CN104123561B - Fuzzy C-mean algorithm remote sensing image automatic classification method based on spatial attraction model - Google Patents

Abstract

An automatic classification method of fuzzy C-means remote sensing images based on a spatial gravity model, which is suitable for automatic segmentation and classification of remote sensing images, medical images and other images. The steps are: determine the number of pixels in the remote sensing digital image, and use the standard FCM model to cluster the image to obtain initialization, and then calculate the space gravity and space constraint penalty between each pixel and other pixels in its neighborhood window in turn factor, and finally get the fuzzy factor By adding the fuzz factor Add the standard FCM model to get a new clustering objective function, and loop to find the fuzzy matrix and cluster centers Until the cluster center does not continue to change or the operation reaches the maximum number of iterations, then use the finally obtained fuzzy membership matrix U={u _ki } _c×N , and use the criterion of maximizing the membership matrix to process each pixel of the remote sensing image Points are marked by category to determine the category described by each pixel, forming a remote sensing image classification thematic map, thereby realizing automatic classification of remote sensing digital images. The method is simple, highly automated, less affected by image noise, and has high image segmentation and classification accuracy.

Description

Fuzzy C-Means Remote Sensing Image Automatic Classification Method Based on Spatial Gravity Model

technology neighborhood

The invention relates to an automatic classification method for remote sensing images, in particular to a fuzzy C-mean automatic classification method for remote sensing images based on a spatial gravity model used in the automatic segmentation and classification of remote sensing images, medical images and other images.

Background technique

Remote sensing data classification is an important technology for extracting thematic category data from remote sensing data, which provides a rich source of data for information extraction in various industries. The current classification methods are mainly divided into supervised classification and unsupervised classification methods. Compared with supervised classification methods, unsupervised classification methods can extract information from remote sensing data without prior knowledge. classification has a very important position.

The existing unsupervised classification methods mainly include methods such as ISODATA, KNN, K-means, Markov random field (MRF) and fuzzy C-means clustering algorithm (FCM). Among them, FCM is a prototype-based clustering algorithm, which has the characteristics of simplicity, high efficiency, and strong data adaptability. Compared with the hard classification method, FCM can obtain the membership degree of each pixel belonging to each category, and can be used as much as possible. It is possible to retain as much image information as possible, and it is more suitable for representing remote sensing images with a large number of mixed pixels. The standard FCM algorithm was proposed by Dune and later promoted by Bezdek, which is an iterative optimization method. However, the influence between adjacent pixels is not considered in the standard FCM clustering process, which will cause large errors when segmenting images with low signal-to-noise ratios. In order to overcome the problem that the standard FCM algorithm is highly sensitive to noise, many researchers have added image space constraints to the standard FCM model objective function, and proposed many improved FCM clustering algorithms, such as FCM_S, FCM_S1, FCM_S2, BCFCM, GG-FCM, EnFCM, FGFCM, etc. These improved algorithms improve the performance of standard FCM to a certain extent. However, the effects of these methods are seriously affected by parameters such as window size and scale factor, and the selection of these parameters has great uncertainty. Therefore, the generality of these algorithms sex awaits further verification. At present, there is a fuzzy local information C-means clustering algorithm FLICM, which adopts local constraints that can integrate local spatial information and pixel features at the same time, and can be calculated without additional parameters. However, when the image noise is large, the segmentation accuracy of the FLICM algorithm is still low. The main reason is that in FLICM, the spatial distance between the central pixel and the neighboring pixels and the membership value of the neighboring pixels are simply considered. The membership degree value of the central pixel is not considered, and the proposed fuzzy factor calculation method has no physical meaning. In the segmentation and classification results of FLICM, the edge of the region is over-smoothed, and a large amount of detailed information is lost. Therefore, there is an urgent need for an efficient FCM algorithm that not only preserves edge detail information, but also considers local spatial information.

Contents of the invention

Technical problem: the main purpose of the present invention is to overcome the deficiencies in the prior art, and provide a fuzzy C-mean value based on a spatial gravity model with a simple method, a high degree of automation, little influence by image noise, and high image segmentation and classification accuracy Automatic classification method of remote sensing images.

