An automated cleaning method and system for large-scale remote sensing samples

By introducing digital elevation model and normalized cloud index adaptive threshold segmentation into large-scale remote sensing sample cleaning, candidate windows of different sizes are generated and weighted fusion is performed, which solves the problems of repair failure and fusion distortion of neighboring similar pixel interpolation algorithm under complex terrain, and improves the integrity and consistency of data.

CN122820706BActive Publication Date: 2026-11-17GUIZHOU PROVINCIAL SURVEYING & MAPPING DATA ARCHIVES (GUIZHOU PROVINCIAL BASIC GEOGRAPHIC INFORMATION CENT) +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611278107.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-17
Estimated Expiration
2046-08-21

AI Technical Summary

Technical Problem

Existing neighbor-similar-cell interpolation algorithms fail to repair due to insufficient effective cells in high-altitude or cloud-dense areas, while they tend to introduce heterogeneous cells with large spectral differences in complex terrain areas, making it difficult to meet the differentiated requirements of neighborhood search windows under different terrain conditions.

Method used

By introducing a digital elevation model to extract the terrain feature vectors of pixels for clustering, and combining the normalized cloud index for adaptive threshold segmentation, candidate windows of different sizes are generated. Spatial distribution features and spectral discrete features are quantified, and weighted fusion is performed by combining temporal proximity and spectral consistency. The optimal search window is then selected for cloud area repair.

Benefits of technology

It significantly improves the accuracy of cloud mask recognition, ensures the data integrity and spatiotemporal consistency after cleaning large-scale remote sensing samples, and solves the problems of repair failure and fusion distortion in complex terrains caused by traditional methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122820706B_ABST
    Figure CN122820706B_ABST
Patent Text Reader

Abstract

This application relates to the field of image data processing technology, specifically to an automated cleaning method and system for large-scale remote sensing samples. The method includes: first, introducing a digital elevation model to extract terrain feature vectors and performing clustering, combined with normalized cloud index adaptive segmentation, effectively eliminating the interference of terrain factors on the spectrum and significantly improving the recognition accuracy of cloud masks under complex terrain. Second, generating multi-size candidate windows for the pixels to be repaired, quantifying spatial distribution and spectral discrete features, and dynamically weighting and fusing them according to window size. This mechanism balances information abundance and surface homogeneity, effectively overcoming the contradiction between insufficient information in small windows and heterogeneous distortion in large windows. Finally, considering the temporal proximity and spectral consistency of reference images, a weighted average of the initial quality scores of each temporal phase is performed to select the optimal search window. This application aims to ensure the data integrity and spatiotemporal consistency after cleaning large-scale remote sensing samples.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of image data processing technology, specifically to an automated cleaning method and system for large-scale remote sensing samples. Background Technology

[0002] In large-scale automated cleaning of remote sensing samples, cloud inpainting is a crucial step in addressing the loss of surface information and the generation of invalid, low-quality pixels caused by cloud cover. Without effective inpainting, cloud-covered samples are highly susceptible to noise interference. By applying cloud inpainting algorithms to restore the true ground feature information of obscured areas, not only can the integrity and consistency of the samples be significantly improved, but the reliability and usability of the cleaned dataset can also be ensured, thus providing solid data support for high-precision intelligent remote sensing interpretation.

[0003] The Nearest Neighbor Similarity Interpolation (NSPI) algorithm, a commonly used cloud infilling method, primarily searches for spectrally similar neighboring pixels by setting a fixed window and weights them based on spatial distance and spectral differences to calculate missing values. However, due to the use of a fixed neighborhood search window, this algorithm has significant limitations in practical applications: in information-sparse areas such as high altitudes or densely clouded areas, it often fails to repair due to insufficient effective pixels; while in areas with complex terrain features, it easily introduces heterogeneous pixels with large spectral differences, causing fusion distortion, and it is difficult to take into account the differentiated requirements of the neighborhood search window under different terrain conditions. Summary of the Invention

[0004] In view of the above, it is necessary to provide an automated cleaning method and system for large-scale remote sensing samples to solve the above problems.

