A visible light remote sensing image thin cloud correction method, device, equipment and medium

Abstract

The application discloses a visible light remote sensing image thin cloud correction method, comprising the following steps: obtaining a to-be-corrected thin cloud image and a cloud-free reference image, and pre-processing the two images; constructing a difference field based on the pre-processed to-be-corrected thin cloud image and the cloud-free reference image, independently analyzing the difference field data, and obtaining a plurality of independent components and a mixing coefficient matrix; calculating the low-frequency energy proportion of each independent component, identifying a thin cloud independent component according to the low-frequency energy proportion, and extracting a cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix; performing thin cloud stripping processing on the to-be-corrected thin cloud image based on the cross-band coefficient vector, and obtaining a cloud-free clear image after thin cloud correction. The application effectively solves the problems of image blurring and ground information distortion caused by thin clouds, significantly improves the clarity and usability of thin cloud images, and improves the quality and application value of remote sensing images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, apparatus, device, and medium for correcting thin clouds in visible light remote sensing images. Background Technology

[0002] Visible light remote sensing images are widely used in scenarios such as agricultural monitoring, urban fine mapping, change detection, and target recognition. However, they are significantly affected by cloud cover. Currently, the proportion of cloud cover globally remains at a high level for a long time, making cloud and fog removal a high-frequency and essential requirement in remote sensing preprocessing. Compared to the loss of surface information caused by thick cloud cover, thin clouds (including thin fog and thin cirrus clouds) have semi-transparent characteristics. In visible light, they usually appear as radiation disturbances caused by scattering of atmospheric suspended particles, such as high brightness, decreased contrast, color shift, and blurred texture. Moreover, they are spatially smooth and have no obvious boundaries, making quantitative correction of thin clouds more challenging.

[0003] Currently, multiple technical approaches have been developed for thin cloud correction in visible light remote sensing images. However, there are still significant shortcomings in areas such as stable separation of thin cloud radiation disturbances, cross-source and cross-resolution adaptability, and the interpretability and controllability of results. It is difficult to obtain reliable and consistent correction results simultaneously under complex scenarios and diverse data conditions. Summary of the Invention

[0004] This invention provides a method for correcting thin clouds in visible light remote sensing images, which effectively solves the problems of image blurring and surface information distortion caused by thin clouds, significantly improves the clarity and usability of thin cloud images, and enhances the quality and application value of remote sensing images.

[0005] In a first aspect, embodiments of the present invention provide a method for thin cloud correction in visible light remote sensing images, comprising:

[0006] Acquire thin cloud images to be corrected and cloudless reference images, and preprocess both images;

[0007] A difference domain is constructed based on the preprocessed thin cloud image to be corrected and the cloudless reference image. Independent component analysis is performed on the difference domain data to obtain multiple independent components and mixing coefficient matrices.

[0008] Calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix;

[0009] Based on the cross-band coefficient vector, the thin cloud image to be corrected is subjected to thin cloud stripping processing to obtain a clear, cloud-free image after thin cloud correction.

[0010] Furthermore, the preprocessing of the two images includes:

[0011] The cloudless reference image is resampled to make the spatial resolution of the resampled cloudless reference image consistent with that of the thin cloud image to be corrected.

[0012] The resampled cloudless reference image is subjected to band-by-band radiometric homogenization to ensure that the homogenized cloudless reference image is in the same band as the thin cloud image to be corrected.

[0013] Furthermore, the band-by-band radiometric homogenization processing of the resampled cloudless reference image includes:

[0014] Based on the thin cloud image to be corrected and the resampled cloudless reference image, calculate the total difference of each pixel across all bands;

[0015] Based on the total difference, a stable set of pixels is selected, and the least squares method is used to calculate the radiometric uniformity parameters within the stable pixel region.

[0016] The resampled cloudless reference image is corrected band by band using the radiometric homogenization parameters to obtain a radiometrically homogenized cloudless reference image.

[0017] Furthermore, a difference domain is constructed based on the preprocessed thin cloud image to be corrected and the cloudless reference image. Independent component analysis is then performed on the difference domain data to obtain multiple independent components and mixing coefficient matrices, including:

[0018] Band-by-band difference is performed on the preprocessed thin cloud image to be corrected and the cloudless reference image to obtain a multi-band difference image.

[0019] Construct the observation matrix of the multi-band differential image; wherein each column of the observation matrix corresponds to the multi-band differential vector of a pixel;

[0020] Independent component analysis was performed on the observation matrix to obtain three independent components and a mixing coefficient matrix; wherein the three independent components are thin cloud radiation disturbance, actual ground change and system residual, respectively.

[0021] Furthermore, the calculation of the low-frequency energy proportion of each independent component includes:

[0022] Perform a two-dimensional Fourier transform on each independent component image to obtain the corresponding spectrum and calculate the power spectrum;

[0023] A low-frequency window is defined based on the center of the spectrum, and the energy within the low-frequency window is compared with the total energy in the entire frequency domain.

[0024] The ratio of low-frequency energy to total energy across the entire frequency domain is taken as the proportion of low-frequency energy in this independent component.

[0025] Furthermore, the step of identifying independent components of thin clouds based on the low-frequency energy ratio includes:

[0026] The three independent components are sorted from high to low according to the proportion of low-frequency energy.

[0027] If there exists a unique component with the largest proportion of low-frequency energy, it is identified as an independent component of thin clouds.

[0028] If there are multiple components with the largest proportion of low-frequency energy, calculate the gradient energy of each independent component and select the independent component with the smallest gradient energy as the thin cloud independent component.

