CN114066816B - Unsupervised change detection method for SAR images based on static wavelet transform extraction - Google Patents

Abstract

The invention discloses a SAR image unsupervised change detection method based on static wavelet transform extraction, and the realization steps are as follows: firstly, a preprocessing stage. The Lee filter is used to filter and smooth the two SAR images; the second step is to generate a difference map. Use the logarithmic ratio operator to generate a difference map; the last stage is to analyze the difference map. First use SWT2 (db2), (two-dimensional static wavelet transform, db2 is the wavelet basis function) to decompose the difference map to obtain approximate image, horizontal detail image, vertical detail image and diagonal detail image. Then use EM_GGM to decompose the four groups of images respectively, and finally select the probability of the second-order neighborhood window to form the feature vector, and obtain the result through K-means.

Description

Unsupervised change detection method for SAR images based on static wavelet transform extraction

technical field

The invention belongs to the technical field of image change detection, and relates to an unsupervised change detection method for SAR images based on static wavelet transform extraction.

Background technique

SAR is an active side-looking radar system, and the imaging geometry belongs to the oblique-range projection type. Therefore, SAR images are quite different from optical images in terms of imaging mechanism, geometric features, and radiation features. SAR uses the principle of synthetic aperture to realize high-resolution microwave imaging, and has various characteristics such as all-day, all-weather, high-resolution, and large width. Therefore, the SAR system has unique advantages in the application of disaster monitoring, environmental monitoring, ocean monitoring, resource exploration, crop yield estimation, surveying and mapping, and military applications, and can play a role that other remote sensing methods are difficult to play. attention.

The basic flow paradigm of single polarization SAR image change detection is: 1) Preprocessing 2) Generate difference map 3) Analyze difference map. The two steps of generating difference map and analyzing difference map are the key research directions of SAR image change detection in recent years. The main purpose is to reduce the influence of coherent speckle noise on SAR images.

1) Preprocessing is the basic preparation for generating the difference map.

2) The purpose of generating the difference map is to initially distinguish between the unchanged class and the changed class in the two SAR images. The generation of the difference map is actually to find a matrix that can represent the distance between the two SAR images. After the matrix is visualized It is the difference map. We can construct a difference map with the same size as the two through some difference operation. The improved difference map generation algorithm is as follows: 1. Log-ratio (LR) operator operation, on the basis of the ratio difference map, an additional logarithmic operation is added. This method converts the coherent speckle noise in the SAR image into additive noise, and after logarithmic transformation, the difference image is nonlinearly shrunk, which enhances the contrast between the change class and the non-change class. The nature of the logarithmic operation itself can reduce the large difference caused by the ratio operation, so it can further reduce the influence of wild points in the unchanged background part. It is more effective when the changed area is smaller than the unchanged area, but because the logarithm The operation shrinkage is strong, so the pixel values in the edge area are easily blurred. 2. The mean ratio (Mean-ratio, MR) operator operation, the object of comparison is no longer the corresponding isolated pixel point, but the mean value of the neighborhood where the pixel point is located, which has a certain degree of inhibition on the wild points that appear alone , but due to the lack of scaling transformation, if the noise does not appear in the form of points but in the form of patches, it is difficult to effectively suppress its influence. 3. Combined Difference Image (Combined Difference Image, CDI), this method performs parameter weighting on the difference difference image and the LR difference image to obtain a new difference image. In the CDI method, the difference difference map and the LR difference map are subjected to mean filtering and median filtering respectively to initially remove noise interference and wild points, and then use artificially weighted parameters to obtain the final fusion difference map. This method is simple and easy to implement, and is suitable for parallel processing, and the speed is fast; however, it contains artificial parameters and requires multiple tests to obtain the optimal parameter value, and it is not easy to automatically select according to the nature of the image itself. ) 4. Neighborhood-based Ratio (NR), the NR operator is a weighted average of the contrast difference map and the MR difference map. This weight can represent whether the position of the central pixel is in a homogeneous area or a heterogeneous area, a low value corresponds to a homogeneous area, and a high value corresponds to a heterogeneous area. This method fully combines the gray information and spatial information of pixels, and the weighting parameters are completely determined by the properties of the image itself, which improves the robustness of the difference map construction. 5. Wavelet Fusion (WF) method, firstly perform wavelet transform on the generated LR and MR difference maps, and then extract the low-frequency band of the MR difference map and the high-frequency band of the LR difference map, that is, extract the MR difference map The overall information of the graph and the detail information of the LR difference graph. Then, LL, LH, HL and HH are fused according to the fusion rules based on neighborhood to generate a new wavelet transform image. Finally, the wavelet inverse transform is performed to obtain the WF fusion difference map. This method combines the properties of the wavelet transform to combine the advantages of both LR and MR difference maps. 6. Combining the texture and intensity features of the SAR image to construct the difference map (Intensity and Texture, IT), decompose the two input SAR images with sparse and low-rank coefficients, and obtain the corresponding intensity and texture information respectively. Different information is used to construct a difference map and then fused. This method not only extracts the main change area in the SAR image, but also prevents speckle noise from affecting the performance of the difference map, especially it has strong robustness in maintaining this performance.

3) It is the analysis of the difference map. After the difference map is generated, it needs to be analyzed, and finally a black and white binary map is generated. There are four commonly used analysis methods: threshold analysis, cluster analysis, graph cut analysis and level set analysis. The threshold analysis method is to find an optimal threshold value through a certain threshold value selection method, and then divide the difference image into two categories based on the threshold pixel value; the cluster analysis method is to obtain the unchanged class and The two cluster centers of the changed class are then divided into two classes by the nearest neighbor method; the graph-cut analysis method is another binary classification method of the image, which essentially classifies the labels of the unchanged class and the changed class to the pixels , the method continuously optimizes the energy of the given constraint function. When the energy reaches the minimum, the image pixel can correspond to the optimal label; the level set analysis method uses the curve evolution to transform the evolution problem of the two-dimensional closed curve into three-dimensional The implicit method of surface evolution of the level set function in space is used to solve it, that is, to construct a three-dimensional level set function, and then calculate the deconstructed curve set whose value is zero, so as to obtain the image segmentation result.

Contents of the invention

The purpose of the present invention is to provide an unsupervised change detection method for SAR images based on static wavelet transform extraction, which can accurately separate SAR images.

The technical scheme adopted in the present invention is based on the SAR image unsupervised change detection method extracted by static wavelet transform, which specifically includes the following steps:

Step 1, preprocessing the two SAR images;

Step 2, based on the results obtained in step 1, use the logarithmic ratio operator to generate a difference map;

Step 3, analyze the difference map generated in step 2.

The present invention is also characterized in that:

The specific process of step 1 is: use the Lee filter to filter and smooth the two SAR images.

The specific process of step 2 is:

Assuming that the two SAR images are I ₁ and I ₂ respectively, the difference map D of the images I ₁ and I ₂ is:

D=|log(I ₁ )-log(I ₂ )| (1).

The specific process of step 3 is:

Step 3.1, using the db2 wavelet basis function to perform single-layer static wavelet transform on the difference map D obtained in step 2;

Step 3.2, using the EM-GGM algorithm to classify the images obtained in step 3.1;

In step 3.3, perform K-means clustering on the images classified in step 3.2.

The specific process of step 3.3 is: perform vhad feature selection for each image pixel in the second-order neighborhood, and finally use the K-means algorithm for clustering.

The beneficial effects of the present invention are: firstly, the present invention uses a Lee filter for filtering processing, eliminates noise, and minimizes the influence of noise points on change detection results. The difference map is then generated using the log ratio operator. Finally, the difference map is first decomposed by SWT2 (db2), (two-dimensional static wavelet transform, db2 is the wavelet basis function), and then decomposed by EM_GGM, and the window probability is selected to form the feature vector, and the result is obtained by K-means. The detection method provided by the present invention has the following advantages: (A) minimize the impact of speckle noise in the SAR image; (B) when analyzing the difference map, the SWT2 (db2) algorithm is introduced, and low-pass filtering can be used to reduce noise in the frequency domain Impact. Introducing the EM_GGM algorithm, GGM is a more robust and flexible model for describing changed and unchanged pixels in the difference map.