[0005] The first aspect of this application provides an automated cleaning method for large-scale remote sensing samples, the method comprising: Acquire remote sensing images, multi-temporal reference images, and corresponding digital elevation model data of the target area; The terrain feature vectors of pixels are extracted using digital elevation model data. The pixels are clustered, and the normalized cloud index is calculated within each cluster for adaptive threshold segmentation to determine cloudless pixels and pixels to be repaired. Centered on each pixel to be repaired, candidate windows of different sizes are preset; for each temporal reference image, the spatial distribution characteristics of cloudless pixels and the discrete characteristics of spectral vectors within each candidate window are analyzed, and the initial quality score of each candidate window in each temporal phase is obtained by combining the size of the candidate window. Based on the temporal proximity of the acquisition time of each phase reference image and the spectral consistency of spectral features, the fusion weight of each phase is obtained. The initial quality scores of each phase are weighted and averaged to obtain the final quality score of each candidate window, and the optimal search window is obtained. Based on cloudless pixels within the optimal search window, the pixel value is interpolated and repaired using a neighboring similar pixel interpolation algorithm.

[0006] Preferably, the terrain feature vector includes the elevation, slope, aspect sine, and aspect cosine of each pixel.

[0007] Preferably, the specific calculation process of the normalized cloud index is as follows: The sum and difference between the reflectance of each pixel in the blue band and the reflectance in the red band are calculated separately. The negative correlation mapping result of the obtained sum and the obtained difference are positively fused to obtain the normalized cloud index.

[0008] Preferably, the specific process for determining cloudless pixels and pixels to be repaired is as follows: A threshold segmentation algorithm is applied to the normalized cloud index of all pixels in each cluster to obtain a segmentation threshold. The normalized cloud index of pixels in the remote sensing image to be repaired and the reference images of each time phase is compared with the segmentation threshold of the cluster in which the pixel belongs. If the normalized cloud index is greater than or equal to the segmentation threshold, it is determined to be a cloud-polluted pixel; otherwise, it is determined to be a cloudless pixel. The cloud-polluted pixels in the remote sensing image to be repaired are marked as pixels to be repaired.

[0009] Preferably, obtaining the initial quality score of each candidate window at each time phase specifically involves: Analyze the proportion of cloudless pixels in each candidate window and the spatial distance between them and the pixels to be repaired to calculate information availability; analyze the deviation of the spectral vector of cloudless pixels in each candidate window from its mean vector to calculate surface homogeneity. Based on the size of the candidate window, information availability and surface uniformity are weighted and fused to obtain the initial quality score of each candidate window in each time phase.

[0010] Preferably, the availability of the computational information specifically includes: Under the reference images of each time phase, the number of cloudless pixels in the candidate window is counted, and the proportion of cloudless pixels is calculated. The corresponding pixel in each temporal reference image of the pixel to be repaired is taken as the reference pixel. The average Euclidean distance from each cloudless pixel in the candidate window to the reference pixel is calculated and normalized to obtain the normalized average distance. The negative correlation mapping result of the normalized average distance is positively fused with the proportion of cloudless pixels to obtain the information availability of the candidate window under the reference image of each time phase.

[0011] Preferably, the calculation of surface homogeneity specifically involves: In the reference images of each time phase, the spectral values ​​of each cloudless pixel in each candidate window in the preset spectral band are extracted to form a spectral vector. The mean vector of all spectral vectors obtained in the candidate window is calculated. The average Euclidean distance from the spectral vector of all cloudless pixels in the candidate window to the mean vector is calculated and substituted into the negative exponential function to obtain the surface uniformity of the candidate window under the reference images of each time phase.

[0012] Preferably, the specific process for obtaining the fusion weights for each time phase is as follows: Calculate the time interval between the acquisition time of the remote sensing image to be repaired and the acquisition time of each temporal reference image, and use the negative correlation mapping result of the time interval as the temporal proximity weight of each temporal reference image. Calculate the average of the mean vectors of all temporal reference images, and use it as the global mean vector; Calculate the Euclidean distance between the mean vector of each temporal reference image and the global mean vector, and substitute the Euclidean distance into a negative exponential function to obtain the spectral consistency weight of each temporal reference image; The fusion weights for each time phase are obtained by multiplying the temporal proximity weights of the reference images for each time phase by the corresponding spectral consistency weights.

