Rapid remote sensing image change detection method based on superpixels

Abstract

The invention discloses a rapid remote sensing image change detection method based on superpixels, and the method comprises the steps: obtaining remote sensing images of the same region in different periods, dividing the remote sensing images into a T1 period and a T2 period, and carrying out the preprocessing of the remote sensing images in the two periods; overlapping and fusing spectral features of the two periods of images to form a new data block so as to obtain super-pixel blocks with the same scale in different periods; performing principal component analysis on the obtained superpixel blocks, and extracting representative features in each superpixel block; and carrying out binary clustering analysis on the extracted features by using a k multi-mean algorithm, and dividing the extracted features into changed image blocks and unchanged image blocks to realize change detection processing. Compared with the prior art, the rapid remote sensing image change detection method based on the superpixels has the advantages that the spatial information of the image is reserved, the calculation complexity is reduced, and the change detection accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image change detection, and in particular to a superpixel-based rapid remote sensing image change detection method. Background Art

[0002] Remote sensing image change detection refers to the analysis of remote sensing images, satellite images and other data. It is usually used to detect remote sensing images taken in the same area at different times to derive changes in the area over a period of time. Change detection is of great significance for monitoring natural disasters, urban expansion, land use changes, etc.

[0003] At present, remote sensing image change detection methods can be divided into pixel-based change detection methods and object-oriented change detection methods according to the different basic units of detection. Among them, the pixel-based change detection method has advantages in simple processing and reliable results, but these methods are all based on pixels or pixel statistical attributes for change detection.

[0004] In addition, existing high-resolution remote sensing images still have complex spectra and many "same object, different spectra, different objects, same spectrum" phenomena, which also leads to smaller spectral differences between different types of objects and larger spectral differences between the same type of objects, making threshold setting more difficult. At the same time, the method based on pixels as the basic unit will produce more fragmented results, and the time required for post-processing will also increase significantly. Moreover, when the spatial resolution of remote sensing images becomes higher, the accuracy of these methods will decrease significantly.

[0005] Many scholars have introduced object-oriented change detection methods based on the characteristics of high spatial resolution remote sensing images and in order to better analyze the overall surface and obtain rich spatial information.

[0006] Object-oriented change detection is a technology that detects changes based on feature comparison of remote sensing image objects. This method first performs image segmentation, organizes pixels into image blocks with similar features, uses the segmented image blocks as basic units, and then detects changes by comparing the attributes between image blocks. At the same time, object-oriented methods can better capture changes at the ground or object level, have more advantages for comprehensive scenes and large-scale change detection, can better integrate spectral information and spatial information, and have better robustness to illumination changes, noise, etc.

[0007] However, object-oriented change detection methods still have many shortcomings, such as: the current remote sensing image change detection algorithm still has some missed detection problems, especially in complex situations, such as multi-temporal data, different sensor data, etc.; the segmentation algorithm may produce over-segmentation; the clustering method used in traditional change detection is difficult to obtain good accuracy; and the computational complexity is also more complex than the pixel-based method. Therefore, further improvement is needed. Summary of the invention

[0008] The technical problem to be solved by the present invention is to overcome the above technical defects and provide a superpixel-based fast remote sensing image change detection method that retains the spatial information of the image, reduces the complexity of calculation and improves the accuracy of change detection.

[0009] In order to solve the above technical problems, the present invention provides a technical solution: a superpixel-based rapid remote sensing image change detection method, comprising the following steps:

[0010] S1: Acquire remote sensing images of different time phases of the same geographical area, including T1 and T2, and perform preprocessing to obtain image pairs and change label images of the two time phase remote sensing images;

[0011] S2: Overlay and fuse the spectral features of the two images to form a new data block, and perform SLIC superpixel segmentation to obtain superpixel blocks of the same scale in different periods;

[0012] S3: Perform principal component analysis on the superpixel blocks in S2 to extract representative features in each superpixel block;

[0013] S4: Use the k-means algorithm to perform binary clustering analysis on the features extracted in S3:

[0014] The blocks are divided into changed blocks and unchanged blocks for change detection.

