CN107610114A - Optical satellite remote sensing image cloud snow mist detection method based on SVMs - Google Patents

Abstract

The invention discloses a satellite remote sensing image cloud, snow and fog detection method based on a support vector machine, comprising the following steps: first, collecting a large number of different types of surface objects and cloud, snow and fog sample image data images as a training set, and obtaining The grayscale and texture features of the image form a feature set; machine learning is performed on the feature set of all samples by the method of support vector machine to obtain cloud, snow, and fog image classifiers. Secondly, use the obtained cloud, snow, and fog image classifier to determine the category of the image to be tested, and perform morphological closing operation and overlapping area correction to determine the type of target area in the remote sensing image; finally, re-select training samples to obtain a new image classification The device performs secondary detection on the remote sensing image of the satellite to be measured, and compares it with the first detection, and finally determines the judgment results of cloud, snow, and fog in the remote sensing image to be measured. Experimental results show that the method of the present invention can obtain higher detection accuracy.

Description

Detection method of cloud, snow and fog in optical satellite remote sensing images based on support vector machine

technical field

The invention belongs to the field of quality detection of satellite remote sensing images, in particular to a method for detecting cloud, snow and fog of satellite remote sensing images based on a support vector machine.

Background technique

In optical satellite remote sensing images, remote sensing information is often affected by clouds, fog and snow. Clouds and snow will cover the surface information of the area where the image is located. Fog and haze will cover up many characteristic information in remote sensing images. Therefore, it is necessary to detect cloud, snow, and fog areas in remote sensing images, and to eliminate corresponding image data with excessive coverage of invalid information, so as to improve the utilization rate of optical satellite remote sensing images.

The current remote sensing image cloud, snow, and fog detection methods mainly focus on the detection of clouds or fog, and the detection and identification of clouds, fog, and clouds and snow. It mainly includes threshold method, feature extraction method and so on. The method of remote sensing image cloud detection mainly uses the spectral reflectance under different bands to set the spectral threshold to judge whether it is a cloud, or extracts image features and performs cloud detection according to feature classification. The method of remote sensing image fog detection is mainly the study of typical cases, using remote sensing data to extract features for monitoring research on heavy fog. The cloud and snow detection method mainly utilizes the characteristics of clouds and snow that are similar in the visible light band and have great differences in short-wave infrared, and identify snow by constructing the cloud and snow contrast enhancement factors, or by calculating the fractal dimension of the texture features of the panchromatic image. Realize the identification of clouds and snow. The cloud, snow, and fog detection methods are usually superimposed using the above methods.

After searching the existing literature, it is found that the current cloud, snow, and fog detection methods have the following problems: First, the existing methods are difficult to detect clouds, snow, and fog at the same time. The detection method is affected by the type of detection, and a single detection method is difficult to adapt to a variety of detection requirements; the existing threshold method is not reliable, because the detection results are affected by the space-time type, it is difficult to extend to the more common detection; the feature extraction method selects The feature information of the image is insufficient, and the detection accuracy is not high enough. Second, the detection efficiency of cloud, snow, and fog detection methods is not high, and the complexity of the algorithm is high. It is difficult to quickly detect and identify the impact of big data. There are certain requirements for remote sensing data sources, and the universality is poor.

Contents of the invention

The purpose of the invention is to enhance the timeliness of remote sensing image quality inspection, improve the utilization rate of remote sensing images, so that it can be applied to the quality inspection system of domestic satellite image products such as Ziyuan No. 1, Zi Zi No. 3, Tianhui No. 1 and Gaofen No. 1 .

In order to realize this object, the present invention proposes a kind of satellite remote sensing image cloud, snow, fog detection method based on support vector machine, and the specific realization of technical scheme of the present invention comprises the following steps:

Step 1, collect a large number of sample image data of clouds, snow, fog and ground features;

Step 2, extracting grayscale features and texture features of various sample images to form feature vectors;

Step 3, use the support vector machine to train the feature vector of the sample image, and obtain the cloud image classifier, snow image classifier and fog image classifier composed of decision functions respectively;

Step 4, perform down-sampling processing on the original image of the satellite remote sensing image to be measured to obtain a thumbnail image, perform image segmentation on the thumbnail image to obtain sub-images, and calculate feature vectors composed of grayscale features and texture features of all sub-images;