[0013] Preferably, the optimal search window is the candidate window with the highest final quality score.

[0014] Secondly, embodiments of this application also provide an automated cleaning system for large-scale remote sensing samples, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of any of the methods described above.

[0015] The beneficial effects of the above scheme are as follows: This application introduces a digital elevation model (DEM) to extract the topographic feature vector of pixels and performs clustering. Based on this, a normalized cloud index is calculated within each cluster for adaptive threshold segmentation. This effectively eliminates the interference of topographic factors such as elevation and slope on spectral features, avoids false detections and false negatives caused by shadows or bright surfaces under complex terrain, and thus significantly improves the recognition accuracy of cloud masks.

[0016] Furthermore, candidate windows of different sizes are generated centered on each pixel to be repaired, and their spatial distribution characteristics (information availability) and spectral discrete characteristics (surface homogeneity) are quantified respectively. By combining the window size to perform weighted fusion of multidimensional quality indicators, this mechanism effectively solves the contradiction of traditional fixed windows in complex scenes—it overcomes the problem of insufficient effective information in small windows and avoids the fusion distortion caused by introducing heterogeneous pixels in large windows.

[0017] Finally, taking into account the temporal proximity and spectral consistency of the reference images, a weighted average of the initial quality scores for each temporal phase is calculated to adaptively select the optimal search window. This strategy significantly improves cloud area restoration in complex terrain and areas with drastic surface changes, ensuring data integrity and spatiotemporal consistency after large-scale remote sensing sample cleaning. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the steps of an automated cleaning method for large-scale remote sensing samples, provided as an embodiment of this application. Detailed Implementation

[0019] The following, with reference to the accompanying drawings, details a specific scheme for an automated cleaning method and system for large-scale remote sensing samples provided in this application. Please refer to the accompanying drawings. Figure 1 The diagram illustrates a flowchart of an automated cleaning method for large-scale remote sensing samples according to an embodiment of this application. The method includes the following steps: Step 1: Acquire remote sensing images, multi-temporal reference images, and corresponding digital elevation model data for the target area.

[0020] Remote sensing images of the target area are acquired periodically (once every 3 days in this embodiment) using a spaceborne optical satellite. For each time period, images with cloud cover below a preset percentage (10% in this embodiment) within a previously preset time period (set to 3 months) are selected as reference images for the corresponding time period.

[0021] Obtain a digital elevation model (DEM) covering the target remote sensing image area. Extract blue and red band reflectance from the target area remote sensing image to calculate the cloud index, and download it through a ground receiving station or cloud platform. After radiometric calibration, atmospheric correction, and geometric registration, ensure that the data spatial alignment and resolution are consistent.

[0022] Step 2: Extract the terrain feature vector of the pixels using digital elevation model data, cluster the pixels, calculate the normalized cloud index within each cluster for adaptive threshold segmentation, and determine the cloudless pixels and the pixels to be repaired.

[0023] In the automated cleaning process for large-scale remote sensing samples, cloud masks are typically obtained by detecting clouds using the cloud index. The core assumption of the cloud index is that clouds strongly reflect blue light, while the Earth's surface reflects blue light weakly. The expression for the normalized cloud index is: In the formula, The normalized cloud index represents the pixel. , These represent the reflectance of pixels in the blue and red bands, respectively. As can be seen from the formula, the normalized cloud index is related to the reflectance in the blue and red bands. However, this difference is not only reflected in clouds and the land surface, but also in different terrain variations.

[0024] As topography changes, the reflectivity of the underlying surface of clouds also changes, leading to significant differences in reflectivity between cloud-polluted areas and the land surface due to different land surface types. For example, as the land cover type transitions from vegetation / water bodies to bare rock / snow, the reflectivity of the underlying surface (land surface) increases significantly in the blue and near-infrared bands, resulting in a smaller difference in reflectivity between clouds and the land surface. Furthermore, slope and aspect alter the solar incidence angle and the angle at which sensors observe the ground, causing significant differences in apparent reflectivity for the same feature on sunny and shady slopes. Therefore, using a globally fixed cloud index threshold and ignoring the influence of topography on the cloud index may result in cloud masks with a certain degree of topographic interference.