[0015] Preferably, S1 includes:

[0016] MATLAB is used to select and crop the remote sensing images of the same geographical area and different phases, and the pixel-aligned image pairs are obtained after normalization.

[0017] The label images corresponding to the original images of the T1 and T2 phases are selected and cropped according to the regions of the registered remote sensing images of different phases to obtain the corresponding label images, and pixel-by-pixel comparison is performed to obtain the change label images of the remote sensing images of the T1 and T2 phases.

[0018] Preferably, obtaining the super pixel block in S2 comprises the following steps:

[0019] S2.1: evenly divide the T1 remote sensing image and the T2 remote sensing image into initial superpixel grids, and determine the initial clustering centers of the T1 remote sensing image and the T2 remote sensing image;

[0020] S2.2: Perform iterative clustering and update the cluster centers;

[0021] S2.3: Optimize the results in S2.2 to obtain super-pixel images of the T1 remote sensing image and the T2 remote sensing image.

[0022] Preferably, S2.2 includes: assigning surrounding pixels to the nearest cluster center, updating the position of each cluster center through all assigned pixels, repeating the iteration, and generating a plurality of super pixels;

[0023] The S2.3 includes: retrieving small-size superpixel blocks from the segmentation results of the T1 remote sensing image and the T2 remote sensing image, and merging the small-size superpixel blocks into adjacent and similar superpixels to obtain superpixel images of the T1 remote sensing image and the T2 remote sensing image.

[0024] Preferably, S3 includes the following steps:

[0025] S3.1: Linear transformation is applied to the T1 phase super-pixel image and the T2 phase super-pixel image to obtain low-dimensional data of the original data;

[0026] S3.2: Based on the low-dimensional data in S3.1, representative features of the T1 phase superpixel image and representative features of the T2 phase superpixel image are obtained.

[0027] Preferably, the S3.1 includes: respectively calculating the covariance matrix of the T1 phase super-pixel image data and the T2 phase super-pixel image data, performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors;

[0028] S3.2 includes selecting principal components according to the eigenvalues ​​and eigenvectors in S3.1, and constructing a projection matrix, multiplying the superpixel image data of the T1 period and the T2 period by the projection matrix to obtain the reduced dimension data, that is, to obtain the first representative feature and the second representative feature.

[0029] Preferably, the eigenvalues ​​obtained in S3.1 are sorted in descending order, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected, where k is the dimension after dimensionality reduction, and the eigenvector is the principal component, and the selected k eigenvectors are combined into a projection matrix.

[0030] Preferably, the S4 further comprises the following steps:

[0031] S4.1: randomly select m cluster centers of representative features of T1 superpixel images and T2 superpixel images, calculate the distance from the data point to the cluster center according to the Euclidean distance, and assign each point to the cluster center with the closest distance;

[0032] S4.2: Determine the similarity matrix based on the distance from the data point to the cluster center and the objective function of the algorithm, where the objective function is:

[0033]

[0034] In the formula, x i is the i-th data in the data matrix X, a j is the jth center in the cluster center matrix A, S is the similarity matrix, γ is the regularization parameter, λ is the trade-off parameter, is the normalized Laplacian matrix, in D is a diagonal matrix, and the elements on the diagonal are ii Equal to the sum of the elements in the corresponding row i of the matrix P, F = [U, V] T ,U,V are The first k left and right singular vectors of ;

[0035] S4.3: Update the new cluster center according to the data points assigned to each cluster center, and repeat this process to update the cluster center and similarity matrix until the position of the cluster center no longer changes;

[0036] S4.4: Determine the clustering results of the representative features of the T1 superpixel image and the representative features of the T2 superpixel image according to the updated cluster centers and similarity matrix;

[0037] S4.5: Compare each cluster center in the clustering results of period T1 and period T2:

[0038] That is, by calculating the Euclidean distance of the cluster centers and comparing the intra-class differences and inter-class differences, the final change results of the remote sensing images in the first period and the second period are determined.