Step 5, classify the sub-images of the satellite remote sensing image to be measured, including the following sub-steps,

Step 5.1, input the feature vectors extracted in step 4 into the cloud, snow and fog image classifiers obtained in step 3 for prediction and classification;

Step 5.2, divide all sub-images into cloud area, fog area, snow area and object area according to the type of target area;

Step 5.3, according to the cloud and ground object area, fog and ground object area, snow and ground object area are divided into three binarized images, wherein the ground object area in each image takes the same zero value, cloud, snow, fog Take different image values for the area;

Step 6, performing a morphological "close" operation on the classification results obtained in step 5;

Step 7: Compare the three binarized image values at the same location to obtain the detection results of clouds, snow, and fog in the satellite remote sensing image to be tested.

Further, the implementation of step 7 is as follows,

Step 7.1, compare the three binarized image values at the same position, if the image values of the same position of the three images are the same, then determine that this position is a feature area; if there are two identical values at the same position, then determine that this position is The category area represented by the third image value; if the image values of the three images at the same location are all different, it is determined that the location is an overlapping area with clouds, snow, and fog, and the points with zero values in it are marked as two overlaps area, mark the points without zero value as triple overlapping areas;

Step 7.2, repeat step 7.1, compare all the image values of the three binarized images, obtain the discrimination results of cloud, snow, fog area, ground object area and overlapping area, and correct the overlapping area; first judge whether the overlapping area contains In other areas, if it is included, replace the overlapping area with other areas; secondly, determine the category of the two overlapping areas, if it is circumscribed with a certain category area, then determine the category of the overlapping area as the category after the overlap area is removed from the determined circumscribed category , if not, confirm after the category of the three-overlapping area is determined; for the three-overlapping area, if it is circumscribed with an area of a certain category, it is determined that the overlapping area is the second-overlapping area after the circumscribed category is removed, and if it is circumscribed with the second-overlapping area, Then it is determined that the category of the overlapping area is the category of the overlapping area and its circumscribed two overlapping area categories; finally it is determined that the two overlapping areas are only circumscribed by different two overlapping areas, and these areas are respectively determined as categories after removing the common category, and finally Get the judgment result.

In step 7.3, the judgment result is subjected to morphological "closed" operation to obtain the detection result of cloud, snow and fog in the satellite remote sensing image to be tested.

Further, it also includes step 8, reselecting appropriate amount of cloud and ground objects, fog and ground objects, snow and ground objects samples as training samples, repeating steps 2-7 to perform secondary detection on the remote sensing image of the satellite to be measured, and the secondary detection The detection result is compared with the first detection result. If the two detection results of the same position are the same, it is determined that the category of the position is the category obtained by any one detection result. If the two detection results of the same position are different, the category of the position is determined The category is ground objects, and the detection results are finally obtained.

Further, the implementation of step 2 is as follows,

Step 2.1, calculating the grayscale features of the sample image, including the grayscale mean, grayscale variance, first-order difference and histogram information entropy of the sample image;

The formula for calculating the gray mean value is:

Wherein, f(i,j) is the gray value at (i,j), S=M×N, M is the width of the sample image, and N is the height of the sample image;

The formula for calculating the variance of the gray level is,

The formula for calculating the first-order difference is,

The formula for calculating the histogram information entropy is,

Among them, h[g] is the histogram of the sample image, h[g](i) is the percentage of pixels at a certain gray level in the entire sample image, and M is the maximum gray level;

Step 2.2, calculate the texture features of the sample image, including the gradient standard deviation, mixing entropy, inverse moment,

texture fractal;

The formula for calculating the standard deviation of the gradient is,

G(i,j;d,θ)=#{(x ₁ ,y ₁ )(x ₂ ,y ₂ )|f(x ₁ ,y ₁ )=i,f(x ₂ ,y ₂ )=j, |(x ₁ ,y ₁ )-(x ₂ ,y ₂ )|＝d,∠((x ₁ ,y ₁ ),(x ₂ ,y ₂ ))＝θ} where d represents the distance between two pixels , θ represents the orientation angle between pixels, f(x ₁ ,y ₁ ) and f(x ₂ ,y ₂ ) represent the gray values of (x ₁ ,y ₁ ) and (x ₂ ,y ₂ ) respectively, ∠ represents The angle between the pixel point and the horizontal position, # indicates the number of pixel pairs obtained according to the constraints in the set, ∑#L _x L _y indicates the sum of the number of all pixel pairs under a specific position relationship; L _g indicates The maximum value of the gray level, L represents the maximum value of the gradient;

