CN111353371A

CN111353371A - Coastline extraction method based on satellite-borne SAR image

Info

Publication number: CN111353371A
Application number: CN201911153838.3A
Authority: CN
Inventors: 刘晓霞; 魏曦; 李静; 刘坡
Original assignee: Electronic Science Research Institute of CTEC; Chinese Academy of Surveying and Mapping
Current assignee: Electronic Science Research Institute of CTEC; Chinese Academy of Surveying and Mapping
Priority date: 2019-11-22
Filing date: 2019-11-22
Publication date: 2020-06-30

Abstract

A coastline extraction method based on a satellite-borne SAR image comprises the following steps: s1: firstly, selecting an SAR image map, and reading the SAR image; s2: preprocessing an image, amplifying an original SAR image, and intercepting a part of image with a coastline; s3: performing Kmeans algorithm processing on the intercepted image by using a Kmeans clustering segmentation method; s4: combining morphological processing, performing morphological closing operation, selecting structural elements, expanding and corroding; s5: carrying out coastline extraction by using a Canny boundary extraction operator; s6: and analyzing the extracted result, overlapping the extracted coastline with the original map, and verifying the correctness of the coastline extraction result. Compared with other similar edge extraction algorithms, the coastline extracted by the method is good in continuity, excellent in smoothness and high in goodness of fit with a real coastline.

Description

Coastline extraction method based on satellite-borne SAR image

Technical Field

The invention relates to a remote sensing image processing method, in particular to a coastline extraction method based on a satellite-borne SAR image.

Background

The satellite airborne synthetic aperture radar is a basic earth observation tool for satellite remote sensing at present, and compared with the traditional optical remote sensing and hyperspectral remote sensing, the satellite airborne synthetic aperture radar has all-weather and high-resolution imaging capability, and can shoot images particularly when the number of clouds is large.

With the progress of technology, future satellite-borne synthetic aperture radars certainly realize more functions, realize all high-resolution broadband multimode microwave imaging, and realize the miniaturization and cost reduction of artificial synthetic aperture radars.

Although the technical problem of satellite onboard synthetic aperture radar breakthrough still exists in order to obtain as much information as possible at the minimum cost, the satellite onboard synthetic aperture radar inevitably enters a new development period along with the continuous development of the technology.

The synthetic aperture radar has the advantage that the synthetic aperture radar has penetrability so that an image capable of reflecting the scattering characteristics of the target microwave can be obtained, and the synthetic aperture radar is an important method for acquiring the ground object information. In addition, the SAR is coherent imaging, the SAR image can be synthesized in an aperture mode due to the characteristic of the coherent imaging, the SAR image obtained by the method has high resolution, and detailed map information can be provided.

The detection of coastlines is very important in coastal zones and can be used for various activities such as geographic maps, automatic navigation, coast erosion and monitoring. Coastline extraction of Synthetic Aperture Radar (SAR) images is becoming increasingly popular due to its wide coverage and all-weather capabilities. However, high accuracy shoreline extraction remains a challenging problem due to strong ocean currents, shadows, or specific coastal types, i.e., sandy coasts, etc., caused by speckle and contrast deficiencies. In recent decades, many methods have been proposed by many people both at home and abroad for coastline detection of SAR images. The following are roughly classified: boundary tracking, active contour, Level Set, Markovian segmentation, wavelet transform, etc. In the algorithms, the comprehensive evaluation of the detection speed and the detection effect of the horizontal cut-set algorithm is higher than that of other algorithms, and although people improve the algorithms to improve the detection speed, the realization principle is quite complex, the detection speed is relatively slow, and the requirements of actual engineering cannot be met. There is also a method for extracting a coastline specifically for an SAR image, for example, in "a method for extracting a coastline in an SAR coastline image" in a domestic patent with application number CN201610621676.1, a coastline is obtained by confirming a geometric center of an ocean area, making a ray with the point as a starting point, determining a coastline boundary point on the ray, and sequentially connecting the points. The method is suitable for large-scale images, is not suitable for the situation that the images are divided, and needs to be further researched.

