Conditional random field with auxiliary elements for high-resolution remote sensing image change detection method

Abstract

The conditional random field change detection method with auxiliary elements for high-resolution remote sensing images comprises the following steps: step 1, original two-stage high-resolution remote sensing images are registered and radiation corrected for pretreatment, the influence of external factors on the noise of the images is reduced, and a difference image is obtained; step 2, the obtained difference image is superpixel segmented, the obtained superpixel is regarded as an auxiliary element, and the auxiliary element is taken as a to-be-classified object and a pixel in the difference image to form a to-be-classified set; step 3, fuzzy information of elements in the to-be-classified set is obtained for subsequent model construction; step 4, a conditional random field model with auxiliary elements is constructed; and step 5, the labels in the to-be-classified set are obtained, the labels of the auxiliary elements and the pixel labels in the set are used for joint decision to determine a change detection map. The present application aims at more effectively utilizing the rich information in high-resolution images, combining the respective advantages of pixel-oriented and object-oriented, and improving the change detection precision.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remote sensing, and relates to an unsupervised change detection method, in particular to a high-resolution remote sensing image change detection method of conditional random field with auxiliary elements. BACKGROUND

[0002] Change detection technology is a research hotspot in the field of remote sensing and has a wide range of applications in life and military. In recent years, with the continuous progress of sensor technology, the spatial resolution of remote sensing images gradually increases, making the image detail information more abundant and the ground object information more clear. The emergence of high-resolution images greatly expands the application field of change detection. Pixel-oriented analysis takes pixels as the research unit, and the pixel retains the most original and basic spectral characteristics in the image. The pixel-oriented analysis is simple to operate and can distinguish the ground object boundary to a large extent. In change detection, the detection boundary is relatively complete, but in high-resolution images, the spectral heterogeneity is strong, resulting in a large number of false detections in the detection. The overall detection result is poor. Object-oriented analysis takes a group of pixels with similar features as an object as a whole, and takes the whole as the research unit. The object can express the local image features in the high-resolution image, and to a certain extent, it can reduce the number of false detections caused by spectral heterogeneity of pixels. However, the object block ignores the original feature details contained in the image, and the object-oriented analysis is too dependent on the results of the image segmentation algorithm. If the selected segmentation algorithm is suitable for the current image scene, the satisfactory object block can be extracted as the analysis unit. However, if the segmentation effect is poor, the detection accuracy will be greatly reduced. Since there is no segmentation algorithm that can be applied to all scenes, the object-oriented analysis algorithm cannot maintain a good ground object detection boundary and is too affected by the segmentation effect. The algorithm is relatively unstable.

[0003] In order to improve the change detection accuracy of high-resolution remote sensing images, it is necessary to combine the advantages of pixel-oriented and object-oriented analysis, retain the original features of pixels in the image, and use local feature information in the graph to reduce the detection noise caused by spectral heterogeneity, so as to improve the change detection accuracy. SUMMARY

[0004] The purpose of the present application is to more effectively utilize the rich information in high-resolution images, combine the advantages of pixel-oriented and object-oriented analysis, and improve the change detection accuracy. A high-resolution remote sensing image change detection method of conditional random field with auxiliary elements is provided.

[0005] In order to solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0006] The conditional random field with auxiliary elements high-resolution remote sensing image change detection method comprises the following steps:

[0007] Step 1: Register and radiometrically correct the original two-phase high-resolution remote sensing images to reduce the noise impact of external factors on the images, and use the change vector analysis algorithm to obtain the difference image;

[0008] Step 2: Perform superpixel segmentation on the acquired difference image. The obtained superpixels are regarded as auxiliary elements. The auxiliary elements are used as objects to be classified and form a set to be classified with the pixels in the difference image.

[0009] Step 3: Cluster the pixels and auxiliary elements in the set to be classified obtained in Step 2 to obtain fuzzy information of the elements in the set to be classified for subsequent model construction;

[0010] Step 4: Establish a new conditional random field space structure and construct a conditional random field model with auxiliary elements;

[0011] Step 5: Obtain the labels in the set to be classified, and use the auxiliary meta labels and pixel labels in the set to make a joint decision to determine the change detection map.

[0012] In step 3, the pixels and auxiliary elements in the set to be classified obtained in step 2 are clustered using a fuzzy clustering algorithm.

