A method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images

Abstract

The present invention provides a method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images. The method includes: performing median filtering on the single-polarization SAR images, and obtaining superpixel images and edge images; generating a seed point discrimination criterion driven jointly by the edge images and the superpixel images, and automatically obtaining sea and land seed points based on the seed point discrimination criterion; constructing a double-layer graph based on the single-polarization SAR images and the superpixel images; respectively constructing a single-polarization SAR image, a superpixel image, and an edge weight function between the two images on the double-layer graph based on a Gaussian kernel function with the minimum mean ratio, and further constructing a weight similarity metric based on the edge weight function; constructing a random walk model with a double-layer graph structure, and obtaining a binary sea-land segmentation by solving the random walk model with the double-layer graph structure, and finally obtaining a shoreline detection result. The present invention can effectively detect shorelines in single-polarization SAR images and also has strong robustness for aquaculture shorelines that are difficult to detect.

Description

Technical Field

[0001] The present invention relates to the technical field of marine remote sensing image processing, and more particularly, to a method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images. Background Art

[0002] The coastal zone is the most active and concentrated area for human life and economic development. At the same time, it is also a typical ecologically fragile zone and a sensitive area for environmental changes. The shoreline is the most important geographical element for the land-sea demarcation of the coastal zone and is also one of the 27 surface elements recognized by the International Committee on Geographic Data. Achieving dynamic monitoring of the shoreline plays an important role in applications such as topographic surveying, ship navigation, and maritime emergencies, and has important practical significance for promoting the sustainable management and development and utilization of coastal zone resources and the environment. Compared with traditional on-site surveys, manual interpretation, and optical remote sensing, synthetic aperture radar (SAR) images have the advantages of all-weather and all-time, and are more suitable for rapid detection of shorelines. The polarization mode is an important attribute of SAR images. Compared with multi-polarization images, single-polarization SAR images have the advantages of easy acquisition, strong selectivity, and large coverage, and are currently more widely used.

[0003] Coastline detection for single-polarization SAR images can be divided into six categories of methods: (1) Edge detection-based methods. The main idea of this type of method is based on the intensity discontinuity between adjacent pixels in the image. Although the theory is simple, it usually faces problems such as discontinuous edges, false edges, noise sensitivity, and threshold selection. It is very difficult to operate in practice to achieve a compromise between noise reduction and boundary preservation, and careful parameter adjustment may be required to achieve the optimal result. (2) Threshold-based methods. This type of method aims to select the optimal segmentation threshold of methods such as image histograms and constant false alarm rates. The threshold method is suitable for rapid detection of large-scale coverage images, but inevitably faces problems such as threshold sensitivity and manual intervention. (3) Region merging-based methods. The region merging mechanism mainly includes seed point selection and merging criteria, and its segmentation accuracy depends on the quality of the seed points and merging criteria, and often fails in scenarios with low contrast between land and sea. (4) Partial differential equation-based methods. Partial differential equation methods usually include active contour models and level set algorithms. This type of method can obtain continuous detection results and easily integrate statistical information into the speckle noise model. However, it still faces problems such as complex calculations, slow evolution, and sensitivity to processing on large-scale coverage images. (5) Machine learning-based methods. Machine learning methods include classification methods such as support vector machines and fuzzy C-means. First, these methods were used, and deep learning methods such as pulse-coupled neural networks, multi-scale deep models, and deep convolutional neural networks have also been used in the coastline detection of SAR images. However, there are still major problems in terms of a large number of sample labels, complicated training processes, and algorithm transferability. (6) Markov random field-based methods. This type of method can effectively capture and use the spatial information of the image, but it is not robust to low-contrast regions between land and sea and requires setting relevant parameters. Compared with the above methods, the Random Walk (RW) method converts the energy optimization problem into solving a combinatorial Dirichlet problem, can achieve fast calculation without iteration, and has been proven to have advantages in terms of noise robustness and weak edge response.

[0004] Currently, coastline detection for single-polarization SAR images still faces the following problems: (1) The inherent speckle noise in SAR images makes the intensity values show considerable variability even in adjacent regions of uniform areas; (2) The distorted echo signals caused by variable sea conditions and confusing ground objects in the intertidal zone result in a lack of contrast between land and sea. Especially for aquaculture shorelines, the backscattering coefficients of aquaculture ponds are often low and close to the sea surface, resulting in extremely low distinguishability between land and sea. Most extraction methods are not robust at aquaculture coasts; (3) Semi-supervised learning methods represented by the RW method are sensitive to the initialization seed points given manually, and the original additive noise model is not applicable to SAR images with speckle noise. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides a method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images, which can effectively detect shorelines in single-polarization SAR images and also has strong robustness for aquaculture shorelines that are difficult to detect.

