A Coastline Detection Method Based on Markov Field Segmentation of Embedded Volume Scattering Components
By decomposing the coherence matrix of polarimetric SAR images into volume scattering and non-volume scattering components, establishing a joint distribution model and embedding a Markov field segmentation method, the accuracy problem of coastline detection in complex scenes is solved, and accurate coastline extraction is achieved.
Patent Information
- Application Number
- CN202410047308.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-12
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-01-12
AI Technical Summary
In complex scenarios, existing coastline detection methods based on regional statistical characteristics or edge energy image segmentation struggle to accurately segment sea and land areas, especially under conditions of strong speckle noise, sea conditions, and strong scattering interference, which can easily lead to incorrect segmentation.
A Markov field segmentation method based on embedded volume scattering components is adopted. By decomposing the coherence matrix of the polarimetric SAR image into volume scattering and non-volume scattering components, a joint distribution model is established and embedded into the Markov field segmentation model. Combined with an iterative algorithm, the image is segmented into three regions and the coastline is extracted.
It enables accurate detection of coastlines in complex coastal environments, reduces erroneous segmentation, and improves the accuracy and robustness of detection.
Smart Images

Figure CN118091660B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of coastline detection technology in remote sensing images, and in particular to a method for coastline detection in polarimetric synthetic aperture radar images based on Markov field segmentation of embedded scattering components. Background Technology
[0002] The coastline is the boundary between the sea and land, typically spanning large-scale coastal remote sensing images. Therefore, coastline detection is a fundamental step in interpreting coastal features in remote sensing images. In SAR (synthetic aperture radar) images, the sea surface appears as a uniform region with low scattering intensity, while coastal features, primarily vegetation including some beaches and man-made structures, have significantly higher overall scattering intensity than the sea surface. The sea and land exhibit significant differences in scattering intensity and polarization characteristics; therefore, coastline detection in SAR images is usually achieved by segmenting the sea and land regions. In simple scenarios, classic image segmentation methods such as thresholding, region merging, Markov random field (MRF) segmentation, and level set segmentation can achieve accurate sea-land segmentation. However, in complex scenarios, due to strong speckle noise in SAR images, coastal sea conditions, intertidal zones, and strong scattering secondary sidelobe interference, accurate and robust coastline detection remains extremely challenging.
[0003] To improve the accuracy and applicability of coastline detection in complex scenarios, existing methods employ a two-step approach combining coarse detection and fine-tuning. First, a coarse coastline extraction is achieved through region segmentation of the overall area, followed by fine-tuning of the coastline in local areas. The applicant has also conducted extensive prior research, such as the paper "Liu C, Yin JJ, Yang J. Coastline detection in polarimetric SAR images based on mixed edge detection[J]. Systems Engineering and Electronics, 2016, 38(6): 1262-1267." which proposes a binary hypothesis testing model for coastline edges combining polarization characteristics. A mixed-body scattering component ratio edge detector is used for coarse coastline extraction, followed by fine-tuning of the detection results using an active contour model based on edge energy. The paper "Liu C., Yang J., Yin J., et al. Coastline detection in SAR images using a hierarchicallevel set segmentation[J]. IEEE J.Sel.Topics Appl.Earth Observ.Remote" is also relevant. The paper "Sens., 2016, 9(11): 4908-4920" proposes a coastline detection method based on multi-level level set segmentation. It constructs a scale space of polarimetric SAR images with different smooth windows and combines smooth boundary offset quantization analysis to perform two-region level set segmentation on images at different scales to achieve coastline detection from coarse to fine. The paper "Liu C., Xiao Y., Yang J..Acoastline detection method in polarimetric SAR images mixing the region-based and edge-based active contour models[J].IEEE Trans.Geosci.Remote Sens., 2017, 55(7): 3735-3747" proposes a dual-scale active contour model for coastline detection by utilizing the overall and local boundary characteristics of the coastline. In the overall image region, a level set segmentation method based on regional statistical characteristics is used for coarse detection of the coastline. Then, by utilizing the monotonicity of likelihood ratio edge detection, an active contour model based on edge energy is used in the local region to achieve fine detection of the detection results.Furthermore, in the literature "Modava M., Akbarizadeh G., Soroosh M.. Integration of spectral histogram and level set for coastline detection in SAR images[J].IEEE Trans.Aerosp.Electron.Syst.,2019,55(2):810-819.", coarse coastline detection is performed by determining the spectral histogram of the local region, and then refined using the level set segmentation method.