The formula for calculating the mixing entropy is,

The formula for calculating the inverse gap is,

Using the fractal Brownian random field method to solve the texture fractal dimension of the sample image, the expression of the fractal dimension D of the image is,

D=n+1-H

Among them, n refers to the dimension of the sample image space, and H is the self-similarity parameter;

In step 2.3, the above grayscale features and texture features are combined into an 8-dimensional feature vector.

Further, the implementation of step 3 is as follows,

Step 3.1, select some cloud samples and object samples as training samples, and use the feature vectors of the training samples as the training set for training the image classifier T={(x ₁ ,y ₁ )...( _xi ,y _i )} , i=1…N, y _i ∈ ψ={-1, 1}, where 1 represents the positive class, which represents the cloudy area category, -1 represents the negative class, which represents the object area category, x _i ∈ R ⁿ , x _i is the feature vector, N is the number of samples;

Step 3.2, using the support vector machine of the C-SVC model to construct a classification hyperplane, and calculating the Gaussian kernel function,

Among them, x _i and x _j refer to the feature vectors of samples i and j respectively, || _xi -x _j || ² is the square of the Euclidean distance, σ is the variance, i=1...N, j=1...N, N is the number of samples;

Step 3.3, using the convex quadratic programming method to solve the Lagrange multiplier vector of the optimal classification hyperplane in the feature space,

Among them, 0≤α _i ≤C, i=1,2...N, is the Lagrangian multiplier vector, α _i and αj represent the i-th and j-th Lagrangian multipliers respectively, x _i , x _j represents the eigenvectors of the i-th sample and the j-th sample respectively, y _i and y _j represent the category of the i-th sample and the j-th sample respectively, C is the penalty parameter, N is the number of samples, and the Rag The optimal solution of the Langerian multiplier vector is:

α ^* ＝(α ₁ ^* ,α ₂ ^* ,...α _i ^* ...α _N ^* ) ^T

Among them, α _i ^* represents the optimal solution of the i-th Lagrangian multiplier;

Step 3.4, solve the intercept of the optimal classification hyperplane in the feature space, the calculation formula is,

Among them, α _i ^* is the optimal solution of the i-th Lagrangian multiplier, y _i is the category of the i-th sample, and N is the number of samples;

In step 3.5, substitute the obtained Gaussian kernel function, Lagrangian optimal solution, and hyperplane intercept into the decision function,

In step 3.6, the remaining cloud samples and surface object samples are used as test samples to test the decision function, optimize the decision function, and obtain the corresponding cloud image classifier;

Step 3.7, repeat steps 3.1-3.6 to obtain snow image classifier and fog image classifier respectively.

Further, in the step 4, if the remote sensing image to be tested is a panchromatic image, down-sampling is directly performed; if the remote sensing image to be tested is a multispectral image, RGB three-band down-sampling is used.

Further, the implementation of step 6 is to select a structural element whose structural shape is a square and whose structural size is 3×3, respectively perform expansion operations on the three binarized images, and then use the same Structural elements are corroded.

Compared with the prior art, the advantages of the present invention: the method of the present invention can be used for multiple detections after one training, and the image classifier is obtained through a large number of image trainings, and only needs to be used again during detection, and the support vector machine algorithm can predict and classify The time complexity of the stage is low, and the detection of the area type can be carried out quickly; after testing, the method of the present invention is applicable to both panchromatic images and n-channel multi-spectral images. The remote sensing images of multiple domestic satellites including Tianhui No. 1 and Gaofen No. 1 were used for cloud detection, and the accuracy reached 94.8%, 96.4%, 93.2% and 95.2% respectively.

Description of drawings

FIG. 1 is a flow chart of the implementation of the embodiment of the present invention.

