x.

[0054] Optionally, the system can also divide the image into several non-overlapping pixel blocks, each containing a preset number of pixels. For each pixel block, a local coefficient matrix and a local observation vector are constructed. The number of rows in the local coefficient matrix is ​​the number of pixels within the block, and the number of columns is the image width. The original brightness value is filled into the column position corresponding to the row of each pixel within the block, and the rest are 0. The local observation vector corresponds to the reference brightness value of the superpixel to which the pixel in the block belongs. All local coefficient matrices and local observation vectors are merged to form a global initial coefficient matrix and a global initial... The initial observation vector is used; an adaptive regularized diagonal matrix is ​​constructed, and the diagonal elements are adjusted according to the brightness fluctuation variance of each column of pixels. The larger the variance, the larger the diagonal elements (square root of the regularization parameter) are; the adaptive regularized diagonal matrix and the augmented vector are concatenated to the global matrix and vector to form the augmented matrix and augmented vector; the augmented matrix is ​​decomposed using the singular value decomposition (SVD) method to remove components with too small singular values ​​and avoid numerical instability; the linear equation system is solved based on the decomposition results to obtain the target column correction coefficient vector, and the coefficient vector is smoothed to remove individual outlier coefficients.

[0055] S105. Based on the target column correction coefficient vector, perform column-dimensional gain correction on the hyperspectral remote sensing image to be corrected to obtain the corrected hyperspectral remote sensing image.

[0056] Among them, column-dimensional gain correction refers to a correction method that uses a corresponding correction coefficient to multiply and scale the brightness values ​​of all pixels in each column of the image to eliminate the difference in response between columns.

[0057] Specifically, the system obtains the target column correction coefficient vector, the length of which is the same as the width of the image to be corrected, and each element corresponds to the correction gain coefficient of a certain column of the image. During the correction process, the system traverses the image to be corrected column by column. For each column, it extracts the corresponding correction coefficient, and then multiplies the original brightness value of all pixels in that column with the correction coefficient to obtain the corrected brightness value of each pixel. Since the target column correction coefficient vector is obtained based on the superpixel uniformity constraint, it can accurately compensate for the non-uniform response of each column detector. Therefore, through this multiplication scaling operation, the radiometric distortion caused by bright and dark stripes (bad lines) and the Smile effect along the column direction can be effectively eliminated. At the same time, due to the existence of the superpixel region constraint, the real texture and edge information of the ground features are preserved, and finally, a radiometrically uniform and detailed corrected hyperspectral remote sensing image is generated.

[0058] It should be noted that, since hyperspectral remote sensing images typically contain tens to hundreds of spectral bands, steps S101 to S105 in this embodiment are executed independently for each band to be corrected in the hyperspectral remote sensing image. When performing the superpixel segmentation and aggregation operation in step S101, the system preferentially selects any of the following methods to determine the feature data upon which the segmentation is based: Method 1: Calculate the arithmetic mean of the brightness values ​​of all bands or preset representative bands (such as the four typical bands of blue, green, red, and near-infrared) to generate a multi-band mean image, and then determine spatial adjacency and spectral similarity based on the mean image; Method 2: Perform principal component analysis on the hyperspectral image, and select the feature image composed of the top principal components whose cumulative contribution rate reaches a preset proportion (such as 95%) as the segmentation basis; Method 3: Directly calculate the spectral angular distance between pixels based on the full-band spectral vector for segmentation. Regardless of the method used, the superpixel segmentation results generated from the same image are shared across all bands. That is, all bands use the same set of superpixel labels, but each band independently performs the reference brightness value calculation in step S102 and the optimization solution and gain correction process in steps S103 to S105, ultimately obtaining the target column correction coefficient vector corresponding to each band.

[0059] In the above embodiments, the system divides the image into physically meaningful superpixel regions using spectral feature similarity and spatial adjacency, making pixels within the same region tend to belong to the same land cover category. Next, the system constructs constraints within each superpixel region, solving for these constraints to make the corrected pixel values ​​approximate the reference brightness value of that region. This process utilizes the assumption of land cover uniformity within the superpixel to distinguish between column-dimensional non-uniform responses (striped noise) and actual land cover texture variations. By solving the objective optimization function to obtain correction coefficients and applying them, vertical striped noise can be removed, image radiometric uniformity improved, and the spatial structure and texture details of the image preserved, thus mitigating edge blurring.