[0006] The technical means adopted by the present invention are as follows:

[0007] A method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images, the method comprising:

[0008] Obtain a single-polarization SAR image;

[0009] Perform median filtering on the single-polarization SAR image; perform median filtering on the single-polarization SAR image; respectively perform superpixel segmentation processing and edge detection processing on the median-filtered image to obtain a superpixel image and an edge image;

[0010] Driven jointly by the edge image and the superpixel image, generate a seed point discrimination criterion, and automatically obtain sea and land seed points based on the seed point discrimination criterion;

[0011] Construct a double-layer graph based on the single-polarization SAR image and the superpixel image; respectively construct a single-polarization SAR image, a superpixel image, and an edge weight function between the two images on the double-layer graph based on a Gaussian kernel function with the minimum mean ratio, and then construct a weight similarity metric based on the edge weight function;

[0012] Construct a random walk model with a double-layer graph structure, and by solving the random walk model with the double-layer graph structure, obtain a binary sea-land segmentation, and finally obtain the shoreline detection result.

[0013] Further, driven jointly by the edge image and the superpixel image, generating a seed point discrimination criterion includes:

[0014] Process the filtered image using the Ostu threshold method to obtain an adaptive rough sea-land segmentation threshold;

[0015] Automatically obtain initial seed points for the sea and land according to the number of edge points and the superpixel intensity values inside the current superpixel and its adjacent superpixels:

[0016] If there are no edges inside the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is lower than the rough sea-land segmentation threshold, it is determined as a sea seed point,

[0017] If there are edges inside the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is higher than the rough sea-land segmentation threshold, it is determined as a land seed point,

[0018] Other cases are determined as land-sea mixed areas.

[0019] Further, based on the Gaussian kernel function of the minimum mean ratio, a single-polarization SAR image, a superpixel image, and a weight similarity metric between the two images are constructed respectively, including:

[0020] Obtain the edge weight function between two pixels on the single-polarization SAR image of the bilayer graph as:

[0021]

[0022]

[0023] where γ is the first free parameter, g i represents the scattering intensity value at pixel i, and g j represents the scattering intensity value at pixel j;

[0024] Obtain the edge weight function between two superpixels on the superpixel image of the bilayer graph as:

[0025]

[0026] where β is the second free parameter, c i represents the intensity value of superpixel i, and c j represents the intensity value of superpixel j;

[0027] Obtain the edge weight function between a pixel on the single-polarization SAR image of the bilayer graph and a superpixel on the superpixel image as:

[0028]

[0029] where g i represents the scattering intensity value at pixel i, and c j represents the intensity value of superpixel j.

[0030] Further, based on the Gaussian kernel function of the minimum mean ratio, a single-polarization SAR image, a superpixel image, and a weight similarity metric between the two images are constructed respectively, and it also includes:

[0031] According to the edge weight function between two pixels on the single-polarization SAR image, the edge weight function between two superpixels on the superpixel image, and the edge weight function between a pixel on the single-polarization SAR image and a superpixel on the superpixel image, construct the following weight similarity metric:

[0032]

[0033] where V represents the single-polarization SAR image on the bilayer graph, and S represents the superpixel image on the bilayer graph.

[0034] Furthermore, a random walk model with a double-layer graph structure is constructed. By solving the random walk model of the double-layer graph structure, a binary sea-land segmentation is obtained, including:

[0035] According to the weight similarity measure, a weight matrix of the double-layer graph is constructed, and the weight matrix is rearranged in the order of marked points and unmarked points to obtain a new weight matrix;

[0036] Based on the sum of the weights of the edges associated with the image nodes or superpixel nodes, the degree matrix of the double-layer graph is obtained;

[0037] Based on the weight matrix and the degree matrix, the Laplacian matrix of the double-layer graph is obtained;

[0038] The Laplacian matrix is substituted into the objective function of the sea-land segmentation, and the objective function of the sea-land segmentation is solved to obtain the probabilities of the unmarked nodes reaching the sea and land seed points for the first time.