[0029] Furthermore, the step of performing thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector to obtain a clear, cloud-free image after thin cloud correction includes:

[0030] Based on the thin cloud independent components and the corresponding cross-band coefficient vectors, calculate the thin cloud radiation disturbance increment for each band;

[0031] The thin cloud radiation disturbance increment is subtracted pixel by pixel from the corresponding band of the thin cloud image to be corrected to obtain a clear, cloudless image after thin cloud correction.

[0032] Secondly, embodiments of the present invention provide a thin cloud correction device for visible light remote sensing images, comprising:

[0033] The data acquisition module is used to acquire the thin cloud image to be corrected and the cloudless reference image, and to preprocess the two images.

[0034] The independent component analysis module is used to construct a difference domain based on the preprocessed thin cloud image to be corrected and the cloudless reference image, and to perform independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices.

[0035] The independent component discrimination module is used to calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix.

[0036] The thin cloud correction output module is used to perform thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector, so as to obtain a clear image without clouds after thin cloud correction.

[0037] Thirdly, embodiments of the present invention provide an electronic device, comprising:

[0038] Memory, used to store computer programs;

[0039] A processor for executing the computer program;

[0040] Wherein, when the processor executes the computer program, it implements the visible light remote sensing image thin cloud correction method described in any of the first aspects above.

[0041] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program that, when executed, implements the visible light remote sensing image thin cloud correction method described in any of the first aspects above.

[0042] Compared with existing technologies, the present invention provides a method for thin cloud correction of visible light remote sensing images, which has the following advantages: It acquires a thin cloud image to be corrected and a cloudless reference image, and preprocesses both images; it constructs a difference domain based on the preprocessed thin cloud image to be corrected and the cloudless reference image, performs independent component analysis on the difference domain data to obtain multiple independent components and a mixing coefficient matrix; it calculates the low-frequency energy proportion of each independent component, identifies the thin cloud independent components based on the low-frequency energy proportion, and extracts the cross-band coefficient vectors corresponding to the thin cloud independent components from the mixing coefficient matrix; it performs thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vectors to obtain a cloudless clear image after thin cloud correction. This invention effectively solves the problems of image blurring and surface information distortion caused by thin clouds, significantly improves the clarity and usability of thin cloud images, and enhances the quality and application value of remote sensing images. Attached Figure Description

[0043] To more clearly illustrate the technical features of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating a method for correcting thin clouds in visible light remote sensing images provided in an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram illustrating the construction of the difference domain and independent component analysis of a method for thin cloud correction in visible light remote sensing images provided in an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram illustrating the independent component discrimination of thin clouds in a method for correcting thin clouds in visible light remote sensing images provided in an embodiment of the present invention;

[0047] Figure 4 This is a schematic diagram of the structure of a thin cloud correction device for visible light remote sensing images provided in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0052] In a first aspect, embodiments of the present invention provide a method for thin cloud correction in visible light remote sensing images, see [link to previous section]. Figure 1 This is a flowchart illustrating an embodiment of a method for correcting thin clouds in visible light remote sensing images provided by the present invention.

[0053] like Figure 1 As shown, the method includes the following steps:

[0054] S1: Acquire the thin cloud image to be corrected and the cloudless reference image, and preprocess the two images;

[0055] S2: Based on the preprocessed thin cloud image to be corrected and the cloudless reference image, a difference domain is constructed. Independent component analysis is performed on the difference domain data to obtain multiple independent components and mixing coefficient matrices.

[0056] S3: Calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix;

[0057] S4: Based on the cross-band coefficient vector, perform thin cloud stripping processing on the thin cloud image to be corrected to obtain a clear, cloudless image after thin cloud correction.

[0058] In practice, the thin cloud image to be corrected is first acquired and preprocessed. and cloudless reference images Let the number of visible light bands be . The band index is If the spatial location of a pixel is x, then the multi-band observations at pixel x in two time phases can be expressed as:

[0059] ;

[0060] in, The reflectance value of the pixel at spatial location x in all B visible light bands is affected by thin clouds, while the reference reflectance value of the pixel at spatial location x in all B visible light bands is cloudless.

[0061] To ensure that images from different sources have uniform physical dimensions, it is preferable to input them in the form of reflectance. If the input is a digital quantization value or radiance, it is first converted into reflectance data through radiometric calibration and atmospheric correction. This invention does not limit the product level of reflectance; it can be either apparent reflectance or surface reflectance. At the same time, it ensures that the product level of the output reflectance is consistent with that of the input: when the input is surface reflectance, the output is the surface reflectance after thin cloud correction; when the input is apparent reflectance, the output is the apparent reflectance after thin cloud correction. This ensures the consistency and reliability of the correction results in quantitative applications.

[0062] After preprocessing the thin cloud image to be corrected and the cloudless reference image, differential domain data is constructed, and independent component analysis is performed on the multi-band data in the differential domain to obtain multiple independent components and mixing coefficient matrices. This allows for the statistical decomposition of major components such as thin cloud radiation disturbance, surface differences, and system residuals, providing input for subsequent thin cloud independent component discrimination and thin cloud radiation disturbance increment estimation.

[0063] The low-frequency energy ratio is calculated for each independent component. The low-frequency energy ratio is used to reflect whether the frequency domain energy of the independent component is dominated by low frequency. Based on the low-frequency energy ratio, the thin cloud component is identified, the independent component of the thin cloud is determined, and its corresponding cross-band coefficient vector is extracted from the mixing coefficient matrix for subsequent calculation and correction of thin cloud radiation disturbance increment.