[0060] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating a hyperspectral remote sensing image correction method based on superpixel uniformity constraints in an embodiment of this application.

[0061] S201. Traverse every pixel of the hyperspectral remote sensing image to be processed and identify dead pixels whose brightness values ​​are lower than the preset response threshold or are zero.

[0062] Among them, the preset response threshold refers to the brightness threshold value set in advance by the system to determine whether a pixel is a dead pixel. This value is determined based on the normal response range of the detector of the hyperspectral imager. A dead pixel refers to a pixel whose brightness value is abnormally low (below the preset response threshold) or zero due to detector failure, manufacturing defects, or other reasons, and therefore cannot properly reflect the radiation information of ground objects.

[0063] This step is performed before superpixel segmentation and aggregation, and belongs to the preprocessing stage of hyperspectral remote sensing image correction. It is applicable to all hyperspectral remote sensing image correction scenarios with dead pixel interference.

[0064] Specifically, the system acquires basic parameters such as the total number of pixels and the number of rows and columns of the hyperspectral remote sensing image to be processed. Since the presence of dead pixels disrupts the uniformity of superpixel regions, leading to distorted segmentation results and biasing the statistical results of reference brightness values, dead pixel identification is necessary first. The system visits each pixel in the image one by one according to a preset traversal order (e.g., from left to right, from top to bottom), extracting the original brightness value of each pixel. For each pixel, the system compares its brightness value with a preset response threshold and determines whether it is zero. If the pixel brightness value is lower than the preset response threshold, or the brightness value is directly zero, the pixel is determined to be a dead pixel, and its spatial coordinates (row index and column index) are recorded.

[0065] S202. Replace the original brightness value of the dead pixel with the arithmetic mean of the brightness values ​​of the non-dead pixels that are spatially adjacent to the dead pixel.

[0066] Among them, non-dead pixels that are spatially adjacent refer to normal pixels that are directly adjacent to dead pixels in two-dimensional spatial coordinates (such as in the four directions of up, down, left, and right) or indirectly adjacent to dead pixels (such as in the eight directions of diagonal adjacency) and have not been marked as dead pixels; the arithmetic mean refers to the average value obtained by summing the brightness values ​​of adjacent non-dead pixels and dividing by the number of adjacent non-dead pixels.

[0067] The system obtains the coordinate information of all dead pixels marked in step S201. For each dead pixel, the system needs to find its spatially adjacent non-dead pixels, because adjacent normal pixels usually have similar radiation characteristics to the ground objects to which the dead pixel belongs, and their brightness values ​​can approximate the true brightness value of the dead pixel.

[0068] Specifically, the system searches for non-dead pixels surrounding the dead pixel based on a preset neighborhood range (e.g., 4-neighborhood, 8-neighborhood) and collects the original brightness values ​​of these non-dead pixels. If there are a sufficient number of non-dead pixels in the neighborhood (e.g., no less than 3), the arithmetic mean of these brightness values ​​is calculated; if the number of non-dead pixels in the neighborhood is small, the neighborhood range can be appropriately expanded until enough normal pixel brightness values ​​are collected. Finally, the calculated arithmetic mean is used to replace the original abnormal brightness value of the dead pixel, ensuring that the repaired pixel brightness value is consistent with the surrounding environment.

[0069] In some embodiments, the system can determine the pixel coordinates of its 4-neighborhood (up, down, left, and right directions) for each identified dead pixel; determine whether each pixel in the 4-neighborhood is a non-dead pixel, and collect the brightness values ​​of all non-dead pixels; if the number of collected non-dead pixels is ≥3, calculate the arithmetic mean of these brightness values ​​and use this average to replace the original brightness value of the dead pixel; if the number of collected non-dead pixels is <3, expand the neighborhood range to 8-neighborhood (adding the four diagonal directions), re-collect the brightness values ​​of non-dead pixels and calculate the arithmetic mean, and complete the replacement; for dead pixels at the edge or corner, if it is still impossible to collect enough non-dead pixels after expanding the neighborhood, use the brightness values ​​of adjacent non-dead pixels in the same column or row to perform linear interpolation to obtain the replacement value.