Technical solution: In order to achieve the purpose of the above invention, the fuzzy C-means remote sensing image automatic classification method based on the spatial gravity model of the present invention includes the following steps

Step 1: Obtain the remote sensing digital image to be classified, obtain the number of pixels of the remote sensing digital image according to the image size and the number of bands, determine the number of clusters c and the fuzzy index m of the remote sensing digital image according to the actual classification needs, and use the standard The FCM model of the remote sensing digital image is clustered to obtain the initialized fuzzy membership matrix U ₀ ＝{u _ki } _c×N and the clustering center V ₀ ＝{v _k } _c , N is the number of pixels of the remote sensing digital image to be classified number, u _ki represents the membership degree of the i-th pixel in the remote sensing digital image to be classified belongs to the k-th class, and v _k is the center point of the k-th class of the remote-sensing digital image to be classified;

Step 2: Set the size of the neighborhood window, and determine the pixels in the neighborhood window of the remote sensing digital image with each pixel in the remote sensing digital image as the central pixel according to the set neighborhood window size, through the formula Calculate the spatial gravitational force NA _ij between each pixel in the remote sensing digital image and other pixels in its neighborhood window,

Among them, G is a constant, which is used to represent the contribution of the adjustment space constraint to the clustering objective function. Generally, G=1, u _ki represents the membership degree of the neighborhood window center pixel x _i belonging to the kth category, and u _kj represents the neighborhood The jth pixel x _j in the domain window belongs to the membership degree of the kth category, and R _ij represents the Euclidean space distance between the pixel x _i and the pixel x _j ;

Step 3: Using the spatial gravitational force NA _ij between each pixel in the remote sensing digital image and other pixels in the neighborhood window to which it belongs, through the formula Get the spatial constraint penalty factor w _ij between pixels in the remote sensing digital image,

Among them, w _ij represents the influence weight of the jth pixel x _j in the neighborhood window on the center pixel x _i , N _i represents the pixels in the neighborhood window of the center pixel x _i , and NA _ij is the Calculate the spatial gravitational force between each pixel x _i in the image and the remaining pixels x _j in the neighborhood window with pixel x _i as the central pixel, where | _xi -x _j | represents the pixel x _i The difference between the gray value of the pixel x _j in the neighborhood window with the pixel x _i as the central pixel;

Step 4: Through the formula get fuzz factor

Step 5: Through the formula fuzz factor Add to the standard FCM model to get the clustering objective function

Among them, N is the number of pixels of remote sensing digital images to be classified, c is the number of clusters of classified remote sensing digital images, Indicates the membership degree of the i-th pixel in the remote sensing digital image to be classified belongs to the k-th class, m is the fuzzy index of the remote sensing digital image to be classified, x _i represents the i-th pixel of the remote sensing digital image to be classified, v _k is the Classify the center point of the kth category of remote sensing digital images;

Combining the initial fuzzy matrix U ₀ and cluster center V ₀ obtained in step 1, use the formula Calculate the new cluster center of the remote sensing image to be classified count as Indicates the new cluster center currently obtained;

reuse formula Calculate the new membership degree matrix of remote sensing images to be classified count as Indicates the new fuzzy matrix currently obtained; at this time To obtain through the initial cluster center V ₀ , it means The cluster center obtained last time; is obtained through the initial fuzzy matrix U ₀ , which means The fuzzy matrix obtained last time;

Step 6: Determine the cluster center Whether to continue to change or the operation reaches the maximum number of iterations, set the iteration stop threshold ε=1e-5 as a small positive number, and obtain the tth cluster center of the remote sensing digital image to be classified and the t-1th cluster center for comparison, if or b>T one of the two conditions, then the iteration ends, the initial value of b is 0, T is 100, otherwise, set b=b+1, use the fuzzy matrix currently obtained and cluster centers Replace the initial fuzzy matrix U ₀ and cluster center V ₀ respectively, and set as and Return to step 2 and repeat steps 2 to 6 until the or the condition of either b>T;

Step 7: Use the final fuzzy membership degree matrix U={u _ki } _c×N to determine the category of each pixel x _i according to the following formula, that is, for each pixel x _i , the category c _i it belongs to belongs to The category with the largest degree u _ki :

C _i ＝arg _k {max(u _ki )}, k＝1,2,3,,c

Step 8: Assign different colors to different cluster categories according to the category c _i to which each pixel x _i belongs, and form a remote sensing image classification thematic map, thereby realizing automatic classification of remote sensing digital images.

Beneficial effects: the present invention calculates the spatial constraint penalty factor w _ij between pixels in the remote sensing digital image, and considers the gray relationship of local pixels in the remote sensing digital image, and comprehensively proposes a method that can reflect the spatial context characteristics of each pixel Fuzzy factor based on space gravity model Used to suppress noise. In addition to considering the spatial distance between the central pixel and the neighboring pixels and the membership value of the neighboring pixels, the membership value of the central pixel is also considered, and the introduced fuzzy factor It has physical meaning, not only retains edge detail information, but also considers local spatial information. The present invention uses the spatial gravity NA _ij between the center pixel and other pixels in the corresponding neighborhood window to adaptively determine the degree of influence of the neighborhood pixel on the center pixel without learning experience values; introduces fuzzy based on the spatial gravity model factor Into the objective function, this classification method is robust to the processing of remote sensing digital images containing noise, and it can handle image details well; the method is simple, the required parameters are the same as the standard FCM, no other parameters are required, and it is affected by the image The impact of noise is small, the accuracy of image segmentation and classification is high, and it has a wide range of practicability.

Description of drawings:

Fig. 1 is a flow chart of the present invention;

Fig. 2 is a schematic diagram of the pixel neighborhood window structure of the present invention;

Fig. 3 is a comparison diagram between the segmentation result on the remote sensing image of Embodiment 1 of the present invention and the prior art;

Fig. 4 is a comparison diagram of the segmentation result on the remote sensing image of Embodiment 2 of the present invention and the prior art.

Detailed ways:

Embodiments of the present invention will be further described below in conjunction with the accompanying drawings:

As shown in Figure 1, the fuzzy C-means remote sensing image automatic classification method based on the spatial gravity model of the present invention comprises the following steps:

Step 1: Obtain the remote sensing digital image (P, Q, S) to be classified, where P is the number of rows of the image, Q is the number of columns of the image, and the number of bands of the S image. According to the image size and the number of bands, the remote sensing digital image (P,Q,S) is transformed into a 2-dimensional matrix (S,P×Q) form, that is, only the number of pixels and the number of bands of the remote sensing digital image to be classified are considered, and the remote sensing digital image (S,P ×Q) clustering category number c and fuzzy index m; use the standard FCM model to cluster the remote sensing digital image (S,P×Q) to get the initialized fuzzy membership matrix U ₀ ={u _ki } _c×N and clustering center V ₀ ={v _k } _c , N is the number of pixels in the remote sensing digital image to be classified, u _ki represents the membership degree of the i-th pixel in the remote sensing digital image to be classified belonging to the k-th class, v _k is the center point of the kth class of the remote sensing digital image to be classified;

Step 2: Set the size of the neighborhood window, as shown in Figure 2, a window with a window size of 5, the numbers 1, 2, 4... in the figure represent the neighborhood range of the central pixel x, according to the set neighborhood The window size is determined in the remote sensing digital image (S, P×Q), and each pixel in the remote sensing digital image is regarded as the pixel in the neighborhood window of the central pixel, through the formula Calculate the spatial gravitational force NA _ij between each pixel in the remote sensing digital image and other pixels in the neighborhood window to which it belongs,

Among them, G is a constant, which is used to represent the contribution of the adjustment space constraint to the clustering objective function. Generally, G=1, u _ki represents the membership degree of the neighborhood window center pixel x _i belonging to the kth category, and u _kj represents the neighborhood The jth pixel x _j in the domain window belongs to the membership degree of the kth category, and R _ij represents the Euclidean space distance between the pixel x _i and the pixel x _j ;

Step 3: Using the spatial gravitational force NA _ij between each pixel in the remote sensing digital image and other pixels in its neighborhood window, through the formula Get the spatial constraint penalty factor w _ij between pixels in the remote sensing digital image,

Among them, w _ij represents the influence weight of the jth pixel x _j in the neighborhood window on the center pixel x _i , N _i represents the pixels in the neighborhood window of the center pixel x _i , and NA _ij is the Calculate the spatial gravitational force between each pixel x _i in the image and the remaining pixels x _j in the neighborhood window with pixel x _i as the central pixel, where | _xi -x _j | represents the pixel x _i The difference between the gray value of the pixel x _j in the neighborhood window with the pixel x _i as the central pixel;

Step 4: Through the formula get fuzz factor

Step 5: Through the formula fuzz factor Added to the standard FCM model, the added fuzzy factor is obtained The clustering objective function of

Among them, N is the number of pixels of remote sensing digital images to be classified, that is, N=P×Q; c is the number of clusters of classified remote sensing digital images, Indicates the membership degree of the i-th pixel in the remote sensing digital image to be classified belongs to the k-th class, m is the fuzzy index of the remote sensing digital image to be classified, x _i represents the i-th pixel of the remote sensing digital image to be classified, v _k is the Classify the center point of the kth category of remote sensing digital images;

Combining the initial fuzzy matrix U ₀ and cluster center V ₀ obtained in step 1, use the formula Calculate the new cluster center of the remote sensing image to be classified count as Indicates the new cluster center currently obtained;

reuse formula Calculate the new membership degree matrix of remote sensing images to be classified count as Indicates the new fuzzy matrix currently obtained; at this time To obtain through the initial cluster center V ₀ , it means The cluster center obtained last time; is obtained through the initial fuzzy matrix U ₀ , which means The fuzzy matrix obtained last time;

Step 6: Determine the cluster center Whether to continue to change or the operation reaches the maximum number of iterations, set the iteration stop threshold of the number of operations ε=1e-5 as a small positive number, and obtain the tth clustering center of the remote sensing digital image to be classified and the t-1th cluster center for comparison, if or b>T one of the two conditions, then the iteration ends, the initial value of b is 0, T is 100, otherwise, set b=b+1, use the fuzzy matrix currently obtained and cluster centers Replace the initial fuzzy matrix U ₀ and cluster center V ₀ respectively, and set as and Return to step 2 and repeat steps 2 to 6 until the or the condition of either b>T;

Step 7: Use the final fuzzy membership degree matrix U={u _ki } _c×N to determine the category of each pixel x _i according to the following formula, that is, for each pixel x _i , its category c _i is the membership The category with the largest degree u _ki :

C _i ＝arg _k {max(u _ki )}, k＝1,2,3,,c

Step 8: Assign different colors to different cluster categories according to the category c _i to which each pixel x _i belongs, and form a remote sensing image classification thematic map, thereby realizing automatic classification of remote sensing digital images.

All the algorithms in the present invention are programmed under Matlab7.8, and the remote sensing data of two different regions are selected as verification data, and the verified samples of the classification are subjected to strict geometric correction to the original remote sensing images, and satisfy the correction error Less than 0.5 pixel requirements, and then obtain the required verification samples by visual interpretation on the corrected image. Finally, the accuracy of classification results is evaluated by using the production accuracy, overall accuracy and Kappa coefficient, and the FLNAICM of the present invention is compared with the standard FCM and FLICM algorithms.

Embodiment 1. The QuickBird remote sensing image with a resolution of 0.61 meters and a size of 200×200 pixels including red, green and blue bands is used. This data is located in the urban area of Xuzhou City, Jiangsu Province, China, and the acquisition time is 2005 In August 2009, Fig. 3(a) and (b) are the original classification images and reference data images respectively. The images are divided into 4 categories: structures, bare land, water and vegetation, and the fuzzy index is 2.

Figure 3(c)-(e) represent the classification results of FCM, FLICM and FLNAICM respectively. In Figure 3(c), due to the similarity of the spectrum and the existence of image noise, FCM only uses the spectral characteristics of the image and does not consider Spatial context information has caused many "salt and pepper" phenomena in the classification results. From Figure 3(d) and Figure 3(e), it can be seen that the classification effect of FLICM and FLNAICM is better than that of FCM, and most of the "noise points" in the classification map are removed, forming a better homogeneity category area. It can be seen from Figure 3(d) that there is an over-smoothing phenomenon in the classified plots, forming a classified plot in a larger area. For example, in Figure 3(d) marked A, B, and C, many classified plots are lost Compared with the FLICM method, it can be seen from Figure 3(e) that FLNAICM retains a lot of detailed information, for example, A, B, and C are marked in Figure 3(e), the reason is that FLNAICM uses the central image The spatial gravitational force between the central pixel and the neighboring pixels is regarded as the degree of influence of the neighboring pixels on the central pixel space, and this spatial gravitational force fully considers the local spatial relationship and grayscale relationship between the central pixel and the neighboring pixel pixels, and It can be calculated automatically based on the characteristics of the cell. FLICM simply utilizes the spatial distance and membership degree between the central pixel and neighboring pixels. Table 1 shows the accuracy evaluation results of the classification results of FCM, FLICM and FLNAICM. Compared with FCM and FLICM, FLNAICM gives the highest classification accuracy.

Table 1. Number of validation samples, production precision, overall precision and Kappa coefficient in Implementation 1

Example 2

In Example 2, a QuickBird remote sensing image with a size of 200×200 pixels and a resolution of 0.61 meters including red, green and blue bands was used. This data is located in the suburban area of Xuzhou City, Jiangsu Province, China. The acquisition time For August 2005, Figure 4(a) and Figure 4(b) are the original classified images and reference data images respectively, the images are divided into 4 categories: road, bare land, water and vegetation, and the fuzzy index is 2.

Figure 4(c)-(e) are the classification results of FCM, FLICM and FLNAICM respectively. In Figure 4(c), due to the similarity of the spectrum and the existence of image noise, the classification method of FCM only uses the spectral characteristics of the image. Spatial context information is not considered, and a large number of "salt and pepper" phenomena appear in the classification results. From Figure 4(d) and Figure 4(e), it can be seen that the classification effect of FLICM and FLNAICM is better than that of FCM, and most of the "noise points" in the classification diagram have been removed. There is an over-smoothing phenomenon in Figure 4(d), and the details of many classified plots are lost. For example, A, B, and C are marked in Figure 4(d). Compared with the FLICM method, it can be seen from Figure 4(e) , FLNAICM retains a lot of detailed information, such as A, B, and C in Figure 4(e), the reason is that FLNAICM uses the spatial gravitational force between the central pixel and neighboring pixels as the center The influence degree of the pixel space, and this spatial gravity fully considers the local spatial relationship and gray level relationship between the central pixel and the neighboring pixels, and can be automatically calculated according to the characteristics of the pixel. In FLICM, the spatial distance and membership degree between the central pixel and the neighboring pixels are simply used. Table 2 shows the accuracy evaluation results of the classification results of FCM, FLICM and FLNAICM. Compared with FCM and FLICM, FLNAICM gives the highest classification accuracy.

Table 2. Number of validation samples, production precision, overall precision and Kappa coefficient in Implementation 2

In summary, the present invention proposes a FLNAICM classification method based on the gravitational force of the neighborhood pixel space, which utilizes the gravitational force between the central pixel and the neighborhood pixel to estimate the degree of influence of the neighborhood pixel on the central pixel, and reasonably introduces Spatial context information is added to the standard FCM to improve its classification accuracy and the robustness of the algorithm to image classification with noise.

Claims (1)

1. a kind of fuzzy C-mean algorithm remote sensing image automatic classification method based on spatial attraction model, it is characterised in that comprising following Step：

Step 1：Remote Sensing Digital Image to be sorted is obtained, the pixel of Remote Sensing Digital Image is obtained according to picture size and wave band number Number, according to actual classification it needs to be determined that cluster the classification number c and Fuzzy Exponential m of Remote Sensing Digital Image, and utilizes standard FCM models cluster Remote Sensing Digital Image to obtain initialization fuzzy membership matrix U₀={ u_ki}_c×NWith cluster centre V₀= {v_k}_c, N be Remote Sensing Digital Image to be sorted pixel number, u_kiRepresent that in Remote Sensing Digital Image to be sorted i-th pixel belongs to The degree of membership of kth class, k_vFor the central point of Remote Sensing Digital Image kth class to be sorted；