[0064] After determining the independent components of thin clouds and their cross-band coefficient vectors, the thin cloud radiation disturbance increment corresponding to each band in the difference domain is calculated, and the thin cloud radiation disturbance increment is stripped from the visible light remote sensing image with thin clouds band by band to obtain a clear image without clouds after thin cloud correction.

[0065] It should be noted that, while outputting clear, cloudless images, it can also output the thin cloud radiation disturbance increment and the thin cloud cross-band coefficient vector, which can be used for the quantitative characterization, quality assessment, or subsequent quantitative applications of thin cloud radiation disturbance.

[0066] In summary, this invention addresses multi-source, multi-temporal, and multi-band visible light remote sensing imagery. It introduces temporal images with thin clouds and reference temporal images without clouds, constructing differential images through geometric registration, radiometric consistency, and resolution consistency. Furthermore, it performs independent component analysis on the multi-band differential data in the differential domain, separating independent components such as thin cloud radiation disturbance, surface differences, and system residuals. The statistical characteristics of thin cloud radiation disturbance in the spatial and frequency domains are used to identify independent thin cloud components. Subsequently, the mixing coefficients of these independent components in each band are used to estimate the increment of thin cloud radiation disturbance in each band. This increment is then stripped band by band from the temporal image with thin clouds to obtain the reflectance output after thin cloud correction. This invention achieves precise removal of thin cloud interference, effectively solving the problems of image blurring and surface information distortion caused by thin clouds, significantly improving the clarity and usability of thin cloud images, and enhancing the quality and application value of remote sensing images. It can be widely applied in fields such as agricultural monitoring, environmental monitoring, and urban planning that rely on high-quality remote sensing images.

[0067] In one optional implementation, the preprocessing of the two images includes:

[0068] The cloudless reference image is resampled to make the spatial resolution of the resampled cloudless reference image consistent with that of the thin cloud image to be corrected.

[0069] The resampled cloudless reference image is subjected to band-by-band radiometric homogenization to ensure that the homogenized cloudless reference image is in the same band as the thin cloud image to be corrected.

[0070] Specifically, when the thin cloud image to be corrected and the cloudless reference image come from different sensors, there are differences in the center wavelength and spectral response function of the two in the visible light band. Direct comparison will introduce systematic errors. Therefore, it is necessary to perform band correspondence and comparability constraints. Select the set of visible light bands that are covered by both images as the bands for subsequent processing. To address issues such as band center wavelength shift and inconsistent spectral response, inter-band linear normalization, regression mapping, statistical matching, or a combination of the above methods are used to achieve spectral consistency constraints between the bands of multi-source images, so that the visible light bands of different sensors are comparable.

[0071] After completing geometric registration, the two images are made resolution and radiometrically consistent to further eliminate differences in spatial and radiometric scales.

[0072] When the spatial resolution of the cloudless reference image and the thin cloud image to be corrected are inconsistent, resampling is performed on the cloudless reference image, and the resampling operator is denoted as follows. The expression for resampling is:

[0073] ;

[0074] in, This indicates that the resampled cloudless reference image is at the same grid and resolution as the thin cloud image to be corrected.

[0075] It should be noted that resampling can employ bilinear interpolation, bicubic interpolation, Lanczos interpolation, guided interpolation, or edge-preserving upsampling, etc., to ensure that excessive blur is not introduced during the resolution improvement process and to preserve the spatial details corresponding to thin clouds.

[0076] Because there are systemic differences such as imaging geometry, spectral response, and radiometric calibration deviations between different time phases and different sensors, directly performing differential operations on the image would cause these systemic differences to overlap with the radiation disturbances of thin clouds, severely interfering with the separation of thin cloud signals. Therefore, based on resolution uniformity, band-by-band radiometric uniformity is performed on the cloudless reference image. Radiometric uniformity can be achieved by linear gain / bias normalization, histogram matching, regression normalization based on stable ground features, or a combination thereof. This invention does not limit the specific implementation method.

[0077] This embodiment uses the above preprocessing operations to make the two images comparable in terms of spatial location, spatial resolution, band correspondence, and radiometric scale, providing a unified data foundation for subsequent differential domain construction and independent component analysis.

[0078] In one optional implementation, the band-by-band radiometric homogenization processing of the resampled cloudless reference image includes:

[0079] Based on the thin cloud image to be corrected and the resampled cloudless reference image, calculate the total difference of each pixel across all bands;

[0080] Based on the total difference, a stable set of pixels is selected, and the least squares method is used to calculate the radiometric uniformity parameters within the stable pixel region.

[0081] The resampled cloudless reference image is corrected band by band using the radiometric homogenization parameters to obtain a radiometrically homogenized cloudless reference image.

[0082] Specifically, a linear uniformity model is used for radiation correction, expressed as:

[0083] ;

[0084] in, Let reflectance be the cloud-free reference pixel in the k-th visible light band after radiation homogenization. This is the scaling factor (gain), used to adjust the reflectivity amplitude. This is the offset coefficient (bias), used to adjust the reflectivity baseline.

[0085] In solving for the radiometric homogenization parameters, the differences are first calculated pixel-by-pixel and band-by-band based on the thin cloud image to be corrected and the resampled cloudless reference image. The differences in each band are then accumulated to obtain the total difference for each pixel across the entire visible light spectrum. The formula is as follows:

[0086] ;

[0087] After obtaining the total difference of each pixel, a subset of pixels with relatively small total differences are selected as a stable pixel set. The stable pixel set represents reliable pixels that have not undergone significant changes in the land surface between two time phases, are not contaminated by thin clouds, and can be used for radiometric correction. Within the stable pixel set, using the radiometric value of the thin cloud image to be corrected as a reference, the least squares method is used to perform linear fitting on the resampled cloudless reference image to obtain the band-by-band radiometric uniformity parameters. The formula is as follows:

[0088] ;

[0089] in, To stabilize the pixel set, the reflectance of the stable pixels in the thin cloud image and the corrected reference image should be as close as possible. The solution obtained at this time... , It can most accurately compensate for the system bias in this band.