[0039] The advantages of the present invention compared with the prior art are:

[0040] 1. The present invention uses the SLIC superpixel segmentation method to segment the remote sensing image into superpixels, which can better preserve the spatial information of the image and reduce the complexity of calculation;

[0041] 2. The present invention uses principal component analysis to effectively eliminate noise and redundant information in the data, further improving the accuracy of change detection;

[0042] 3. The present invention uses the k-means clustering algorithm to realize clustering, which can more accurately reflect the clustering characteristics of data points, so as to obtain better clustering results for comparison, and thus can better detect the changes of two remote sensing images in different phases, obtain better change detection accuracy, and also improve the accuracy of change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of the steps of a superpixel-based rapid remote sensing image change detection method provided by the present invention;

[0044] Figure 2 A superpixel segmentation method provided by the present invention: a flow chart of the SLIC method;

[0045] Figure 3 The present invention provides a flowchart of the k-multi-means clustering method. DETAILED DESCRIPTION

[0046] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0047] In recent years, object-oriented remote sensing image change detection has received more and more attention, and many object-oriented change detection methods have been derived. Therefore, the improvement of the object-oriented post-classification comparison method has also attracted much attention. Therefore, starting from this aspect, the present invention proposes a fast remote sensing image change detection method based on superpixels to improve object-oriented change detection. At the same time, it can perform change detection with high robustness, high precision and high efficiency, and achieve a higher change detection accuracy.

[0048] The present invention will be described in more detail later in conjunction with the accompanying drawings. The present invention provides a method for rapid remote sensing image change detection based on superpixels, and its flow chart is as follows: Figure 1 As shown, the following steps are included:

[0049] Step 1: Acquisition and preprocessing of remote sensing images at different phases;

[0050] Step 1.1: First, data preparation includes searching for remote sensing images of the same geographical area and different phases, where the original images used have high spatial resolution. Then, the remote sensing images of the same geographical area and different phases are subjected to region selection and cropping using MATLAB software, and normalization is performed at the same time, thereby obtaining pixel-aligned image pairs, namely, aligned T1 phase remote sensing images and T2 phase remote sensing images.

[0051] Step 1.2, the label images corresponding to the original images of the two time phases are respectively selected and cropped according to the regions of the remote sensing images of different time phases registered above to obtain corresponding label images, and pixel-by-pixel comparison is performed to obtain the change label images of the remote sensing images of the two time phases;

[0052] Step 2: Overlay and fuse the spectral features of the two images to form a new data block, and perform SLIC superpixel segmentation to obtain superpixel blocks of the same scale at different periods. At this time, the superpixel blocks have time-space-spectral characteristics. The spectral characteristics of unchanged objects are the same, while the spectra of changed objects are different.

[0053] The present invention adopts the SLIC method to perform superpixel segmentation, and its flow chart is as follows: Figure 2 As shown, the superpixel segmentation process of the first remote sensing image and the second remote sensing image will be described in detail below, including:

[0054] Step 2.1, dividing the T1 phase remote sensing image and the T2 phase remote sensing image into initial super-pixel grids;

[0055] The process first specifies the desired number of superpixels K and calculates the average size of superpixels. Then, the first remote sensing image and the second remote sensing image are divided into a roughly uniform grid, where each grid cell corresponds to a superpixel.

[0056] Step 2.2, determining the initial clustering centers of the T1 remote sensing images and the T2 remote sensing images;

[0057] Selecting a seed point for each grid unit of the initial superpixel grid, wherein the seed point is a pixel with a strong edge selected by calculating a gradient, and then using each seed point as an initial clustering center, thereby obtaining the initial clustering centers of the T1 remote sensing image and the T2 remote sensing image;

[0058] Step 2.3, perform iterative clustering and update the cluster center;

[0059] Calculate the feature distance of pixels within a certain range of the cluster center, and assign each pixel to the cluster center that is closest to it; for each cluster, calculate the average position and color value of all pixels assigned to it, and then update the position of the cluster center to this average position and color; repeat the steps of pixel assignment and cluster center update until the convergence condition is reached, that is, a more reasonable division is achieved;

[0060] Step 2.4, performing subsequent optimization processing on the segmentation results;

[0061] The segmentation results of the T1 remote sensing image and the T2 remote sensing image are retrieved into small-sized super-pixel blocks, and these small-sized super-pixel blocks are merged into adjacent and similar super-pixels to achieve further optimization results, that is, super-pixel images of the T1 remote sensing image and the T2 remote sensing image are obtained;