Disclosure of Invention

In view of the above, the present invention provides a coastline extraction method based on a satellite-borne SAR image, which is used to solve the problem of coastline extraction, and the technical solution is as follows:

a coastline extraction method based on a satellite-borne SAR image comprises the following steps:

s1: firstly, selecting an SAR image map, and reading the SAR image;

s2: preprocessing an image, amplifying an original SAR image, and intercepting a part of image with a coastline;

s3: performing Kmeans algorithm processing on the intercepted image by using a Kmeans clustering segmentation method;

s4: combining morphological processing, performing morphological closing operation, selecting structural elements, expanding and corroding;

s5: carrying out coastline extraction by using a Canny boundary extraction operator;

s6: and analyzing the extracted result, overlapping the extracted coastline with the original map, and verifying the correctness of the coastline extraction result.

Further, in step S3, in the Kmeans clustering algorithm, the selected K value is 10.

Further, in step S3, the algorithm processing further includes performing segmentation binarization according to the clustering pattern of the 10 th cluster of the K-means cluster map, and performing filling and denoising processing on the binarized image.

In step S4, the structural element is selected to be 4.

In step S4, dilation and erosion are calculated for a region in the image, relating to the features of lines and points, in which dilation is a region that expands around and erosion is a region that reduces around from the same time.

In step S5, the Canny boundary extraction operator performs coastline extraction, and includes the following steps:

① denoising, wherein the original image data must be convolved with a two-dimensional Gaussian filter template;

② gradient calculation the use of a derivative operator gives the derivative G of the grey scale image in each of the two directions_XAnd G_YFrom the derivative obtained, the magnitude | G | and direction θ of the gradient can be calculated:

③ gradient direction determination, calculating the direction of the edge and dividing the gradual change direction of the edge by a plurality of angles to find out the adjacent pixels in the pixel direction;

④ traversing the whole image, when the gray values of two pixels are not the maximum value of the gray values of the pixels before and after the gradient direction, the gray value of the pixel is 0, i.e. not the edge;

⑤ derives two thresholds by accumulating histograms, above which it must be an edge and below which it should not be, if the detection is in the middle of the two thresholds, then whether the edge pixels in the pixels' neighbours are judged by the edge pixels not having a higher threshold, if so, an edge, otherwise, no edge.

The method can be used for conveniently extracting the coastline from the SAR image with the complex scene and large geographic coverage, and has strong automation and adaptability.

Drawings

FIG. 1 is a technical roadmap for the present method;

FIG. 2 is an original image of Sentinel-1A;

FIG. 3 is a cut out portion of a shoreline;

FIG. 4 is a result of processing by the kmeans algorithm;

FIG. 5 is a divided binary graph;

FIG. 6 is a random matrix;

FIG. 7 is a first step in clustering;

FIG. 8 is a clustering iteration step;

FIG. 9 is a cluster demonstration result;

FIG. 10 is the denoising result;

FIG. 11 is a result of morphological processing;

FIG. 12 is a pre-noise reduced image;

FIG. 13 is a partial enlargement of an image before noise reduction;

FIG. 14 shows the canny operator extraction result;

fig. 15 is a coastline extraction end result.

Detailed Description

The coastline extraction method based on the satellite-borne SAR image, as shown in figure 1, comprises the following steps: 1) firstly, selecting an SAR image map, and reading the SAR image; 2) preprocessing an image, amplifying an original SAR image, and intercepting a part of image with a coastline; 3) performing Kmeans algorithm processing on the intercepted image by using a Kmeans clustering segmentation method; the processing also comprises segmentation and binarization according to a clustering mode of the 10 th cluster of the K-means clustering graph, and filling and denoising processing is carried out on the binarized image; 4) combining morphological processing, performing morphological closing operation, selecting a structural element of four, expanding and corroding; 5) carrying out coastline extraction by using a Canny boundary extraction operator; 6) and analyzing the extracted result, namely overlapping the extracted coastline with the original map, and verifying the correctness of the coastline extracted result. The following is a detailed description of the steps:

a

As shown in fig. 2, the original SAR image to be processed is an image of Sentinel-1A in the sea area of yellow sea, the imaging time is 2017.12.05, the polarization mode is VV, the pixel size is 15 ×. the information of the SAR image is mainly affected by backscattering, the brightness of the image represents backscattering intensity, the coarser the inner surface of the pixel, the stronger the backscattering, the specular reflection is generated on the smooth surface, the backscattering is very weak, and the brightness is low.