[0004] However, existing coastline detection methods based on regional statistical characteristics or edge energy image segmentation have limitations in situations with complex sea conditions and coastal features. In complex coastal environments, the sea surface and land areas are not homogeneous; the sea surface may contain areas with different sea conditions, and the land may contain areas with different vegetation, urban buildings, and bare land. When the area of a strong scattering body on land is too large, it may misclassify features into strong scattering body areas and other non-strong scattering areas. Due to imaging range ambiguity, large areas of strong scattering along the coast can create secondary sidelobe interference on the sea surface, causing the scattering intensity near the shore to be much higher than that of calm sea surfaces, thus leading to incorrect segmentation as land. Summary of the Invention
[0005] To address the need for coastline detection in complex coastal environments and under strong noise interference, especially when coastal areas include sea surfaces affected by strong scattering sidelobes and intertidal zones with varying water content, SAR image coastline detection models based solely on coherence matrices or single-channel scattering intensity statistical distributions often suffer from segmentation errors. This invention proposes a coastline detection method based on Markov field segmentation with embedded volume scattering components. Utilizing the low scattering power of strong scattering interference bodies and the symmetry of the scattering coherence matrix of natural features, the scattering coherence matrix is divided into volume scattering and non-volume scattering components. A joint probability distribution model of the volume and non-volume scattering component scattering matrices is established and embedded into the Markov field segmentation method, achieving image segmentation into strong, medium, and low scattering regions. Coastline detection is then extracted from the low-scattering regions.
[0006] The technical solution of this invention is as follows:
[0007] A coastline detection method based on Markov field segmentation of embedded scattering components includes the following steps:
[0008] Step 1: Perform initial segmentation on the polarimetric SAR image containing the coastal area:
[0009] The coherence matrix of the polarimetric SAR image is decomposed to obtain the secondary scattering and volume scattering components of the ground objects. The volume scattering power is thresholded to extract low scattering regions, and the secondary scattering power is thresholded to extract strong scattering regions, thus obtaining the initial segmentation results.
[0010] Step 2: Perform joint distribution embedding Markov random field segmentation:
[0011] Step 2.1: Based on the reflection symmetry of natural features, the total polarization coherence matrix of the features is divided into volume scattering and non-volume scattering coherence matrix parts;
[0012] Step 2.2: Establish a joint distribution model for the coherence matrix parts of volume scattering and non-volume scattering, and embed the joint distribution into the objective function of the Markov random field partitioning model;
[0013] Step 2.3: Use an iterative algorithm to implement Markov random field segmentation, dividing the ground features into low-scattering, strong-scattering, and other regions;
[0014] Step 3: Extract the coastline:
[0015] Step 3.1: Extract the low scattering region from the Markov random field segmentation results obtained in Step 2;
[0016] Step 3.2: Using the largest connected domain as the main water area, merge regions according to their outline distances at a set distance to determine the land-sea separation result, and define the boundary of the binary map as the coastline.
[0017] Furthermore, in step 1, the Freeman ternary decomposition method is used to decompose the coherence matrix T of the polarimetric SAR image into the sum of three scattering components: surface scattering, secondary scattering, and volume scattering.
[0018] T = P s T s +P d T d +P v T v
[0019] Where T s T d T v These are the surface scattering, secondary scattering, and volume scattering models of the target, respectively, P s P d P v These represent the power levels of the corresponding components;
[0020] For P respectively d P v Thresholding segmentation is used to determine strong scattering regions and low scattering regions:
[0021] For each pixel in the polarimetric SAR image, the following operation is performed: a w×w sliding window region is determined with one pixel as the center;
[0022] Use P v When dividing the sea surface, select P from the sliding window area. v The region with the lowest average power is the sea surface sample area W. v Use P d When segmenting the strong scattering region, select P from the sliding window region. d The region with the highest average power is the strong scattering sample region W. d ;
[0023] If region W v Average P v Power is μ v dB, W d Average P d Power is μ d dB, based on sea level P v and strong scattering P d The fluctuation range is determined by a constant A, and the low scattering region segmentation threshold t1 is obtained as μ. v +A, the strong scattering region segmentation threshold t2 is μ d -A, thus obtaining the initial segmentation result.