[0062] Step 3: Perform principal component analysis on the superpixel blocks obtained in step (2) to extract representative features from each superpixel block;

[0063] Step 3.1, principal component analysis is to convert the correlated variables in the original data into a set of new variables that are uncorrelated with each other through linear changes, so that the original data can be replaced by fewer variables, thereby extracting better representative features; first, the covariance matrix of the super-pixel image data of the T1 period and the super-pixel image data of the T2 period are calculated respectively, and then the eigenvalue decomposition of the covariance matrix is ​​performed to obtain the eigenvalue and the corresponding eigenvector;

[0064] Step 3.2, use the calculated eigenvalues ​​and eigenvectors to select the principal components and construct the projection matrix;

[0065] The obtained eigenvalues ​​are sorted in descending order, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected, where k is the dimension after dimensionality reduction, and the eigenvector is the principal component; the selected k eigenvectors are combined into a projection matrix;

[0066] Step 3.3, multiplying the superpixel image data of the two periods by the projection matrix to obtain the data after dimensionality reduction, that is, obtaining the first representative feature and the second representative feature. This process effectively eliminates noise and redundant information, and also reduces the computational complexity;

[0067] Step 4: Use the k-means algorithm to perform binary cluster analysis on the features extracted in step (3), dividing them into blocks that have changed and blocks that have not changed, to achieve change detection processing.

[0068] The flow chart of the k-means clustering algorithm of the present invention is as follows: Figure 2 As shown, the following is the detailed process of clustering using this algorithm;

[0069] Step 4.1: Process the first feature and the second feature respectively. First, randomly select m data points as the initial cluster centers. Then, use the Euclidean distance to calculate the clustering of each data point to the cluster center, and assign each data point to the cluster center with the closest distance.

[0070] Step 4.2, using the cluster center matrix A, according to the objective function of the k-means algorithm, calculate the similarity matrix S;

[0071] The objective function is:

[0072]

[0073] In the formula, x i is the i-th data in the data matrix X, a jis the jth center in the cluster center matrix A, S is the similarity matrix, γ is the regularization parameter, λ is the trade-off parameter, is the normalized Laplacian matrix, in D is a diagonal matrix, and the elements on the diagonal are ii Equal to the sum of the elements in the corresponding row i of the matrix P, F = [U, V] T ,U,V are The first k left and right singular vectors of ;

[0074] According to the objective function, the similarity matrix S can be calculated to achieve a clustering;

[0075] Step 4.3, update F according to the calculated similarity matrix S, thereby updating the similarity matrix S and the cluster center matrix A; that is, use the data points assigned to each cluster center to update the new cluster center, where this is updated according to the average value of all points assigned to each center; repeat the update of the cluster center and the similarity matrix until the position of the cluster center no longer changes;

[0076] Step 4.4, obtain the corresponding clustering results according to the updated similarity matrices of the two periods, that is, obtain the clustering results of period T1 and the clustering results of period T2 respectively;

[0077] Step 4.5: Compare each cluster center in the clustering results of period T1 and period T2; by calculating the Euclidean distance of the cluster centers and comparing their intra-class differences and inter-class differences, the final change results of the remote sensing images of period T1 and period T2 can be obtained.

[0078] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A superpixel-based fast remote sensing image change detection method, characterized in that: The following steps are involved: S1: Acquire remote sensing images of different time phases of the same geographical area, including T1 and T2, and perform preprocessing to obtain image pairs and change label images of the two time phase remote sensing images; S2: Overlay and fuse the spectral features of the two images to form a new data block, and perform SLIC superpixel segmentation to obtain superpixel blocks of the same scale in different periods; S3: Perform principal component analysis on the superpixel blocks in S2 to extract representative features in each superpixel block; S4: Use the k-means algorithm to perform binary clustering analysis on the features extracted in S3: The blocks are divided into changed blocks and unchanged blocks for change detection.