II

A part of the shoreline is cut out on the basis of fig. 2, i.e. fig. 3. Image segmentation undoubtedly plays an important role as one of the key techniques for digital image processing. Image segmentation provides a solid foundation for subsequent image processing, recognition, analysis and understanding by extracting meaningful feature locations (e.g., image edges and image regions) in an image. Currently, many different methods for extracting edges or image segmentation have been developed for various images. However, none of these methods can be applied to all cases, and it is one of the main reasons why the research of image segmentation is still one of the hot spots of image processing.

III

As shown in fig. 4, the captured image in fig. 3 is processed by a Kmeans algorithm, after the processing by the Kmeans algorithm, the distinctiveness of the ocean and land intensities is increased, and then the segmentation and binarization are performed according to the clustering pattern of the 10 th cluster of the K-means clustering map to obtain fig. 5.

In the simplest clustering algorithm, there is no doubt that the Kmeans clustering segmentation is local. This is a very typical unsupervised learning algorithm. It is mainly used to automatically group similar samples into one category. The clustering algorithm is the biggest difference from the classification algorithm in that the clustering algorithm is an unsupervised learning algorithm which can be automatically processed. The classification algorithm belongs to a supervised learning algorithm and needs human intervention.

The difficulty of the Kmeans algorithm is that different k values need to be set to obtain different clustering results. The uncertainty of the k value is however a disadvantage of this algorithm. Often, to achieve good experimental results, multiple attempts are required to select the optimal k value. In this experiment, the selected K value was 10. When K is large or small, the clustering effect is not as good as 10, and is comparatively dispersed.

The basic idea of Kmeans is to concentrate k points into space and classify the object that is closest to it. Through an iteration method, the value of each clustering center is continuously updated, and the whole updating iteration process can not be stopped until the optimal clustering result appears. If formulated, assume that the cluster is divided into (C)₁，C₂，┄C_i) Then, it is expressed as:

wherein C is_iIs the first cluster, x is C_iSample point of (1), u_iIs C_iCenter of mass (C)_iMean of all samples), SSE is the clustering error of all sample samples, and represents how good the clustering effect is. For a simple understanding of this process, a small experiment is performed herein to demonstrate the operational process.

First, three normally distributed random matrices are generated using the randn function, as shown in fig. 6, and they are referred to as original images.

K random cluster center points are then generated, and for simplicity, k is set to be the same as the number of matrices, i.e., k is 3. The next step is to calculate the distance of each data point to these centers separately. Each data point takes the central point with the shortest distance as its own category. At this time, each data point has its own corresponding center point, however, as can be seen from fig. 7, the clustering at this time is not accurate.

At this time, the position of the center point is recalculated. The centers of all blue dots are calculated and the blue center dot is moved to the calculated position. The green dot in fig. 8 is now classified as the new blue center dot and is thus colored blue. The center points of green and red are also moved to the new position in the same way.

And (4) after the steps are repeated continuously until the iteration of the central point tends to be stable, stopping the iteration and finishing the algorithm. The clustering result is obtained as shown in fig. 9.

The samples with the same color are gathered into a cluster, and finally three clusters are formed, so that the samples in the same cluster have high similarity, and the difference between different clusters is high.

Immediately after clustering the images, binarization processing is performed on the processed images. The simple points of the image binarization are all points of the whole image expressed as 0 or 255, namely the effect of non-black or white, and the whole image is changed into a distinct white-white effect. In other words, the original grey value image is converted to a black and white binary image, and by selecting an appropriate threshold, global and local features can be seen. Points with a gray level greater than or equal to the threshold are determined as one type, the gray level is represented as 255, and the point with a gray level below that threshold is determined as another type, the gray level is represented as 0 (i.e., the background or something don't matter).

The binaryzation of the image can be also dynamic by dynamically adjusting the threshold value, so that the effect on the segmentation image can be observed, and the dynamic analysis is realized. By defining non-overlapping portions using closed or connected boundaries, a more ideal binary image can be obtained. This is a critical step for digital image processing, where binary images are of great importance, especially in some aspects of image processing, where many systems require processing through binary images.

To process and analyze a binary image, binarizing the grayscale image is obviously the first step. The binary image obtained by conversion only relates to black or white points when the next image processing is carried out, and the multilevel property of the pixels does not need to be considered, so that the processing process can be simplified.