[0039] Furthermore, obtaining the binary sea-land segmentation includes:

[0040] Define a binary image with the same size as the original single-polarization SAR image, and the pixel values of the binary image are obtained according to the following method:

[0041]

[0042] where Q is the binary image, v i is the pixel of the binary image pixel point, and f i land is the probability that the pixel point reaches the land seed point for the first time, and f i sea is the probability that the pixel point reaches the sea seed point for the first time.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] 1. The present invention can effectively detect the aquaculture shoreline of the single-polarization SAR image and adapt to complex coastal types. The double-layer random walk method with the minimum mean ratio proposed by the present invention can effectively overcome the influence of speckle noise, so as to realize the accurate extraction of the aquaculture shoreline.

[0045] 2. The present invention takes the lead in introducing the random walk method into the shoreline detection of SAR images and improves the way of selecting seed points. Automatically selecting sea and land seed points can overcome the sensitivity problem of seed point selection. At the same time, the constructed double-layer random walk model can make full use of spatial information and has good robustness. Brief Description of the Drawings

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0047] Figure 1 This is the flowchart of the coastline detection method of the present invention.

[0048] Figures 2a - 2i This is a schematic diagram of the experimental results of coastline detection for single-polarization Sentinel-1 SAR images, where:

[0049] Figure 2a This is the original SAR image of experimental area 1 in the embodiment;

[0050] Figure 2b This is the land-sea binary image of experimental area 1 in the embodiment;

[0051] Figure 2c This is the coastline detection result (white line) of experimental area 1 in the embodiment;

[0052] Figure 2d This is the original SAR image of experimental area 2 in the embodiment;

[0053] Figure 2e This is the land-sea binary image of experimental area 2 in the embodiment;

[0054] Figure 2f This is the coastline detection result (white line) of experimental area 2 in the embodiment;

[0055] Figure 2g This is the original SAR image of experimental area 3 in the embodiment;

[0056] Figure 2h This is the land-sea binary image of experimental area 3 in the embodiment;

[0057] Figure 2i This is the coastline detection result (white line) of experimental area 3 in the embodiment. Detailed implementation manners

[0058] To enable those skilled in the art to better understand the solution of the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0059] As Figure 1 shown, the present invention provides a method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images, mainly including: automatic seed point initialization driven by the combination of superpixels and edges, mainly including median filtering, SLIC superpixel segmentation, Canny edge detection, and seed discrimination criteria; construction of similarity metrics with the minimum mean ratio, mainly including similarity metrics for the original SAR image, similarity metrics for the superpixel image, and similarity metrics between two images; solving random walks with a double-layer graph structure, including double-layer graph construction, double-layer graph model solving, and image binarization.

[0060] 1. Automatic seed point initialization driven by the combination of superpixels and edges

[0061] As a semi-supervised learning method, the seed point initialization of random walks is set manually, providing guiding prior knowledge for subsequent random walks. However, this method usually has a large degree of subjectivity. To solve the above problems, the present invention adopts a seed point initialization strategy driven by the combination of superpixels and edges.

[0062] Without loss of generality, we choose the well-known median filter to filter the image, choose the simple linear iterative clustering (SLIC) method to generate superpixels, choose the Canny edge detection method to obtain edges, and choose the Otsu threshold method to obtain an adaptive rough land-sea segmentation threshold.

[0063] Assume that S = [sp i m×1 represents the m superpixels obtained by SLIC, where sp i = [c i , e i T is a vector, c i represents the intensity value of the center point of sp i , e i represents the number of edge points obtained by Canny in sp i , and T sea_land represents the threshold obtained by the Otsu method. According to the number of edge points and the superpixel intensity values inside the current superpixel and its adjacent superpixels, the initial seed points of the ocean and land can be automatically obtained:

[0064]

[0065] where i~j indicates that the superpixels sp i and sp j are adjacent superpixels. There are no edges inside the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is lower than the threshold T sea_land ​​, it is determined as a marine seed point; there are edges inside both the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is higher than the threshold T sea_land , it is determined as a land seed point; in other cases, it is determined as a land-sea mixed area.