[0024] Furthermore, in step 1, the range of w is selected as 10 to 30.
[0025] Furthermore, in step 2.1, the coherence matrix T is decomposed into... and P v Two parts.
[0026] Furthermore, in step 2.2, the image plane is represented as Ω, and the data set is {T}. x |x∈Ω}, the ideal segmentation result is represented as {ω} x |x∈Ω}, where ω x If the set of value categories is {1,2,…,K}, then the segmentation model is modeled as a Markov field model under known observation data T, where the latent variable ω follows a Markov field. The optimal segmentation model is expressed as:
[0027]
[0028] Where U1 is the likelihood energy and U2 is the prior energy;
[0029]
[0030]
[0031] L is the number of views, and λ is the weighting coefficient. For category ω x The mean of matrix T′ in the region, where tr(.) is the trace of the matrix. For category ω x Average power in the region; δ(·,·) is the impulse function, d∈N x N x It represents the 8-neighborhood of pixel x.
[0032] Furthermore, in step 2.3, based on the initial segmentation results and setting the λ value, the optimal segmentation model established in step 2.2 is iteratively solved to obtain the segmentation results of the three-region Markov field segmentation model.
[0033] Furthermore, in step 3, the region with the minimum average volume scattering power in the Markov random field segmentation results is determined as the water area M, and the maximum width of the bridge is set as W. The region with the largest area in M is taken as the main sea surface region, and the region of M is merged under the distance threshold W to determine the final sea surface region R3; the boundary line of the sea surface region is determined as the final coastline detection result.
[0034] Beneficial effects
[0035] This invention proposes a coastline detection method based on Markov field segmentation with embedded volume scattering components. Utilizing the low scattering power of strong scattering interference and the symmetry of the scattering coherence matrix of natural features, the method divides the scattering coherence matrix into two parts: volume scattering components and non-volume scattering components. A joint probability distribution model of the volume scattering and non-volume scattering component scattering matrices is established, and this new distribution model is embedded into the Markov field segmentation method to achieve image segmentation into strong, medium, and low scattering regions. Coastline detection is then performed using the low-scattering region. Specifically, it has the following characteristics:
[0036] 1. The polarization 3*3 coherence matrix is decomposed into a volume scattering component and a non-volume scattering 2*2 coherence matrix. The joint distribution of the two is established and embedded into the MRF segmentation model, thus realizing the effective fusion of polarization scattering information.
[0037] 2. By introducing the weighting coefficients of the volume scattering component and the non-volume coherence matrix, a balance between the two is achieved by adjusting the weights, thus making it suitable for different coastal scenarios and interference levels;
[0038] 3. Threshold segmentation was achieved using the secondary scattering and volume scattering components from the three-component decomposition, resulting in accurate initial segmentation.
[0039] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0040] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:
[0041] Figure 1 : Flowchart of the method of the present invention;
[0042] Figure 2 The results of RADARSAT-2 detection of the coastline in Singapore include: (a) Pauli pseudo-color map; (b) true distribution of the coastline; (c) sample area selection; (d) segmentation results based on volume scattering and secondary scattering power thresholds; (e) segmentation results when the proposed method λ is set to 0.1; (f) segmentation of water areas; (g) land-sea segmentation results; (h) coastline extraction results, where the red line represents the detected coastline; and (i) comparison between the detection results and the true distribution results.
[0043] Figure 3 Comparison of RADARSAT-2 coastline detection results in Singapore region. (a) Coherence matrix MRF segmentation results of two regions. (b) Level set segmentation results. (c) Results of the proposed method. Detailed Implementation
[0044] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0045] like Figure 1 As shown, the coastline detection method based on Markov field segmentation with embedded volume scattering components proposed in this invention mainly consists of three stages: polarization decomposition initial segmentation, joint distribution embedded Markov random field (MRF) segmentation, and coastline extraction. In the polarization decomposition and initial segmentation stage, the Freeman ternary component decomposition method is first used to decompose the coherence matrix of the polarimetric SAR image to obtain the secondary scattering and volume scattering components of the ground objects. Then, threshold segmentation is performed on the volume scattering power to extract low-scattering regions, and threshold segmentation is performed on the secondary scattering power to extract strong-scattering regions, obtaining the initial segmentation result. In the joint distribution embedded MRF segmentation stage, based on the reflection symmetry of natural ground objects, the fully polarimetric coherence matrix of the ground objects is first divided into volume scattering and non-volume scattering coherence matrix parts. Then, a joint distribution model of the volume scattering and non-volume scattering coherence matrix parts is established, and the joint distribution is embedded into the objective function of the MRF segmentation model. Finally, an iterative algorithm is used to implement MRF segmentation to divide the ground objects into low-scattering, strong-scattering, and other regions. In the coastline extraction stage, low-scattering regions are first extracted from the MRF segmentation results. Then, the largest connected region is used as the main water area, and the land-sea segmentation results are determined by merging regions according to their outline distances at a set distance. The boundary of the binary map is then determined as the coastline.
[0046] The specific process is as follows:
[0047] 1. Initial Segmentation
[0048] Typically, the scattering intensity of the sea surface and land differs significantly, and the overall distribution of scattering intensity in coastal SAR images exhibits a bimodal distribution. However, in reality, if the area of the strong scattering region is too large or the area of the sea surface region is too small, the segmentation of the two regions may divide the image into a strong scattering region and a non-strong scattering region. Furthermore, the scattering intensity of water surface areas along the coast that are subject to strong scattering interference is higher than that of calm sea surfaces, and the segmentation of the two regions may also incorrectly classify the interfering water areas as land.
[0049] Considering that coastal landforms typically include: built-up areas with strong scattering, vegetated areas, and bare land and sea surface areas with low scattering, and that the secondary scattering power of built-up areas is much higher than that of non-strong scattering areas, and the volume scattering power of sea surface areas is much lower than that of non-sea surface areas, threshold segmentation can be performed on the secondary scattering power and volume scattering power calculated by the three-component decomposition, dividing the high secondary scattering power areas into strong scattering areas and the low volume scattering power areas into low scattering areas, thereby achieving the initial three-region segmentation.
[0050] Here, the Freeman ternary decomposition method is used to decompose the coherence matrix T of the polarimetric SAR image into the sum of three scattering components: surface scattering, secondary scattering, and volume scattering.
[0051] T = P s T s +P d T d +P v T v (1)
[0052] Where T s T d T v These are the surface scattering, secondary scattering, and volume scattering models of the target, respectively, P s P d P v These represent the power levels of the corresponding components.
[0053] For P respectively d P v Thresholding segmentation is used to determine strong scattering and low scattering regions. However, commonly used thresholding methods, such as the OTU method, can determine incorrect thresholds when the areas of strong or low scattering regions are too large or too small. Therefore, considering that the difference in scattering power between the sea surface area and the non-sea surface area, as well as between strong and weak scattering regions, falls within a defined range at a specific incident angle, the threshold can be determined by selecting a typical sample area within this defined range.
[0054] For each pixel in the polarimetric SAR image, the following operation is performed: a w×w sliding window region is determined with one pixel as the center. To ensure the effectiveness of the sample region, w should not be too small or too large, and the range is selected from 10 to 30.
[0055] Use P v When segmenting the sea surface, considering the low scattering intensity of the sea surface region, P is selected from the sliding window region. v The region with the lowest average power is the sea surface sample area W. v .
[0056] Use P d When segmenting strong scattering regions, considering that the secondary scattering power in strong scattering building areas is much higher than in other areas, P is selected from the sliding window region. d The region with the highest average power is the strong scattering sample region W. d .
[0057] If region W v Average P v Power is μ v dB, W d Average P d Power is μ d dB, based on sea level P v and strong scattering P d The fluctuation range is determined by a constant A, and the low scattering region segmentation threshold t1 is obtained as μ. v +A, the strong scattering region segmentation threshold t2 is μ d -A, from which the initial three-region segmentation result R1 can be obtained.
