Cloud layer detection algorithm based on combination of two strategies

Abstract

The invention discloses a cloud layer detection algorithm based on the combination of two strategies, and the algorithm comprises the steps: S1, obtaining a cloud-containing image, carrying out the fine segmentation of the image based on MST, and obtaining a segmentation pattern spot; s2, performing cloud detection on the segmented image spots by adopting an SVM image classification strategy; s3, performing cloud detection on the segmented pattern spots by adopting a spectrum multi-factor detection strategy; and S4, performing D-S decision fusion by using an SVM image classification cloud detection result and a spectrum multi-factor cloud detection result to obtain an optimal cloud-containing image cloud detection strategy. According to the method, D-S decision fusion is carried out on the SVM image classification cloud detection result and the spectrum multi-factor cloud detection result, the optimal cloud detection result of the cloud-containing image is obtained, the cloud detection accuracy of the cloud-containing image is high, and the effect is better.

Description

Technical Field

[0001] The present invention relates to the technical field of cloud detection, and in particular to a cloud detection algorithm based on a combination of two strategies. Background Art

[0002] Clouds appear differently on images, but they can be roughly divided into radiation features, geometric features, elevation features, and temporal features. Based on the above cloud features, there are currently many cloud detection technologies, mainly including: 1) Physical threshold method. The physical threshold method is the most common and simplest algorithm in image cloud detection technology. This type of method mainly uses the radiation characteristics of clouds to set thresholds, and the detection effect is good; the problem with the method is that the algorithm model is limited to specific scenes, and it is difficult to judge the existence of clouds in certain environments. For example, under conditions of snow and thin clouds, similar reflectivity makes it difficult to distinguish between clouds and ground objects. The early methods of the physical threshold method include the ISCCP method and the GTM method, but both rely on a priori threshold settings; in summary, the physical threshold method mostly uses the multispectral bands of the image to calculate the radiation value characteristics of its bands, such as high reflectivity, low temperature, etc. The key to the physical threshold method is how to select the threshold according to the physical characteristics of the image, and to reduce the amount of calculation in the calculation and reduce hardware requirements. 2) The detection method based on texture and spatial characteristics mainly uses the radiation characteristics and texture characteristics of the image to count the grayscale value of the image. Generally, clouds are brighter than the underlying surface, the grayscale of clouds is higher than the grayscale of the underlying surface, the edges are sharper, and there will be more obvious mutation characteristics in the grayscale value. However, due to the large number of cloud types, different shapes, and different types of cloud features are distributed differently in different feature spaces, and the ground features are also similar, so only using texture features for cloud detection will have limited accuracy improvement.

[0003] The methods listed above all have detection limitations to varying degrees. In order to obtain better cloud detection results for cloud images, a better detection method is needed. This method can integrate the cloud detection algorithms under a multi-strategy combination and use decision fusion technology to comprehensively vote on the cloud detection results obtained by the two strategies to achieve a higher detection accuracy.

[0004] Patent document No. CN112669298A provides a ground-based cloud image cloud detection method based on model self-training. Its solution is a simple detection method based on machine learning, which is greatly affected by cloud feature selection and training sets, and the detection accuracy is not easy to guarantee.

[0005] Patent document CN108399363B provides a cloud detection method based on polarization images, which essentially adopts a physical threshold method. This method is difficult to determine the existence of clouds in certain environments. For example, under conditions of thick snow and thin clouds, similar reflectivities make it difficult to distinguish between clouds and ground objects, and detection accuracy is also difficult to guarantee. Summary of the invention

[0006] The purpose of the present invention is to provide a cloud detection algorithm based on a combination of two strategies, and to use the DS decision fusion technology to comprehensively vote on the cloud detection results obtained by the two strategies to obtain the optimal cloud detection result.