Detailed ways

In order to facilitate those of ordinary skill in the art to understand and implement the present invention, the present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments. The implementation examples described here are only used to illustrate and explain the present invention, but do not limit the scope of the present invention. protected range.

With reference to Fig. 1, the present invention takes No. 02C of No. 1 resource, panchromatic image data of No. 3 satellite and multispectral remote sensing image data of Tianhui No. 1 as an example, and its realization steps are as follows:

Step 1, sample acquisition

The original image of the sample is down-sampled to a 1024×1024 pixel 8-bit bmp format thumbnail, and the thumbnail image is segmented. If the remote sensing image is a panchromatic image, down-sampling is directly adopted; if the remote sensing image is a multi-spectral image, RGB three-band down-sampling is used. Divide the panchromatic image into 32×32 sample blocks, divide the multispectral image into 16×16 image blocks, and select 1500 ground object samples, 1000 cloud samples, 1000 snow samples, and 1000 fog samples as training sample data.

Step 2, feature extraction: There are many differences in the grayscale information of clouds, snow, fog, and ground objects. Generally speaking, the average brightness of cloud areas in panchromatic and multispectral images is much higher than that of fog areas, and fog The brightness value of the area is greater than that of the ground object area, and the cloud and snow areas have similar spectral characteristics. At the same time, there are obvious differences in the gray distribution and gray level change of the cloud, snow, fog and ground objects. Therefore, The image of the target area and the object image can be distinguished to a certain extent by using the gray feature of the image.

Extract the grayscale feature and texture feature vector value of the sample image to form an 8-dimensional feature vector. The specific implementation steps are as follows:

Step 2.1, calculate the grayscale features of the sample image

Step 2.1.1, to find the average gray value, use the following formula to calculate:

Wherein, f(i,j) is the gray value at (i,j), S=M×N, M is the width of the sample image, and N is the height of the sample image.

Step 2.1.2, calculate the grayscale variance of the sample image:

Step 2.1.3, calculate the first order difference of the sample image:

Step 2.1.4, calculate the histogram information entropy of the sample image:

Among them, h[g] is the histogram of the sample image, h[g](i) is the percentage of pixels at a certain gray level (i) in the entire image, and M is the maximum gray level.

Step 2.2, calculate the texture features of the sample image: From the perspective of the visual characteristics of the human eye, the texture information of clouds, snow, and fog in satellite remote sensing images is often simpler and simpler than that of ground objects. In addition, the texture information of clouds, snow, and fog in images The edge of the area is also relatively blurred and round, while the edge of the ground object is usually sharper and has a large gradient. Therefore, the texture information and gradient information of satellite remote sensing images can be used to detect and divide cloud, snow, fog regions and ground object information. In addition, the texture features of cloud, snow, and fog images are also significantly different. For cloud samples, its texture is random, changeable and difficult to detect, and its appearance is messy and irregular, with thick and blurred edge textures; fog sample texture is relatively uniform, smooth, and edge shape is regular; snow samples are influenced by ground texture. The effect of , has better directionality and large gradient changes. Through the comprehensive information of image grayscale and image gradient, different texture features of clouds, snow, and fog can be distinguished.

Step 2.2.1, calculate the gray gradient co-occurrence matrix G(i,j,d,θ) of the sample image, using the following formula:

G(i,j;d,θ)=#{(x ₁ ,y ₁ )(x ₂ ,y ₂ )|f(x ₁ ,y ₁ )=i,f(x ₂ ,y ₂ )=j, |(x ₁ ,y ₁ )-(x ₂ ,y ₂ )|＝d,∠((x ₁ ,y ₁ ),(x ₂ ,y ₂ ))＝θ} where d represents the distance between two pixels , θ represents the direction angle between pixels, (x, y) represents the coordinates of the pixel point, f(x, y) represents the gray value of the point, ∠ represents the angle between the pixel point and the horizontal position, # represents the value according to the set The number of pixel pairs obtained from the restriction conditions, for example, the pixel values of two points are 1 and 2 respectively, the distance between them is 1, θ=0° is the horizontal direction, traverse the entire image, and count the pixels that meet these conditions How many pixel pairs are formed by two points.