[0066] 2. Construction of Similarity Measure of Minimum Mean Ratio

[0067] First, some important symbolic notations are given and briefly described. Given a two-layer graph G = (P, E, W) = {G X , G Y}, where an image can be constructed into a weighted undirected graph G X = (V, E X , W X ), whose nodes v ∈ V, edges For each node v i represents each image pixel point x i . Each edge connects v i and v j two nodes. The weight of each edge is used to measure the possibility of a random walker passing through this edge, that is, the similarity measure, which directly determines the action direction of the random walker. Usually, a Gaussian kernel function is selected to define:

[0068] w ij = exp(-β(g i - g j ) 2 )

[0069] where g i and g j represent the scattering intensity values at pixels i and j, and β is a free parameter used to adjust the size of the weight. This way of weight function has a good effect on optical images, however, it is not applicable to SAR images with multiplicative speckle noise. To solve this problem, an edge weight function suitable for SAR images is proposed. Here, the logarithmic ratio distance of pixels is used to replace the difference distance between pixels. This method has been proven to be highly robust to speckle noise in SAR imaging. Therefore, the logarithmic ratio distance of pixels can be defined as:

[0070] dist ij = |min(log(g i ) - log(g j ), log(g j ) - log(g i ))|

[0071] The above formula can be simplified to:​

[0072]

[0073] Next, the edge weight function between two pixels on the image can be defined as:

[0074]

[0075] where γ is a preset first free parameter for adjusting the weight size, and g i represents the scattering intensity value at pixel i, and g j represents the scattering intensity value at pixel j.

[0076] Similarly, a superpixel image can be constructed into a weighted undirected graph G Y =(S, E Y , W Y ), where its nodes s ∈ S, and the edge Each node s i represents each image pixel point y i . Each edge connects s i and s j of the two nodes. The weight of each edge is used to measure the possibility of a random walker passing through this edge. As a consistent unit of the region, superpixels can effectively suppress the influence of speckle noise. The edge weight function between two superpixels on the superpixel image can be defined as:

[0077]

[0078] where β is a preset second free parameter, and c i represents the intensity value of superpixel i, and c j represents the intensity value of superpixel j.

[0079] In addition, the edge is used to connect the nodes between graph G X and graph G Y . Each edge connects v i and s j of the two nodes. The weight of each edge is used to measure the possibility of a random walker passing through this edge, and its edge weight function can be defined as:

[0080]

[0081] By summarizing the weight similarities in the above three cases, the following weight similarity metric can be obtained:

[0082]

[0083] 3. Random Walk Solution for Double-Layer Graph Structure

[0084] The transition probability of random walk and the combined Dirichlet problem have been proven to have the same solution, and the result of random walk image segmentation can be obtained by solving the combined Dirichlet problem. The land-sea label seed nodes obtained through Step 1 divide the nodes in the double-layer graph G into two categories: S L represents the labeled points (seed points), and S U and V together form the unlabeled points (non-seed points), where S L ∪S U =S. Assume that the arrival probability of the unlabeled point v i to a certain labeled point s L is f i . The defined combined Dirichlet formula, that is, the objective function for land-sea segmentation of the image is:

[0085]

[0086] where f is a matrix of (m + n)×2, and L is the Laplacian matrix of the graph, which can be calculated from the weight matrix and the degree matrix. Based on the similarity measure obtained in Step 2, the weight matrix of the double-layer graph is constructed and can be expressed as:

[0087]

[0088] Rearranging W in the order of labeled points and unlabeled points, with the sorting rule that labeled points come first and unlabeled points come second, the weight matrix can be expressed as:

[0089]

[0090] Then, the degree matrix D of the nodes on the graph is defined. The degree of the image node v i or the superpixel node s i represents the sum of the weights of all edges associated with the node v i or s i and can be defined as:

[0091]

[0092] where and represent the degrees of the image node and the superpixel node respectively.

[0093] With the weight matrix and the degree matrix, the Laplacian matrix L of the graph can be defined as:

[0094]

[0095] Substitute L into the objective function, and at the same time arrange f in the order of labeled points and unlabeled points, then the objective function can be decomposed as:

[0096]

[0097] where f L and f U correspond to the probabilities of the labeled points and unlabeled points for the two categories of sea and land respectively. Taking the differential of O(f U ) with respect to f U , we can get:

[0098]

[0099] Take the last n rows of f U to get the final result f V = [f land , f sea n×2 , let f i sea and f i land represent the probabilities of the unlabeled nodes reaching the sea and land seed points for the first time respectively. Define a binary image Q with the same size as the original SAR image, then Q can be obtained by the following formula:

[0100]

[0101] Finally, through the binary image, the complete SAR image coastline extraction result can be obtained.

[0102] Next, through specific application examples, the solutions and effects of the present invention will be further described.