[0090] It should be noted that a robust regression method can also be used, which can avoid the interference of a small number of abnormally stable pixels on the parameters. This invention does not make specific limitations here.

[0091] Since the stable pixel set excludes areas of surface change and thin cloud cover, the obtained radiometric uniformity parameters can accurately reflect the system radiometric deviation between the two images, avoiding interference from thin cloud signals and surface changes on parameter estimation, and ensuring the accuracy and robustness of radiometric correction.

[0092] Using the solved band-by-band radiometric homogenization parameters, the resampled cloudless reference image is linearly corrected band by band, unifying its radiometric values ​​to the same reference as the thin cloud image to be corrected. Finally, all B bands are combined with the radiometrically homogenized reference bands to form a complete cloudless reference image. ,as follows:

[0093] ;

[0094] This correction process only adjusts the radiometric scale of the cloudless reference image without changing its spatial structure and relative relationship with ground features. It can eliminate systematic bias to the greatest extent while preserving the true surface information, making the cloudless reference image and the thin cloud image to be corrected strictly comparable.

[0095] In one optional implementation, the step of constructing a difference domain based on the preprocessed thin cloud image to be corrected and a cloudless reference image, and performing independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices, including:

[0096] Band-by-band difference is performed on the preprocessed thin cloud image to be corrected and the cloudless reference image to obtain a multi-band difference image.

[0097] Construct the observation matrix of the multi-band differential image; wherein each column of the observation matrix corresponds to the multi-band differential vector of a pixel;

[0098] Independent component analysis was performed on the observation matrix to obtain three independent components and a mixing coefficient matrix; wherein the three independent components are thin cloud radiation disturbance, actual ground change and system residual, respectively.

[0099] For example, see Figure 2 The diagram shown illustrates the construction of the difference domain and independent component analysis. For thin cloud images to be corrected, The multi-band differential image is obtained by subtracting the cloudless reference image from the thin cloud image to be corrected, which is the resampled cloudless reference image. , The image consists of multiple grayscale images, each corresponding to a visible light band. Independent component analysis is performed on the multi-band difference image, with the number of components K=3, outputting three independent components. and the mixing coefficient matrix ,in For the first The cross-band coefficient vector corresponding to each independent component.

[0100] Specifically, after geometric registration, resolution homogenization, and radiometric homogenization, the thin cloud image to be corrected and the cloudless reference image are subjected to band-by-band, pixel-by-pixel differential operations to obtain a multi-band differential image, denoted as... The pixel value of the differential image at position x in the k-th band can be expressed as:

[0101] ;

[0102] To accommodate subsequent matrix operations, the difference pixel values ​​of all B bands at each pixel x are stacked to form the multi-band difference vector for that pixel, expressed as:

[0103] ;

[0104] in, This indicates that the difference vector is a B-dimensional real vector, consistent with the total number B of visible light bands.

[0105] Furthermore, the spatial locations of the N effective pixels are denoted as... With each valid pixel The corresponding multi-band difference vectors are taken as a column, and the difference vectors of all N effective pixels are stacked sequentially to construct the input observation matrix of the ICA algorithm, expressed as:

[0106] ;

[0107] in, The observation matrix X represents a real matrix with B rows and N columns, where the number of rows B corresponds to the total number of visible light bands and the number of columns N corresponds to the total number of effective pixels.

[0108] It is understandable that each column of the observation matrix X corresponds to a multi-band differential vector of an effective pixel, and each row corresponds to the differential reflectance value of all effective pixels in the visible light band.

[0109] Furthermore, in the difference domain, using the linear mixture assumption of ICA, the constructed observation matrix X is expressed in linear mixture form:

[0110] ;

[0111] in, This is a mixing coefficient matrix, where the number of rows B corresponds to the total number of visible light bands, and the number of columns K corresponds to the number of independent components. For independent source matrices, the number of rows K corresponds to the number of independent components, and the number of columns N corresponds to the total number of effective pixels.

[0112] It is understandable that differential domain multi-band observation (i.e., the column vectors of the observation matrix X) are essentially a superposition of multiple types of difference signals. Considering the application scenario of thin cloud correction in visible light remote sensing images in this invention, the main sources of these difference signals can be summarized into the following three categories:

[0113] (i) Radiation perturbation terms introduced by thin clouds in the visible light band;

[0114] (ii) True surface differences between the two temporal images (including surface changes and differences in surface reflectance caused by observed geometric changes);

[0115] (iii) System residuals: including systemic effects that are difficult to completely eliminate, such as differences in cross-sensor spectral response, radiometric uniformity residuals, registration errors, resolution uniformity errors, and BRDF differences.

[0116] Based on the above three sources of difference, difference domain observation It can be further written as a superposition of three types of principal components, with the mathematical expression as follows:

[0117] ;

[0118] in, This represents the contribution of thin cloud radiation perturbation (i.e., the difference vector corresponding to the thin cloud radiation perturbation term). This represents the contribution of surface differences (i.e., the difference vector corresponding to the actual surface differences). Representing the system residual contribution (i.e., the difference vector corresponding to the system residual term), all three types of vectors are B-dimensional real vectors, and are related to the difference vector. Dimensions are consistent.

[0119] Within the linear mixture framework of the ICA, the three types of principal components mentioned above can be equivalently expressed as a linear mixture of three statistically independent sources, namely:

[0120] ;

[0121] in, For 3-dimensional independent source vectors, , , These correspond to the independent source values ​​of the three types of principal components, respectively.