Step 2：Neighborhood window size is set, is determined remote digital according to the neighborhood window size of setting in Remote Sensing Digital Image Each pixel passes through formula as the pixel in the neighborhood window of center pel in imageCalculate distant Feel the spatial attraction NA in each pixel and its affiliated neighborhood window between other pixels in digitized video_ij,

Wherein, G is constant, for representing to adjust contribution of the space constraint to clustering object function, if G=1, u_kiRepresent i-th Pixel x_iBelong to the degree of membership of k-th of classification, u_kjRepresent pixel x_iNeighborhood window in j-th yuan of x_jBelong to k-th classification Degree of membership, R_ijRepresent pixel x_iWith pixel x_jBetween theorem in Euclid space distance；

Step 3：Drawn using the space between other pixels in Remote Sensing Digital Image in each pixel and its affiliated neighborhood window Power NA_ij, pass through formulaObtain in Remote Sensing Digital Image space constraint punishment between pixel because Sub- w_ij,

Wherein, w_ijRepresent j-th of pixel x in neighborhood window_jTo center pel x_iWeighing factor, N_iRepresent center pel x_i Neighborhood window in pixel, NA_ijFor each pixel x in the image that is calculated in step 2_iWith by pixel x_iAs middle imago Remaining pixel x in the neighborhood window of member_jBetween spatial attraction, | x_i-x_j| represent pixel x_iWith by pixel x_iAs middle imago Pixel x in the neighborhood window of member_jBetween gray value difference；

Step 4：Pass through formulaObtain fuzzy factor

Step 5：Pass through formulaBy fuzzy factorStandard FCM models are added to, Obtain cluster object function

Wherein, N be Remote Sensing Digital Image to be sorted pixel number, c be classification Remote Sensing Digital Image cluster number, u_ki ^mTable Show that in Remote Sensing Digital Image to be sorted i-th pixel belongs to the degree of membership of kth class, m is the fuzzy of Remote Sensing Digital Image to be sorted Index, x_iRepresent i-th of pixel of Remote Sensing Digital Image to be sorted, v_kFor the central point of Remote Sensing Digital Image kth class to be sorted；

With reference to the initially fuzzy matrix U obtained in step 1₀With cluster centre V₀, utilize formulaCalculate to be sorted The new cluster centre of remote sensing imageIt is calculated asRepresent currently available new cluster centre；

Recycle formulaCalculate the new subordinated-degree matrix of remote sensing image to be sortedIt is calculated asRepresent currently available new fuzzy matrix；At this timeFor initial cluster centre V₀Obtain, representIt is preceding once to obtain The cluster centre arrived；For initial fuzzy matrix U₀Obtain, representThe preceding fuzzy matrix once obtained；

Step 6：Judge cluster centreWhether continue change or computing reaches maximum iteration, set iteration stopping threshold epsilon =1e-5 is a small positive number, t-th of cluster centre that Remote Sensing Digital Image to be sorted is tried to achieveWith the t-1 cluster centreIt is compared, if meetingOr the condition of b ＞ T one of both, then iteration terminate, the b is initial It is worth for 0, T 100, otherwise, setting b=b+1, with currently available fuzzy matrixAnd cluster centreSubstitute respectively initial Fuzzy matrix U₀With cluster centre V₀, and be set asWithBack to step 2, step 2~step 6 is repeated, Until meetOr the condition of b ＞ T one of both；

Step 7：Utilize the fuzzy membership matrix U={ u finally obtained_ki}_c×N, each pixel x is determined according to equation below_iInstitute Belong to classification, i.e., for each pixel x_i, its generic c_iFor degree of membership u_kiMiddle that maximum classification；

C_i=arg_k{max(u_ki), k=1,2,3 ..., c

Step 8：According to each pixel x_iGeneric c_iDifferent cluster classifications is assigned to different colors, forms remote sensing image Classification thematic map, so as to fulfill the automatic classification of Remote Sensing Digital Image.