[0070] Furthermore, in this embodiment, the lower limit for a sufficient number of non-dead pixels can be set to 3. The specific termination condition for neighborhood expansion is as follows: the system first searches for non-dead pixels in a 4-neighborhood. If the number collected is less than 3, it expands to an 8-neighborhood. If there are still less than 3 in the 8-neighborhood (for example, dead pixels are located in the four corners or edge areas of the image), it further expands to a 5×5 neighborhood window centered on the dead pixel. If 3 or more non-dead pixels cannot be collected in the 5×5 neighborhood, the brightness value of the nearest non-dead pixel in the same column or row is used for linear interpolation replacement. When there are no available non-dead pixels in the same column or row within a preset search radius (for example, 10 pixels), the system marks the dead pixel position as an "invalid pixel" and skips it in the subsequent superpixel segmentation aggregation and reference brightness value statistics stages to avoid introducing incorrect replacement values ​​that may interfere with the correction results.

[0071] S203. Calculate the global gradient mean of the hyperspectral remote sensing image.

[0072] Among them, the global gradient mean refers to the average value obtained after statistically analyzing the spatial gradient magnitude of all pixels in the hyperspectral remote sensing image; the spatial gradient magnitude refers to the brightness change rate between a pixel and its neighboring pixels, which is used to reflect the intensity of texture change at the pixel's location; the image texture complexity refers to the fineness and frequency of change of the texture of ground features in the image. The more complex the texture, the higher the global gradient mean, and vice versa.

[0073] The system acquires hyperspectral remote sensing images after dead pixel restoration. Since the texture complexity of the image directly affects the superpixel segmentation effect (more refined segmentation is required for areas with complex textures, while coarser segmentation can be used for areas with flat textures), it is necessary to quantify the texture complexity by calculating the global gradient mean.

[0074] Specifically, for each pixel in the image, the system calculates its spatial gradient magnitude. A common method is to use gradient operators (such as the Sobel operator and the Prewitt operator) to calculate the gradient components in the horizontal and vertical directions, and then obtain the spatial gradient magnitude of that pixel by taking the square root of the sum of the squares. A larger value indicates a more drastic change in brightness and richer texture at that location. The system iterates through all pixels in the image, calculates the spatial gradient magnitude for each pixel, sums all gradient magnitudes, and then divides by the total number of pixels in the image to obtain the global gradient mean.

[0075] Taking the Sobel operator as an example, its horizontal convolution kernel... Convolution kernel in the vertical direction They are respectively: , For a pixel with coordinates (i, j) in the image, its horizontal gradient component and vertical gradient components By respectively , The spatial gradient magnitude is obtained by convolving the pixel with its 3×3 neighborhood. The calculation formula is: .

[0076] S204. When the global gradient mean is higher than the first preset threshold, increase the total number of superpixel regions so that the average number of pixels contained in a single superpixel region falls within the first numerical range.

[0077] The first preset threshold is a critical value set by the system in advance to determine whether the image texture belongs to a complex type. This value is determined based on the statistical results of the global gradient mean of a large number of images with different texture complexities. The total number of superpixel regions refers to the number of superpixel regions formed after image segmentation. The average number of pixels contained in a single superpixel region is the value obtained by dividing the total number of pixels in the image by the total number of superpixel regions. The first numerical range is a pre-set range of the average number of pixels in a single superpixel region applicable to images with complex textures. The values ​​in this range are relatively small, which can achieve more refined superpixel segmentation.

[0078] Specifically, the system compares the global gradient mean calculated in step S203 with a first preset threshold. When the global gradient mean is higher than the first preset threshold, it indicates that the terrain texture in the image is complex, with a large amount of detail and edge information. If a small number of superpixel regions is used, a single superpixel region will be too large, potentially containing multiple different terrain features, thus violating the superpixel uniformity assumption and affecting subsequent correction results. Therefore, the system increases the total number of superpixel regions. The system first obtains the total number of pixels in the image, and based on a preset first numerical range (e.g., 50-100 pixels), calculates the required total number of superpixel regions (from the maximum value of the first numerical range to the minimum value of the first numerical range). Then, within this range, the final total number of superpixel regions is determined, ensuring that the average number of pixels contained in a single superpixel region falls within the first numerical range, thereby achieving finer superpixel segmentation and accurately capturing terrain details and edges.