[0025] To achieve adaptive and accurate segmentation of cloud-contaminated areas in remote sensing images under complex terrain, this application first acquires the topographic data of the target area remote sensing image pixel by pixel.

[0026] The digital elevation model (DEM) of the target area remote sensing image is projected onto the same coordinate system as the remote sensing image and resampled to the same pixel size as the remote sensing image; the elevation of pixel i is extracted from the processed DEM. ,slope and slope aspect Align the calculated slope raster with the elevation raster pixel by pixel, and then adjust the slope aspect. Convert to the sine value of slope sin( ) and aspect cosine value cos( This yields the four-dimensional terrain feature vector for each pixel i. This is used to describe the topographic factors at different locations in the remote sensing image of the target area. Elevation, slope, aspect sine, and aspect cosine are standardized using Z-score. A small positive number of 0.01 is added to the denominator of the Z-score standard deviation to prevent the denominator from being zero, thus unifying the dimensions and obtaining the topographically standardized feature vector for each pixel. .

[0027] Based on these standardized terrain feature vectors, specific classification labels can be labeled for pixel terrain, thereby obtaining the dominant terrain type of a local area based on the pixel terrain label. Then, based on the dominant terrain type of the local area, adaptive threshold segmentation is performed on each area to obtain cloud masks on the remote sensing image.

[0028] The terrain-normalized feature vector of pixel i in the remote sensing image after normalization K-means clustering was employed (in this embodiment, K=9 was set to balance the detail of terrain classification with the sufficiency of samples in histogram statistics within a single category. In practice, the K value can be adaptively adjusted according to the actual terrain undulation and image size of the target area). The maximum number of iterations was set to 300 for iterative K-means clustering. After clustering, the category label of each pixel was recorded; the category labels included: Category 1, Category 2, Category 3, ... Category 9; It should be noted that the category labels are only used to distinguish different terrain feature clusters and there is no specific order requirement.

[0029] The normalized cloud index within each pixel is calculated using the normalized cloud index formula. Calculate the grayscale distribution histogram of the cloud index for each category; obtain the corresponding segmentation threshold for the grayscale histogram of each category using the Otsu method.

[0030] Based on the segmentation threshold corresponding to the cluster where pixel i belongs Cloud contamination detection is performed on remote sensing images. For pixels... Normalized cloud index ,like If it is, then it is determined to be a cloud-polluted pixel; otherwise... These are cloudless pixels. Through pixel-by-pixel comparison and labeling across the entire image range, morphological closure operations are performed using structuring elements of a preset size (such as a 3×3 or 5×5 rectangular kernel; in this embodiment, it is set to 3×3) to eliminate internal holes, resulting in a binarized cloud mask image. All pixels within the cloud mask area are then directly labeled as pixels to be repaired. Then proceed to subsequent cloud pollution pixel repair.

[0031] Step 3: Centering on each pixel to be repaired, preset candidate windows of different sizes; for each temporal reference image, analyze the spatial distribution characteristics of cloudless pixels and the discrete characteristics of spectral vectors within each candidate window, and combine the size of the candidate window to obtain the initial quality score of each candidate window in each temporal phase.

[0032] In the process of cloud restoration using the Neighboring Similar Pixel Interpolation (NSPI) algorithm, a neighborhood window needs to be set with the pixel to be restored as the center. Within this window range of the historical reference image, cloud-free pixels with similar spectral characteristics are searched, and the pixel values ​​of these similar pixels are used to interpolate the pixel to be restored.

[0033] When setting the neighborhood search window, two factors need to be considered: First, information availability, i.e., the number of cloudless pixels within the neighborhood window of the center to be repaired; second, the uniformity of the land surface. A more uniform land surface indicates a more homogeneous type of vegetation within the window and less spectral fluctuation. In this case, the probability of similar pixels found within the window being mixed with heterogeneous vegetation (such as vegetation found from water bodies) is lower, resulting in higher repair quality. The neighborhood window is scored based on these two factors: higher information availability and higher land surface uniformity, richer neighboring samples, and more similar characteristics among them. Abundant neighboring samples and the high consistency of the samples themselves together ensure the reliability of similar pixel searches. Under these conditions, the similarity quality between similar pixels found through the neighborhood window and the pixel to be repaired is higher.