It can be seen that there is much noise on the binarized image, the dot pattern and the pinholes are spread over the whole image, and they are first filled and denoised for the next steps. After the denoising process is completed, fig. 10 is obtained.

Fourthly

Obviously, the coastline part still seems to have a problem, then, morphological closed operation is applied, the selection of the structural elements is repeatedly tried for the test effect, and finally, the denoising effect of the rest structural elements is not obvious when the structural element is 4. Expansion followed by erosion gave the figure 11.

It is well known that it is generally impossible to remove the noise completely. However, in order to convert the image from a grayscale image to a binary image, the occurrence of some noise is also inevitable. This situation can cause difficulties for image extraction.

There are a wide variety of manifestations of noise in binary images. Most representative are dots and holes, as shown in FIG. 12. Fig. 13 is a cut-out portion of the image before noise reduction in the present experiment (i.e., fig. 12). Dot patterns and pinholes refer to pixel connection parts and relatively small-area zero-pixel connection parts. Generally, the etching process and the swelling process can effectively remove these connections.

Set theory is the mathematical basis and language used in mathematical morphology. There are several most basic operations in image morphology processing called dilation, erosion, switching, which have their own features in both binary and grayscale images. These basic operations can also be derived and incorporated into various algorithms for mathematical morphology.

Let f (x, y) be the input image, g (i, j) represent the structural element, theta is the corrosion operation symbol,

for the sign of the dilation operation, D_fAnd D_gThe domains of f and g, respectively, then the erosion and expansion of f by g can be expressed as:

f(x,y)Θg(i,j)＝min{f(x+i,y+j)-g(i,j)|(x+i,y+j)∈D_f,(i,j)∈D_g} (2)

is provided with

Represents an open operation,. represents a closed operation, defined as:

wherein the mathematical operation is usually used to remove the bright details smaller than the structural elements, while leaving the large overall bright features unchanged; the close operation is typically used to remove dark details smaller than the structuring element while also keeping the highlighting feature elements unchanged.

Dilation and erosion are features that calculate a region (line and point) in the image. In an image, dilation is a region that expands to the periphery, while erosion is a region that decreases from the periphery at the same time. It is noted that in general, swelling and erosion are not reciprocal, i.e. they can be used in cascade. After dilation and erosion, or after erosion has expanded, it is generally not possible to restore the original image, but it produces a new form of transformation. This is also called a form opening and closing operation.

Obviously, image noise filtering is an integral part of image preprocessing. The morphological noise filter is constructed by combining open and closed operations. The noise blocks around the target in the above-mentioned binary image can be eliminated by performing an on operation on the set using the structural elements; in contrast, noise holes in the target can be removed using a closed operation.

In the above method, the selection of the structural element is important and must be larger than all the noise (including noise blocks and noise holes) in the binary image. In conducting this study, a number of structural elements were tried repeatedly, and finally the selected parameter size was determined to be 4. And when the structural elements take the numerical values, the experiment has better effect.

When image processing is actually performed, on operations are generally used to eliminate sizes smaller than the structural elements. The overall gray value and bright area of the image are kept as large as possible over the structural elements while preserving detail. Instead, the closing operation is used to eliminate black detail smaller than the structural elements. It is of course also necessary to keep the gray values of the image and dark areas larger than the structuring elements.

Through the above description, a conclusion can be simply obtained, the operation of filtering various noises can be achieved by combining the opening operation and the closing operation, and if multiple structural elements can be combined with the opening and closing operation, a great result can be obtained in the aspect of protecting image details.

Five of them

The coastline in fig. 11 is extracted using the Canny operator, as shown in fig. 14. Finally, in order to verify the correctness of the coastline extraction result, the extracted coastline is superimposed with the original map, as shown in fig. 15.

The edge detection technique is very important for processing a digital image because an edge is a boundary line between an object and a background to be extracted, and an edge can be extracted to distinguish the object from the background. In the image, the boundary represents the end of a feature region and the beginning of another feature region. The internal features or attributes of the bounded boundaries are the same, but the internal features or attributes of different regions are different. Certain image features, including grayscale, color, or texture features, facilitate edge detection. Edge detection is actually a kind of detection of changes in image characteristics.

The classic method of edge extraction is to examine the gray scale change of each pixel of an image in a specific region. Edges are detected using simple methods using first or second directional derivatives near the edges. This method is called edge detection.