Description of drawings

Fig. 1 is the flowchart of the SAR image unsupervised change detection method extracted based on static wavelet transform in the present invention;

Figure 2 is a change detection reference map of the Bern dataset;

Fig. 3 is the change detection figure that Bern data set adopts the SAR image unsupervised change detection method that the present invention extracts based on static wavelet transform to obtain;

Figure 4 is a change detection reference map of the Ottawa dataset;

Fig. 5 is the change detection figure that Ottawa data set adopts the SAR image unsupervised change detection method that the present invention extracts based on static wavelet transform to obtain;

Figure 6 is a change detection reference map of the Shihmen Reservoir dataset;

Fig. 7 is a change detection diagram obtained by using the unsupervised change detection method of the SAR image extracted based on static wavelet transform in the Shihmen Reservoir data set of the present invention.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

The present invention is based on the SAR image unsupervised change detection method extracted by static wavelet transform. First, the static wavelet transform (Stationary wavelet transform, SWT) decomposition is performed on the difference map. The static wavelet transform is to overcome the discrete wavelet transform (discrete wavelet transform, DWT) translation. A wavelet transform algorithm designed for the disadvantage of invariance. The static wavelet transform is different from the discrete wavelet transform, mainly in that after each order of high-pass filter and low-pass filter , the filter is upsampled (Upsampling) to replace the downsampling of the discrete wavelet transform after the filter .

Secondly, the four groups of images obtained by SWT decomposition are decomposed by the expectation maximization algorithm of generalized Gaussian model (EM-GGM). The four groups of images are approximation images, horizontal detail images, vertical detail images and diagonal detail images respectively. The EM algorithm based on the generalized Gaussian model is used to fit the histogram of the class conditional distribution of the changed class and the unchanged class. At the beginning, the Set the initialization parameters to generate a lookup table, the E step is to use the Bayesian rule to calculate the posterior estimation of each pixel belonging to the two opposite classes according to the parameters at the time of iteration t, and the M step is to provide the new generalized Gaussian model for the opposite class Estimated values of parameters, iterations, until the difference between the last two posterior probabilities is less than the set minimum value, stop iterations, otherwise return to step E.

Finally, it is classified by K-Means clustering algorithm, and the feature selection is vhad second-order neighborhood. v is a vertical window, h is a horizontal window, a is a full window, and d is a diagonal window. The K-Means clustering algorithm uses the two points of the largest inter-class distance and the smallest intra-class distance to find the appropriate cluster center through iteration. Use the squared Euclidean distance as the basis for clustering samples.

As shown in Figure 1, it specifically includes the following steps:

Step 1, the preprocessing stage. The Lee filter is used to filter and smooth the two SAR images;

Step 2, the phase of generating a difference map. Use the logarithmic ratio operator to generate a difference map;

The specific process of step 2 is:

Assuming that the two SAR images are I ₁ and I ₂ respectively, the difference map D of the images I ₁ and I ₂ is:

D=|log(I ₁ )-log(I ₂ )| (1).

Step 3, analyze the difference map stage. First use SWT2 (db2), (two-dimensional static wavelet transform, db2 is the wavelet basis function) to decompose the difference map to obtain approximate image, horizontal detail image, vertical detail image and diagonal detail image. Then use EM_GGM to decompose the four groups of images respectively, and finally select the probability of the second-order neighborhood window to form the feature vector, and obtain the result through K-means.

The specific process of step 3 is:

Step 3.1, SWT2 (two-dimensional static wavelet transform):

The single-layer static wavelet transform is performed on the image with db2 wavelet basis function. MATLAB provides the function swt2 of the two-dimensional static wavelet transform, because it does not down-sample the decomposition coefficients, so the results of single-layer decomposition and multi-layer decomposition are the same in form.

Step 3.2, using the EM-GGM algorithm to classify the images obtained in step 3.1;

Choose GGM to model the conditional probability density function P( _xi |c), P( _xi |u), c is the label of the changed class, and u is the label of the unchanged class; the probability density function is:

Among them, μ _m , σ _m , γ _m are the mean, standard deviation and shape parameters of the two classes respectively; for the parameter set θ={P(c), P(u), μ _c , μ _u , σ _c , σ Reliable estimates of _u , γ _c , γ _u } can be learned by the EM algorithm, P(c), P(u) are prior probabilities; this algorithm is a maximum likelihood algorithm for learning parameters in incomplete data problems A general method for natural estimation. It includes an expectation step (E step) and a maximization step (M step) in iterations, and iterates until convergence is reached. The description of the EM-GGM algorithm is as follows:

start: initialization parameters:

θ ⁰ ＝{P(c), P(u), μ _c , μ _u , σ _c , σ _u , γ _c , γ _u } (3)

Generate a parameter lookup table and initialize the threshold T as 100:

Step E: Based on the parameters of the t-th iteration, the posterior estimate of each pixel belonging to the two opposite classes is calculated using Bayesian rule as:

P ^t (m| _xi )＝P ^t (m)P ^t ( _xi |m)/P ^t ( _xi ) (3);

P ^t (x _i )=P ^t (c)P ^t (x _i |c)+P ^t (u)P ^t (x _i |u) (4);

M step: update model parameters:

Among them, |X| represents the number of pixels, which is equal to n here, representing n pixels. P( _xi ) is the probability density function of pixel x _i at position i in the difference image. X represents the set of all pixels.

Convergence: Set the number of iterations to 200, and stop if the number of iterations is met. Otherwise proceed to E-step.

In step 3.3, perform K-means clustering on the images classified in step 3.2.

The vhad feature selection is performed on the pixels in each second-order neighborhood, and finally the K-means algorithm is used for clustering.

Example

Effect of the present invention can be specified by simulation experiment:

1. Experimental conditions

The microcomputer CPU used in the experiment is Intel Pentium 43.0GHz memory 1GB, and the programming platform is Matlab 7.0.1. The SAR images used for the experiment are the Bern dataset, the Ottawa dataset, and the Shihmen Reservoir dataset.

2. Experimental content

The first is the preprocessing stage. The Lee filter is used to filter and smooth the two SAR images; the second step is to generate a difference map. Use the logarithmic ratio operator to generate a difference map; the last stage is to analyze the difference map. First use SWT2 (db2), (two-dimensional static wavelet transform, db2 is the wavelet basis function) to decompose the difference map to obtain approximate image, horizontal detail image, vertical detail image and diagonal detail image. Then use EM_GGM to decompose the four groups of images respectively, and finally select the probability of the second-order neighborhood window to form the feature vector, and obtain the result through K-means.

3. Experimental results

Table 1 is the change detection evaluation index of the method proposed by the present invention and the other three methods on three data sets;

Table 1

Fig. 2 is the change detection reference figure of Bern data set; Fig. 3 is the change detection figure that Bern data set adopts the method of the present invention to obtain; Fig. 4 is the change detection reference figure of Ottawa data set; Fig. 5 is Ottawa data set adopts the method of the present invention The change detection figure that obtains; Fig. 6 is the change detection reference figure of Shihmen Reservoir data set; Fig. 7 is the change detection figure that ShihmenReservoir data set adopts the method of the present invention to obtain; Experimental result shows that the method that the present invention proposes can obtain good change detection result.

Claims (2)

1. The SAR image unsupervised change detection method based on static wavelet transformation extraction is characterized by comprising the following steps of: the method specifically comprises the following steps:

step 1, preprocessing two SAR images;

the specific process of the step 1 is as follows: filtering and smoothing the two SAR images by using a Lee filter;

step 2, generating a difference graph by adopting a logarithmic ratio operator based on the result obtained in the step 1;

the specific process of the step 2 is as follows:

let two SAR images be I respectively ₁ 、I ₂ Image I ₁ 、I ₂ The difference diagram D of (2) is:

D＝|log(I ₁ )-log(I ₂ )| (1)；

step 3, analyzing the difference graph generated in the step 2;

the specific process of the step 3 is as follows:

step 3.1, performing single-layer static wavelet transformation on the difference map D obtained in the step 2 by using db2 wavelet basis function;

step 3.2, classifying the images obtained in the step 3.1 by adopting an EM-GGM algorithm;

and 3.3, carrying out K-means clustering on the images classified in the step 3.2.

2. The method for detecting the unsupervised change of the SAR image based on the static wavelet transform extraction as set forth in claim 1, wherein: the specific process of the step 3.3 is as follows: and carrying out vhad feature selection on the image pixels in each second-order neighborhood, and finally clustering by using a K-means algorithm.