[0034] Information availability and surface homogeneity differ depending on the size of the neighborhood search window. Small windows contain fewer effective pixels and have lower information availability, but the surface coverage within the window is relatively uniform, resulting in higher homogeneity. Large windows can provide more effective pixels and improve information availability, but they may include multiple land cover types, leading to an increase in spectral standard deviation and a decrease in homogeneity.

[0035] Therefore, for all cloud pollution pixels to be repaired, this application combines the availability of information for windows of different sizes and the uniformity of the land surface for windows of different sizes to conduct an initial quality score assessment for windows of different sizes, specifically: To select the optimal search window size for different repair centers, it is first necessary to preset a series of search windows of different sizes: For remote sensing images that need to be repaired Extract the reference image of the k-th time phase. Based on the registered coordinate system and spatial projection, the spatial position mapping and positioning of the pixels to be repaired is performed to obtain the location in the reference image. Mid-position of pixels to be repaired The location.

[0036] For each pixel to be repaired ,by Centered on the target pixel, a series of square neighborhood windows are generated according to the preset initial side length (set to 5 in this embodiment), maximum side length (set to 17 in this embodiment), and incremental step size (set to 2 to ensure complete symmetry of the central pixel), forming a set of candidate window sizes. , This represents the side length of the m-th candidate window.

[0037] To evaluate the quality of different candidate window sizes, it is necessary to analyze the information availability of each candidate window size: information availability is first reflected in the proportion of cloudless pixels in each candidate window size. The higher the proportion of cloudless pixels, the richer the sample size of similar pixels, and the more reliable the quality of the obtained similar pixels.

[0038] For reference image Cloud detection was performed on the reference image using the aforementioned adaptive threshold segmentation method combined with terrain, resulting in a cloud mask. Pixels within the cloud mask were then marked as cloud-covered pixels. The remaining pixels are marked as cloudless pixels. .

[0039] Pixels awaiting repair Centered on the reference image at the k-th time phase, the statistical size is... Cloudless pixels in the candidate window Number of The percentage of cloudless pixels in the window is calculated using the following expression: In the formula, This indicates the number of cloudless pixels within the candidate window. This represents the actual total number of pixels within the image area of ​​the window, typically 1. When the image is at the edge, the number of pixels is counted according to the actual number of truncated pixels.

[0040] In addition, information availability is also reflected in the distance between cloudless pixels and the target pixel. The closer a cloudless pixel is to the target pixel, the more similar its spectral information is to the surface attributes of the target pixel, indicating that these cloudless pixels have higher availability. For the k-th phase size is The candidate window will contain the pixels to be repaired. Using the corresponding pixel in the temporal reference image k as the reference pixel, all cloud-free pixels are counted. Euclidean distance from the center of the reference pixel to the pixel center Calculate all cloudless pixels of The average distance, divided by the maximum theoretical distance from the center of the window to the corner point ( ), to obtain the normalized average distance It should be noted that if If so, the information availability score of the window of that size in phase k is directly assigned to 0.

[0041] Because there are no cloud pixels in the window Normalized mean distance from the center to be repaired The richer the information and the closer it is to the center, the greater its usability. Merging these elements yields the pixels to be repaired. In the k-th time phase Information availability within the neighborhood search window The specific formula is as follows:

[0042] In the formula, This represents the percentage of cloudless pixels in time phase k. This represents the normalized average distance from the cloudless pixel in phase k to the restoration center.

[0043] The higher the homogeneity of the surface within the neighborhood window of the pixel to be repaired, the higher the window quality. Higher surface homogeneity indicates smaller spectral variations and simpler textures of the objects within the window. During cloud removal, the available neighboring similar pixels are purer and more consistent, and the fusion repair result is less likely to be mixed with information from heterogeneous objects. Therefore, it is necessary to evaluate the homogeneity of the surface within the window.