In step 2.2.2, the gray level co-occurrence matrix G(i,j,d,θ) is normalized to H(i,j;d,θ), and its calculation formula is as follows:

Among them, ∑#L _x L _y represents the sum of the number of all pixel pairs under a specific positional relationship (that is, the distance d and the included angle θ are the same).

Step 2.2.3, calculate the gradient standard deviation of the sample image, first calculate the gradient average:

L _g represents the maximum value of the gray level, and L represents the maximum value of the gradient.

Substitute the gradient mean T into the following formula to obtain the gradient standard deviation:

Step 2.2.4, calculate the mixing entropy of the sample image, the calculation formula is as follows:

Step 2.2.5, extract the local stationarity features of the sample image, and calculate the inverse gap, the calculation formula is as follows:

In step 2.2.6, the fractal Brownian random field method is used to solve the texture fractal dimension of the sample image;

Solve the constant H(0<H<1), so that the distribution function F(t) satisfies:

F(t) is a distribution function irrelevant to x and Δx, H is a self-similar parameter, f(x) is called a real random function about x, n refers to the dimension of the sample image space, then the expression of the fractal dimension D of the image is :

D=n+1-H

Step 3, image classifier training

Use the support vector machine method to train the feature vectors of the sample images, and obtain cloud image classifiers, snow image classifiers and fog image classifiers composed of decision functions respectively. The specific implementation includes the following sub-steps:

Step 3.1, take 80% of the cloud samples and object samples as training samples, and use the feature vectors of the training samples as the training set for training the image classifier T={(x ₁ ,y ₁ )...(x _i ,y _i )}, i=1...N. Among them, y _i ∈ ψ = {-1, 1} where 1 represents the positive class, that is, the cloudy area category, -1 represents the negative class, that is, the ground object area category, x _i ∈ R ⁿ , x _i is the feature vector, N is the number of samples.

Step 3.2, use the support vector machine of the C-SVC model to construct the classification hyperplane, and calculate the Gaussian kernel function:

Where x _i and x _j refer to the feature vectors of samples i and j respectively, || _xi -x _j || ² is the square of the Euclidean distance, σ is the variance, and the value of σ is generally selected according to the experimental results. Since it is difficult to completely classify the feature vectors correctly during the classification process, the purpose of σ can be understood as setting an error tolerance range and ignoring errors within the range. i=1...N, j=1...N, N is the number of samples.

Step 3.3, use the convex quadratic programming method to solve the Lagrangian multiplier vector of the optimal classification hyperplane in the feature space:

Among them, 0≤α _i ≤C, i=1,2...N, is the Lagrangian multiplier vector, α _i and αj are the i-th and j-th Lagrangian multipliers respectively, x _i , x _j represent the eigenvectors of the i-th sample and the j-th sample respectively, y _i and y _j represent the category of the i-th sample and the j-th sample respectively, C is the penalty parameter, N is the number of samples, and the solution is The optimal solution for the Lagrange multiplier vector is:

α ^* ＝(α ₁ ^* ,α ₂ ^* ,...α _i ^* ...α _N ^* ) ^T

where α _i ^* represents the optimal solution of the ith Lagrangian multiplier.

Step 3.4, solve the intercept of the optimal classification hyperplane in the feature space, the calculation formula is:

Among them, α _i ^* is the optimal solution of the i-th Lagrangian multiplier, y _i is the positive and negative category of the i-th sample, and N is the number of samples.

In step 3.5, substitute the obtained Gaussian kernel function, Lagrangian optimal solution, and hyperplane intercept into the decision function:

In step 3.6, 20% of the cloud samples and ground object samples are used as the test sample set to test the decision function, optimize the decision function, and obtain the corresponding cloud image classifier.

Step 3.7, repeat steps 3.1-3.6 to obtain snow image classifier and fog image classifier respectively.

Step 4, feature extraction of the image to be tested

The original image to be tested is down-sampled to a 1024×1024 pixel 8-bit bmp format thumbnail. If the remote sensing image is a panchromatic image, the down-sampling process is directly adopted; if the remote sensing image is a multi-spectral image, the RGB three-band downsampling Sampling, segmenting the thumbnail images to obtain 1024 32×32 pixel sub-images, extracting feature vectors of all sub-images, including grayscale features and texture feature vectors, the specific extraction can be achieved through step 2.