[0058] 2. Jointly Distributed Embedding MRF Segmentation
[0059] If the image plane is represented by Ω, and the data set is {T} x |x∈Ω}, the ideal segmentation result is represented as {ω} x |x∈Ω}, where ω x If the set of value categories is {1,2,…,K}, then the segmentation model can be modeled as a Markov field model under known observation data T, where the latent variable ω follows a Markov field. By the maximum a posteriori probability criterion, the optimal segmentation model is expressed as:
[0060]
[0061] According to the Hammersley-Clifford theorem, ω being a Markov field is equivalent to P(ω|T) being a Gibbs distribution, from which we can obtain
[0062]
[0063] Where Z is the normalization factor, and C is the clique formed by adjacent pixel categories. Energy term V c (ω) is composed of different groups formed between data and categories, as well as between adjacent categories.
[0064] Furthermore, given P(ω|T)∝P(T|ω)P(ω), we know that U(ω) consists of the likelihood energy U1 of the data under the known category label and the prior energy U2 under the adjacent categories. Therefore, the energy term V... c (ω) is composed of different clusters formed by data and categories, as well as adjacent categories. From (2) and (3), the optimal segmentation model is equivalent to...
[0065]
[0066] As mentioned earlier, directly segmenting the three regions based on the distribution of the coherence matrix has the problem of incorrectly segmenting the sea surface area with strong scattering interference along the coast. Therefore, it is necessary to extract polarization parameters that can effectively reduce the contrast between the interfering and non-interfering sea surfaces for segmentation. However, statistical results show that the secondary sidelobe interference caused by strong scattering structures is mainly composed of secondary scattering components. Extracting volume scattering components for segmentation can effectively suppress the interference components. If volume scattering components are extracted and a volume scattering component distribution is established for MRF segmentation, the problem is that there are low-scattering bare land areas in the coastal zone. The volume scattering power difference between these low-scattering areas and the sea surface area is small, making them easy to misclassify as sea surface areas. Therefore, it is necessary to find a balance between volume scattering components and non-volume scattering components to ensure that the contrast between the interfering and non-interfering water areas is small, while ensuring that the contrast between the coastal bare land areas and the sea surface is high.
[0067] Considering the volume scattering model T in Freeman decomposition v Given a diagonal matrix with a diagonal of (2,1,1), the volume scattering power P is 4T. 33 Matrix T is a 3x3 matrix. 33 It is the (3,3) element of matrix T. For natural features, the coherence matrix T satisfies the following form:
[0068]
[0069] The coastal areas are dominated by natural features, therefore T can be decomposed into... and P v Two parts, at this time f(T|ω)∝f(T′|ω)f(P) v |ω). For different coastal environments, to reduce erroneous segmentation, T′ and P v Since the weights of the components are different, the weight parameter λ is increased, as shown in equation (6).
[0070] f(T|ω)∝f(T′|ω) λ f(P v|ω) 1-λ (6)
[0071] By adjusting λ, P can be increased or decreased. v The proportions are as follows: when λ→1, f(T|ω) is mainly composed of T′; while when λ→0, f(T|ω) is mainly composed of P. v Main ingredients.
[0072] Coherence matrix T 33 Equal to <|S HV | 2 > / 2, where S HV The cross-polarization scattering coefficients follow a gamma distribution with an appearance number of L and a mean of u.
[0073]
[0074] Where "(.)" represents the Gamma function.
[0075] If the coherence matrix distribution follows a Wishart distribution, the average coherence matrix is ∑, and the number of views is L, then T ~ W(∑,L,p):
[0076]
[0077] Where p is the dimension of the Pauli data vector, tr(.) is the trace of the matrix, and K(L,p)=π p(p-1) / 2 "(L)..."(L-p+1).
[0078] At this point, the likelihood energy U1 of the data in U(ω) is
[0079]
[0080] in For category ω x Average power in the region For category ω x The average value of matrix T′ in the region, 0≤λ≤1.
[0081] The prior energy V2 at point x is defined as N x V represents the 8-neighborhood of pixel x. d (ω x )=β(1-δ(ω x ,ω d δ(·,·) is the impulse function, and β is a constant greater than 0, which adjusts the magnitude of the prior energy. The larger β is, the greater the energy when adjacent pixels belong to different categories. Here, it is set to 0.5. The prior energy U2 in U(ω) is...
[0082] U2=∑ x,d β(1-δ(ωx ,ω d (10)
[0083] For coastline detection, the value of λ depends on the area of strong scattering and the proportion of sea surface area. When the area of strong scattering is large, the amount of interfering water increases accordingly, and the value of λ approaches 0. When the sea surface area is large, the value of λ should be increased to effectively distinguish between bare coastal land and sea surface.
[0084] Given a determined λ value, with R1 as the initial segmentation, substitute equations (9) and (10) into equation (4), and use the conditional iteration method to solve for the category that satisfies the minimum U(ω), thus obtaining the segmentation result R2 of the three-region MRF segmentation model.
[0085] 3. Coastline extraction
[0086] The region with the lowest average volume scattering power in the R2 segmentation results is identified as the water area M. Since the sea surface region may be interrupted by a bridge, the maximum bridge width W is set. The region with the largest area in M is designated as the main sea surface region. Region merging is performed on M at a distance threshold W to determine the final sea surface region R3. The boundary line of the binary image R3 is then used as the final coastline detection result.
[0087] Next, experiments were conducted using RADARSAT-2 fully polarimetric SAR data of a portion of Singapore. The single-look data resolution was approximately 5m × 5m, and the data size was 1000 × 1000. A map of the actual coastline distribution was drawn using Google Earth, and the mean bias and standard deviation were used to evaluate the detection performance. Statistically, the threshold segmentation parameter C ranges from 7 to 20 dB; in this experiment, it was set to 13 dB. Based on experimental comparisons, detection performance was significantly better when λ was less than 0.3 than when it was greater than 0.3; therefore, λ was set to 0.1 in this experiment.
[0088] like Figure 2 (a) is the Pauli pseudocolor image of the data. Figure 2 (b) shows the actual distribution of coastlines, with the green line representing the drawn coastline. Observe. Figure 2 (a) It can be observed that the Singapore data is distributed over a large area of urban built-up areas, and the coastline includes both water areas subject to strong scattering interference and bare land areas with high water content, which makes it difficult to model the land-sea separation between the two regions. Figure 2 (c) shows the selected sea surface and strong scattering sample areas, with sea surface samples marked in red and strong scattering samples marked in green. Observation shows that the selection of sea surface and strong scattering sample areas based on regional average volume scattering and secondary scattering power is achieved relatively well. Figure 2(d) shows the initial three-region segmentation results obtained using the proposed threshold segmentation method. It can be observed that the threshold segmentation achieves effective separation of the sea surface and strong scattering, but the problem is that there are many isolated pixels. Figure 2 (e) shows the three-region segmentation results of the proposed method when λ is set to 0.1. It can be seen that the proposed method achieves good segmentation of the three regions and greatly reduces isolated segmented pixels. Figure 2 (f) is the land-water separation map obtained by extracting water areas based on the MRF three-region segmentation results. Figure 2 (g) To obtain the land-sea segmentation result by performing regional merging processing on the segmentation result. Figure 2 (h) represents the final coastline detection result, where the red line represents the detection result. Figure 2 (i) A comparison chart of the detection results and the true distribution results, where the green line represents the true distribution results. Observation Figure 2 (i) It can be seen that the proposed method achieves accurate coastline detection. Calculation of the deviation between the detected coastline and the actual distributed coastline shows that the average deviation is 1.45 pixels and the standard deviation is 3.85 pixels.
[0089] Figure 3 The results show a performance comparison between the proposed method and different coastline detection methods. Figure 3 (a) shows the MRF segmentation results of the two regions based on the 3*3 coherence matrix. Figure 3 (b) shows the result of level set partitioning. Figure 3 (c) shows the results when λ is set to 0.1 in the proposed method. It can be observed that, due to the large area of the strongly scattering building region, both comparison methods, except for the proposed method, segment the image into strongly scattering and non-strongly scattering regions.
[0090] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention without departing from the principles and spirit of the present invention.
Claims
1. A coastline detection method based on Markov field segmentation of embedded scattering components, characterized in that: Includes the following steps: Step 1: Perform initial segmentation on the polarimetric SAR image containing the coastal area: The coherence matrix of the polarimetric SAR image is decomposed to obtain the secondary scattering and volume scattering components of the ground objects. The volume scattering power is thresholded to extract low scattering regions, and the secondary scattering power is thresholded to extract strong scattering regions, thus obtaining the initial segmentation results. Step 2: Perform joint distribution embedding Markov random field segmentation: Step 2.1: Based on the reflection symmetry of natural features, the total polarization coherence matrix of the features is divided into volume scattering and non-volume scattering coherence matrix parts; Step 2.2: Establish a joint distribution model for the coherence matrix parts of volume scattering and non-volume scattering, and embed the joint distribution into the objective function of the Markov random field partitioning model; Step 2.3: Use an iterative algorithm to implement Markov random field segmentation, dividing the ground features into low-scattering, strong-scattering, and other regions; Step 3: Extract the coastline: Step 3.1: Extract the low scattering region from the Markov random field segmentation results obtained in Step 2; Step 3.2: Using the largest connected domain as the main water area, merge regions according to their outline distances at a set distance to determine the land-sea separation result, and define the boundary of the binary map as the coastline.
2. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 1, characterized in that: In step 1, the Freeman ternary decomposition method is used to decompose the coherence matrix T of the polarimetric SAR image into the sum of three scattering components: surface scattering, secondary scattering, and volume scattering. T=P s T s +P d T d +P v T v Where T s T d T v These are the surface scattering, secondary scattering, and volume scattering models of the target, respectively, P s P d P v These represent the power levels of the corresponding components; For P respectively d P v Thresholding segmentation is used to determine strong scattering regions and low scattering regions: For each pixel in the polarimetric SAR image, the following operation is performed: a w×w sliding window region is determined with one pixel as the center; Use P v When dividing the sea surface, select P from the sliding window area. v The region with the lowest average power is the sea surface sample area W. v Use P d When segmenting the strong scattering region, select P from the sliding window region. d The region with the highest average power is the strong scattering sample region W. d ; If region W v Average P v Power is μ v dB, W d Average P d Power is μ d dB, based on sea level P v and strong scattering P d The fluctuation range is determined by a constant A, and the low scattering region segmentation threshold t1 is obtained as μ. v +A, the strong scattering region segmentation threshold t2 is μ d -A, thus obtaining the initial segmentation result.
3. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 2, characterized in that: In step 1, the range of w is 10 to 30.
4. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 2, characterized in that: In step 2.1, the coherence matrix T is decomposed into... and P v Two parts.
5. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 4, characterized in that: In step 2.2, the image plane is represented as Ω, and the data set is {T}. x |x∈Ω}, the ideal segmentation result is represented as {ω} x |x∈Ω}, where ω x If the set of value categories is {1, 2, ..., K}, then the segmentation model is modeled as a Markov field model under known observation data T, where the latent variable ω follows a Markov field. The optimal segmentation model is expressed as: Where U1 is the likelihood energy and U2 is the prior energy; in L is the number of views, and λ is the weighting coefficient. For category ω x The mean of matrix T′ in the region, where tr(.) is the trace of the matrix. For category ω x Average power in the region; δ(·,·) is the impulse function, d∈N x N x It represents the 8-neighborhood of pixel x.
6. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 5, characterized in that: In step 2.3, based on the initial segmentation results and setting the λ value, the optimal segmentation model established in step 2.2 is iteratively solved to obtain the segmentation results of the three-region Markov field segmentation model.
7. The coastline detection method based on Markov field segmentation of embedded scattering components according to claim 1, characterized in that: In step 3, the region with the minimum average volume scattering power in the Markov random field segmentation results is determined as the water area M, and the maximum width of the bridge is set as W. The region with the largest area in M is taken as the main sea surface region. Under the distance threshold W, the region of M is merged to determine the final sea surface region R3. The boundary line of the sea area is determined as the final coastline detection result.
Citation Information
Patent Citations
Target detection apparatus and target detection method
JP2010175383A
Joint Deep Learning for Land Cover and Land Use Classification
US20200065968A1