[0013] In step 3, the specific details are as follows:

[0014] Y D ={p i |i=1,2,3…N} is the difference image, p i For differential image Y D The i-th pixel in the middle, Y represents D Spectral values ​​of mid-pixel, Representing cell p i spectral values, Y SD ={s i |i=1,2,3…M} represents the auxiliary element set, s i Describes the i-th auxiliary element. Y represents SD Spectral values ​​of auxiliary elements, Represents superpixel s i The spectral values, auxiliary elements, and pixels constitute the set to be classified, Y. U ={Y D ,Y SD}, G U ={G D G SD} represents the set Y to be classified. U The spectral characteristics are used to iteratively optimize the objective function J through fuzzy C-means clustering. C To calculate the membership degree of pixels and auxiliary elements in the set to be classified:

[0015]

[0016] in, Represents the set Y to be classified U medium pixel p i The degree of membership in category j. Represents the set Y to be classified. U Middle auxiliary element s i The degree of membership to category j, c j Let q represent the cluster center of category j, q be the weighting index, and C be the number of categories.

[0017] In step 4, on the set to be classified, not only the constraint relationship between pixels is considered, but also the constraint relationship between pixels and auxiliary elements is considered to establish a new conditional random field space structure. The feature information of pixels, the local image feature information contained in the auxiliary elements, and the fuzzy information of the elements in the set to be classified in step 3 are used to construct a conditional random field model with auxiliary elements.

[0018] In step 4, the specific details are as follows:

[0019] remember For the set Y to be classified U The set of medium-sized pixel tags, among which, For pixel p i The tag value. For the set Y to be classified U Auxiliary meta tag set, For auxiliary element s i The label value. For the set to be classified, not only the constraints between pixels are considered, but also the constraints between pixels and auxiliary elements, establishing a new conditional random field spatial structure. The conditional random field model with auxiliary elements is for the set Y to be classified. U Its corresponding label X U ={X p ,X s} Perform posterior probability modeling, the energy function E(X) of the conditional random field with auxiliary elements U |Y U It can be written in the following form: that is:

[0020]

[0021] in, and These are the univariate potential functions of the pixels and auxiliary elements in the set to be classified, respectively. and λ and λ' are the binary potential functions between pixels in the set to be classified, and between a pixel and an auxiliary element, respectively. λ is a non-negative constant, serving as a balance factor in the model.

[0022] In step 5, the labels of the set to be classified are obtained through the graph cut algorithm inference model, and the joint decision of auxiliary elements and pixel labels is used to finally determine the change detection map.

[0023] In step 5, when the pixel label matches the corresponding auxiliary pixel label, the pixel label remains unchanged; otherwise, the label value is 0.

[0024]

[0025] Among them, pixels Belongs to auxiliary elements In the neighborhood system, the label value is either 0 or 1, where 0 represents a change class and 1 represents a change class, thus obtaining the final change detection map.

[0026] Compared with the prior art, the present invention has the following technical effects:

[0027] The technical solution proposed in this invention considers both pixel and auxiliary element feature information, integrating the advantages of object-oriented and pixel-oriented approaches into a single field model classification framework. This preserves the original features of pixels in the image while utilizing local feature information in the image to reduce detection noise caused by spectral heterogeneity. Furthermore, it introduces neighborhood relationships between pixels and auxiliary elements to construct a new field model structure, which can better utilize spatial information. Finally, the final detection result is obtained through the joint decision-making of auxiliary elements and pixel labels, thereby improving the change detection accuracy and algorithm stability.

[0028] This invention proposes a conditional random field (CRF) method for change detection in high-resolution remote sensing images, incorporating auxiliary elements. Instead of directly modeling pixels or objects to obtain the final change detection result, this method treats object blocks as auxiliary variables—that is, auxiliary elements and the entire pixel set—as a classification set. On this set, constraints are established not only between pixels but also between pixels and auxiliary elements. This creates a new architecture for the CRF model, utilizing the original pixel features, local features of objects, and additional spatial information constructed within the model. This results in change detection results with better boundaries, fewer false positives, and higher accuracy. Furthermore, object blocks only serve a dynamic auxiliary classification function, mitigating accuracy fluctuations caused by differences in image segmentation algorithm selection and improving algorithm stability. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0030] Figure 1 This is the experimental data at time T1 of this embodiment of the invention;

[0031] Figure 2 This is the experimental data at time T2 of this embodiment of the invention;

[0032] Figure 3 This is a reference diagram showing the changes in experimental data from an embodiment of the present invention;

[0033] Figure 4 This is a flowchart of the present invention;

[0034] Figure 5 This is a change detection result image based on the fuzzy C-means clustering algorithm;

[0035] Figure 6 This is a change detection result image based on the superpixel blur C-means clustering algorithm;

[0036] Figure 7 This is a change detection result image based on the fully connected conditional random field algorithm;

[0037] Figure 8 This is a graph showing the change detection results based on the principal component analysis and K-means clustering algorithm.

[0038] Figure 9 This is a change detection result image based on the fuzzy C-means conditional random field algorithm;

[0039] Figure 10 This is a change detection result image based on the hybrid conditional random field algorithm;

[0040] Figure 11 This is a diagram showing the change detection results of an embodiment of the present invention. Detailed Implementation

[0041] A method for detecting changes in high-resolution remote sensing images using a conditional random field with auxiliary elements. The technical solution includes the following steps:

[0042] Step 1: Perform preprocessing such as registration and radiometric correction on the original two-phase high-resolution remote sensing images to reduce the noise impact of external factors on the images, and use the change vector analysis algorithm to obtain differential images;

[0043] Step 2: Perform superpixel segmentation on the acquired difference image. The obtained superpixels are regarded as auxiliary elements. The auxiliary elements are used as objects to be classified and form a set to be classified with the pixels in the difference image.

[0044] Step 3: Use a fuzzy clustering algorithm to cluster the pixels and auxiliary elements in the set to be classified obtained in Step 2, and obtain the fuzzy information of the elements in the set to be classified for subsequent model construction;

[0045] Step 4: On the set to be classified, not only the constraint relationship between pixels, but also the constraint relationship between pixels and auxiliary elements are considered to establish a new conditional random field spatial structure. The feature information of pixels, the local image feature information contained in the auxiliary elements, and the fuzzy information of the elements in the set to be classified in Step 3 are used to construct a conditional random field model with auxiliary elements.

[0046] Step 5: Obtain the labels in the set to be classified through the graph cut algorithm reasoning model, and use the auxiliary meta labels and pixel labels in the set to make a joint decision to finally determine the change detection map.

[0047] In step 3, the pixels and auxiliary elements in the set to be classified obtained in step 2 are clustered using a fuzzy clustering algorithm to obtain fuzzy information of the elements in the set to be classified for subsequent model construction.

[0048] Specifically, record Y D ={p i |i=1,2,3…N} is the difference image, p i For differential image Y D The i-th pixel in the middle, Y represents D Spectral values ​​of mid-pixel, Representing cell p i Spectral values ​​of Y. SD ={s i |i=1,2,3…M} represents the auxiliary element set, s i Describes the i-th auxiliary element. Y represents SD Spectral values ​​of auxiliary elements, Represents superpixel s i The spectral values. The set to be classified, composed of auxiliary elements and pixels in step 2, is Y. U ={Y D ,Y SD}, G U ={G D G SD} represents the set Y to be classified. U The spectral characteristics of the target function J are obtained by iteratively optimizing the objective function J using fuzzy C-means clustering. C To calculate the membership degree of pixels and auxiliary elements in the set to be classified:

[0049]

[0050] in, Represents the set Y to be classified U medium pixel p i The degree of membership in category j. Represents the set Y to be classified. UMiddle auxiliary element s i The degree of membership to category j, c j Let q represent the cluster center of category j, q be the weighting index, and C represent the number of categories. In this invention, there are two categories: changed and unchanged, i.e., C = 2.

[0051] In step 4, on the set to be classified, not only the constraint relationship between pixels is considered, but also the constraint relationship between pixels and auxiliary elements is considered to establish a new conditional random field spatial structure. The feature information of pixels and the local image feature information contained in the auxiliary elements, as well as the fuzzy information of the elements in the set to be classified in step 3, are used to construct a conditional random field model with auxiliary elements.