[0122] Based on this, this step sets the number of independent components to K=3, that is, decomposes the main components of the difference domain into three independent components to characterize the main statistical independent source signals in the difference domain. This setting allows ICA decomposition to cover the three main sources of difference in the difference domain, while avoiding the problem that an excessive number of components would further decompose noise and local texture details into redundant components, thereby reducing the stability of thin cloud component discrimination.

[0123] Stack the three independent sources row by row to obtain the independent source matrix:

[0124] ;

[0125] in, , represents the value vector of the j-th independent source over all N valid pixels, corresponding to the j-th row of the independent source matrix S, ensuring that the dimensions of the independent source matrix match those of the observation matrix and the mixing coefficient matrix.

[0126] Furthermore, the observation matrix X undergoes two preprocessing steps: centering and whitening, to ensure that the processed data meets the requirements of ICA solution and improves decomposition accuracy and stability.

[0127] The core purpose of centering is to eliminate the mean shift of the data in each band of the observation matrix X, making the mean of the difference data in each band zero. This highlights the structural differences and statistical independence within the data, and avoids interference from the mean term in the ICA component decomposition. First, the mean vector of the difference vectors of all effective pixels is calculated. The calculation formula is:

[0128] ;

[0129] Where N is the total number of valid pixels, The multi-band difference vector of the i-th effective pixel is finally obtained. It is a B-dimensional mean vector, corresponding to the mean of the difference data for each visible light band.

[0130] The centered observation matrix is ​​obtained by subtracting the outer product of the mean vector μ and the all-1 vector from the observation matrix X. The calculation formula is as follows:

[0131] ;

[0132] The core purpose of whitening is to improve the centered observation matrix. Decorrelation and scaling are performed to ensure that the processed data are linearly independent across all dimensions, with each dimension having a variance of 1. This simplifies the ICA solution process and improves decomposition stability and convergence speed. First, the covariance matrix of the centered observation matrix is ​​calculated using the following formula:

[0133] ;

[0134] For covariance matrix Perform eigenvalue decomposition to obtain the eigenvalue matrix Λ and the eigenvector matrix E, which satisfy... Based on the eigenvalue decomposition results, a whitening matrix is ​​constructed, and the whitened data is obtained:

[0135] ;

[0136] ;

[0137] Wherein, the whitening matrix V is a B×B dimensional matrix, which is compared with the centered observation matrix. Multiplying by (B×N dimensions) yields the whitened data Z in B×N dimensions.

[0138] After whitening, the covariance matrix of Z (I is a B×B dimensional identity matrix), meaning that the data of each band of Z are linearly independent and have uniform variance, which fully meets the solution requirements of the ICA algorithm and can be directly used for subsequent ICA decomposition.

[0139] Finally, based on the whitened data Z, the unmixing matrix is ​​obtained by ICA. The independent component estimates are obtained, expressed as follows:

[0140] ;

[0141] in, This is the independent component estimation matrix. Its j-th row (j=1,2,3) corresponds to the value of the j-th independent component in all N effective pixels, and its i-th column corresponds to the estimated value of the three independent components at the i-th effective pixel.

[0142] Based on the unmixing matrix W and the whitening matrix V, the estimated value of the mixing coefficient matrix  can be recovered. For the sake of simplifying the notation,  will be denoted as A. The mixing coefficient matrix A, the whitening matrix V, and the unmixing matrix W satisfy the following equivalence relation:

[0143] ;

[0144] Due to the independent component estimation matrix As a 3×N two-dimensional matrix, it lacks spatial intuitiveness and cannot be directly used for the identification of independent components of thin clouds. Therefore, it needs to be remapped back to the spatial domain to obtain three two-dimensional independent component images:

[0145] ;

[0146] Finally, the mixing coefficient matrix A is further decomposed, and the mixing coefficient matrix is ​​denoted as A. ,in For the first The cross-band coefficient vector corresponding to each independent component.

[0147] This embodiment achieves effective decoupling of differential domain mixed signals by constructing a differential domain, organizing the observation matrix, and performing ICA preprocessing and solving. This lays a core foundation for subsequent thin cloud identification and improves the stability of thin cloud identification.

[0148] In one optional implementation, calculating the low-frequency energy percentage of each independent component includes:

[0149] Perform a two-dimensional Fourier transform on each independent component image to obtain the corresponding spectrum and calculate the power spectrum;

[0150] A low-frequency window is defined based on the center of the spectrum, and the energy within the low-frequency window is compared with the total energy in the entire frequency domain.

[0151] The ratio of low-frequency energy to total energy across the entire frequency domain is taken as the proportion of low-frequency energy in this independent component.

[0152] For example, see Figure 3 This is a schematic diagram for distinguishing independent components of thin clouds. A two-dimensional Fourier transform is performed on the image of each independent component to obtain its spectrum. The bright white area in the center of the spectrum corresponds to the low frequency domain, representing the smooth large-scale structure in space. The dark black area on the outside corresponds to the high frequency domain, representing the sharp small-scale details in space. The white dashed circle is the low frequency window, used to delineate the statistical range of low frequency energy.

[0153] Specifically, the j-th independent component image is represented as a two-dimensional discrete function. ,in Using row and column coordinates, perform a two-dimensional Fourier transform on them and calculate the power spectrum:

[0154] ;

[0155] ;

[0156] in, Represents the two-dimensional Fourier transform operator. For frequency domain coordinates, The power spectrum is used to characterize the energy distribution of this independent component at different spatial frequencies.