[0044] In the reference image of the k-th time phase, with the reference pixel at the corresponding position as the center, an image of size [size missing] is extracted. The candidate window contains the spectral values ​​of all cloudless pixels in preset spectral bands, including but not limited to: blue band (B), green band (G), red band (R), near-infrared band (NIR), shortwave infrared 1 (SWIR1), and shortwave infrared 2 (SWIR2). If the number of extracted bands is n, then the spectral values ​​of each cloudless pixel i within the window form an n-dimensional spectral vector. .

[0045] Within the window of the k-th time phase, assume there are a total of For each cloudless pixel, calculate the spectral vector of that pixel. The mean vector is calculated. The average Euclidean distance from the spectral vectors of all cloudless pixels within the window to the mean vector is then calculated. This average Euclidean distance is then converted to surface homogeneity using a negative exponential function. The specific formula is as follows: In the formula, σ is a preset adjustment constant (in this embodiment, the average value of the local reflectance variance of each band is taken, and a very small positive number of 0.001 is added to prevent the denominator from being 0); exp() represents an exponential function with the natural constant as the base. This formula indicates that the smaller the spectral deviation of cloudless pixels from the average value, the smaller the distance, and the closer the surface homogeneity is to 1, indicating that the land cover within the window is purer and more uniform. It should be noted that if the number of cloudless pixels in the candidate window... Then the candidate window will be in the time phase Surface homogeneity under reference image The value is directly assigned to 0.

[0046] When the window is small, it is more likely to lack useful information; when the window is large, it is more likely to introduce heterogeneous pixels. Therefore, information availability has a greater weight when the window is small, while uniformity has a greater weight when the window is large. Information availability and surface uniformity are fused based on window size to obtain initial quality scores for windows of different sizes.

[0047] For the candidate window size of the k-th time phase Normalize it by its maximum and minimum values ​​to obtain the window size coefficient. The larger this coefficient value, the larger the current calculation window size. When the window is small, information availability needs to be considered. Weight 1- At the same time, when the window is larger, more weight needs to be given to surface uniformity, and the surface uniformity should be set accordingly. Weight is For each repaired pixel Its different window sizes Next, at time phase k, and The weighted fusion yields the initial quality score as follows: : .

[0048] Since the initial quality score is analyzed only for a single time phase and window size, and the land surface may change in different time phases (e.g., vegetation growth, crop harvesting), the uniformity of the land surface and cloud cover will vary, causing the initial quality score to fluctuate accordingly. Therefore, each window size corresponds to a different initial quality score for each time phase.

[0049] Step 4: Based on the temporal proximity of the acquisition time of each phase reference image and the spectral consistency of the spectral features, obtain the fusion weight of each phase, perform a weighted average of the initial quality scores of each phase, obtain the final quality score of each candidate window, and obtain the optimal search window.

[0050] The reference value of reference images from different time periods for setting the window size of the remote sensing image to be restored also varies. The longer the time interval, the greater the possibility of changes in ground features, and the more likely the spectral characteristics of cloudless pixels in that time period are inconsistent with the actual ground surface at the current restoration time. The smaller the standard deviation of the spectral mean, the more consistent the pixel values ​​within that window are across different time periods, and the higher the reference value.

[0051] Therefore, the initial quality scores of windows at each time phase can be fused for a comprehensive evaluation of window quality. Considering the proximity of each time phase to the current time, the closer the phase is to the current time, the smaller the spectral fluctuation within the window at each time phase, and the higher the weight. This yields the final weight of the fused windows of different sizes for each center to be repaired, thus selecting the optimal search window size for each center to be repaired. Specifically: First, to quantify the reference value of images from different time phases, their fusion weights need to be quantified. The longer the time interval, the greater the likelihood of changes in ground features, and the more likely the spectral characteristics of cloudless pixels at that time phase are inconsistent with the actual ground surface at the current restoration moment. The temporal proximity weight can be obtained by considering the time interval between the reference image and the current image. Obtain the current image to be repaired The acquisition time is denoted as Reference image of the k-th time phase The acquisition time is denoted as Then the time approximation weight of the k-th phase is... ;in, The time decay constant is used in this embodiment. For each day, the time weight of the reference image k is increased; the shorter the time interval, the greater the weight.

[0052] Reference image at the k-th time phase If the number of cloudless pixels in the candidate window is 0, then that time phase is skipped, not included in the calculation of the global mean vector, and its spectral consistency weight is directly assigned to 0. Simultaneously, the total number of valid time phases K participating in the calculation is reduced by 1. If the number of cloudless pixels is greater than 0, then the pixels to be repaired are extracted. The neighborhood search window size is mean vector at time Calculate the spectral consistency weights for each time phase: In the formula, Let K represent the mean vector of the spectral vectors of all pixels within the k-th temporal window, where K represents the total number of valid temporal phases. Represents the computation between vectors The norm is used to quantify the true Euclidean distance in the multidimensional feature space between the spectral mean of the current k-th time phase and the aforementioned global average spectral benchmark. The larger the distance value, the more drastically the spectrum of that time phase deviates from the normal level and the more interference it contains.

[0053] Multiplying the time proximity weight and the spectral consistency weight yields the size of the k-th temporal window. Initial quality score fusion weights The specific formula is as follows: .

[0054] The final window quality score is obtained by weighted fusion of the initial quality scores of each phase. Furthermore, the optimal window for each center to be repaired is obtained through the comprehensive quality score: for any candidate window size... Obtain its initial quality score at each time phase. The initial quality scores from each time phase are merged to obtain the final window quality score. The specific formula is as follows:

[0055] In the formula, This represents the initial quality score for this window at the k-th time phase. Let k be the fusion weight for time phases, where K represents the total number of time phases.

[0056] Treating the repaired pixels Compare the sizes of all candidate windows Final score Take the window size corresponding to the maximum score. If multiple windows have the same size score, the smaller window should be selected first to reduce the amount of computation.

[0057] Step 5: Based on the cloudless pixels within the optimal search window, use the nearest similar pixel interpolation algorithm to interpolate and repair the pixel values ​​of the pixels to be repaired.

[0058] Perform NSPI cloud repair based on the optimal window: First, the pixel to be repaired... The optimal search window obtained in the previous step As the neighborhood boundary, the reference images of each time phase are traversed sequentially. Within the optimal window of each time phase, the pixel at the corresponding position of the pixel to be repaired in the reference image of that time phase is obtained as the reference pixel. Then, the Euclidean distance between each cloudless pixel and the reference pixel in all bands is calculated. The smaller the distance, the more similar the spectral features. Combining the obtained time proximity weight, the top number of cloudless pixels with the highest comprehensive similarity after time weighting are selected. In this embodiment, the preset number is 5. If there are fewer than 5 cloudless pixels in the window, the process falls back to the suboptimal window (i.e., the window size with the second highest final score) and the number of cloudless pixels is rechecked until it meets ≥ 5. If the number of cloudless pixels is still less than 5 after traversing all candidate window sizes, the search is exited, the pixel to be repaired is marked as unrepaired, and the next pixel to be repaired is processed. Based on the selected similar pixels, the NSPI algorithm is used to establish a residual fitting model to calculate the interpolated pixel value of the pixel to be repaired at the current time, thus completing the cloud pollution area repair. Finally, all pixels to be repaired are processed sequentially, and the complete remote sensing image after automated cleaning is output.

[0059] Based on the same inventive concept as the above methods, this application also provides an automated cleaning system for large-scale remote sensing samples, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described automated cleaning methods for large-scale remote sensing samples.

Claims

1. An automated cleaning method for large-scale remote sensing samples, characterized in that, The method includes the following steps: Acquire remote sensing images, multi-temporal reference images, and corresponding digital elevation model data of the target area; The terrain feature vectors of pixels are extracted using digital elevation model data. The pixels are clustered, and the normalized cloud index is calculated within each cluster for adaptive threshold segmentation to determine cloudless pixels and pixels to be repaired. Centered on each pixel to be repaired, candidate windows of different sizes are preset; for each temporal reference image, the spatial distribution characteristics of cloudless pixels and the discrete characteristics of spectral vectors within each candidate window are analyzed, and the initial quality score of each candidate window in each temporal phase is obtained by combining the size of the candidate window. Based on the temporal proximity weight of the acquisition time of each temporal reference image and the spectral consistency weight of the spectral features, the fusion weight of each temporal phase is obtained. The initial quality scores of each temporal phase are weighted and averaged to obtain the final quality score of each candidate window, and the optimal search window is obtained. Based on cloudless pixels within the optimal search window, the pixel value interpolation and repair of the pixels to be repaired is performed using the nearest similar pixel interpolation algorithm. The process of obtaining the initial quality score of each candidate window at each time phase is as follows: Analyze the proportion of cloudless pixels in each candidate window and the spatial distance between them and the pixels to be repaired to calculate information availability; analyze the deviation of the spectral vector of cloudless pixels in each candidate window from its mean vector to calculate surface homogeneity. Based on the size of the candidate window, information availability and surface uniformity are weighted and fused to obtain the initial quality score of each candidate window in each time phase. The specific process for obtaining the fusion weights for each time phase is as follows: In the reference images of each time phase, the spectral values ​​of each cloud-free pixel in each candidate window in the preset spectral band are extracted to form a spectral vector, and the mean vector of all spectral vectors obtained in the candidate window is calculated. Calculate the time interval between the acquisition time of the remote sensing image to be repaired and the acquisition time of each temporal reference image, and use the negative correlation mapping result of the time interval as the temporal proximity weight of each temporal reference image. Calculate the average of the mean vectors of all temporal reference images, and use it as the global mean vector; Calculate the Euclidean distance between the mean vector of each temporal reference image and the global mean vector, and substitute the Euclidean distance into a negative exponential function to obtain the spectral consistency weight of each temporal reference image; The fusion weights for each time phase are obtained by multiplying the temporal proximity weights of the reference images for each time phase by the corresponding spectral consistency weights.

2. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The terrain feature vector includes the elevation, slope, aspect sine, and aspect cosine of each pixel.

3. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The specific calculation process for the normalized cloud index is as follows: The sum and difference between the reflectance of each pixel in the blue band and the reflectance in the red band are calculated separately. The negative correlation mapping result of the obtained sum and the obtained difference are positively fused to obtain the normalized cloud index.

4. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The specific process for determining cloud-free pixels and pixels to be repaired is as follows: A threshold segmentation algorithm is applied to the normalized cloud index of all pixels in each cluster to obtain a segmentation threshold. The normalized cloud index of pixels in the remote sensing image to be repaired and the reference images of each time phase is compared with the segmentation threshold of the cluster in which the pixel belongs. If the normalized cloud index is greater than or equal to the segmentation threshold, it is determined to be a cloud-polluted pixel; otherwise, it is determined to be a cloudless pixel. The cloud-polluted pixels in the remote sensing image to be repaired are marked as pixels to be repaired.

5. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The availability of the computational information is specifically as follows: Under the reference images of each time phase, the number of cloudless pixels in the candidate window is counted, and the proportion of cloudless pixels is calculated. The corresponding pixel in each temporal reference image of the pixel to be repaired is taken as the reference pixel. The average Euclidean distance from each cloudless pixel in the candidate window to the reference pixel is calculated and normalized to obtain the normalized average distance. The negative correlation mapping result of the normalized average distance is positively fused with the proportion of cloudless pixels to obtain the information availability of the candidate window under the reference image of each time phase.

6. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The calculation of surface homogeneity specifically involves: Calculate the average Euclidean distance from the spectral vector of all cloudless pixels within the candidate window to the mean vector, and substitute it into the negative exponential function to obtain the surface homogeneity of the candidate window under the reference images of each time phase.

7. The automated cleaning method for large-scale remote sensing samples as described in claim 1, characterized in that, The optimal search window is specifically the candidate window with the highest final quality score.

8. An automated cleaning system for large-scale remote sensing samples, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-7.

Citation Information

Patent Citations

  • Urban green land coverage rate dynamic monitoring method and device

    CN120071159A

  • Ice and snow identification method and system based on remote sensing image

    CN122289940A