[0007] The object of the present invention can be achieved by the following technical solution: A cloud detection algorithm based on a combination of two strategies comprises the following steps:

[0008] S1. Obtain a cloud-containing image, perform fine segmentation on the image based on MST, and obtain segmentation patches;

[0009] S2, using SVM image classification strategy to detect clouds on the segmented image patches;

[0010] S3, for the segmented image spots, a spectral multi-factor detection strategy is used to detect clouds;

[0011] S4. Use the SVM image classification cloud detection results and the spectral multi-factor cloud detection results to perform DS decision fusion to obtain the optimal cloud detection strategy for cloud-containing images.

[0012] Further: the MST in S1 performs fine segmentation on the image including the following steps:

[0013] S11, initialize each pixel in the cloud-containing image, select the pixel as a vertex of the spanning tree, use the line between two adjacent pixels as the edge weight, and construct an image matrix;

[0014] S12, sort the edge weights, and use the Kruskal algorithm to construct the minimum spanning tree of the image matrix starting from the minimum edge weight;

[0015] S13, determining whether the edge weight of the current pixel is greater than a preset segmentation threshold, and if so, adding the pixel to the spanning tree;

[0016] S14, repeat step S13 until the segmentation of the cloud-containing image is completed, and obtain the segmentation spots of the cloud-containing image.

[0017] Further: the cloud detection step of adopting the SVM image classification strategy in S2 includes:

[0018] S21, extracting spectral features and texture features of the segmented spots;

[0019] S22, performing SVM classification on the segmented image spots according to spectral features and texture features;

[0020] S23. Output cloud detection results based on the SVM classified images.

[0021] Further: the spectral characteristics in S21 include:

[0022] Brightness spectrum characteristics:

[0023]

[0024] Where N is the number of segmented pattern bands, O v is the current segmentation patch calculated, O v ′ is the segmentation pattern O v The number of pixels in the image, (x, y) represents the coordinates of the pixel, n is the band where the pixel is located, P n (x, y) is the grayscale value of the pixel element of the corresponding band and coordinate, (x, y)∈O v For all the segments O v The coordinates of the internal pixels of ;

[0025] Mean spectral characteristics:

[0026]

[0027] Standard deviation spectral characteristics:

[0028]

[0029] Median spectral characteristics:

[0030] When O v When ′ is an odd number, the calculation formula is:

[0031]

[0032] When O v When ′ is an even number, the calculation formula is:

[0033]

[0034] Skewness spectrum characteristics:

[0035]

[0036] Maximum and minimum spectral characteristics:

[0037] Max(n)=Max(P n (x,y)

[0038] Min(n)=Min(P n (x,y)

[0039] The texture features in S21 include:

[0040] Contrast texture features:

[0041]

[0042] Among them, i, j represent the gray level, probability density P i,j is the joint conditional probability density between gray levels, indicating the probability of gray level j appearing when gray level i is the starting point;

[0043] Homogeneous texture features:

[0044]

[0045] Entropy texture features:

[0046]

[0047] Dissimilarity texture features:

[0048]

[0049] Angular second-order moment texture features:

[0050] Mean texture feature:

[0051]

[0052] Variance texture features:

[0053]

[0054] Correlation texture features:

[0055]

[0056] Further: the cloud detection step of adopting the spectral multi-factor detection strategy in S3 includes:

[0057] S31, performing image recognition on the segmented image spots to obtain the normalized difference vegetation index NDVI, the normalized difference water index NDWI, and the normalized difference snow index NDSI;

[0058] S32, respectively using the normalized difference vegetation index NDVI, the normalized difference water index NDWI and the normalized difference snow index NDSI to detect the type of land objects in the segmented image patch;

[0059] S33, classifying the corresponding ground object type according to the set threshold value and distinguishing the cloud;

[0060] S34. Use the Delphi method to filter out the three index division results and output the cloud detection results.