2. The method for rapid remote sensing image change detection based on superpixels according to claim 1, characterized in that: The S1 includes: MATLAB is used to select and crop the remote sensing images of the same geographical area and different phases, and the pixel-aligned image pairs are obtained after normalization. The label images corresponding to the original images of the T1 and T2 phases are selected and cropped according to the regions of the registered remote sensing images of different phases to obtain the corresponding label images, and pixel-by-pixel comparison is performed to obtain the change label images of the remote sensing images of the T1 and T2 phases.

3. The method for rapid remote sensing image change detection based on superpixels according to claim 1, characterized in that: The step of obtaining the super pixel block in S2 comprises the following steps: S2.1: evenly divide the T1 remote sensing image and the T2 remote sensing image into initial superpixel grids, and determine the initial clustering centers of the T1 remote sensing image and the T2 remote sensing image; S2.2: Perform iterative clustering and update the cluster centers; S2.3: Optimize the results in S2.2 to obtain super-pixel images of the T1 remote sensing image and the T2 remote sensing image.

4. The method of rapid remote sensing image change detection based on superpixels according to claim 3 is characterized in that: The step S2.2 includes: assigning surrounding pixels to the nearest cluster center, updating the position of each cluster center by all assigned pixels, repeating the iteration, and generating a plurality of super pixels; The S2.3 includes: retrieving small-size superpixel blocks from the segmentation results of the T1 remote sensing image and the T2 remote sensing image, and merging the small-size superpixel blocks into adjacent and similar superpixels to obtain superpixel images of the T1 remote sensing image and the T2 remote sensing image.

5. The method for rapid remote sensing image change detection based on superpixels according to claim 1, characterized in that: The S3 includes the following steps: S3.1: Linear transformation is applied to the T1 phase super-pixel image and the T2 phase super-pixel image to obtain low-dimensional data of the original data; S3.2: Based on the low-dimensional data in S3.1, representative features of the T1 phase superpixel image and representative features of the T2 phase superpixel image are obtained.

6. The method of rapid remote sensing image change detection based on superpixels according to claim 5, characterized in that: The S3.1 includes: calculating the covariance matrix of the T1 phase super pixel image and the T2 phase super pixel image data respectively, performing eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; S3.2 includes selecting principal components according to the eigenvalues ​​and eigenvectors in S3.1, and constructing a projection matrix, multiplying the superpixel image data of the T1 period and the T2 period by the projection matrix to obtain the reduced dimension data, that is, to obtain the first representative feature and the second representative feature.

7. The method of rapid remote sensing image change detection based on superpixels according to claim 6, characterized in that: The eigenvalues ​​obtained in S3.1 are sorted in descending order, and the eigenvectors corresponding to the largest k eigenvalues ​​are selected, where k is the dimension after dimensionality reduction, and the eigenvector is the principal component. The selected k eigenvectors are combined into a projection matrix.

8. The method for rapid remote sensing image change detection based on superpixels according to claim 1, characterized in that: The S4 also includes the following steps: S4.1: randomly select m cluster centers of representative features of T1 superpixel images and T2 superpixel images, calculate the distance from the data point to the cluster center according to the Euclidean distance, and assign each point to the cluster center with the closest distance; S4.2: Determine the similarity matrix based on the distance from the data point to the cluster center and the objective function of the algorithm, where the objective function is: In the formula, x i is the i-th data in the data matrix X, a j is the jth center in the cluster center matrix A, S is the similarity matrix, γ is the regularization parameter, λ is the trade-off parameter, is the normalized Laplacian matrix, in D is a diagonal matrix, and the elements on the diagonal are ii Equal to the sum of the elements in the corresponding row i of the matrix P, F = [U, V] T ,U,V are The first k left and right singular vectors of ; S4.3: Update the new cluster center according to the data points assigned to each cluster center, and repeat this process to update the cluster center and similarity matrix until the position of the cluster center no longer changes; S4.4: Determine the clustering results of the representative features of the T1 superpixel image and the representative features of the T2 superpixel image according to the updated cluster centers and similarity matrix; S4.5: Compare each cluster center in the clustering results of period T1 and period T2: That is, by calculating the Euclidean distance of the cluster centers and comparing the intra-class differences and inter-class differences, the final change results of the remote sensing images in the first period and the second period are determined.

Images