The basic idea of edge detection is to determine whether a pixel is located at an edge position of an object by detecting the state of each pixel and its neighbors. If each pixel is located on the boundary of the object, the gray values of the neighboring cells will change more.

The boundaries extracted by the Canny operator are relatively continuous and smooth. Thus, the Canny operator is selected herein as the edge operator for the coastline extraction.

The Canny edge operator is optimal as an edge detection operator. This is widely used in many image processing fields. Canny checks for edges has several requirements for the detector: the error rate is low: real edge points are not lost as far as possible, so that the edge is prevented from being judged as a non-edge point; high position accuracy: the detected edge should be as close as possible to the true edge; each edge point has a unique response, forming a single pixel wide edge. The Canny edge operator is specifically realized by the following steps:

① De-noising the edge detection algorithm cannot process the raw image unprocessed and therefore must first convolve the raw data with a Gaussian smoothing template.

② gradient calculation the use of a derivative operator makes it easy to derive the respective derivatives G of a grey scale image in two directions_XAnd G_YFrom the derivative obtained, the magnitude and direction of the gradient can be calculated:

G_Xand G_YRepresenting the derivative of the grey scale image in each of the x and y directions.

③ the gradient direction determines calculating the direction of the edge and dividing the direction of the gradual transition of the edge by a number of angles (e.g., 0, 45, 90, and 135) to find neighboring pixels in the pixel direction.

④ traverse the entire image, the gray value of a pixel is 0, i.e. not an edge, when the gray values of two pixels are not the maximum of the gray values of the pixels before and after the gradient direction.

⑤ derive the two thresholds by means of cumulative histograms, it is clear that above the threshold must be an edge and below the threshold should not be an edge if the detection is in the middle of the two thresholds then whether an edge pixel in the neighborhood of the pixel is judged by the edge pixel not having a higher threshold, if so, an edge, otherwise not an edge.

Sixthly,

Observing fig. 15, it can be seen that the extracted coastline is good in continuity and excellent in smoothness, and substantially coincides with the coastline although the accuracy is not high enough in some places.

Compared with the commonly used boundary tracking method, the Kmeans cluster segmentation method is slightly inferior to the boundary tracking method in precision, but is much stronger than the former in applicability. Since the boundary tracking method can only detect the coastline under the condition of a given SAR image, if the complete coastline is segmented by the image, the use of the boundary tracking method can cause an embarrassing situation that only half of the coastline can be traced and the other half of the coastline cannot be displayed. The Kmeans method has no limitation, has stronger adaptability to the division of the coastline by the image, and is simpler and more convenient than a boundary tracking method.

The method for extracting the coastline by combining the Kmeans clustering segmentation and the morphology is suitable for automatically extracting the coastline in the SAR image. Experimental results show that the influence of speckle noise can be effectively relieved by combining Kmeans and a morphological processing method. Furthermore, the approach presented herein has advantages in maintaining the continuity of the shoreline in complex scenarios.

Claims

1. A coastline extraction method based on a satellite-borne SAR image comprises the following steps:

s1: firstly, selecting an SAR image map, and reading the SAR image;

2. The coastline extraction method based on spaceborne SAR images as claimed in claim 1, wherein: in step S3, the K value selected in the Kmeans clustering algorithm is 10.

3. The coastline extraction method based on spaceborne SAR images as claimed in claim 2, wherein: in step S3, the algorithm processing further includes segmenting and binarizing according to the clustering pattern of the 10 th cluster of the K-means cluster map, and performing filling and denoising processing on the binarized image.

4. The coastline extraction method based on spaceborne SAR images as claimed in claim 1, wherein: in step S4, the structural element is selected to be 4.

5. The coastline extraction method based on spaceborne SAR images as claimed in claim 1, wherein: in step S4, dilation and erosion are calculated for a region in the image, relating to the features of lines and points, in which dilation is a region that expands around and erosion is a region that reduces around from the same time.

6. The coastline extraction method based on spaceborne SAR images as claimed in claim 1, wherein: in step S5, the Canny boundary extraction operator performs coastline extraction, and includes the following steps:

② gradient calculation of derivative operatorUsing derivation of respective derivatives G in two directions of the gray-scale image_XAnd G_YFrom the derivative obtained, the magnitude | G | and direction θ of the gradient can be calculated: