Cloud Detection Method in Remote Sensing Images Based on Region Growing

Abstract

The invention discloses a remote sensing image cloud detection method based on region growth, which belongs to radar remote sensing or image processing technology and mainly solves the problems of low detection rate, high false alarm rate and holes and breaks in detection results when remote sensing image cloud detection lacks thermal infrared frequency band data. The implementation steps are as follows: determining a remote sensing image to be input and first performing superpixel processing; then establishing a multi-resolution model; performing multiplication and subtraction operations on superpixel patches with the same number in different images of superpixelized remote sensing images of different resolutions, and normalizing the results; using an outlier detection method to respectively detect original significant superpixel blocks and original significant region superpixel blocks; and detecting cloud targets using a weighted similarity method based on superpixel blocks. The invention can realize the detection of cloud areas in high-resolution remote sensing images when there is a lack of thermal infrared frequency band data, and uses superpixels as detection units to avoid holes, breaks and the like in the detection results, and filters out false alarms by using the weighted similarity between significant superpixel blocks and significant region superpixel blocks, thereby effectively improving the detection rate of cloud target detection in remote sensing images and reducing the false alarm rate of detection.

Description

Technical Field

[0001] The present invention belongs to radar remote sensing or image processing technology, that is, using image processing technology to analyze radar observation information, and specifically involves the application of superpixel processing, multi-resolution model, and weighted similarity method in remote sensing image cloud target detection. Background Art

[0002] With the rapid development of remote sensing image acquisition technology, the application of remote sensing image data has become increasingly widespread, encompassing areas such as geographic surveying and mapping, environmental monitoring, and reconnaissance monitoring. However, this rapid development also presents challenges to remote sensing image data interpretation, such as the dramatic increase in data volume and the increasing resolution of images. When interpreting large amounts of remote sensing imagery, cloud occlusion often impacts image interpretation. For example, the presence of cloud can severely impact interpretation results in processes such as ground target detection and object classification. This is particularly true for high-resolution remote sensing image interpretation, where higher resolution provides richer detail in object information. This richness in cloud detail also creates greater interference with interpretation. Therefore, accurately detecting cloud regions within large amounts of high-resolution remote sensing image data is crucial for remote sensing image interpretation.

[0003] At present, the most mature cloud region detection technology abroad is the automatic cloud cover assessment (ACCA) method based on U.S. Earth Resources Satellite data. This method relies heavily on the thermal infrared frequency band in U.S. Earth Resources Satellite data. However, most domestic high-resolution satellite sensors do not have this frequency band, which limits the application of this method. In addition, in the case of high resolution, the cloud region information in the image data is richer and the fluctuation of local information is more dramatic. The cloud region detection method based on pixel points will cause holes and breaks to appear in the detection results, resulting in problems such as missed detection and increased false alarm rate in the detection results. In response to the above problems, the present invention proposes a method for cloud detection in high-resolution remote sensing images based on superpixel region growth. This method can detect cloud regions in high-resolution remote sensing images in the absence of thermal infrared frequency band data, and uses superpixels as detection units to avoid holes and breaks in the detection results. False alarms are filtered out by weighted similarity between significant superpixel blocks and significant region superpixel blocks, effectively improving the detection rate of cloud target detection in remote sensing images and reducing the false alarm rate of detection. Summary of the Invention

[0004] The detailed technical solutions of the present invention are as follows:

[0005] A cloud detection method for remote sensing images based on region growing, such as Figure 1 As shown, the following steps are included:

[0006] Step 1: Superpixel the original remote sensing image containing clouds. Use the optimized SLIC (Simple Linear Iterative Clustering) algorithm to perform superpixel processing on the original remote sensing image to obtain a remote sensing image I containing k superpixel blocks. Each superpixel is numbered from 1 to k from left to right and from top to bottom.

[0007] Step 2: Obtain cloud target saliency map and cloud area positioning saliency map based on the information extraction model for the super-pixelated remote sensing image;

[0008] 2.1: Establish a multi-resolution model based on superpixels for the remote sensing image obtained in step 1. The specific method is as follows: first, perform SVD decomposition on the remote sensing image I to obtain a diagonal matrix Σ and two unitary matrices U and V; then sort the diagonal elements in the diagonal matrix Σ from large to small, and calculate the mean of the non-zero diagonal elements in the diagonal matrix; secondly, filter out the diagonal elements in the diagonal matrix that are less than the mean of the diagonal elements to obtain a new diagonal matrix; finally, combine the new diagonal matrix with the two unitary matrices to obtain a lower resolution image I1; repeat the above steps to obtain remote sensing images I with different resolutions i , the value range of variable i is [1,m], and variable m corresponds to remote sensing image I m , the number of non-zero elements in its diagonal matrix is ​​1;

[0009] 2.2: Perform multiplication of the superpixel patches with the same number in different images on the m superpixel remote sensing images of different resolutions obtained in step 2.1, and normalize the results to finally obtain k superpixel patches, which are then used to form the cloud target saliency map I in the order of their numbers. TR , the method is shown in formula (1);

[0010]

[0011] Among them, variable I i (g) represents the superpixel block numbered g in the i-th remote sensing image, and the value of the variable g is [1, k]; the symbol Φ k (·) means that k superpixel blocks are restored to an image in the order in which they are numbered when the superpixels are generated;

[0012] 2.3: Using outlier detection method to detect cloud target saliency map I TR The significance mean of the k super-pixel blocks in the image is detected to obtain the significant super-pixel block SP(f), f∈[1,k], and the number of the super-pixel block SP(f) is mapped to the remote sensing image I to obtain the original significant super-pixel block ISP(f) in I;

[0013] 2.4: Perform inter-layer subtraction operation on the super-pixel patches of the same number in different images of the m super-pixelized remote sensing images with different resolutions obtained in step 2.1. The specific method is as follows: first, select the pth layer as the center, s as the expansion distance, select the qth layer as the edge, and the variable s=2,3,q=p±s;Secondly, the same numbered superpixel blocks in the p layer and its corresponding q layer image are subtracted, and the results of the subtraction operation are summed and normalized to finally obtain k superpixel blocks, which are then used to form the cloud area localization saliency map I in the order of numbers. CP , the method is shown in formula (2):

[0014]

[0015] 2.5: Computing Cloud Area Localization Saliency Map I CP The saliency mean of each superpixel block in the image is obtained, and all saliency means are detected using the outlier detection method to find the cloud area positioning saliency map I CP The salient area in the image is obtained, and the super pixel block SR(h) in the salient area is obtained, h∈[1,k], and the number of the super pixel block SR(h) is mapped to the remote sensing image I to obtain the original salient area super pixel block ISR(h) in I.

[0016] Step 3: Perform cloud target detection based on superpixel block weighted similarity on the original salient superpixel block ISP(f) and the original salient region superpixel block ISR(h) obtained in step 2.

[0017] 3.1: Take the original salient superpixel block ISP(f) obtained in step 2 as the central block and the original salient region superpixel block ISR(h) as the peripheral block; traverse the central block and calculate the relative entropy D between each central block ISP(f) and its peripheral block ISR(h) within the search range R. K (f,h), the method is shown in formula (3):

[0018]

[0019] Among them, the variable P ISP(f) represents the probability distribution model of the central block ISP(f), the variable P ISR(h) represents the probability distribution model of the surrounding block ISR(h); the search range R is a rectangular area with ISP(f) as the center and 2r superpixel blocks as the side length; Where Z is the cloud area positioning saliency map I CP The maximum size of the connected salient region, S is the superpixel block size, and α is a constant, which is 1.2 here;

[0020] 3.2: Take the centroid C of the central block ISP(f) ISP(f)and the centroid C of the surrounding block ISR(h) ISR(h) The Euclidean distance between each central block ISP(f) and its surrounding blocks ISR(h) within the search range R is obtained by S (f,h), the method is shown in formula (4):

[0021]

[0022] in, are the coordinates of the center of mass of the central block and the peripheral blocks in the image respectively;

[0023] 3.3: Through relative entropy D K (f,h) and relative distance D S (f, h) is used to calculate the weighted similarity D(f, h) between the central block and the surrounding blocks. The method is shown in formula (5):

[0024]

[0025] Among them, the variable M A is the maximum value of the maximum relative entropy of all central blocks ISP(f) and the background area B in image I. The background area B is defined as the other areas in image I excluding the central block and peripheral blocks; the variable M RM is the maximum value of the maximum Euclidean distance between the centroid of all center blocks ISP(f) and the centroid of the edge superpixels within their corresponding search range R;

[0026] 3.4: Take the average relative entropy D between all central blocks ISP(f) KT The maximum Euclidean distance D corresponding to each central block ST (f) is used to calculate the weighted similarity threshold T(f), as shown in formula (6):

[0027]

[0028] 3.5: By comparing the weighted similarity D(f,h) between the f-th central block and its corresponding h-th surrounding block with the weighted similarity threshold T(f) corresponding to the central block, if D(f,h) is less than or equal to T(f), the surrounding block ISR(h) is judged as a cloud target and retained. If D(f,h) is greater than T(f), the surrounding block ISR(h) is judged as a non-cloud target and removed. After traversing all central blocks, if a surrounding block is retained and removed for different central blocks, it is retained. Finally, the central block and the retained surrounding blocks are combined to obtain the target saliency map, and the location of the cloud target is detected. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of the remote sensing image cloud detection method based on region growing of the present invention; DETAILED DESCRIPTION

[0030] The specific implementation of the present invention is as follows:

[0031] A cloud detection method for remote sensing images based on region growing, such as Figure 1 As shown, the following steps are included:

[0032] The optimized SLIC (Simple Linear Iterative Clustering) algorithm is used to perform superpixel processing on the original remote sensing image to obtain a remote sensing image I containing k superpixel blocks. Each superpixel is numbered from 1 to k from left to right and from top to bottom.

[0033] Step 2: Obtain cloud target saliency map and cloud area positioning saliency map based on the information extraction model for the super-pixelated remote sensing image;

[0034] 2.1: Establish a multi-resolution model based on superpixels for the remote sensing image obtained in step 1. The specific method is as follows: first, perform SVD decomposition on the remote sensing image I to obtain a diagonal matrix Σ and two unitary matrices U and V; then sort the diagonal elements in the diagonal matrix Σ from large to small, and calculate the mean of the non-zero diagonal elements in the diagonal matrix; secondly, filter out the diagonal elements in the diagonal matrix that are less than the mean of the diagonal elements to obtain a new diagonal matrix; finally, combine the new diagonal matrix with the two unitary matrices to obtain a lower resolution image I1; repeat the above steps to obtain remote sensing images I with different resolutions i , the value range of variable i is [1,m], and variable m corresponds to remote sensing image I m , the number of non-zero elements in its diagonal matrix is ​​1;

[0035] 2.2: Perform multiplication of the superpixel patches with the same number in different images on the m superpixel remote sensing images of different resolutions obtained in step 2.1, and normalize the results to finally obtain k superpixel patches, which are then used to form the cloud target saliency map I in the order of their numbers. TR , the method is shown in formula (1);

[0036]

[0037] Among them, variable I i (g) represents the superpixel block numbered g in the i-th remote sensing image, and the value of the variable g is [1, k]; the symbol Φ k (·) means that k superpixel blocks are restored to an image in the order in which they are numbered when the superpixels are generated;

[0038] 2.3: Using outlier detection method to detect cloud target saliency map I TRThe significance mean of the k super-pixel blocks in the image is detected to obtain the significant super-pixel block SP(f), f∈[1,k], and the number of the super-pixel block SP(f) is mapped to the remote sensing image I to obtain the original significant super-pixel block ISP(f) in I;

[0039] 2.4: Perform inter-layer subtraction operation on the super-pixel patches of the same number in different images of the m super-pixelized remote sensing images with different resolutions obtained in step 2.1. The specific method is as follows: first, select the pth layer as the center, s as the expansion distance, select the qth layer as the edge, and the variable s=2,3,q=p±s;Secondly, the same numbered superpixel blocks in the p layer and its corresponding q layer image are subtracted, and the results of the subtraction operation are summed and normalized to finally obtain k superpixel blocks, which are then used to form the cloud area localization saliency map I in the order of numbers. CP , the method is shown in formula (2):

[0040] 2.5: Computing Cloud Area Localization Saliency Map I CP The saliency mean of each superpixel block in the image is obtained, and all saliency means are detected using the outlier detection method to find the cloud area positioning saliency map I CP The salient region in the image is obtained, and the super pixel block SR(h) in the salient region is obtained, h∈[1,k], and the number of the super pixel block SR(h) is mapped to the remote sensing image I to obtain the original salient region super pixel block ISR(h) in I;

[0041] Step 3: Perform cloud target detection based on superpixel block weighted similarity on the original salient superpixel block ISP(f) and the original salient region superpixel block ISR(h) obtained in step 2.

[0042] 3.1: Take the original salient superpixel block ISP(f) obtained in step 2 as the central block and the original salient region superpixel block ISR(h) as the peripheral block; traverse the central block and calculate the relative entropy D between each central block ISP(f) and its peripheral block ISR(h) within the search range R. K (f,h), the method is shown in formula (3):

[0043]

[0044] Among them, the variable P ISP ( f) represents the probability distribution model of the central block ISP(f), the variable P ISR ( h) represents the probability distribution model of the surrounding block ISR(h); the search range R is a rectangular area with ISP(f) as the center and 2r superpixel blocks as the side length; Where Z is the cloud area positioning saliency map I CP The maximum size of the connected salient region, S is the superpixel block size, and α is a constant, which is 1.2 here;

[0045] 3.2: Take the centroid C of the central block ISP(f) ISP ( f) and the centroid C of the surrounding block ISR(h) ISR ( h) The Euclidean distance between each central block ISP(f) and its surrounding blocks ISR(h) within the search range R is obtained by S (f,h), the method is shown in formula (4):

[0046]

[0047] in, are the coordinates of the center of mass of the central block and the peripheral blocks in the image respectively;

[0048] 3.3: Through relative entropy D K (f,h) and relative distance D S (f, h) is used to calculate the weighted similarity D(f, h) between the central block and the surrounding blocks. The method is shown in formula (5):

[0049]

[0050] Among them, the variable M A is the maximum value of the maximum relative entropy of all central blocks ISP(f) and the background area B in image I. The background area B is defined as the other areas in image I excluding the central block and peripheral blocks; the variable M RM is the maximum value of the maximum Euclidean distance between the centroid of all center blocks ISP(f) and the centroid of the edge superpixels within their corresponding search range R;

[0051] 3.4: Take the average relative entropy D between all central blocks ISP(f) KT The maximum Euclidean distance D corresponding to each central block ST (f) is used to calculate the weighted similarity threshold T(f), as shown in formula (6):

[0052]

[0053] 3.5: By comparing the weighted similarity D(f,h) between the f-th central block and its corresponding h-th surrounding block with the weighted similarity threshold T(f) corresponding to the central block, if D(f,h) is less than or equal to T(f), the surrounding block ISR(h) is judged as a cloud target and retained. If D(f,h) is greater than T(f), the surrounding block ISR(h) is judged as a non-cloud target and removed. After traversing all central blocks, an OR operation is performed on the judgment results. That is, if a surrounding block is retained and removed with respect to different central blocks, it is retained. Finally, the central block and the retained surrounding blocks are combined to obtain the target saliency map, and the location of the cloud target is detected.

Claims

1. A remote sensing image cloud detection method based on region growing includes the following steps: Step 1: Superpixelate the original remote sensing image containing clouds; The optimized SLIC (Simple Linear Iterative Clustering) algorithm is used to perform superpixel processing on the original remote sensing image to obtain a remote sensing image I containing k superpixel blocks. Each superpixel is numbered from 1 to k from left to right and from top to bottom. Step 2: Obtain cloud target saliency map and cloud area positioning saliency map based on the information extraction model for the super-pixelated remote sensing image; 2.1: Establish a multi-resolution model based on superpixels for the remote sensing image obtained in step 1. The specific method is as follows: first, perform SVD decomposition on the remote sensing image I to obtain a diagonal matrix Σ and two unitary matrices U and V; then sort the diagonal elements in the diagonal matrix Σ from large to small, and calculate the mean of the non-zero diagonal elements in the diagonal matrix; then, filter out the diagonal elements in the diagonal matrix that are less than the mean of the diagonal elements to obtain a new diagonal matrix; finally, combine the new diagonal matrix with the two unitary matrices to obtain a lower resolution image I1; repeat the above steps to obtain remote sensing images I with different resolutions i , the value range of variable i is [1,m]. When the variable i takes the value m, the corresponding remote sensing image is I m , at this time I m The number of non-zero elements in the corresponding diagonal matrix is ​​1; 2.2: Perform multiplication of the superpixel patches with the same number in different images on the m superpixel remote sensing images of different resolutions obtained in step 2.1, and normalize the results to finally obtain k superpixel patches, which are then used to form the cloud target saliency map I in the order of their numbers. TR , the method is shown in formula (1); Among them, variable I i (g) represents the superpixel block numbered g in the i-th remote sensing image, and the value of the variable g is [1, k]; the symbol Φ k (·) means that k superpixel blocks are restored to an image in the order in which they are numbered when the superpixels are generated; 2.3: Using outlier detection method to detect cloud target saliency map I TR The significance mean of the k super-pixel blocks in the image is detected to obtain the significant super-pixel block SP(f), f∈[1,k], and the number of the super-pixel block SP(f) is mapped to the remote sensing image I to obtain the original significant super-pixel block ISP(f) in I; 2.4: Perform inter-layer subtraction operation on the super-pixel patches of the same number in different images of the m super-pixelized remote sensing images with different resolutions obtained in step 2.

1. The specific method is as follows: first, select the pth layer as the center, s as the expansion distance, select the qth layer as the edge, and the variable s=2,3,q=p±s; Then, the same numbered superpixel blocks in the p-th layer and its corresponding q-th layer image are subtracted, and the results of the subtraction operation are summed and normalized to finally obtain k superpixel blocks, which are then used to form the cloud area positioning saliency map I in the order of numbers. CP , the method is shown in formula (2): 2.5: Computing Cloud Area Localization Saliency Map I CP The saliency mean of each superpixel block in the image is obtained, and all saliency means are detected using the outlier detection method to find the cloud area positioning saliency map I CP The salient area in the image is obtained, and the super pixel block SR(h) in the salient area is obtained, h∈[1,k], and the number of the super pixel block SR(h) is mapped to the remote sensing image I to obtain the original salient area super pixel block ISR(h) in I; Step 3: Perform cloud target detection based on superpixel block weighted similarity on the original salient superpixel block ISP(f) and the original salient region superpixel block ISR(h) obtained in step 2. 3.1: Take the original salient superpixel block ISP(f) obtained in step 2 as the central block and the original salient region superpixel block ISR(h) as the peripheral block; traverse the central block and calculate the relative entropy D between each central block ISP(f) and its peripheral block ISR(h) within the search range R. K (f,h), the method is shown in formula (3): in, variable P ISP(f) represents the probability distribution model of the central block ISP(f), the variable P ISR(h) represents the probability distribution model of the surrounding block ISR(h); the search range R is a rectangular area with ISP(f) as the center and 2r superpixel blocks as the side length; Where Z is the cloud area positioning saliency map I CP The maximum size of the connected salient region, S is the superpixel block size, and α is a constant; 3.2: Take the centroid C of the central block ISP(f) ISP ( f) and the centroid C of the surrounding block ISR(h) ISR ( h) The Euclidean distance between each central block ISP(f) and its surrounding blocks ISR(h) within the search range R is obtained by S (f,h), the method is shown in formula (4): in, are the coordinates of the center of mass of the central block and the peripheral blocks in the image respectively; 3.3: Through relative entropy D K (f,h) and relative distance D S (f, h) is used to calculate the weighted similarity D(f, h) between the central block and the surrounding blocks. The method is shown in formula (5): Among them, the variable M A is the maximum value of the maximum relative entropy of all central blocks ISP(f) and the background area B in image I. The background area B is defined as the other areas in image I excluding the central block and peripheral blocks; the variable M RM is the maximum value of the maximum Euclidean distance between the centroid of all center blocks ISP(f) and the centroid of the edge superpixels within their corresponding search range R; 3.4: The weighted similarity threshold T(f) is calculated using the average relative entropy between all center blocks ISP(f) and the maximum Euclidean distance corresponding to each center block. The method is shown in formula (6): 3.5: By comparing the weighted similarity D(f,h) between the f-th central block and its corresponding h-th surrounding block with the weighted similarity threshold T(f) corresponding to the central block, if D(f,h) is less than or equal to T(f), the surrounding block ISR(h) is judged as a cloud target and retained. If D(f,h) is greater than T(f), the surrounding block ISR(h) is judged as a non-cloud target and removed. After traversing all central blocks, if a surrounding block is retained and removed for different central blocks, it is retained. Finally, the central block and the retained surrounding blocks are combined to obtain the target saliency map, and the location of the cloud target is detected.

Images