Step 5, classification of images to be tested

In step 5.1, input the feature vectors extracted in step 4 into the corresponding cloud, snow, and fog image classifiers obtained in step 3 for prediction and classification, and classify the feature vectors through the decision function.

Step 5.2, repeat step 3 until all sub-images are classified, and divide all sub-images into cloud area, fog area, snow area and non-cloud, snow and fog area (ie object area) according to the category of the target area.

Step 5.3, according to the cloud and ground object area, the fog and ground object area, and the snow and ground object area, it is divided into three binarized images, wherein the ground object area between each image takes the same zero value, and the cloud, snow , different image values are taken between fog regions.

Step 6, Morphological closing operation

Select the structural elements whose structural shape is square and whose structural size is 3×3, respectively perform expansion operations on the three binarized images, and then use the same structural elements to perform corrosion operations on the obtained processed images to convert clouds, snow, fog The regions are joined together, and the noisy regions at the edges are finally eliminated.

Step 7, overlapping area correction

Step 7.1, compare the three binarized image values at the same position: if the image values of the same position of the three images are the same, it is determined that the position is a feature area; if there are two identical values, it indicates that the position is the third If the image values are different, it indicates that there are overlapping areas of clouds, snow, and fog in the sub-image, and the points with zero values in them will be marked as two overlapping areas, and the points without zero values will be marked as for three overlapping regions.

Step 7.2, repeat step 7.1, compare all the image values of the three binarized images, obtain the discrimination results of cloud, snow, fog area, ground object area and overlapping area, and correct the overlapping area; first judge whether the overlapping area contains In other areas, if it is included, replace the overlapping area with other areas (including overlapping areas and definite category areas); secondly, determine the category of the two overlapping areas, if it is circumscribed with a certain definite category area, then determine the category of the overlapping area as overlapping The category of the area after the determinate circumscribed category is removed, if there is no category, it will be confirmed after the category of the three-overlapping area is determined; for the triple-overlapped area, if it is circumscribed with a certain category area, then it is determined that the overlapping area is the two-overlapped area after the circumscribed category is removed area, if it is circumscribed with the two overlapping areas, then it is determined that the category of the overlapping area is the category of the overlapping area and its circumscribed two overlapping area categories; finally it is determined that the two overlapping areas only circumscribe different two overlapping areas, and these areas are judged respectively In order to remove the category after the common category, the final judgment result is obtained. For example, if the periphery of the cloud and snow overlapping area is a definite fog area, then the cloud and snow area is judged as a fog area; If the area is circumscribed, it is judged that the area is a snow area; if the cloud area determined by the cloud, snow and fog area is circumscribed, the area is determined to be a snow and fog area; ; If the cloud and snow area is circumscribed with the cloud and fog area, it is judged that the cloud and snow area and the cloud and fog area are snow area and fog area respectively.

In step 7.3, the judgment result is subjected to morphological "closed" operation, and the operation method is as described in step 6 to obtain the final cloud, snow, and fog detection results.

Step 8, secondary detection

Reselect 500 samples of cloud and ground objects, fog and ground objects, snow and ground objects to make a support vector machine classifier, and select highlighted samples for ground objects. Perform "secondary detection" on the sample to be tested, and compare the detection results of the secondary detection with the first detection results. If the two detection results at the same position are the same, then determine the category of the position as the category obtained from any one detection result , if the two detection results of the same position are different, it is determined that the category of the position is a ground object, and finally the detection result is obtained. For example, if the result of the second inspection is cloud or snow, and the result of the first inspection is fog, the area is determined to be a ground object. Only when the results of the first inspection and the second inspection are both clouds can the area be determined as a cloud.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (7)

1. The method for detecting the cloud and the snow fog of the optical satellite remote sensing image based on the support vector machine is characterized by comprising the following steps of:

step 1, collecting a large amount of cloud, snow, fog and ground object sample image data;

step 2, extracting gray features and texture features of various sample images to form feature vectors;

step 3, training the feature vectors of the sample images by using a support vector machine to respectively obtain a cloud image classifier, a snow image classifier and a fog image classifier which are formed by decision functions;

step 4, performing down-sampling processing on an original image of the satellite remote sensing image to be detected to obtain a thumbnail, performing image segmentation on the thumbnail to obtain sub-images, and calculating a feature vector consisting of gray features and texture features of all the sub-images;

step 5, classifying the sub-images of the remote sensing image of the satellite to be detected, comprising the following sub-steps,

step 5.1, respectively inputting the feature vectors extracted in the step 4 into the cloud, snow and fog image classifiers obtained in the step 3 for prediction classification;

step 5.2, dividing all the sub-images into a cloud area, a fog area, a snow area and a ground object area according to the types of the target areas;

step 5.3, dividing the cloud and ground feature area, the fog and ground feature area and the snow and ground feature area into three binary images, wherein the ground feature area in each image takes the same zero value, and the cloud, snow and fog areas take different image values;

step 6, performing morphological 'closing' operation on the classification result obtained in the step 5;

and 7, comparing three binary image values at the same position to obtain detection results of cloud, snow and fog in the remote sensing image of the satellite to be detected.

2. The method for detecting the cloud and snow fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 1, wherein: the implementation of said step 7 is as follows,

step 7.1, comparing three binary image values at the same position, and if the image values at the same position of the three images are the same, judging that the position is a ground object area; if two same values exist at the same position, the position is judged to be a category area represented by a third image value; if the image values of the three images at the same position are different, judging that the position is an overlapping area with cloud, snow and fog, recording a point with a zero value as the overlapping area, and recording a point without the zero value as a triple overlapping area;

7.2, repeating the step 7.1, comparing all image values of the three binary images to obtain discrimination results of cloud, snow and fog areas, ground object areas and overlapping areas, and correcting the overlapping areas; firstly, judging whether the overlapped area is contained in other areas, and if so, replacing the overlapped area with other areas; secondly, judging the type of the overlapping area, if the overlapping area is externally connected with a certain determined type area, judging the type of the overlapping area to be the type of the overlapping area except the determined externally connected type, and if not, confirming after the type of the overlapping area is judged to be finished; if the type of the overlapping area is circumscribed with the overlapping area, the type of the overlapping area is judged to be different from the type of the overlapping area circumscribed with the overlapping area; finally, judging the condition that the overlapping areas are only externally connected with different overlapping areas, respectively judging the areas as the types after the common types are removed, and finally obtaining a judgment result;

and 7.3, performing morphological 'closed' operation on the judgment result to obtain detection results of cloud, snow and fog in the remote sensing image of the satellite to be detected.

3. The method for detecting the cloud and fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 1 or 2, wherein: and 8, selecting a proper amount of cloud and ground object samples, fog and ground object samples, snow and ground object samples as training samples, repeating the steps 2-7 to carry out secondary detection on the satellite remote sensing image to be detected, comparing the detection result of the secondary detection with the first detection result, determining the type of the position as the type obtained by any one detection result if the two detection results of the same position are the same, and determining the type of the position as the ground object if the two detection results of the same position are different, thereby finally obtaining the detection result.

4. The method for detecting the cloud and the fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 3, wherein: the implementation of said step 2 is as follows,

step 2.1, calculating gray level characteristics of the sample image, including a gray level mean value, a gray level variance, a first order difference and a histogram information entropy of the sample image;

wherein the calculation formula of the gray level mean value is,

wherein f (i, j) is the gray value at (i, j), S = M × N, M is the width of the sample image, N is the height of the sample image;

the gray-scale variance is calculated by the formula,

the first order difference is calculated by the formula,

the calculation formula of the information entropy of the histogram is,

wherein, h [ g ] is the histogram of the sample image, h [ g ] (i) is the percentage of the pixel under a certain gray level to the whole sample image, M is the maximum gray level;

step 2.2, calculating texture features of the sample image, including gradient standard deviation, mixed entropy, inverse difference moment and texture fractional dimension of the sample image;

wherein the standard deviation of the gradient is calculated by the formula,

G(i,j；d,θ)＝#{(x ₁ ,y ₁ )(x ₂ ,y ₂ )|f(x ₁ ,y ₁ )＝i,f(x ₂ ,y ₂ )＝j,|(x ₁ ,y ₁ )-(x ₂ ,y ₂ )|＝d,∠((x ₁ ,y ₁ ),(x ₂ ,y ₂ ) Where d represents the distance between two pixels, theta represents the direction angle between pixels, f (x) ₁ ,y ₁ ) And f (x) ₂ ,y ₂ ) Are respectively provided withRepresents (x) ₁ ,y ₁ ) And (x) ₂ ,y ₂ ) The gray value of (c) represents the angle between the pixel point and the horizontal position, # represents the number of pixel pairs obtained according to the limiting conditions in the set, # L _x L _y Represents the sum of the numbers of all pixel pairs in a specific positional relationship; l is a radical of an alcohol _g Represents the maximum of the gray level, L represents the maximum of the gradient;

the calculation formula of the mixed entropy is that,

the calculation formula of the inverse difference is that,

the fractal Brownian random field method is used for solving the texture fractal dimension of the sample image, and the expression of the fractal dimension D of the image is,

D＝n+1-H

wherein n refers to the spatial dimension of the sample image, and H is a self-similarity parameter;

and 2.3, forming 8-dimensional feature vectors by the gray features and the texture features.

5. The method for detecting the cloud and the fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 4, wherein:

the implementation of said step 3 is as follows,

step 3.1, selecting partial cloud samples and ground feature samples as training samples, and using the feature vectors of the training samples as training sets T = { (x) of the training image classifier ₁ ,y ₁ )...(x _i ,y _i )}，i＝1…N，y _i E psi = { -1,1}, wherein 1 represents a positive class representing a cloud region class, -1 represents a negative class representing a ground feature region class, and x _i ∈R ⁿ ，x _i Is a feature vector, and N is the number of samples;

step 3.2, constructing a classification hyperplane by adopting a support vector machine of a C-SVC model, calculating a Gaussian kernel function,

wherein x is _i And x _j Respectively refer to the feature vectors of samples i and j, | | x _i -x _j || ² The square of Euclidean distance, sigma is variance, i =1 \ 8230, N, j =1 \ 8230, N, N is the number of samples;

step 3.3, solving Lagrange multiplier vector of the optimal classification hyperplane of the feature space by adopting a convex quadratic programming method,

wherein, 0 is less than or equal to alpha _i C, i =1,2 \8230n, lagrange multiplier vector, alpha _i And α j denote the ith and jth Lagrangian multipliers, x, respectively _i ,x _j Feature vectors, y, representing the ith and jth samples, respectively _i And y _j Respectively representing the types of the ith sample and the jth sample, wherein C is a penalty parameter, N is the number of samples, and the optimal solution of the Lagrange multiplier vector is obtained by:

α ^* ＝(α ₁ ^* ,α ₂ ^* ,...α _i ^* ...α _N ^* ) ^T

wherein alpha is _i ^* Represents an optimal solution for the ith Lagrangian multiplier;

step 3.4, solving the intercept of the optimal classification hyperplane of the feature space, wherein the calculation formula is as follows,

wherein alpha is _i ^* Is the optimal solution of the ith Lagrangian multiplier, y _i The type of the ith sample is shown, and N is the number of the samples;

step 3.5, substituting the obtained Gaussian kernel function, the Lagrange optimal solution and the hyperplane intercept into a decision function,

step 3.6, taking the rest cloud samples and ground object samples as test samples, testing the decision function, optimizing the decision function, and simultaneously obtaining corresponding cloud image classifiers;

and 3.7, repeating the steps 3.1-3.6 to respectively obtain a snow image classifier and a fog image classifier.

6. The method for detecting the cloud and fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 5, wherein: in the step 4, if the remote sensing image to be detected is a panchromatic image, the down-sampling processing is directly adopted, and if the remote sensing image to be detected is a multispectral image, the down-sampling is carried out by adopting RGB three-band.

7. The method for detecting the cloud and snow fog of the optical satellite remote sensing image based on the support vector machine as claimed in claim 6, wherein: the implementation manner of the step 6 is to select the structural elements with the square structural shape and the 3 x 3 structural size, respectively perform expansion operation on the three binary images, and perform corrosion operation on the obtained processed images by using the same structural elements.