Remote sensing image classification method based on improved Markov random field

Abstract

The invention discloses a remote sensing image classification method based on an improved Markov random field, and belongs to the technical field of remote sensing image object classification. The invention comprises the following steps: step one: determining the classification number of remote sensing images, and randomly classifying the remote sensing images; step two: based on the classical Markov random field model, considering the relationship between pixels in the observation field and the distance relationship between each pixel and the central pixel in the spatial neighborhood potential group, improving the prior energy function formula in the classical MRF algorithm; step three: based on the classical Markov random field model, considering the relationship between adjacent elements in the observation field, improving the likelihood energy function formula in the classical MRF algorithm; step four: using the improved prior energy and likelihood energy function formulas in step two and step three, solving the label field corresponding to the minimum global energy, and updating the classification label of the central pixel according to the solution result; step five: judging whether the pixel value mean of the classification label updated in step four converges.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image ground object classification, in particular to a remote sensing image classification method based on an improved Markov random field. Background Art

[0002] Remote sensing image object classification is the process of assigning an object category to each pixel in the image. It has always been a key and hot issue in remote sensing image analysis. Markov random field refers to a random variable that is only related to its adjacent random variables and has nothing to do with those non-adjacent random variables.

[0003] With the continuous development of remote sensing images with sub-meter pixels, the ground object information in the images is richer, the trend of different spectra for the same object and the same spectrum for different objects is strengthened, and the uncertainty of pixel categories is increased, resulting in great challenges for traditional spectral and pixel-by-pixel classification methods. The classification results often show misclassification and salt-and-pepper phenomena. Existing remote sensing image classification technologies are mainly divided into two categories: supervised classification and unsupervised classification. The defect of supervised classification is that it requires manual participation in sample selection and model training, the degree of automation is not high, and the classification accuracy depends on the number and quality of selected samples; unsupervised classification has a high degree of automation and does not require manual participation, but existing unsupervised classification algorithms such as the K-means algorithm and the fuzzy clustering algorithm do not consider the spatial information in remote sensing images, and the accuracy of remote sensing image classification is not high. Summary of the invention

[0004] The purpose of the present invention is to provide a remote sensing image classification method based on an improved Markov random field to solve the problems raised in the above background technology.

[0005] In order to solve the above technical problems, the present invention provides the following technical solution: the method comprises the following steps:

[0006] Step 1: Determine the number of remote sensing image classifications and randomly classify the remote sensing images;

[0007] Step 2: Based on the classical Markov random field model, the relationship between the pixels in the observation field and the distance relationship between each pixel in the spatial neighborhood potential group and the central pixel are considered, and the prior energy function formula in the classical MRF algorithm is improved, where MRF represents Markov random field;

[0008] Step 3: Based on the classic Markov random field model, the relationship between adjacent elements of the observation field is considered and the likelihood energy function formula in the classic MRF algorithm is improved;

[0009] Step 4: Use the improved prior energy and likelihood energy function formulas in steps 2 and 3 to solve the label field corresponding to the minimum global energy, and update the classification label of the central pixel according to the solution result;

[0010] Step 5: Determine whether the mean pixel value of the classification label updated in step 4 converges. If not, repeat steps 2 to 5. Otherwise, output the classification result.

[0011] Furthermore, in the step 1, the number of remote sensing image classifications is determined, and the specific method for randomly initializing the remote sensing image classification is as follows:

[0012] Step 1 (I). Select an N×M remote sensing image and classify it into L regions. The pixel set of the remote sensing image is S = {(i, j)|1≤i≤N,1≤j≤M}, where i represents the pixel height of the remote sensing image, j represents the pixel width of the remote sensing image, and (i, j) represents the region size of the intercepted remote sensing image;

[0013] Step 1 (II). Use the label field and observation field in the MRF model to describe the remote sensing image, then the label field X = {x s |s∈S}, observation field Y={y s |s∈S}, where x s Represents the pixel value of the pixel in the remote sensing image label field s area, y s It represents the central pixel value in the remote sensing image observation field s area. The label field uses the prior distribution to describe the local correlation of pixels, and the observation field uses the conditional distribution function to describe the distribution of observation data.

[0014] Step 1 (III). Determine the number of remote sensing image categories according to actual needs, limit the value range generated by the random function rend based on the number of categories, use the random function rand to randomly generate a value, and assign the generated value to a pixel in the labeling field until all pixels in the labeling field are labeled.

[0015] Furthermore, the specific method for improving the prior energy function formula in the classic MRF algorithm in step 2 is:

[0016] Each pixel in the neighborhood of the remote sensing image has the same influence on the central pixel. The prior energy function in the classic MRF algorithm does not consider the distance between each pixel in the neighborhood and the central pixel. According to the first law of geography, objects with similar spatial distances are more likely to belong to the same category, and the closer the neighborhood pixel is to the central pixel, the greater its influence on the central pixel. Therefore, the improved parameter β introduces the distance weight coefficient d. At the same time, the prior energy function in the classic MRF algorithm only considers the relationship between the label field, and does not consider the relationship between the pixels in the observation field. The improved parameter β also needs to introduce the influence coefficient α of the observation field pixels on the central pixel label.

[0017] Step 2 (I). Introduce the distance weight coefficient d to improve the parameters on the spatial neighborhood potential group in the classic MRF algorithm. The specific formula of the distance weight coefficient d is:

[0018]

[0019] The parameter β on the improved spatial neighborhood potential group is:

[0020] β = β′*d;

[0021] Wherein, z[i] represents the neighborhood pixel set of the central pixel, i = 1, 2, 3…, represents the number of neighborhood pixels of the central pixel, β′ represents the parameter on the spatial neighborhood potential group in the classic MRF algorithm, β′>0;

[0022] Step 2 (II). Introduce the influence coefficient α of the observation field pixel on the center pixel label, and improve the prior energy function formula in the classic MRF algorithm. The specific formula of the influence coefficient α is:

[0023]

[0024] Among them, y z[i] represents the pixel value of the central pixel neighborhood, i = 1, 2, ..., n, represents the number of pixels in the central pixel neighborhood, y s represents the central pixel value in the remote sensing image observation field s area, s represents the area in the remote sensing image, |y Z[i] -y s | represents the absolute value of the difference between the i-th neighborhood pixel value and the central pixel value, Represents the sum of the differences between n neighboring pixel values ​​and the central pixel value;

[0025] Step 2 (III). Based on the improved parameters of step 2 (I) and step 2 (II), the prior energy function formula in the classic MRF algorithm is improved. The improved prior energy function formula U1(x) is:

[0026]

[0027]

[0028] Among them, V c (x) represents the potential energy function defined on the group C, x represents the label field, C represents the group set on the remote sensing image, represents the prior energy function defined on the set in δ, δ represents the set of universal neighborhood systems defined on the remote sensing image pixel set, c represents a group, all elements in c belong to the remote sensing image pixel set S, and each pair of different positions in c is always adjacent, x s Represents the pixel value of the pixel in the remote sensing image label field s area, xt Represents the pixel value of the pixel in the remote sensing image label field t area, s, t represent the area in the remote sensing image, t is the adjacent area of ​​s, when x s =x t When x s ≠x t When , it means that the energy between the s and t regions depends on d, α, and β′.

[0029] Furthermore, the likelihood energy function formula U′2(x,y) in the classic MRF algorithm is constructed. The specific construction method is:

[0030] Step 1: According to Bayes' rule, we can get: P X|Y (x|y)=P X (x)P Y|X (y|x) / P Y (y), due to P Y (y) is a constant given by the observation field in the Markov random field, so it can be discarded, then P X|Y (x|y) depends on P X (x)P Y|X (y|x);

[0031] Step 2: Based on the equivalence principle between MRF and Gibbs random field, Gibbs random field represents Gibbs distribution, which refers to the form of a product of a series of non-negative functions on the largest group (s) of G in the undirected graph model. It can be seen that:

[0032] P X (x) = Z -1 [exp(-U1(x) / T)];

[0033] Where T is a constant, is the normalization constant, is the prior energy function defined on the set in δ, V c (x) represents the potential energy function defined on the group C, where C represents the group set on the remote sensing image;

[0034] Step 3: Based on Step 1 and Step 2, the Gaussian function is used to describe the grayscale attributes of the remote sensing image. The specific Gaussian function expression P(y s |x s )for:

[0035]

[0036] in, Represents the mean value of the pixel values ​​corresponding to each class label, Represents the standard deviation of the pixel values ​​corresponding to each class label, ys represents the central pixel value in the remote sensing image observation field s, x s represents the pixel value of the pixel in the area of ​​the label field s;

[0037] Step 4: Based on the Gaussian function expression in Step 3, under a given label field x, the classical likelihood energy function formula U′2(x,y) is obtained as follows:

[0038]

[0039] Furthermore, the specific method for improving the likelihood energy function formula in the classic MRF algorithm in step 3 is:

[0040] Regarding the improvement of the likelihood energy function, the classic Markov likelihood energy function only considers the relationship between the corresponding elements of the observation field and the label field, and does not consider the relationship between the adjacent elements of the observation field. For example, when the likelihood energy of a certain feature class is very small, it does not mean that the reliability of this feature is very high. It is also necessary to refer to the difference between its pixel point and the surrounding pixel points. This difference needs to be added as a parameter to the energy term to participate in the solution to describe the reliability of the pixel point.

[0041] Step 3 (I). Use the Laplace operator value to express the relationship between the elements of the observation field. The relationship between the elements of the observation field is expressed by for:

[0042]

[0043] Among them, s represents the area in the remote sensing image, I s represents the Laplace operator value at point s, H bench represents the Laplace operator reference value, Represents the relationship value between the elements of the observation field, which is used to describe the reliability of the pixels in the observation field;

[0044] Step 3 (II). Based on the parameters obtained in step 3 (I), the classical likelihood energy function formula U′2(x, y) obtained in step 4 is improved. The improved likelihood energy function formula U2(x, y) is:

[0045]

[0046] in, Represents the mean value of the pixel values ​​corresponding to each class label, It represents the standard deviation of the pixel values ​​corresponding to each category of labels, and S represents the pixel set of the remote sensing image.

[0047] Furthermore, in step 4, the label field corresponding to the minimum global energy is solved, and the specific method for updating the classification label of the central pixel according to the solution result is:

[0048] Step 4 (I). Since the correlation between pixel labels depends on the neighborhood system defined on S, in order to estimate the best label from the observation field Y The image classification problem is transformed into a maximum a posteriori probability problem, and a solution model for the optimal label field is constructed based on this principle.

[0049] Step 4 (II). Based on Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), the model Convert to solve the minimum value model;

[0050] Step 4 (III). Compare the solution results of step 4 (II), and replace the label of the central pixel with the label corresponding to the minimum value according to the comparison result.

[0051] Furthermore, the optimal labeling field solution model is constructed in step 4 (I) and step 4 (II) And the model The specific method for converting to solving the minimum value model is:

[0052] Constructing the optimal labeling field solution model Model The specific formula is:

[0053]

[0054] in, represents the best label field in remote sensing images, argmax represents the function of finding the parameters of the function;

[0055] From Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), we can know that P X|Y The (x|y) formula can be expressed as:

[0056] P X|Y (x|y)=P X (x)P Y|X (y|x)=Z -1 [exp(-U1(x) / T)]*[exp(-U2(x,y)];

[0057] right Taking the logarithm of both sides gives:

[0058]

[0059] Solved:

[0060]

[0061] Among them, argmin represents the variable value when the objective function U(x,y) takes the minimum value, U1(x) represents the improved prior energy function, U2(x,y) represents the improved likelihood energy function, and x represents the label of the label field.

[0062] Furthermore, the specific method for determining whether the updated pixel value mean of the classification label converges in step 5 is:

[0063] Step 5 (I). Calculate the difference between the mean values ​​of the pixels of each category label before and after, and compare the calculated result with the set threshold. If the calculated result is within the threshold interval, it is determined that the updated mean value of the pixels of the category label has converged, and the classification result is output. Otherwise, it does not converge. The set threshold is [0,1];

[0064] Step 5 (II). Determine whether the mean pixel value of each category label has changed. If not, determine whether the mean pixel value of the updated category label has converged. At this time, output the classification result. Otherwise, it has not converged.

[0065] Furthermore, the neighborhood types in the remote sensing image include four nearest neighbors and eight nearest neighbors.

[0066] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0067] 1. Compared with the supervised classification method, the present invention does not require manual participation in sample selection, and the classification is fully automated with a high degree of automation. Compared with traditional unsupervised classification methods such as the K-means method and the fuzzy clustering method, the present invention makes full use of the spatial information in the remote sensing image, making the image classification accuracy higher.

[0068] 2. Compared with the classical Markov classification method, the present invention not only considers the prior information between the pixels in the Markov label field and the likelihood information between the pixels in the observation field and the label field, but also makes full use of the relationship between the neighboring pixels in the observation field, enhances the interference to the noise, achieves better classification effect, and further shortens the solution time.

[0069] 3. The present invention converts the image classification problem into a maximum a posteriori probability problem to solve the optimal labeling of the observation field, and combines the improved likelihood energy function and the prior energy function to convert the optimal labeling field model into a minimum value model. There is no need to manually compare the solution results, and the error probability is reduced, thereby further improving the remote sensing image classification effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0071] Figure 1 It is a workflow diagram of the remote sensing image classification method based on improved Markov random field of the present invention;

[0072] Figure 2 It is a schematic diagram of a relatively simple original image of a building image of the remote sensing image classification method based on an improved Markov random field of the present invention;

[0073] Figure 3 It is a relatively simple reference schematic diagram of building image classification based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0074] Figure 4 It is a schematic diagram of the fuzzy classification results of a relatively simple building image based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0075] Figure 5 It is a schematic diagram of the classic MRF classification results of a relatively simple building image based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0076] Figure 6 It is a schematic diagram of improved Markov classification results of a relatively simple building image based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0077] Figure 7 It is a schematic diagram of the original image of a relatively complex building image of the remote sensing image classification method based on the improved Markov random field of the present invention;

[0078] Figure 8 It is a reference schematic diagram of complex building image classification based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0079] Fig. 9 It is a schematic diagram of the fuzzy classification results of a relatively complex building image based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0080] Fig.10 It is a schematic diagram of the classical MRF classification results of a relatively complex building image based on the remote sensing image classification method of the present invention based on the improved Markov random field;

[0081] Fig.11 It is a schematic diagram of improved Markov classification results of a relatively complex building image based on the remote sensing image classification method of the present invention based on the improved Markov random field. DETAILED DESCRIPTION

[0082] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0083] See also Figure 1 The present invention provides a technical solution: the method comprises the following steps:

[0084] Step 1: Determine the number of remote sensing image classifications and randomly classify the remote sensing images. The specific method is as follows:

[0085] Step 1 (I). Select an N×M remote sensing image and classify it into L regions. The pixel set of the remote sensing image is S = {(i, j)|1≤i≤N,1≤j≤M}, where i represents the pixel height of the remote sensing image, j represents the pixel width of the remote sensing image, and (i, j) represents the region size of the intercepted remote sensing image;

[0086] Step 1 (II). Use the label field and observation field in the MRF model to describe the remote sensing image, then the label field X = {x s |s∈S}, observation field Y={y s |s∈S}, where x s Represents the pixel value of the pixel in the remote sensing image label field s area, y s It represents the central pixel value in the remote sensing image observation field s area. The label field uses the prior distribution to describe the local correlation of pixels, and the observation field uses the conditional distribution function to describe the distribution of observation data.

[0087] Step 1 (III). Determine the number of remote sensing image classifications according to actual needs, limit the value range generated by the random function rend based on the number of classifications, use the random function rand to randomly generate a value, and assign the generated value to a pixel in the labeling field until all pixels in the labeling field are labeled;

[0088] Step 2: Based on the classical Markov random field model, the relationship between the pixels in the observation field and the distance relationship between each pixel and the central pixel in the spatial neighborhood potential group are considered, and the prior energy function formula in the classical MRF algorithm is improved. MRF represents the Markov random field. The specific method of improvement is:

[0089] Each pixel in the neighborhood of the remote sensing image has the same influence on the central pixel. The prior energy function in the classic MRF algorithm does not consider the distance between each pixel in the neighborhood and the central pixel. According to the first law of geography, objects with similar spatial distances are more likely to belong to the same category, and the closer the neighborhood pixel is to the central pixel, the greater its influence on the central pixel. Therefore, the improved parameter β introduces the distance weight coefficient d. At the same time, the prior energy function in the classic MRF algorithm only considers the relationship between the label field, and does not consider the relationship between the pixels in the observation field. The improved parameter β also needs to introduce the influence coefficient α of the observation field pixels on the central pixel label.

[0090] Step 2 (I). Introduce the distance weight coefficient d to improve the parameters on the spatial neighborhood potential group in the classic MRF algorithm. The specific formula of the distance weight coefficient d is:

[0091]

[0092] The parameter β on the improved spatial neighborhood potential group is:

[0093] β = β′*d;

[0094] Wherein, z[i] represents the neighborhood pixel set of the central pixel, i = 1, 2, 3…, represents the number of neighborhood pixels of the central pixel, β′ represents the parameter on the spatial neighborhood potential group in the classic MRF algorithm, β′>0;

[0095] Step 2 (II). Introduce the influence coefficient α of the observation field pixel on the center pixel label, and improve the prior energy function formula in the classic MRF algorithm. The specific formula of the influence coefficient α is:

[0096]

[0097] Among them, y z[i] represents the pixel value of the central pixel neighborhood, i = 1, 2, ..., n, represents the number of pixels in the central pixel neighborhood, y s represents the central pixel value in the remote sensing image observation field s area, s represents the area in the remote sensing image, |y Z[i] -y s | represents the absolute value of the difference between the i-th neighborhood pixel value and the central pixel value, Represents the sum of the differences between n neighboring pixel values ​​and the central pixel value;

[0098] Step 2 (III). Based on the improved parameters of step 2 (I) and step 2 (II), the prior energy function formula in the classic MRF algorithm is improved. The improved prior energy function formula U1(x) is:

[0099]

[0100]

[0101] Among them, V c (x) represents the potential energy function defined on the group C, x represents the label field, C represents the group set on the remote sensing image, represents the prior energy function defined on the set in δ, δ represents the set of universal neighborhood systems defined on the remote sensing image pixel set, c represents a group, all elements in c belong to the remote sensing image pixel set S, and each pair of different positions in c is always adjacent, x s Represents the pixel value of the pixel in the remote sensing image label field s area, x t Represents the pixel value of the pixel in the remote sensing image label field t area, s, t represent the area in the remote sensing image, t is the adjacent area of ​​s, when x s =x t When x s ≠x t When , it means that the energy between the s and t regions depends on d, α, and β′;

[0102] Step 3: Based on the classic Markov random field model, considering the relationship between adjacent elements of the observation field, the likelihood energy function formula in the classic MRF algorithm is improved. The specific method is:

[0103] Construct the likelihood energy function formula U′2(x,y) in the classic MRF algorithm. The specific construction method is:

[0104] Step 1: According to Bayes' rule, we can get: P X|Y (x|y)=P X (x)P Y|X (y|x) / P Y (y), due to P Y (y) is a constant given by the observation field in the Markov random field, so it can be discarded, then P X|Y (x|y) depends on P X (x)P Y|X (y|x);

[0105] Step 2: Based on the equivalence principle between MRF and Gibbs random field, Gibbs random field represents Gibbs distribution, which refers to the form of a product of a series of non-negative functions on the largest group (s) of G in the undirected graph model. It can be seen that:

[0106] P X (x) = Z -1 [exp(-U1(x) / T)];

[0107] Where T is a constant, is the normalization constant, is the prior energy function defined on the set in δ, V c (x) represents the potential energy function defined on the group C, where C represents the group set on the remote sensing image;

[0108] Step 3: Based on Step 1 and Step 2, the Gaussian function is used to describe the grayscale attributes of the remote sensing image. The specific Gaussian function expression P(y s |x s )for:

[0109]

[0110] in, Represents the mean value of the pixel values ​​corresponding to each class label, Represents the standard deviation of the pixel values ​​corresponding to each class label, y s represents the central pixel value in the remote sensing image observation field s, x s represents the pixel value of the pixel in the area of ​​the label field s;

[0111] Step 4: Based on the Gaussian function expression in Step 3, under a given label field x, the classical likelihood energy function formula U′2(x,y) is obtained as follows:

[0112]

[0113] The specific method to improve the likelihood energy function formula in the classic MRF algorithm is:

[0114] Regarding the improvement of the likelihood energy function, the classic Markov likelihood energy function only considers the relationship between the corresponding elements of the observation field and the label field, and does not consider the relationship between the adjacent elements of the observation field. For example, when the likelihood energy of a certain feature class is very small, it does not mean that the reliability of this feature is very high. It is also necessary to refer to the difference between its pixel point and the surrounding pixel points. This difference needs to be added as a parameter to the energy term to participate in the solution to describe the reliability of the pixel point.

[0115] Step 3 (I). Use the Laplace operator value to express the relationship between the elements of the observation field. The relationship between the elements of the observation field is expressed by for:

[0116]

[0117] Among them, s represents the area in the remote sensing image, I s represents the Laplace operator value at point s, H bench represents the Laplace operator reference value, Represents the relationship value between the elements of the observation field, which is used to describe the reliability of the pixels in the observation field;

[0118] Step 3 (II). Based on the parameters obtained in step 3 (I), the classical likelihood energy function formula U′2(x, y) obtained in step 4 is improved. The improved likelihood energy function formula U2(x, y) is:

[0119]

[0120] in, Represents the mean value of the pixel values ​​corresponding to each class label, It represents the standard deviation of the pixel values ​​corresponding to each type of label, and S represents the pixel set of the remote sensing image;

[0121] Step 4: Use the improved prior energy and likelihood energy function formulas in steps 2 and 3 to solve the label field corresponding to the minimum global energy, and update the classification label of the central pixel according to the solution. The specific method is:

[0122] Step 4 (I). Since the correlation between pixel labels depends on the neighborhood system defined on S, in order to estimate the best label from the observation field Y The image classification problem is transformed into a maximum a posteriori probability problem, and a solution model for the optimal label field is constructed based on this principle. Model The specific formula is:

[0123]

[0124] in, represents the best label field in remote sensing images, argmax represents the function of finding the parameters of the function;

[0125] Step 4 (II). Based on Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), the model Convert to solve the minimum value model, the specific method is:

[0126] From Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), we can know that P X|Y The (x|y) formula can be expressed as:

[0127] P X|Y (x|y)=P X (x)P Y|X (y|x)=Z -1 [exp(-U1(x) / T)]*[exp(-U2(x,y)];

[0128] right Taking the logarithm of both sides gives:

[0129]

[0130] Solved:

[0131]

[0132] Among them, argmin represents the variable value when the objective function U(x,y) takes the minimum value, U1(x) represents the improved prior energy function, U2(x,y) represents the improved likelihood energy function, and x represents the label of the label field;

[0133] Step 4 (III). Compare the solution results of step 4 (II), and replace the label of the central pixel with the label corresponding to the minimum value according to the comparison result;

[0134] Step 5: Determine whether the mean pixel value of the classification label updated in step 4 converges. If not, repeat steps 2 to 5. Otherwise, output the classification result. The specific method is:

[0135] Step 5 (I). Calculate the difference between the mean values ​​of the pixels of each category label before and after, and compare the calculated result with the set threshold. If the calculated result is within the threshold interval, it is determined that the updated mean value of the pixels of the category label has converged, and the classification result is output. Otherwise, it does not converge. The set threshold is [0,1];

[0136] Step 5 (II). Determine whether the mean pixel value of each category label has changed. If not, determine whether the mean pixel value of the updated category label has converged. At this time, output the classification result. Otherwise, it has not converged.

[0137] Embodiment 1: A remote sensing image is selected and divided into three categories. A pixel and its neighboring pixels in the selected remote sensing image are solved as an example. The label field and the corresponding observation field are shown in the following table:

[0138]

[0139] Calculate the prior energy and likelihood energy of the first category, i.e., the remote sensing image labeled 1. When calculating the prior energy of the image labeled 1, assume that the center pixel is labeled 1 instead of 3. Then compare the 8 neighboring pixels with the center pixel, based on the prior energy calculation formula: Combining the display distribution of the numbered field and the observation field, we can see that:

[0140] The labels of the neighboring pixels in the northwest direction, the neighboring pixels in the north direction, the neighboring pixels in the east direction, and the neighboring pixels in the southwest direction are all 1, which is the same as the label of the central pixel (the label of the central pixel is assumed to be 1 at this moment), then:

[0141] V c (1) = 0;

[0142] The pixel number of the northeast neighbor is 3, which is different from the center pixel number 1. Then:

[0143]

[0144]

[0145] Among them, 0.5 is a fixed value of β, is the value of the distance weight coefficient d. Since the neighboring pixel is located in the northeast direction, the value is 0.23 is the influence coefficient α of the observation field pixel on the center pixel label;

[0146] Similarly, the pixel number of the neighboring pixel in the west direction is 2, which is different from the center pixel number 1. Then:

[0147]

[0148] V c (1)=d*α*β′=0.5*1*0.16;

[0149] The south-facing neighbor pixel number 2 is different from the center pixel number 1, then:

[0150]

[0151] V c (1)=d*α*β′=0.5*1*0.176;

[0152] The southeast neighbor pixel number is 3, which is different from the center pixel number 1. Then:

[0153]

[0154]

[0155] In summary, the prior energy labeled 1 is:

[0156]

[0157] Based on the likelihood energy calculation formula, the likelihood energy of the number 1 is calculated, then:

[0158] First calculate the Laplace operator value I s , and then calculate the mean pixel value of the pixels labeled 1 in the entire remote sensing image and variance Substitute the calculation results into the likelihood energy calculation formula: We get U2(1,y)=0.4221;

[0159] Similarly, the prior energy and likelihood energy of labels 2 and 3 are calculated as follows: U1(2) = 0.4551, U2(2,y) = 0.4362; U1(3) = 0.4453, U2(3,y) = 0.4437, then the global energy of labels 1, 2, 3 is:

[0160] U(1,y)=U1(1)+U2(1,y)=0.8533;

[0161] U(2,y)=U1(2)+U2(2,y)=0.8913;

[0162] U(3,y)=U1(3)+U2(3,y)=0.889;

[0163] Compare the three global energies. When the global energy is the smallest, the label is 1, and the label of the central pixel is updated from the original 3 to 1.

[0164] Example 2: Please refer to Figure 2-11 ,In order to verify the performance of the remote sensing image classification method based on the improved Markov random field model, ,real remote sensing images were selected for experiments. ,Two images in the Merced Land Use Dataset scene classification dataset were selected for ,verification, and the remote sensing image classification method of the improved Markov ,random field model was compared with the classical Markov classification method ,and the fuzzy clustering method. ,Merced Land Use Dataset is a 21-class land use image dataset for ,research purposes, with 100 images in each class. ,These images are manually extracted from the large images of the USGU ,National Map Urban Area Image Set. ,These images are used in urban areas across the country. ,The pixel resolution of this public dataset is about 0.3m, and the size of each ,image is 256x256 pixels;

[0165] Experiment 1: A relatively simple building image was selected, and the fuzzy clustering, classic MRF, and improved MRF algorithms were used to classify the selected relatively simple building image. From the classification calculation results shown in Table 1, it can be seen that the three algorithms have good classification of buildings and roads, and the user accuracy and producer accuracy are both above 90%. For buildings, the improved MRF algorithm has higher producer accuracy and user accuracy than the fuzzy clustering and classic MRF algorithms. The user accuracy is improved by 0.6% and 0.3% respectively compared with the results obtained by the fuzzy clustering and classic MRF algorithms. , the producer accuracy is improved by 0.9% and 0.7% respectively; for roads, the improved MRF algorithm has higher classification accuracy than fuzzy clustering and classic MRF algorithms, and the user accuracy is improved by 1.8% and 6.7% respectively compared with the results of fuzzy clustering and classic MRF algorithms, and the producer accuracy is improved by 1.1% and 0.5% respectively; for building shadows, the user accuracy of the improved MRF algorithm is improved by 2.7% and 1.9% respectively compared with fuzzy clustering and classic MRF algorithms, and the producer accuracy is improved by 3.4% and 2.6% respectively, as shown in Table 1:

[0166] Table 1 Confusion matrix of classification results of Experiment 1

[0167]

[0168] Among them, PA represents producer accuracy and UA represents user accuracy;

[0169] Experimental verification shows that for simpler building images, the improved MRF algorithm outperforms the fuzzy clustering and classic MRF algorithms in both user accuracy and producer accuracy.

[0170] Experiment 2: A more complex building image is selected, and the fuzzy clustering, classic MRF and improved MRF algorithms are used to classify the selected complex building image. From the classification calculation results shown in Table 2, it can be seen that for buildings, the improved MRF algorithm has better producer accuracy and user accuracy than the other two algorithms. Compared with the fuzzy clustering and classic MRF algorithms, the user accuracy is improved by 1.2% and 0.49%, and the producer accuracy is improved by 3.22% and 2.46%, respectively; for roads, the improved MRF algorithm has higher classification accuracy than the fuzzy clustering and classic MRF algorithms. Compared with the fuzzy clustering and classic MRF algorithms, the user accuracy is improved by 16.81% and 11.94%, respectively, and the producer accuracy is improved by 8.99% and 3.81%, respectively; for vegetation, the improved MRF algorithm improves the user accuracy by 1.65% and 0.68% compared with the fuzzy clustering and classic MRF algorithms, and the producer accuracy is improved by 4.61% and 3.08%, respectively. Table 2 is shown below:

[0171] Table 2 Confusion matrix of classification results of Experiment 2

[0172]

[0173] Experimental verification shows that for more complex building images, the improved MRF algorithm is superior to the fuzzy clustering and classic MRF algorithms in terms of both user accuracy and producer accuracy.

[0174] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0175] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A remote sensing image classification method based on an improved Markov random field, characterized by: The method comprises the following steps: Step 1: Determine the number of remote sensing image classifications and randomly classify the remote sensing images; Step 2: Improve the prior energy function formula in the classic MRF algorithm. The improved prior energy function formula U1(x) is: V c (x) represents the potential energy function defined on the group C, x represents the label field, C represents the group set on the remote sensing image, represents the prior energy function defined on the set in δ, δ represents the set of universal neighborhood systems defined on the remote sensing image pixel set, c represents a group, all elements in c belong to the remote sensing image pixel set S, and each pair of different positions in c is always adjacent. s =x t When x s ≠x t When , it means that the energy between the s and t regions depends on d, α, and β′; s and t represent the regions in the remote sensing image, t is the adjacent region of s, and x s Represents the pixel value of the pixel in the remote sensing image label field s area, x t represents the pixel value of the pixel in the remote sensing image label field t area, d represents the weight coefficient, β′ represents the parameter on the spatial neighborhood potential group in the classic MRF algorithm and β′>0, α represents the influence coefficient of the observation field pixel on the central pixel label, z[i] represents the neighborhood pixel set of the central pixel point, i=1, 2, 3…, represents the number of neighborhood pixels of the central pixel point, y z[i] represents the pixel value of the central pixel neighborhood, i = 1, 2, ..., n, represents the number of pixels in the central pixel neighborhood, y s represents the central pixel value in the remote sensing image observation field s area, |y Z[i] -y s | represents the absolute value of the difference between the i-th neighborhood pixel value and the central pixel value, It represents the sum of the differences between the values ​​of n neighboring pixels and the central pixel value, and α represents the influence coefficient of the introduced observation field pixels on the central pixel label; Step 3: Improve the likelihood energy function formula in the classic MRF algorithm. The improved likelihood energy function formula U2(x,y) is: Among them, x s represents the pixel value of the pixel in the area of ​​the label field s, Represents the mean value of the pixel values ​​corresponding to each class label, represents the standard deviation of the pixel values ​​corresponding to each class of labels, S represents the pixel set of the remote sensing image, Represents the relationship value between the elements of the observation field, which is used to describe the reliability of the pixels in the observation field. s represents the Laplace operator value at point s, H bench represents the Laplace operator reference value; Step 4: Use the improved prior energy and likelihood energy function formulas in steps 2 and 3 to solve the label field corresponding to the minimum global energy, and update the classification label of the central pixel according to the solution result; Step 5: Determine whether the mean pixel value of the classification label updated in step 4 converges. If not, repeat steps 2 to 5. Otherwise, output the classification result.

2. The remote sensing image classification method based on improved Markov random field according to claim 1, characterized in that: In the step 1, the number of remote sensing image classifications is determined, and the specific method for randomly initializing the remote sensing image classification is as follows: Step 1 (I). Select an N×M remote sensing image and classify it into L regions. The pixel set of the remote sensing image is S = {(i, j)|1≤i≤N,1≤j≤M}, where i represents the pixel height of the remote sensing image, j represents the pixel width of the remote sensing image, and (i, j) represents the region size of the intercepted remote sensing image; Step 1 (II). Use the label field and observation field in the MRF model to describe the remote sensing image, then the label field X = {x s |s∈S}, observation field Y={y s |s∈S}, where x s Represents the pixel value of the pixel in the remote sensing image label field s area; Step 1 (III). Determine the number of remote sensing image categories according to actual needs, limit the value range generated by the random function rend based on the number of categories, use the random function rand to randomly generate a value, and assign the generated value to a pixel in the labeling field until all pixels in the labeling field are labeled.

3. The remote sensing image classification method based on improved Markov random field according to claim 2, characterized in that: Construct the likelihood energy function formula U2′(x,y) in the classic MRF algorithm. The specific construction method is: Step 1: According to Bayes' rule, we can get: P X|Y (x|y)=P X (x)P Y|X (y|x) / P Y (y), due to P Y (y) is a constant given by the observation field in the Markov random field, so it can be discarded, then P X|Y (x|y) depends on P X (x)P Y|X (y|x); Step 2: Based on the principle that MRF and Gibbs random field are equivalent, we can know that: P X (x)=Z -1 [exp(-U1(x) / T)]; Where T is a constant, is the normalization constant, is the prior energy function defined on the set in δ; Step 3: Based on Step 1 and Step 2, the Gaussian function is used to describe the grayscale attributes of the remote sensing image. The specific Gaussian function expression P(y s |x s )for: in, Represents the mean value of the pixel values ​​corresponding to each class label; Step 4: Based on the Gaussian function expression in Step 3, under a given label field x, the classical likelihood energy function formula U2′(x,y) is obtained as follows: based on The classic likelihood energy function formula U2′(x,y) obtained in Step 4 is improved, and the improved likelihood energy function formula is U2(x,y).

4. The remote sensing image classification method based on improved Markov random field according to claim 3 is characterized in that: The specific method of solving the label field corresponding to the minimum global energy in step 4 and updating the classification label of the central pixel according to the solution result is: Step 4 (I). Since the correlation between pixel labels depends on the neighborhood system defined on S, in order to estimate the best label from the observation field Y The image classification problem is transformed into a maximum a posteriori probability problem, and a solution model for the optimal label field is constructed based on this principle. Step 4 (II). Based on Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), the model Convert to solve the minimum model; Step 4 (III). Compare the solution results of step 4 (II), and replace the label of the central pixel with the label corresponding to the minimum value according to the comparison result.

5. The remote sensing image classification method based on improved Markov random field according to claim 4, characterized in that: The solution model for constructing the optimal labeling field in step 4 (I) and step 4 (II) And the model The specific method for converting to solving the minimum value model is: Constructing the optimal labeling field solution model Model The specific formula is: in, represents the best label field in remote sensing images, argmax represents the function of finding the parameters of the function; From Step 1, Step 2 and the improved likelihood energy function formula U2(x,y), we can know that P X|Y The (x|y) formula can be expressed as: P X|Y (x|y)=P X (x)P Y|X (y|x)=Z -1 [exp(-U1(x) / T)]*[exp(-U2(x,y)]; right Taking the logarithm of both sides gives: Solved: Among them, arg min represents the variable value when the objective function U(x,y) takes the minimum value, U1(x) represents the improved prior energy function, U2(x,y) represents the improved likelihood energy function, and x represents the label of the label field.

6. The remote sensing image classification method based on improved Markov random field according to claim 5, characterized in that: The specific method for judging whether the updated pixel value mean of the classification label converges in step 5 is: Step 5 (I). Calculate the difference between the mean values ​​of the pixels of each category label before and after, and compare the calculated result with the set threshold. If the calculated result is within the threshold interval, it is determined that the updated mean value of the pixels of the category label has converged, and the classification result is output. Otherwise, it does not converge. The set threshold is [0,1]; Step 5 (II). Determine whether the mean pixel value of each category label has changed. If not, determine whether the mean pixel value of the updated category label has converged. At this time, output the classification result. Otherwise, it has not converged.

7. The remote sensing image classification method based on improved Markov random field according to claim 6, characterized in that: The neighborhood types in the remote sensing image include four nearest neighbors and eight nearest neighbors.

Images