[0052] Specifically, remember For the set Y to be classified U The set of medium-sized pixel tags, among which, For pixel p i The tag value. For the set Y to be classified U Auxiliary meta tag set, For auxiliary element s i The label value. For the set to be classified, not only the constraints between pixels are considered, but also the constraints between pixels and auxiliary elements, establishing a new conditional random field spatial structure. The conditional random field model with auxiliary elements is for the set Y to be classified. U Its corresponding label X U ={X p ,X s} Perform posterior probability modeling, the energy function E(X) of the conditional random field with auxiliary elements U |Y U It can be written in the following form: that is:

[0053]

[0054] in, and These are the univariate potential functions of the pixels and auxiliary elements in the set to be classified, respectively. and λ and λ' are the binary potential functions between pixels in the set to be classified, and between a pixel and an auxiliary element, respectively. λ is a non-negative constant, serving as a balance factor in the model.

[0055] Univariate potential function These represent the sets Y to be classified. U lower pixel p i Assign tags Costs and auxiliary elements i Assign tags The cost, one yuan Defined as follows:

[0056]

[0057]

[0058] in, Representing pixels p respectively i Category Membership degree and auxiliary element s i Category The membership degree is obtained from step 3.

[0059] The binary potential function models spatial relationships in an image, encouraging elements in the same neighborhood to be assigned the same label, thereby reducing false tests during detection. Unlike traditional conditional random field (CRF) processing methods, in this patent, the classification set is no longer composed solely of pixels or objects, but rather includes superpixels as auxiliary variables along with the original pixels. In this case, the classification set contains not only the basic features of the original pixels but also the local graph features implied in the auxiliary pixels. In the classification set Y... U To establish a binary potential function relationship, it is first necessary to construct the neighborhood relationship between elements in the set. For the set Y to be classified... U medium pixel p i Local neighborhood system The construction utilizes the original differential image Y D The construction of eight-neighbor relationships for a single pixel. Since the relationship between a pixel and an auxiliary pixel is analogous to that of a local and global element, and an auxiliary pixel is composed of pixels, auxiliary pixels and the pixels within them should be encouraged to be assigned the same label. Therefore, the classification set Y... U Middle auxiliary element s i local neighborhood Includes all in auxiliary element s i The internal cells, that is, all the cells in the auxiliary cell s i Each internal cell is its neighborhood. In this way, our model considers both the interaction between cells and the interaction between auxiliary cells and cells, making full use of spatial information and integrating cell-oriented and object-oriented ideas into a single framework.

[0060] Binary potential function Represent the set Y to be classified respectively U Quantification of interactions between pixels and between pixels and auxiliary pixels; binary potential encourages neighboring pixels and neighboring pixels to acquire the same label:

[0061]

[0062]

[0063] Among them, dist(p i ,pj ), dist(s i ,p j ) represent adjacent pixels p i ,p j Euclidean distance and auxiliary element s between i Its neighboring pixel p j European distance, Representing pixels p respectively i Pixel p j and auxiliary elements s i spectral values, For all The mean, and Similarly.

[0064] In step 5, the set Y to be classified is obtained by reasoning on the model. U The tag X U By utilizing the joint decision-making of auxiliary elements and pixel labels, the final change detection map is determined.

[0065] Specifically, when the pixel label matches the corresponding auxiliary pixel label, the pixel label remains unchanged; if they are different, the label value is 0.

[0066]

[0067] Among them, pixels Belongs to auxiliary elements In the neighborhood system, the label value is either 0 or 1, where 0 indicates a change class and 1 indicates a change class, thus obtaining the final change detection map.

[0068] To facilitate a better understanding of the present invention by those skilled in the art, the following embodiments are provided:

[0069] This embodiment uses two high-resolution remote sensing images taken by the SPOT 5 satellite in April 2008 and February 2009, respectively. The spatial resolution of the images is 2.5m, and the image size is 1000×1000 pixels. The corresponding location is Tianjin, China. Figure 1 , Figure 2 and Figure 3 The images at time T1 and T2 and their change reference diagrams are shown respectively, with the change reference diagrams obtained through visual interpretation.

[0070] Figure 4 This is a flowchart of the technology proposed in this invention. The technical solution adopted in this invention, a conditional random field high-resolution remote sensing image change detection with the introduction of auxiliary elements, includes the following steps:

[0071] Step 1: Perform preprocessing such as registration and radiometric correction on the original two-phase high-resolution remote sensing images to reduce the noise impact of external factors on the images, and use the change vector analysis algorithm to obtain differential images.

[0072] In this embodiment, radiometric and geometric corrections are performed on the high-resolution remote sensing images Y1 and Y2 acquired at times T1 and T2 to eliminate errors caused by external factors in the detection results. A difference image is generated from the processed images Y1 and Y2 using a vector transformation analysis algorithm; the resulting difference image is denoted as Yi. D The differential image Y D The specific implementation of the calculation process is as follows:

[0073]

[0074] Where Y t,b (t=1,2) represents image Y t The pixel value of the b-th band (1≤b≤B). In this embodiment, B=3.

[0075] Step 2: Perform superpixel segmentation on the acquired difference image. The obtained superpixels are regarded as auxiliary elements. The auxiliary elements are used as objects to be classified and form a set to be classified with the pixels in the difference image.

[0076] In this embodiment, Y is denoted as D ={p i |i=1,2,3…N} is the difference image, p i For differential image Y D The i-th pixel in the differential image Y. D Superpixel segmentation yields an auxiliary set, denoted as Y. SD ={s i |i=1,2,3…M},s i Let Y represent the i-th auxiliary element. The auxiliary elements and the pixels together form the set to be classified, denoted as Y. U ={Y D ,Y SD In this embodiment, the number of auxiliary elements M = 4000.

[0077] Step 3: Use a fuzzy clustering algorithm to cluster the pixels and auxiliary elements in the set to be classified obtained in Step 2, and obtain the fuzzy information of the elements in the set to be classified for subsequent model construction.

[0078] Specifically, remember Y represents D Spectral values ​​of mid-pixel, Representing cell p i Spectral values. Y represents SD Spectral values ​​of auxiliary elements, Represents superpixel s i The spectral values. The set to be classified, composed of auxiliary elements and pixels in step 2, is Y. U ={Y D ,Y SD}, G U ={G D G SD} represents the set Y to be classified. U The spectral characteristics of the target function J are obtained by iteratively optimizing the objective function J using fuzzy C-means clustering. C To calculate the membership degree of pixels and auxiliary elements in the set to be classified:

[0079]

[0080] in, Represents the set Y to be classified U medium pixel p i The degree of membership in category j. Represents the set Y to be classified. U Middle auxiliary element s i The degree of membership to category j, c j Let q represent the cluster center of category j, q be the weighting index, and C represent the number of categories. In this invention, there are two categories: changed and unchanged, i.e., C = 2.

[0081] Step 4: On the set to be classified, not only the constraint relationship between pixels, but also the constraint relationship between pixels and auxiliary elements are considered to establish a new conditional random field spatial structure. The feature information of pixels, the local image feature information contained in the auxiliary elements, and the fuzzy information of the elements in the set to be classified in Step 3 are used to construct a conditional random field model with auxiliary elements.

[0082] Specifically, remember For the set Y to be classified U The set of medium-sized pixel tags, among which, For pixel p i The tag value. For the set Y to be classified U Auxiliary meta tag set, For auxiliary element s i The label value. For the set to be classified, not only the constraints between pixels are considered, but also the constraints between pixels and auxiliary elements, establishing a new conditional random field spatial structure. The conditional random field model with auxiliary elements is for the set Y to be classified. U Its corresponding label X U ={X p ,X s} Perform posterior probability modeling, the energy function E(X) of the conditional random field with auxiliary elements U |Y UIt can be written in the following form: that is:

[0083]

[0084] in, and These are the univariate potential functions of the pixels and auxiliary elements in the set to be classified, respectively. and λ and λ' are the binary potential functions between pixels in the set to be classified, and between a pixel and an auxiliary element, respectively. λ is a non-negative constant, serving as a balance factor in the model; in this embodiment, λ = 3.6.

[0085] Univariate potential function These represent the sets Y to be classified. U lower pixel p i Assign tags Costs and auxiliary elements i Assign tags The cost, one yuan Defined as follows:

[0086]

[0087]

[0088] in, Representing pixels p respectively i Category Membership degree and auxiliary element s i Category The membership degree is obtained from step 3.

[0089] The binary potential function models spatial relationships in an image, encouraging elements in the same neighborhood to be assigned the same label, thereby reducing false tests during detection. Unlike traditional conditional random field (CRF) processing methods, in this patent, the classification set is no longer composed solely of pixels or objects, but rather includes superpixels as auxiliary variables along with the original pixels. In this case, the classification set contains not only the basic features of the original pixels but also the local graph features implied in the auxiliary pixels. In the classification set Y... U To establish a binary potential function relationship, it is first necessary to construct the neighborhood relationship between elements in the set. For the set Y to be classified... U medium pixel p i Local neighborhood system The construction utilizes the original differential image Y D The construction of eight-neighbor relationships for a single pixel. Since the relationship between a pixel and an auxiliary pixel is analogous to that of a local and global element, and an auxiliary pixel is composed of pixels, auxiliary pixels and the pixels within them should be encouraged to be assigned the same label. Therefore, the classification set Y... U Middle auxiliary element s ilocal neighborhood Includes all in auxiliary element s i The internal cells, that is, all the cells in the auxiliary cell s i Each internal cell is its neighborhood. In this way, our model considers both the interaction between cells and the interaction between auxiliary cells and cells, making full use of spatial information and integrating cell-oriented and object-oriented ideas into a single framework.

[0090] Binary potential function Represent the set Y to be classified respectively U Quantification of interactions between pixels and between pixels and auxiliary pixels; binary potential encourages neighboring pixels and neighboring pixels to acquire the same label:

[0091]

[0092]

[0093] Among them, dist(p i ,p j ), dist(s i ,p j ) represent adjacent pixels p i ,p j Euclidean distance and auxiliary element s between i Its neighboring pixel p j European distance, Representing pixels p respectively i Pixel p j and auxiliary elements s i spectral values, For all The mean, and Similarly.

[0094] Step 5: Use the graph cut algorithm to reason about the model and obtain the set Y to be classified. U The tag X U By utilizing the joint decision-making of auxiliary elements and pixel labels, the final change detection map is determined.

[0095] Specifically, when the pixel label matches the corresponding auxiliary pixel label, the pixel label remains unchanged; if they are different, the label value is 0.

[0096]

[0097] Among them, pixels Belongs to auxiliary elements In the neighborhood system, the label value is either 0 or 1, where 0 indicates a change class and 1 indicates a change class, thus obtaining the final change detection map.

[0098] like Figures 5-11 As shown, change detection graphs for the following algorithms are presented: fuzzy C-means clustering algorithm (A), superpixel fuzzy C-means clustering algorithm (B), fully connected conditional random field algorithm (C), principal component analysis K-means clustering algorithm (D), fuzzy C-means conditional random field (E), mixed conditional random field (F), and the present invention (G). Table 1 presents the statistical results of the change detection graphs for the above different change detection methods.

[0099] Table 1 Statistical comparison of results from different change detection methods

[0100]

[0101] contrast Figures 5-11 The change detection graphs and statistical results presented in Table 1 for different change detection methods show that the change detection performance of the present invention is significantly better than other comparative change detection algorithms. The detection results of the present invention simultaneously exhibit the smallest overall error and the highest Kappa coefficient. The overall error of the present invention is 59371, which is significantly lower than that of the following algorithms: fuzzy C-means clustering algorithm, superpixel fuzzy C-means clustering algorithm, fully connected conditional random field algorithm, principal component analysis K-means algorithm, fuzzy C-means conditional random field algorithm, and hybrid conditional random field algorithm. The method reduces 72409, 20635, 30452, 32069, 27531, and 12096 pixels, respectively. The Kappa coefficient of the detection result of this invention is 0.7658, which is 18.72%, 5.81%, 9.33%, 9.56%, 7.16%, and 4.54% higher than the fuzzy C-means clustering algorithm, superpixel fuzzy C-means clustering algorithm, fully connected conditional random field algorithm, principal component analysis K-means algorithm, fuzzy C-means conditional random field algorithm, and hybrid conditional random field algorithm, respectively.

[0102] This invention proposes a high-resolution remote sensing image change detection method using a conditional random field (CRF) with auxiliary elements. Super-pixels are introduced as auxiliary elements into the original observation field, forming a new observation field with auxiliary elements. This method fully considers the interactions between pixels and between pixels and auxiliary elements, using auxiliary elements to dynamically constrain pixels. It integrates object-oriented and pixel-oriented approaches into the same random field framework. Furthermore, under the new observation field, spectral constraints and label constraints are considered simultaneously, constructing a CRF model more suitable for high-resolution image scenarios. This model can reduce "salt-and-pepper noise" in change detection and improve the accuracy of change detection results.

Claims

1. A method for detecting changes in high-resolution remote sensing images using conditional random fields with auxiliary elements, characterized in that, It includes the following steps: Step 1: Perform registration and radiometric correction preprocessing on the original two high-resolution remote sensing images to reduce the noise impact of external factors on the images and obtain differential images; Step 2: Perform superpixel segmentation on the acquired difference image. The obtained superpixels are regarded as auxiliary elements. The auxiliary elements are used as objects to be classified and form a set to be classified with the pixels in the difference image. Step 3: Cluster the pixels and auxiliary elements in the set to be classified obtained in Step 2 to obtain fuzzy information of the elements in the set to be classified for subsequent model construction; Step 4: Establish a new conditional random field space structure and construct a conditional random field model with auxiliary elements; Step 5: Obtain the labels in the set to be classified, and use the auxiliary meta labels and pixel labels in the set to make a joint decision to determine the change detection map; In step 4, on the set to be classified, not only the constraint relationship between pixels is considered, but also the constraint relationship between pixels and auxiliary elements is considered to establish a new conditional random field spatial structure. The feature information of pixels and the local image feature information contained in the auxiliary elements, as well as the fuzzy information of the elements in the set to be classified in step 3, are used to construct a conditional random field model with auxiliary elements. In step 4, the specific details are as follows: remember X p ={ } is the set to be classified Y U The set of medium-sized pixel tags, among which, For pixels p i The tag value, X s ={ } is the set to be classified Y U Auxiliary meta tag set, as auxiliary element s i The label values, on the set to be classified, consider not only the constraints between pixels but also the constraints between pixels and auxiliary elements, establishing a new conditional random field spatial structure. The conditional random field model with auxiliary elements is for the set to be classified. Y U Its corresponding tags X U ={ X p , X s } Perform posterior probability modeling, energy function of conditional random field with auxiliary elements E ( X U | Y U Write it in the following form: （2）； in, and These are the univariate potential functions of the pixels and auxiliary elements in the set to be classified, respectively. and These are the binary potential functions between pixels in the set to be classified, and between a pixel and an auxiliary element, respectively. It is a non-negative constant, used as a balance factor in the model.

2. The method according to claim 1, characterized in that, In step 3, the pixels and auxiliary elements in the set to be classified obtained in step 2 are clustered using a fuzzy clustering algorithm.

3. The method according to claim 1 or 2, characterized in that, In step 3, the specific details are as follows: remember Y D ={ p i | i =1,2,3… N } represents a differential image. p i Differential image Y D The Middle i One pixel, G D ={ }express Y D Spectral values ​​of mid-pixel, Represents a pixel p i spectral values, Y SD ={ s i | i =1,2,3… M } represents the auxiliary element set. s i Indicates the first i One auxiliary element, G SD ={ }express Y SD Spectral values ​​of auxiliary elements, Superpixel s i The spectral values, auxiliary elements, and pixels constitute the set to be classified. Y U ={ Y D , Y SD }, G U ={ G D , G SD } represents the set to be classified. Y U Spectral characteristics, utilizing fuzzy C Mean clustering optimizes the objective function through iteration. J C To calculate the membership degree of pixels and auxiliary elements in the set to be classified: （1）； in, Represents the set to be classified Y U medium pixel p i Category j The degree of subordination, Represents the set to be classified Y U Middle auxiliary element s i Category j The degree of subordination, c j Indicates category j Cluster centers q As a weighted index, C Indicates the number of categories.

4. The method according to claim 1, characterized in that, In step 5, the labels of the set to be classified are obtained through the graph cut algorithm inference model, and the joint decision of auxiliary elements and pixel labels is used to finally determine the change detection map.

5. The method according to claim 4, characterized in that, In step 5, when the pixel label matches the corresponding auxiliary pixel label, the pixel label remains unchanged; otherwise, the label value is 0. （3）； Among them, pixels Belongs to auxiliary elements In the neighborhood system, the label value is either 0 or 1, where 0 indicates a change class and 1 indicates a change class, thus obtaining the final change detection map.

Images