[0061] Further: the normalized difference vegetation index NDVI in S31 is:

[0062] NDVI = (NIR-R) / (NIR+R)

[0063] The normalized difference water index NDWI is:

[0064] NDWI=(G-NIR) / (G+NIR)

[0065] The normalized difference snow index NDSI is:

[0066] NDSI=(G-SWIR) / (G+SWIR)

[0067] Among them, NIR represents the spectral mean of the image object in the near-infrared band, SWIR represents the spectral mean of the image object in the short-wave infrared band, R represents the spectral mean of the image object in the red band, and G represents the spectral mean of the image object in the green band.

[0068] Further: the DS decision fusion is performed using the SVM image classification cloud detection result and the spectral multi-factor cloud detection result in S5 to obtain the optimal cloud detection strategy for cloud-containing images, including the following steps:

[0069] S51, according to the SVM image classification strategy cloud detection result, obtain the probability allocation function BPAF of each patch of the segmented image patch after SVM image classification, including:

[0070]

[0071] in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α S is the confidence of a piece of evidence in the judgment;

[0072] S52, according to the cloud detection result of the spectral multi-factor detection strategy, obtain the probability allocation function BPAF of each patch of the segmented image after the spectral multi-factor detection, including:

[0073]

[0074] in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α D is the confidence of a piece of evidence in the judgment;

[0075] S53, combining the two methods,j (h u ), m j (h c ), m j (h u ,h c ) performs DS decision fusion to obtain the optimal segmentation pattern cloud detection result.

[0076] Beneficial effects of the present invention:

[0077] 1. The present invention obtains the optimal cloud detection result of cloud-containing images by performing DS decision fusion on the SVM image classification cloud detection result and the spectral multi-factor cloud detection result. The cloud detection accuracy of cloud-containing images is high and the effect is better.

[0078] 2. The present invention performs fine segmentation of cloud-containing images based on MST, obtains segmentation patches with similar grayscale, and classifies the types of objects and clouds obtained according to the patches of segmentation patches more accurately and precisely, providing an operating basis for SVM image classification cloud detection and spectral multi-factor cloud detection, which is beneficial to improving the accuracy of cloud detection of various strategies.

[0079] 3. The SVM image classification cloud detection method of the present invention classifies the spectral features and texture features of the segmented spots in detail, fully extracts features, and improves the accuracy of SVM image classification cloud detection.

[0080] 4. The spectral multi-factor cloud detection method of the present invention adopts three spectral indices, namely NDVI, NDWI and NDSI, to respectively detect the ground object type and cloud of the segmented image patches; and through the Delphi method screening, the judgment results of the ground object type and cloud of the segmented image patches are more accurate, thereby improving the accuracy of spectral multi-factor cloud detection.

[0081] 5. The DS decision fusion of the present invention adopts the probability allocation function BPAF for the patch type of the segmented patch obtained according to the two strategies, obtains the judgment of the patch type of the segmented patch by the two strategies, obtains the optimal detection output result, and obtains higher accuracy of cloud detection of cloud-containing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 The figure is a flow chart of a cloud detection algorithm based on a combination of two strategies according to the present invention. DETAILED DESCRIPTION

[0083] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0084] like Figure 1 As shown, the present invention discloses a cloud detection algorithm based on a combination of two strategies, comprising the following steps:

[0085] S1. Obtain a cloud-containing image, perform fine segmentation on the image based on MST, and obtain segmentation patches;

[0086] S2, using SVM image classification strategy to detect clouds on the segmented image patches;

[0087] S3, for the segmented image spots, a spectral multi-factor detection strategy is used to detect clouds;

[0088] S4. Use the SVM image classification cloud detection results and the spectral multi-factor cloud detection results to perform DS decision fusion to obtain the optimal cloud detection strategy for cloud-containing images.

[0089] The MST fine segmentation of the image includes the following steps:

[0090] S11, initialize each pixel in the cloud-containing image, select the pixel as a vertex of the spanning tree, use the line between two adjacent pixels as the edge weight, and construct an image matrix;

[0091] S12, sort the edge weights, and use the Kruskal algorithm to construct the minimum spanning tree of the image matrix starting from the edge with the smallest weight;

[0092] S13, determining whether the edge weight of the current pixel is greater than a preset segmentation threshold, and if so, adding the pixel to the spanning tree;

[0093] S14, repeat step S13 until the segmentation of the cloud-containing image is completed, and obtain the segmentation spots of the cloud-containing image.

[0094] The MST image segmentation algorithm adopts the minimum spanning tree segmentation method. By initializing each pixel point in the cloud-containing image, some pixels are selected as spanning tree vertices, and the connection between two adjacent pixels is regarded as the edge weight (the positional relationship between two adjacent pixels can be a four-neighborhood or an eight-neighborhood). The size of the edge weight can be calculated by calculating the dissimilarity measure of the feature vectors of the two vertices according to the gray value of the pixel point and the set neighborhood range, thereby constructing the image matrix; the Kruskal algorithm is used to construct the minimum spanning tree, and the obtained edge weights are sorted and constructed from the smallest edge weight; it is determined whether the current edge weight meets the preset segmentation threshold. If it does, it is added to the spanning tree, otherwise it is not added; it is repeated until the current edge weight does not meet the criterion, and the segmentation patches of the cloud-containing image are obtained. The edge weights of each patch are close and can be identified as the same type of ground objects or clouds.

[0095] Strategy 1: Using SVM image classification for cloud detection steps include:

[0096] S21, extracting spectral features and texture features of the segmented spots;

[0097] S22, performing SVM classification on the segmented image spots according to spectral features and texture features;

[0098] S23. Output cloud detection results based on the SVM classified images.

[0099] The spectral characteristics of an image are the most intuitive image characteristics. They are usually the brightness or grayscale values ​​on the image, which are quantified by the radiation value. This value is a relative measure of the ground object, mainly including:

[0100] Brightness spectrum characteristics, this value reflects the overall intensity of the image object, the specific formula is:

[0101]

[0102] Where N is the number of segmented pattern bands, O v is the current segmentation patch calculated, O v ′ is the segmentation pattern O v The number of pixels in the image, (x, y) represents the coordinates of the pixel, n is the band where the pixel is located, P n (x, y) is the grayscale value of the pixel element of the corresponding band and coordinate, (x, y)∈O v For all the segments O v The coordinates of the internal pixels of .

[0103] Mean spectral feature, this value reflects the average intensity of the image object in the nth band. The specific formula is:

[0104]

[0105] Standard deviation spectral feature. This value reflects the degree of spectral dispersion when the image object is located in the nth band. The specific formula is:

[0106]

[0107] The median spectral feature is the grayscale value of the image object located in the middle of all the sorted pixel grayscale values. It reflects the grayscale representative value of the image object. Because it is a value in the middle position, it avoids the influence of grayscale maximum and minimum values ​​caused by external factors such as image exposure accidents inside the object. It has strong robustness. The specific formula is:

[0108] When O v When ′ is an odd number, the calculation formula is:

[0109]

[0110] When O v When ′ is an even number, the calculation formula is:

[0111]

[0112] Skewness spectral feature, this value reflects the asymmetry of the image object in the nth band. The specific formula is:

[0113]

[0114] The maximum and minimum spectral features are the statistics of the maximum and minimum values ​​of the image object when it is located in the nth band. The specific formula is:

[0115] Max(n)=Max(P n (x,y)

[0116] Min(n)=Min(P n (x,y)

[0117] Texture features are global features used to describe the surface properties of objects in an image area. The difference from spectral features is that the statistical features are not limited to a certain pixel value, but require corresponding statistical analysis in a certain area. Texture features are based on statistical features and therefore have strong noise resistance.

[0118] The gray-level co-occurrence matrix reflects the overall information implicit in the image by calculating the correlation between any pair of pixels in a certain direction and distance in the image. The formula of the gray-level co-occurrence matrix is:

[0119] G(i,j,d,k)={[(x,y),(x+Δx,y+Δy)]|f(x,y)=i,f(x+Δx,y+Δy)=j; (x,y)∈O v}

[0120] In the formula, i, j represent the grayscale level, (x, y) represents the image object O v The internal pixel coordinate value of f(x,y) represents the probability that a pixel (x,y) with a grayscale level i and a pixel (x+Δx,y+Δy) with a grayscale level j appear at the same time inside the image object. The distance between the two pixels is represented by d in the formula, and the angle between the line between the two pixels and the x-axis in a clockwise direction is represented by the symbol k.

[0121] Texture features mainly include:

[0122] Contrast texture feature, contrast texture feature is a description of the contrast between adjacent pixels, which can be generally understood as the clarity of the image and the strength of the image texture. The clarity of the texture is positively correlated with the contrast. When the contrast value is larger, the texture primitives are more contrasting, and the texture effect is more obvious: specifically, when the contrast value is small, the texture effect is not obvious; when the contrast value is equal to 0, the image is completely uniform and there is no texture expression. The specific calculation formula is:

[0123]

[0124] Among them, i, j represents the gray level, and the probability density P i,j is the joint conditional probability density between gray levels, indicating the probability of gray level j appearing when gray level i is the starting point;

[0125] Homogeneity texture feature: The regularity of the texture, or the uniformity, can be expressed by the homogeneity texture feature. When the homogeneity value is small, the regularity of the image texture feature is difficult to find and describe; when the homogeneity value is large, it means that the image texture has obvious regularity and is easy to describe. The specific calculation formula is:

[0126]

[0127] Entropy texture feature: The size of the texture feature entropy value reflects the amount of information contained in the image. The entropy value expresses the complexity of the texture and is a measure of the randomness of the image content. When the entropy value is large, it means that the image has a high degree of complexity; when the entropy value is small or equal to 0, it means that the image is less complex at this time. The specific calculation formula is:

[0128]

[0129] Dissimilarity texture feature: Dissimilarity texture feature: Dissimilarity is similar to contrast texture feature. The clarity of image texture is reflected in both. When the dissimilarity value is larger, the texture is more prominent; when the value is smaller, the texture is more blurred; when the value is equal to 0, the image is expressed as completely uniform and without texture. The specific calculation formula is:

[0130]

[0131] Angular second-order moment texture feature: The angular second-order moment feature mainly reflects the fineness of the image texture. The specific calculation formula is:

[0132]

[0133] Mean texture feature: The mean feature mainly reflects the regularity of the image texture. If the texture is messy and difficult to describe, this value is small; if the texture is regular and easy to describe, this value is large. The specific calculation formula is:

[0134]

[0135] Variance texture feature: The variance feature is mainly a measure of the deviation between the pixel value and the mean. When the grayscale changes in the image are large, the variance value is large. The specific calculation formula is:

[0136]

[0137] Correlation texture feature: The correlation feature mainly describes the similarity between row or column elements in the grayscale co-occurrence matrix, reflecting the extension length of a certain grayscale value along a certain direction. The longer the extension, the greater the correlation. It is a measure of the linear relationship of grayscale. The specific calculation formula is:

[0138]

[0139] Based on the above features, the solution uses the SVM method to classify images based on different feature combinations extracted from the above features. SVM image classification is a powerful machine learning method developed from statistics and is often used in classification and regression analysis. The expression formula for classifying images using SVM can be:

[0140]

[0141] Among them, x represents the feature point after sampling (that is, the feature combination selected from the spectral feature and the texture feature), represents the transformed feature space, w represents an n-dimensional vector, and b is a real value.

[0142] When using the support vector machine (SVM) algorithm, the solution maximizes the boundary of the data set by constructing an optimal hyperplane, and for each data point, there is a target label t i (t∈{-1,1}).