[0157] Let the size of the power spectrum be Set the spectrum center as (usually taken) Using the spectrum center as a reference, a low-frequency window function is constructed. .

[0158] In this embodiment, a circular window is preferably used to construct the low-frequency window function. The circular window is divided based on the condition that "the distance from the frequency domain coordinates to the center of the spectrum does not exceed a set radius threshold". ", among which, radius threshold The window function can be set proportionally to the power spectrum size, preferably 5% to 15% of the spectrum width. The mathematical expression for the window function is:

[0159] ;

[0160] in, As an indicator function, its value takes the following rules: when the condition in the parentheses is true, the value is 1, indicating that the frequency domain coordinate belongs to the low frequency region; when the condition in the parentheses is false, the value is 0, indicating that the frequency domain coordinate belongs to the high frequency region.

[0161] Based on this low-frequency window, calculate the... The proportion of low-frequency energy of each independent component is expressed as follows:

[0162] ;

[0163] in, , The larger the value, the more concentrated the frequency domain energy of the independent component is at low frequencies, and the smoother it is spatially.

[0164] This embodiment provides a precise and objective quantitative basis for the identification of independent components in thin clouds through two-dimensional Fourier transform, power spectrum calculation, low-frequency window construction, and low-frequency energy ratio quantification.

[0165] In one optional implementation, identifying the independent components of thin clouds based on the low-frequency energy ratio includes:

[0166] The three independent components are sorted from high to low according to the proportion of low-frequency energy.

[0167] If there exists a unique component with the largest proportion of low-frequency energy, it is identified as an independent component of thin clouds.

[0168] If there are multiple components with the largest proportion of low-frequency energy, calculate the gradient energy of each independent component and select the independent component with the smallest gradient energy as the thin cloud independent component.

[0169] Specifically, thin cloud radiation disturbances exhibit a significantly smooth and gradually varying structure in space, without obvious edges and detailed textures. This spatial characteristic is reflected in the frequency domain as its energy is mainly concentrated in the low-frequency region, with the proportion of low-frequency energy being significantly higher than the actual changes on the surface and the independent components corresponding to the system residuals. This is the core theoretical basis of the discrimination rule of this invention, ensuring the scientificity and rationality of the discrimination logic.

[0170] Based on the above, the three independent components are sorted from high to low according to the proportion of low-frequency energy. When there are cases where they are tied or the proportions of low-frequency energy are close, making it difficult to distinguish them, spatial statistics can be introduced as an auxiliary criterion, such as the gradient energy of the independent component images.

[0171] ;

[0172] in, Representing independent component images The gradient value in the x-direction, Representing independent component images The gradient value in the y-direction.

[0173] Selecting independent components with higher low-frequency energy proportions and lower gradient energies as independent components of thin clouds further improves the stability and accuracy of the independent component discrimination results for thin clouds.

[0174] In one optional implementation, after determining the independent components of thin clouds, column vectors corresponding to the independent components of thin clouds are extracted from the mixing coefficient matrix A as thin cloud cross-band coefficient vectors:

[0175] ;

[0176] Where, represents a matrix The List, An index for the independent components of thin clouds. This indicates that the thin cloud independent component is in the first... Linear mixing coefficients in the visible light band.

[0177] Due to the sign uncertainty in independent component analysis, in order to ensure the consistency of the thin cloud radiation disturbance increment in the sign direction, it is also necessary to perform sign unification processing on the independent components of the thin cloud and their cross-band coefficient vectors.

[0178] Calculate the mean of the independent components of thin clouds on the effective cell set Ω, determine the sign of the mean, and perform a sign flipping operation when the following conditions are met:

[0179] ;

[0180] The expression for the sign-flipping operation is:

[0181] ;

[0182] The above processing does not change the subsequent estimation of the thin cloud radiation disturbance increment. The product form is used only to unify the expression of thin cloud components in the symbol direction, thereby improving the consistency of results when processing across scenarios and data batches.

[0183] In one optional implementation, the step of performing thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector to obtain a cloud-free, clear image after thin cloud correction includes:

[0184] Based on the thin cloud independent components and the corresponding cross-band coefficient vectors, calculate the thin cloud radiation disturbance increment for each band;

[0185] The thin cloud radiation disturbance increment is subtracted pixel by pixel from the corresponding band of the thin cloud image to be corrected to obtain a clear, cloudless image after thin cloud correction.

[0186] Specifically, within the ICA linear mixing framework, the contribution of thin clouds to the difference domain can be expressed as the product of the independent components of the thin clouds and their coefficient vectors. Therefore, the estimated increment of thin cloud radiative perturbation in each band of the difference domain is:

[0187] ;

[0188] in, For thin cloud cross-band coefficient vectors, Let x be the value of the thin cloud independent component at pixel position x.

[0189] For the For each visible light band, the estimated increment of thin cloud radiative perturbation at pixel location x is:

[0190] ;

[0191] in Characterization in the The differential component contribution introduced by thin cloud radiation disturbance in each visible light band.

[0192] It should be noted that the increments of thin cloud radiation perturbations in all B bands... By stacking the bands, the complete estimation result of the thin cloud radiative perturbation increment can be obtained, denoted as . .

[0193] Furthermore, the calculated thin cloud radiation disturbance increments for each band are pixel-by-pixel stripped from the corresponding band of the thin cloud image to be corrected, removing the contribution of thin cloud interference and restoring the true reflectivity information of the surface. Finally, a clear, cloudless image after thin cloud correction is obtained. Specifically, for the k-th visible light band, the correction result is denoted as... The calculation formula is:

[0194] ;

[0195] After completing the thin cloud stripping operation band by band and pixel by pixel, the corrected reflectance of all B bands is stacked by band to obtain the complete thin cloud-corrected visible light remote sensing image, whose vector representation is as follows:

[0196] ;

[0197] To ensure the physical validity of the output image, a range constraint can be imposed on the corrected reflectance value, for example, by setting a range for the corrected reflectance value. The constraint is limited to the preset effective reflectance range [0, 1]. This constraint is used to suppress non-physical values ​​caused by extreme noise or anomalous pixels, without changing the basic calculation process of thin cloud correction.

[0198] This embodiment achieves precise removal of thin cloud interference by cross-band mapping calculation of thin cloud radiation disturbance increment, band-by-band thin cloud stripping correction, and physical rationality constraints, significantly improving the clarity and usability of thin cloud imagery, and enhancing the quality and application value of remote sensing imagery.

[0199] Secondly, embodiments of the present invention provide a thin cloud correction device for visible light remote sensing images, see [link to related document]. Figure 4 This is a schematic diagram of an embodiment of a thin cloud correction device for visible light remote sensing images provided by the present invention.

[0200] like Figure 4 As shown, the device includes:

[0201] The data acquisition module 21 is used to acquire the thin cloud image to be corrected and the cloudless reference image, and to preprocess the two images.

[0202] Independent component analysis module 22 is used to construct a difference domain based on the preprocessed thin cloud image to be corrected and the cloudless reference image, and to perform independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices.

[0203] The independent component discrimination module 23 is used to calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix.

[0204] Thin cloud correction output module 24 is used to perform thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector to obtain a clear, cloud-free image after thin cloud correction.

[0205] In one optional implementation, the preprocessing of the two images includes:

[0206] The cloudless reference image is resampled to make the spatial resolution of the resampled cloudless reference image consistent with that of the thin cloud image to be corrected.

[0207] The resampled cloudless reference image is subjected to band-by-band radiometric homogenization to ensure that the homogenized cloudless reference image is in the same band as the thin cloud image to be corrected.

[0208] In one optional implementation, the band-by-band radiometric homogenization processing of the resampled cloudless reference image includes:

[0209] Based on the thin cloud image to be corrected and the resampled cloudless reference image, calculate the total difference of each pixel across all bands;

[0210] Based on the total difference, a stable set of pixels is selected, and the least squares method is used to calculate the radiometric uniformity parameters within the stable pixel region.

[0211] The resampled cloudless reference image is corrected band by band using the radiometric homogenization parameters to obtain a radiometrically homogenized cloudless reference image.

[0212] In one optional implementation, the step of constructing a difference domain based on the preprocessed thin cloud image to be corrected and a cloudless reference image, and performing independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices, including:

[0213] Band-by-band difference is performed on the preprocessed thin cloud image to be corrected and the cloudless reference image to obtain a multi-band difference image.

[0214] Construct the observation matrix of the multi-band differential image; wherein each column of the observation matrix corresponds to the multi-band differential vector of a pixel;

[0215] Independent component analysis was performed on the observation matrix to obtain three independent components and a mixing coefficient matrix; wherein the three independent components are thin cloud radiation disturbance, actual ground change and system residual, respectively.

[0216] In one optional implementation, calculating the low-frequency energy percentage of each independent component includes:

[0217] Perform a two-dimensional Fourier transform on each independent component image to obtain the corresponding spectrum and calculate the power spectrum;

[0218] A low-frequency window is defined based on the center of the spectrum, and the energy within the low-frequency window is compared with the total energy in the entire frequency domain.

[0219] The ratio of low-frequency energy to total energy across the entire frequency domain is taken as the proportion of low-frequency energy in this independent component.

[0220] In one optional implementation, identifying the independent components of thin clouds based on the low-frequency energy ratio includes:

[0221] The three independent components are sorted from high to low according to the proportion of low-frequency energy.

[0222] If there exists a unique component with the largest proportion of low-frequency energy, it is identified as an independent component of thin clouds.

[0223] If there are multiple components with the largest proportion of low-frequency energy, calculate the gradient energy of each independent component and select the independent component with the smallest gradient energy as the thin cloud independent component.

[0224] In one optional implementation, the step of performing thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector to obtain a cloud-free, clear image after thin cloud correction includes:

[0225] Based on the thin cloud independent components and the corresponding cross-band coefficient vectors, calculate the thin cloud radiation disturbance increment for each band;

[0226] The thin cloud radiation disturbance increment is subtracted pixel by pixel from the corresponding band of the thin cloud image to be corrected to obtain a clear, cloudless image after thin cloud correction.

[0227] It should be noted that the visible light remote sensing image thin cloud correction device provided in this embodiment of the invention is used to execute all the process steps of the visible light remote sensing image thin cloud correction method in the above embodiment. The working principle and beneficial effect of the two are one-to-one, so they will not be described again.

[0228] Thirdly, embodiments of the present invention provide an electronic device, see [link to previous document]. Figure 5 The diagram shown is a structural schematic of an electronic device provided in an embodiment of the present invention.

[0229] like Figure 5 As shown, the device includes:

[0230] Memory 31 is used to store computer programs;

[0231] Processor 32 is used to execute the computer program;

[0232] When the processor 32 executes the computer program, it implements the visible light remote sensing image thin cloud correction method as described in any of the above embodiments.

[0233] For example, the computer program may be divided into one or more modules / units, which are stored in the memory 31 and executed by the processor 32 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the electronic device.