[0079] In some embodiments, the system can statistically analyze the global gradient mean distribution based on a large amount of sample data of complex textured images, and set the upper quartile of the distribution as a first preset threshold; calculate the total number of pixels in the current image, with a preset first numerical range of 60-90 pixels; when the global gradient mean is higher than the first preset threshold, calculate the minimum (total number of pixels ÷ 90) and maximum (total number of pixels ÷ 60) of the total number of superpixel regions; if the calculated total number is an integer range, select the middle value of the range as the final total number of superpixel regions; if it is a non-integer range, round up the minimum value and round down the maximum value, and then select the middle integer as the total number; verify whether the adjusted average number of pixels in a single superpixel region falls within the first numerical range, and if not, fine-tune the total number until the requirements are met.

[0080] Optionally, the system can also employ an adaptive threshold determination method. Based on the global gradient mean distribution of the current image, a first preset threshold is dynamically calculated (e.g., the median of the global gradient mean plus 1.5 times the interquartile range). The initial range of the preset first numerical interval is 50-100 pixels, and this interval is adjusted according to the spatial resolution of the image. The higher the resolution, the smaller the interval value (e.g., when the resolution is 10m, the interval is adjusted to 40-80 pixels). When the global gradient mean is higher than the first preset threshold, the total number of superpixel regions is initially set to the total number of pixels ÷ 70 (the median value of the first numerical interval). The average number of pixels corresponding to the initially set total number is calculated. If this number falls within the adjusted first numerical interval, the total number is determined. If it is higher than the maximum value of the interval, the total number of superpixel regions is further increased. If it is lower than the minimum value of the interval, the total number is appropriately reduced. Combining the distribution of edge pixels in the image, the total number of superpixel regions is finally fine-tuned to ensure more precise superpixel segmentation of the edge regions.

[0081] S205. When the global gradient mean is lower than the second preset threshold, reduce the total number of superpixel regions so that the average number of pixels contained in a single superpixel region falls within the second numerical range.

[0082] The second preset threshold is a critical value set by the system in advance to determine whether the image texture belongs to the flat type. It is determined based on the statistical results of the global gradient mean of a large number of flat texture images. The second numerical range is a range of average pixels in a single superpixel region that is set in advance and applicable to flat texture images. The value of this range is greater than that of the first numerical range, which can achieve coarser-grained superpixel segmentation.

[0083] Specifically, the system compares the global gradient mean with a second preset threshold. When the global gradient mean is lower than this threshold, it indicates that the terrain texture in the image is flat, with less detail and edge information. If a large number of superpixel regions are used, individual superpixel regions will be too small, resulting in insufficient sample points and affecting the statistical accuracy of the reference brightness value. Therefore, the system reduces the total number of superpixel regions. The system obtains the total number of pixels in the image and calculates the required total number of superpixel regions (from the maximum value of the second value range to the minimum value of the second value range) based on a preset second numerical range (e.g., 150-250 pixels). The final total number is then determined within this range to ensure that the average number of pixels in a single superpixel region falls within the second value range. This allows for the acquisition of a sufficient number of sample points through larger superpixel regions, improving the reliability of the reference brightness estimation.

[0084] In some embodiments, the system can statistically analyze the global gradient mean distribution based on a large number of flat texture image samples, and set the lower quartile of the distribution as a second preset threshold; calculate the total number of pixels in the current image, with the preset second numerical range being 180-220 pixels; when the global gradient mean is lower than the second preset threshold, calculate the minimum (total number of pixels ÷ 220) and maximum (total number of pixels ÷ 180) of the total number of superpixel regions; if the total number is in a non-integer range, round up the minimum value and round down the maximum value, and select the middle integer of the range as the initial total number; verify whether the average number of pixels corresponding to the initial total number falls within the second numerical range, and if not, fine-tune the total number until the requirement is met.

[0085] Furthermore, the system pre-sets a first preset threshold value greater than a second preset threshold value. When executing the above judgment logic, if the calculated global gradient mean is between the first and second preset thresholds (including cases equal to either threshold), it indicates that the texture complexity of the current hyperspectral remote sensing image to be processed is moderate; it lacks both abundant edge details and extremely flat regions. In this case, the system determines that the existing initial or default total number of superpixel regions is sufficient to balance the needs of detail preservation and sample statistics. Therefore, the system directly skips the step of adjusting the number of superpixel regions, i.e., it does not adjust the total number of superpixel regions, but maintains the initial parameters for subsequent segmentation and aggregation operations. This processing mechanism not only ensures the robustness of reference brightness value calculation in normal scenarios but also eliminates unnecessary parameter optimization calculations, improving the overall processing efficiency of the system.

[0086] S206. Calculate the spatial gradient magnitude of each pixel within the current superpixel region, and sort the pixels in ascending order based on the spatial gradient magnitude.

[0087] The current superpixel region refers to a single superpixel region that is currently calculating the reference brightness value; the spatial gradient magnitude of a pixel is used to reflect the degree of brightness change between that pixel and its neighboring pixels. The smaller the gradient magnitude, the more stable the brightness at the pixel's location.

[0088] The system processes each superpixel region individually. While pixels within a superpixel region theoretically belong to the same feature, they may exhibit local texture edges or noise interference, resulting in larger gradient magnitudes that can affect the accuracy of the reference brightness value. Therefore, the system needs to calculate the spatial gradient magnitude of each pixel within the superpixel to distinguish between stable pixels and edge / noise pixels.

[0089] Specifically, the system uses gradient calculation operators (such as the Sobel operator and the Prewitt operator) to calculate the gradient components in the horizontal and vertical directions for each pixel within the current superpixel region. These components are then combined to obtain the spatial gradient magnitude of the pixel. After calculation, the system sorts all pixels within the current superpixel region in ascending order of their spatial gradient magnitudes, placing stable pixels with smaller gradient magnitudes at the front and edge / noise pixels with larger gradient magnitudes at the back.

[0090] Optionally, the system can extract all pixels and their coordinate information in the current superpixel region to construct the local pixel matrix of the superpixel; use the Sobel operator to construct 3×3 convolution kernels in the horizontal and vertical directions respectively; traverse each non-edge pixel in the local pixel matrix and perform convolution operations with the 3×3 neighborhood of the pixel using the two convolution kernels respectively to obtain the horizontal gradient component and the vertical gradient component; calculate the spatial gradient magnitude of each pixel by taking the square root of the sum of squares; collect the gradient magnitudes and corresponding pixel identifiers of all pixels in the current superpixel region, sort them in ascending order of gradient magnitude, and generate a sorted pixel list.

[0091] S207. Select the target pixels that rank first by a preset percentage to form a stable pixel subset.

[0092] Among them, the preset percentage refers to the proportion threshold set in advance by the system for filtering stable pixels, which is determined based on the statistical law of pixel gradient distribution in the superpixel area (usually taken as 30%-60%); the target pixel refers to the pixel that is in the top preset percentage after sorting, which has a small spatial gradient amplitude and stable brightness change; the stable pixel subset refers to the set composed of all target pixels, which can reflect the true radiance level of the main ground objects in the current superpixel area.

[0093] After sorting the pixels within the current superpixel region by gradient magnitude, the system needs to select the subset of pixels with the most stable brightness changes from the sorting results. The preset percentage is set based on the assumption that "most pixels within a superpixel are stable pixels, with only a small number being edge or noise pixels." By selecting the preset percentage of pixels, edge pixels with large gradient magnitudes and abnormal pixels affected by noise can be effectively eliminated.

[0094] Specifically, the system determines the total number of pixels within the current superpixel region and calculates the number of target pixels to be selected based on a preset percentage (total number of pixels × preset percentage). Then, from the sorted pixel list, the system selects the top target pixel count as target pixels and integrates these target pixels to form a stable pixel subset. The pixel brightness variations within this subset are gradual, accurately reflecting the radiation characteristics of the main features within the superpixel region.

[0095] S208. Calculate the median of the brightness values ​​of all pixels in the stable pixel subset, and set the median as the reference brightness value of the corresponding band of the current superpixel region.

[0096] The median refers to the value in the middle position after sorting the brightness values ​​of all pixels in the stable pixel subset from smallest to largest (if the number of pixels in the subset is even, the average of the two middle values ​​is taken); the reference brightness value for the corresponding band refers to the baseline brightness value of the current superpixel region in a specific spectral band, which is used to construct correction constraints in the future.

[0097] Specifically, the system processes a subset of stationary pixels for each spectral band within each superpixel region. Since a small amount of slight noise or outliers may still exist within the stationary pixel subset, using the median instead of the mean to calculate the reference brightness value better resists interference from outliers and ensures that the reference brightness value more closely reflects the true radiation level of the ground object. The system extracts the brightness values ​​of all pixels in the current band from the stationary pixel subset, forming a brightness value list. This list is then sorted in ascending order, and the median is determined based on the parity of the number of pixels in the subset: if the number of pixels is odd, the brightness value in the middle position after sorting is taken as the median; if it is even, the arithmetic mean of the two middle brightness values ​​is taken as the median. Finally, the calculated median is directly set as the reference brightness value for the corresponding band of the current superpixel region.

[0098] The hyperspectral remote sensing image correction system of this invention is applied to electronic devices. Figure 3 A schematic diagram of the architecture of an electronic device suitable for implementing embodiments of the present invention is shown.

[0099] It should be noted that, Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0100] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by instructions (computer programs), or by instructions (computer programs) controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor. The electronic device of this embodiment includes a storage medium and a processor, wherein the storage medium stores multiple instructions that can be loaded by the processor to execute any step of the method provided in the embodiments of the present invention.

[0101] Specifically, the storage medium and the processor are electrically connected directly or indirectly to enable data transmission or interaction. For example, these components can be electrically connected to each other via one or more signal lines. The storage medium stores computer-executable instructions that implement data access control methods, including at least one software functional module that can be stored in the storage medium in the form of software or firmware. The processor executes various functional applications and data processing by running the software program and module stored in the storage medium. The storage medium can be, but is not limited to, Random Access Memory (RAM), Read-Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The storage medium stores the program, and the processor executes the program after receiving the execution instructions.

[0102] Furthermore, the software programs and modules within the aforementioned storage medium may also include an operating system, which may include various software components and / or drivers for managing system tasks (e.g., memory management, storage device control, power management, etc.) and can communicate with various hardware or software components to provide an operating environment for other software components. The processor may be an integrated circuit chip with signal processing capabilities. The aforementioned processor may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc., which can implement or execute the methods, steps, and logic flowcharts disclosed in this embodiment. The general-purpose processor may be a microprocessor or any conventional processor.

[0103] Since the instructions stored in the storage medium can execute the steps in any of the methods provided in the embodiments of the present invention, the beneficial effects of any of the methods provided in the embodiments of the present invention can be achieved, as detailed in the preceding embodiments, and will not be repeated here.

[0104] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A hyperspectral remote sensing image correction method based on superpixel uniformity constraints, applied to a hyperspectral remote sensing image correction system, characterized in that, The method includes: Pixels that are spatially adjacent and have similar spectral features in the hyperspectral remote sensing image to be processed are segmented and aggregated to generate multiple superpixel regions, each of which corresponds to a superpixel label. Statistical analysis is performed on the pixel brightness distribution of multiple superpixel regions to determine the reference brightness value corresponding to each superpixel region; A target optimization function is constructed and solved based on the pixel value of each superpixel region and the reference brightness value corresponding to the pixel value to obtain the target column correction coefficient vector. After the hyperspectral remote sensing image is corrected by the target column correction coefficient vector, the overall error between each pixel value and the corresponding reference brightness value is minimized. Based on the target column correction coefficient vector, the hyperspectral remote sensing image to be processed is subjected to column-dimensional gain correction to obtain the corrected hyperspectral remote sensing image.

2. The method according to claim 1, characterized in that, The step of constructing and solving a target optimization function based on the pixel value of each superpixel region and the reference brightness value corresponding to the pixel value to obtain the target column correction coefficient vector specifically includes: Construct a target optimization function that includes a data fidelity term and a regularization term. The data fidelity term is used to characterize the minimization of the difference between the pixel value after correction by the column correction coefficient vector and the corresponding reference brightness value. The regularization term is used to constrain the minimization of the deviation between the column correction coefficient vector and the unit vector. The objective optimization function is transformed into a linear least squares model for solution, yielding the objective column correction coefficient vector. The linear least squares model is as follows: , where A is the coefficient matrix, b is the observation vector, and x is the target column correction coefficient vector to be solved.

3. The method according to claim 2, characterized in that, The objective optimization function is: Where x is the target column correction coefficient vector to be solved, and p is the number of pixels. For the k-th superpixel region, The column index where pixel p is located. This is the original brightness value of pixel p. This is the reference brightness value for the k-th superpixel region. Here, W is the regularization parameter, W is the image width, and K is the total number of superpixel regions. Let x be the correction coefficient of the j-th column in the target column correction coefficient vector. This is the correction coefficient corresponding to the column where pixel p is located.

4. The method according to claim 3, characterized in that, Before the step of transforming the objective optimization function into a linear least squares model for solution to obtain the objective column correction coefficient vector, the method further includes: The original brightness value of each pixel in the hyperspectral remote sensing image to be processed is assigned to the specified column element of the corresponding row in the coefficient matrix. The position of the specified column element is determined by the column index of the pixel. Assign the reference brightness value of the superpixel region corresponding to the pixel to the row element in the observation vector that matches the corresponding row; After the assignment is completed, a diagonal matrix is ​​appended to the bottom of the coefficient matrix. The values ​​of the elements on the main diagonal of the diagonal matrix are all the square roots of the regularization parameter. The row and column dimensions of the diagonal matrix are consistent with the width of the hyperspectral remote sensing image. An extended vector is appended to the bottom of the observation vector after the assignment is completed. All elements of the extended vector are the square root of the regularization parameter, and the length of the extended vector is the same as the width of the hyperspectral remote sensing image.

5. The method of claim 1, wherein, Before the step of segmenting and aggregating spatially adjacent pixels with similar spectral characteristics in the hyperspectral remote sensing image to be processed to generate multiple superpixel regions, the method further includes: Traverse every pixel of the hyperspectral remote sensing image to be processed and identify dead pixels whose brightness values ​​are below the preset response threshold or are zero. The original brightness value of the dead pixel is replaced with the arithmetic mean of the brightness values ​​of the non-dead pixels that are spatially adjacent to the dead pixel.

6. The method of claim 1, wherein, Before the step of segmenting and aggregating spatially adjacent pixels with similar spectral characteristics in the hyperspectral remote sensing image to be processed to generate multiple superpixel regions, the method further includes: Calculate the global gradient mean of the hyperspectral remote sensing image, which is used to characterize the image texture complexity of the hyperspectral remote sensing image to be processed; When the global gradient mean is higher than the first preset threshold, the total number of superpixel regions is increased so that the average number of pixels contained in a single superpixel region falls within the first numerical range. When the global gradient mean is lower than the second preset threshold, the total number of superpixel regions is reduced so that the average number of pixels contained in a single superpixel region falls within the second numerical range, where the value of the first numerical range is less than the value of the second numerical range, and the first preset threshold is greater than the second preset threshold. When the global gradient mean is between the first preset threshold and the second preset threshold, the total number of superpixel regions is not adjusted.

7. The method according to claim 1, characterized in that, The step of statistically analyzing the pixel brightness distribution of multiple superpixel regions to determine the reference brightness value corresponding to each superpixel region specifically includes: Calculate the spatial gradient magnitude of each pixel within the current superpixel region, and sort the pixels in ascending order based on the spatial gradient magnitude. Select target pixels that rank in the top preset percentage to form a stable pixel subset; Calculate the median of the brightness values ​​of all pixels in the stable pixel subset, and set the median as the reference brightness value of the corresponding band of the current superpixel region.

8. A hyperspectral remote sensing image correction system, characterized in that, The system includes: one or more processors and memory; The memory is coupled to the one or more processors, the memory being used to store computer program code, the computer program code including computer instructions, the one or more processors invoking the computer instructions to cause the system to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are executed on a hyperspectral remote sensing image correction system, the system performs the method as described in any one of claims 1-7.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are run on the hyperspectral remote sensing image correction system, the system performs the method as described in any one of claims 1-7.

Images