[0143] When classifying images at the object level, the object features used are often nonlinearly separable. In such cases, the soft margin SVM algorithm is usually used. By introducing the relaxation variable ζ, the following formula is used for constraints:

[0144]

[0145] For t n = 1, we need to make y(x n )>0, for t n = -1, we must make y(x n)<0, from the data point x n The distance to the decision boundary is a hard margin given by:

[0146]

[0147] Strategy 2: Using spectral multi-factor detection to detect clouds. The steps include:

[0148] S31, performing image recognition on the segmented image spots to obtain the normalized difference vegetation index NDVI, the normalized difference water index NDWI, and the normalized difference snow index NDSI;

[0149] S32, using the normalized difference vegetation index NDVI, the normalized difference water index NDWI and the normalized difference snow index NDSI respectively to detect the ground object type and cloud of the segmented image patch;

[0150] S33. Use the Delphi method to screen the detected ground object types and clouds, and output the cloud detection results.

[0151] Three parameter indicators, namely the normalized difference vegetation index NDVI, the normalized difference water index NDWI and the normalized difference snow index NDSI, were selected for segmentation patch recognition.

[0152] The Normalized Difference Vegetation Index (NDVI) can evaluate the growth and coverage of vegetation in image objects. It is widely used in change detection for forest coverage. Its calculation formula is:

[0153] NDVI = (NIR-R) / (NIR+R)

[0154] The normalized difference water index NDWI can highlight water information in image objects and evaluate the water content in vegetation. It is used in fields such as drought monitoring. The calculation formula is:

[0155] NDWI=(G-NIR) / (G+NIR)

[0156] The Normalized Difference Snow Index (NDSI) uses a combination of visible light (mainly green light) and short-wave infrared bands to highlight the snow coverage in the image. The calculation formula is:

[0157] NDSI=(G-SWIR) / (G+SWIR)

[0158] Among them, NIR represents the spectral mean of the image object in the near-infrared band, SWIR represents the spectral mean of the image object in the short-wave infrared band, R represents the spectral mean of the image object in the red band, and G represents the spectral mean of the image object in the green band.

[0159] The thresholds are determined by the normalized difference vegetation index NDVI, the normalized difference water index NDWI and the normalized difference snow index NDSI, and the corresponding ground object types are divided and clouds are distinguished.

[0160] The Delphi method is used to screen the ground object types and clouds divided by the three indexes, and output the cloud detection results; the Delphi method screening is based on expert scoring, and the detection results obtained under the above multi-factor situation are used for voting and scoring. The results with high scores are regarded as the detection results under this method.

[0161] The DS decision fusion is performed using the SVM image classification cloud detection results and the spectral multi-factor cloud detection results to obtain the optimal cloud detection strategy for cloud-containing images, which includes the following steps:

[0162] S51, according to the SVM image classification strategy cloud detection result, obtain the probability allocation function BPAF of each patch of the segmented image patch after SVM image classification, including:

[0163]

[0164] in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α S is the confidence of a piece of evidence in the judgment.

[0165] S52, according to the cloud detection result of the spectral multi-factor detection strategy, obtain the probability allocation function BPAF of each patch of the segmented image after the spectral multi-factor detection, including:

[0166]

[0167] in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α D is the confidence of a piece of evidence in the judgment.

[0168] S53, yes Objects and The DS decision fusion is performed on the objects to obtain the optimal cloud detection result.

[0169] The following is an explanation of the DS decision fusion principle:

[0170] For example: Bel1, Bel2, …, Bel n is the belief function corresponding to Θ, m1,m2,…,m n They are Bel1, Bel2, ..., Bel n The corresponding basic credit distribution is Existence and Then we have:

[0171]

[0172] in

[0173]

[0174] The above formula is the Dempster rule composed of multiple belief functions.

[0175] There are two extreme cases for the normalization coefficient K. One is when a certain attribute of multiple propositions is completely irrelevant, and there is a complete dissimilarity between their evidence theories, resulting in a final synthesis result of zero. The other is completely opposite, which means that when the supporting evidence behind multiple propositions is completely consistent, K n =1, which makes the final synthesis result equal to the cumulative product of each sub-proposition, which is also the assumption of Bayesian theory. Generally speaking, the supporting evidence behind multiple propositions partially overlaps, but also has its own independent contribution. In other words, there is correlation between different propositions.

[0176] DS fusion obeys the basic probability allocation function BPAF, which combines the probability of each spot recognition result mapping to measure the size of the event. Assume that there is a hypothesis space in the cloud spot recognition application, represented by Θ, where Θ is the hypothesis set about belonging / not belonging, and its power set is 2 Θ of.

[0177] Assume A is 2 Θ A non-empty subset of , m(A) represents the degree of subset A. Θ →[0,1] Based on the constraint condition of Equation 25

[0178]

[0179] For a certain class of methods, there are n independent mappings of cloud layer identification numbers, m i (B i ) represents the multi-mapping i(1≤i≤n) and B i ∈2 Θ , Therefore, the BPAF of A is expressed by fusing the probability of the mapping, that is, the calculation formula of m(A) is as follows:

[0180]

[0181] Through DS fusion of the results from the above two strategies, the final fused cloud recognition image result is obtained and output. In the image cloud recognition problem, the assumed recognition space Θ is equal to {hu, hc}, and the existence of hu means that it does not belong to, and hc means that it belongs to these two propositions. Therefore, 2 Θ The three non-empty subsets of are {hu}, {hc} and {h u ,h c}, which means not belonging, belonging and uncertain.

[0182]

[0183] Table 1 DS fusion table of two strategies

[0184] As shown in Table 1, different detection methods are combined with segmented images to calculate the BPAF of the above two subsets. For each object j in the image, {hu}, {hc} and {h u ,h c The BPAF of} is defined as the formula:

[0185]

[0186] in and They represent the number of judgments that object j does not belong to or does not belong to in PBCD mapping i, respectively, and Nt j represents the total number of recognitions counted by object j, α i is the confidence of a piece of evidence in the judgment. , it will decide whether the object belongs to the cloud layer.

[0187] Example:

[0188]

[0189] calculate:

[0190] K=1-(hu1*hc2+hc1*hu2)=0.49

[0191] Mhu=(hu1*hu2+hu1*{hu2,hc2}+{hu1,hc1}*hu2) / K=0.265

[0192] Mhc=(hc1*hc2+hc1*{hu2,hc2}+{hu1,hc1}*hc2) / K=0.714

[0193] Mhu,hc=({hu1,hc1}*{hu2,hc2}) / K=0.02

[0194]

[0195] (0.714>0.265>0.02).

[0196] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

[0197] It should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0198] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0199] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection, an electrical connection, or communication with each other; it can be a direct connection, or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0200] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

Claims

1. A cloud detection algorithm based on a combination of two strategies, characterized in that: The following steps are involved: S1. Obtain a cloud-containing image, perform fine segmentation on the image based on MST, and obtain segmentation patches; S2, using SVM image classification strategy to detect clouds on the segmented image patches; S3, for the segmented image spots, a spectral multi-factor detection strategy is used to detect clouds; S4. Use the SVM image classification cloud detection results and the spectral multi-factor cloud detection results to perform DS decision fusion to obtain the optimal cloud detection strategy for cloud-containing images.

2. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: The MST in S1 performs fine segmentation of the image, including the following steps: S11, initialize each pixel in the cloud-containing image, select the pixel as a vertex of the spanning tree, use the line between two adjacent pixels as the edge weight, and construct an image matrix; S12, sort the edge weights, and use the Kruskal algorithm to construct the minimum spanning tree of the image matrix starting from the minimum edge weight; S13, determining whether the edge weight of the current pixel is greater than a preset segmentation threshold, and if so, adding the pixel to the spanning tree; S14, repeat step S13 until the segmentation of the cloud-containing image is completed, and obtain the segmentation spots of the cloud-containing image.

3. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: The cloud detection steps using the SVM image classification strategy in S2 include: S21, extracting spectral features and texture features of the segmented spots; S22, performing SVM classification on the segmented image spots according to spectral features and texture features; S23. Output cloud detection results based on the SVM classified images.

4. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: The spectral characteristics in S21 include: Brightness spectrum characteristics: Where N is the number of segmented pattern bands, O v is the current segmentation patch calculated, O v ′ is the segmentation pattern O v The number of pixels in the image, (x, y) represents the coordinates of the pixel, n is the band where the pixel is located, P n (x, y) is the grayscale value of the pixel element of the corresponding band and coordinate, (x, y)∈O v are the coordinates of all internal pixels belonging to the segmentation patch Ov; Mean spectral characteristics: Standard deviation spectral characteristics: Median spectral characteristics: When O v When ′ is an odd number, the calculation formula is: When O v When ′ is an even number, the calculation formula is: Skewness spectrum characteristics: Maximum and minimum spectral characteristics: Max(n)=Max(P n (x,y)) Min(n)=Min(P n (x,y)) The texture features in S21 include: Contrast texture features: Among them, i, j represent the gray level, probability density P i,j is the joint conditional probability density between gray levels, indicating the probability of gray level j appearing when gray level i is the starting point; Homogeneous texture features: Entropy texture features: Dissimilarity texture features: Angular second-order moment texture features: Mean texture feature: Variance texture features: Correlation texture features:

5. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: The cloud detection steps using the spectral multi-factor detection strategy in S3 include: S31, performing image recognition on the segmented image spots to obtain the normalized difference vegetation index NDVI, the normalized difference water index NDWI, and the normalized difference snow index NDSI; S32, respectively using the normalized difference vegetation index NDVI, the normalized difference water index NDWI and the normalized difference snow index NDSI to detect the type of land objects in the segmented image patch; S33, classifying the corresponding ground object type according to the set threshold value and distinguishing the cloud; S34. Use the Delphi method to filter out the three index division results and output the cloud detection results.

6. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: The normalized difference vegetation index NDVI in S31 is: NDVI = (NIR-R) / (NIR+R) The normalized difference water index NDWI is: NDWI=(G-NIR) / (G+NIR) The normalized difference snow index NDSI is: NDSI=(G-SWIR) / (G+SWIR) Among them, NIR represents the spectral mean of the image object in the near-infrared band, SWIR represents the spectral mean of the image object in the short-wave infrared band, R represents the spectral mean of the image object in the red band, and G represents the spectral mean of the image object in the green band.

7. The cloud detection algorithm based on the combination of two strategies according to claim 1 is characterized in that: In S4, the DS decision fusion is performed using the SVM image classification cloud detection result and the spectral multi-factor cloud detection result to obtain the optimal cloud detection strategy for cloud-containing images, which includes the following steps: S51, according to the SVM image classification strategy cloud detection result, obtain the probability allocation function BPAF of each patch of the segmented image patch after SVM image classification, including: in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α S is the confidence of a piece of evidence in the judgment; S52, according to the cloud detection result of the spectral multi-factor detection strategy, obtain the probability allocation function BPAF of each patch of the segmented image after the spectral multi-factor detection, including: in represents the number of judgments that do not belong to object j in the allocation function BPAF mapping i, Nt represents the number of judgments that belong to object j in the allocation function BPAF mapping i, j represents the total number of recognitions counted by object j, α D is the confidence of a piece of evidence in the judgment; S53, combining the two methods, j (h u ), m j (h c ), m j (h u ,h c ) performs DS decision fusion to obtain the optimal segmentation pattern cloud detection result.

Images