[0234] The processor 32 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0235] The memory 31 can be used to store the computer programs and / or modules. The processor 32 implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory 31 and calling the data stored in the memory 31. The memory 31 may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created according to the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory 31 may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital card (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0236] It should be noted that the aforementioned electronic devices include, but are not limited to, processors and memory, as will be understood by those skilled in the art. Figure 5 The structural diagram is merely an example of the electronic device described above and does not constitute a limitation on the electronic device. It may include more components than shown in the diagram, or combine certain components, or use different components.

[0237] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed, implements the visible light remote sensing image thin cloud correction method described in any of the above embodiments.

[0238] It should be understood that the present invention can implement all or part of the processes in the above-described visible light remote sensing image thin cloud correction method by instructing related hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above-described visible light remote sensing image thin cloud correction method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0239] 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. It should be noted that, for those skilled in the art, several equivalent obvious modifications and / or equivalent substitutions can be made without departing from the technical principles of the present invention, and these obvious modifications and / or equivalent substitutions should also be considered within the scope of protection of the present invention.

Claims

1. A method for correcting thin clouds in visible light remote sensing images, characterized in that, include: Acquire thin cloud images to be corrected and cloudless reference images, and preprocess both images; A difference domain is constructed based on the preprocessed thin cloud image to be corrected and the cloudless reference image. Independent component analysis is performed on the difference domain data to obtain multiple independent components and mixing coefficient matrices. Calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix; Based on the cross-band coefficient vector, the thin cloud image to be corrected is subjected to thin cloud stripping processing to obtain a clear, cloud-free image after thin cloud correction.

2. The method for thin cloud correction in visible light remote sensing images as described in claim 1, characterized in that, The preprocessing of the two images includes: The cloudless reference image is resampled to make the spatial resolution of the resampled cloudless reference image consistent with that of the thin cloud image to be corrected. The resampled cloudless reference image is subjected to band-by-band radiometric homogenization to ensure that the homogenized cloudless reference image is in the same band as the thin cloud image to be corrected.

3. The method for thin cloud correction in visible light remote sensing images as described in claim 2, characterized in that, The band-by-band radiometric homogenization process for the resampled cloudless reference image includes: Based on the thin cloud image to be corrected and the resampled cloudless reference image, calculate the total difference of each pixel across all bands; Based on the total difference, a stable set of pixels is selected, and the least squares method is used to calculate the radiometric uniformity parameters within the stable pixel region. The resampled cloudless reference image is corrected band by band using the radiometric homogenization parameters to obtain a radiometrically homogenized cloudless reference image.

4. The method for thin cloud correction in visible light remote sensing images as described in claim 1, characterized in that, The method involves constructing a difference domain based on the preprocessed thin cloud image to be corrected and a cloudless reference image, and performing independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices, including: Band-by-band difference is performed on the preprocessed thin cloud image to be corrected and the cloudless reference image to obtain a multi-band difference image. Construct the observation matrix of the multi-band differential image; wherein each column of the observation matrix corresponds to the multi-band differential vector of a pixel; Independent component analysis was performed on the observation matrix to obtain three independent components and a mixing coefficient matrix; wherein the three independent components are thin cloud radiation disturbance, actual ground change and system residual, respectively.

5. The method for thin cloud correction in visible light remote sensing images as described in claim 1, characterized in that, The calculation of the low-frequency energy percentage of each independent component includes: Perform a two-dimensional Fourier transform on each independent component image to obtain the corresponding spectrum and calculate the power spectrum; A low-frequency window is defined based on the center of the spectrum, and the energy within the low-frequency window is compared with the total energy in the entire frequency domain. The ratio of low-frequency energy to total energy across the entire frequency domain is taken as the proportion of low-frequency energy in this independent component.

6. The method for thin cloud correction in visible light remote sensing images as described in claim 1, characterized in that, The process of identifying independent components of thin clouds based on the low-frequency energy ratio includes: The three independent components are sorted from high to low according to the proportion of low-frequency energy. If there exists a unique component with the largest proportion of low-frequency energy, it is identified as an independent component of thin clouds. If there are multiple components with the largest proportion of low-frequency energy, calculate the gradient energy of each independent component and select the independent component with the smallest gradient energy as the thin cloud independent component.

7. The method for thin cloud correction in visible light remote sensing images as described in claim 1, characterized in that, The process of performing thin cloud stripping on the thin cloud image to be corrected based on the cross-band coefficient vector to obtain a clear, cloud-free image after thin cloud correction includes: Based on the thin cloud independent components and the corresponding cross-band coefficient vectors, calculate the thin cloud radiation disturbance increment for each band. The thin cloud radiation disturbance increment is subtracted pixel by pixel from the corresponding band of the thin cloud image to be corrected to obtain a clear, cloudless image after thin cloud correction.

8. A thin cloud correction device for visible light remote sensing images, characterized in that, include: The data acquisition module is used to acquire the thin cloud image to be corrected and the cloudless reference image, and to preprocess the two images. The independent component analysis module is used to construct a difference domain based on the preprocessed thin cloud image to be corrected and the cloudless reference image, and to perform independent component analysis on the difference domain data to obtain multiple independent components and mixing coefficient matrices. The independent component discrimination module is used to calculate the low-frequency energy ratio of each independent component, identify the thin cloud independent component based on the low-frequency energy ratio, and extract the cross-band coefficient vector corresponding to the thin cloud independent component from the mixing coefficient matrix. The thin cloud correction output module is used to perform thin cloud stripping processing on the thin cloud image to be corrected based on the cross-band coefficient vector, so as to obtain a clear image without clouds after thin cloud correction.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program; The processor executes the computer program to implement the visible light remote sensing image thin cloud correction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed, implements the visible light remote sensing image thin cloud correction method as described in any one of claims 1 to 7.

Images