[0103] This embodiment targets the Sentinel-1 SAR image of the Dalian area. Figure 2 is a schematic diagram of the experimental results of the coastline detection of the single-polarization Sentinel-1 SAR image. Among them, Figure 2a is the original SAR image of experimental area 1 in the embodiment, Figure 2b is the sea-land binary map of experimental area 1 in the embodiment, Figure 2c is the coastline detection result (white line) of experimental area 1 in the embodiment; Figure 2d is the original SAR image of experimental area 2 in the embodiment, Figure 2e is the sea-land binary map of experimental area 2 in the embodiment, Figure 2f is the coastline detection result (white line) of experimental area 2 in the embodiment; Figure 2g is the original SAR image of experimental area 3 in the embodiment, Figure 2h ​It is the land-sea binary map of experimental area 3 in the embodiment. Figure 2i It is the coastline detection result (white line) of experimental area 3 in the embodiment. It can be seen from the experimental results that the method of the present invention can effectively extract the aquaculture shoreline in the single-polarization SAR image. Through the automatic minimum mean ratio double-layer random walk detection method, the problem of difficult detection caused by speckle noise and weak land-sea contrast of aquaculture coast in the SAR image is effectively solved.

[0104] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images, characterized in that, the method includes: Obtain single-polarization SAR images; Perform median filtering on the single-polarization SAR images; perform superpixel segmentation and edge detection on the median-filtered images respectively to obtain a superpixel image and an edge image; Driven jointly by the edge image and the superpixel image, generate a seed point discrimination criterion, including: perform Ostu threshold method on the filtered image to obtain an adaptive rough sea-land segmentation threshold, and automatically obtain the initial seed points of the sea and land according to the number of edge points and the superpixel intensity values inside the current superpixel and its adjacent superpixels. If there are no edges inside the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is lower than the rough sea-land segmentation threshold, it is determined as a sea seed point; if there are edges inside the current superpixel and its adjacent superpixels, and the scattering intensity of the current superpixel is higher than the rough sea-land segmentation threshold, it is determined as a land seed point; otherwise, it is determined as a sea-land mixed area. Automatically obtain sea and land seed points based on the seed point discrimination criterion; Construct a bilayer graph based on the single-polarization SAR image and the superpixel image; construct the single-polarization SAR image, the superpixel image and the edge weight function between the two images on the bilayer graph respectively based on the Gaussian kernel function with the minimum mean ratio, and then construct a weight similarity metric based on the edge weight function. The formula for the weight similarity metric is: where V represents the single-polarization SAR image on the bilayer graph, and S represents the superpixel image on the bilayer graph; Construct the weight matrix of the bilayer graph according to the weight similarity metric, rearrange the weight matrix in the order of marked points and unmarked points to obtain a new weight matrix, obtain the degree matrix of the bilayer graph based on the sum of the weights of the edges associated with the image nodes or superpixel nodes, obtain the Laplacian matrix of the bilayer graph based on the weight matrix and the degree matrix, substitute the Laplacian matrix into the objective function of sea-land segmentation, solve the objective function of sea-land segmentation, so as to obtain the probabilities of the unmarked nodes reaching the sea and land seed points for the first time, and obtain the shoreline detection result based on the probabilities.

2. The method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images according to claim 1, characterized in that, Constructing the weight similarity metrics of the single-polarization SAR image, the superpixel image and between the two images respectively based on the Gaussian kernel function with the minimum mean ratio includes: Obtain the edge weight function between two pixels on the single-polarization SAR image on the bilayer graph as: where γ is the first free parameter, and g i represents the scattering intensity value at pixel i, and g j represents the scattering intensity value at pixel j; Obtain the edge weight function between two superpixels on the superpixel image on the bilayer graph as: where β is the second free parameter, c i represents the intensity value of superpixel i, and c j represents the intensity value of superpixel j; Obtain the edge weight function between a pixel on the single-polarization SAR image on the bilayer graph and a superpixel on the superpixel image as: where, g i represents the scattering intensity value at pixel i, and c j represents the intensity value of superpixel j.

3. The method for detecting aquaculture shorelines based on single-polarization SAR Sentinel-1 images according to claim 1, characterized in that, Obtaining binary sea-land segmentation includes: Define a binary image with the same size as the original single-polarization SAR image, and the pixel values of the binarization are obtained according to the following method: Among them, Q is a binary image, and v i is the pixel of a binary image pixel point, and f i land is the probability that the pixel point first reaches the land seed point, and f i sea is the probability that the pixel point first reaches the ocean seed point.

Images