A port detection method based on context features

Abstract

The application provides a port detection method based on context features, which is used for polarimetric SAR images, and uses the differences in polarimetric scattering characteristics of water areas and artificial buildings. The water areas and strong scattering built-up areas are extracted by segmenting different polarimetric parameters extracted by polarimetric decomposition. Then, the coastal narrowband area is extracted and the large-area strong scattering area in the band is identified to realize the extraction of the port region of interest by using the context features of the port water area and land. Finally, the port is identified according to the co-polarization and cross-polarization scattering correlation power distribution of the region of interest by using the asymmetric scattering characteristics of the port context artificial building. The implementation results show that the application realizes the port detection under the conditions of complex coastal environment, various port forms and serious noise interference, realizes the simultaneous detection of different sizes and types of ports, effectively reduces the missed detection and false alarm, and is far superior to the existing methods in the detection rate and false alarm rate.

Description

Technical Field

[0001] This invention relates to a method for detecting ports in remote sensing images, and more particularly to a method for detecting port targets in synthetic aperture radar images based on contextual features. Background Technology

[0002] Ports are crucial water and land transportation hubs located along sea and river coasts. Automatic port detection in SAR images is essential for port area monitoring, disaster prevention, and ship inspection. Due to variations in coastal topography and tides across different regions, ports can be classified as natural or artificial, as well as open, closed, and mixed ports. The significant differences in morphology, structure, and composition among different ports present challenges for port detection methods, whether traditional feature-based approaches or the widely used deep learning-based object detection algorithms.

[0003] Geometrically, port breakwaters and jetties often appear as narrow, elongated strips, straight lines, or arcs with low curvature, making their geometric features distinct relative to the natural coastal topography. Utilizing this characteristic, traditional feature-based port detection methods, based on the separation of land and sea, detect ports by extracting feature points, lines, and combinations of lines along the coastline.

[0004] As proposed in the applicant's previous research (Liu Chun, Yin Junjun, Yang Jian. Small Port Detection Based on Shoreline Feature Point Merging in Polarimetric SAR Images [J]. Journal of Tsinghua University (Natural Science Edition), 2015, 55(08): 849-853. DOI: 10.16511 / j.cnki.qhdxxb.2015.08.006.), a port detection method based on shoreline feature point extraction is proposed. Based on the accurate sea-land segmentation achieved by using the level set segmentation method based on regional statistical characteristics, the method utilizes the characteristic that the unevenness of the port outline forms multiple feature points to achieve port detection based on shoreline feature point extraction. In addition, the applicant also proposed a port detection method based on parallel curve extraction in the literature (C. Liu, Y. Xiao, J. Yang, and J. Yin, "Harbor Detection in Polarimetric SAR Images Based on the Characteristics of Parallel Curves," IEEE Geosci. RemoteSens. Lett., vol. 13, pp. 1400-1404, 2016.), utilizing the parallel outlines on both sides of the port breakwater and seawall; and in the literature (Liu Chun, Xie Chunhua, An Wentao, et al. Artificial Port Detection Based on Polarimetric SAR Images Using Multi-directional Breakwater Scanning [J]. Systems Engineering and Electronics Technology, 2017, 39(02): 291-297.), utilizing the characteristic that the port breakwater is surrounded by water, the applicant proposed a port detection method based on multi-directional breakwater scanning.

[0005] However, statistical analysis of actual ports reveals significant differences in their geometric shapes. Some ports lack breakwaters or their breakwaters are integrated with coastal buildings and docked ships, making it difficult to extract geometric structural features and lacking universality. Therefore, port detection methods based on geometric structural features are prone to missed detections and false alarms.

[0006] Deep learning-based port detection methods learn from a vast number of ports of different shapes, avoiding the need for manually designed feature extraction processes. For example, the paper (Q. Ye, H. Huo, T. Zhu and T. Fang, "Harbor Detection in Large-Scale Remote Sensing Images Using Both Deep-Learned and Topological Structure Features," 2017 10th International Symposium on Computational Intelligence and Design (ISCID), Hangzhou, China, 2017, pp. 218-222.) proposes a port detection method based on deep residual networks and single-step neural networks. First, ResNet is used to extract the coastal zone region, and then the SSD network is used to identify ports in the coastal zone region. The paper (R. Wang, F. Xu, Q. Zhang, J. Pei, Y. Huang, and J. Yang, "Harbor Detection in SAR Images Based on Multidirectional One-Dimensional Scanning," IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium, Waikoloa, HI, USA, 2020, pp. 1-4.) proposes a port detection method that combines geometric features and deep learning. First, multidirectional pier scanning is used to extract the region of interest in the port, and then a multi-layer neural network is constructed to classify the region of interest.

[0007] Although deep learning-based port detection methods do not rely on precise port contour feature extraction, the problem is that the diversity of port morphology requires a massive number of samples to train the neural network model, and there is a lack of reasonable interpretation and processing methods for the detected false alarms. Summary of the Invention

[0008] Ports are large-scale man-made structures located at the junction of land and sea. Due to differences in coastal topography and marine environment, port morphologies are highly complex. However, the contextual environment of ports shares commonalities, such as one side being surrounded by water and the other side being a land area containing numerous man-made structures. Both the water areas and man-made structures related to the port area's context exhibit strong scattering characteristics. Therefore, this invention addresses the problem that traditional methods based on spatial geometric features are insufficient for detecting ports in complex coastal environments due to their diverse morphologies, proposing a port detection method based on contextual features. This method targets polarimetric SAR images and utilizes the differences in polarimetric scattering characteristics between water areas and man-made structures. First, it segments the water areas and strongly scattering built-up areas by extracting different polarization parameters from polarimetric decomposition. Then, it uses the contextual features of the port's water and land areas to extract narrow coastal zones and identify large-area strongly scattering regions within these zones, thus extracting the port's region of interest (ROI). Finally, it utilizes the asymmetric scattering characteristics of the man-made structures within the port's context to identify the port based on the power distribution related to co-polarization and cross-polarization scattering within the ROI.

[0009] The technical solution of this invention is as follows:

[0010] A port detection method based on contextual features includes the following steps:

[0011] Step 1: Acquire the polarimetric SAR image of the target area; perform three-region segmentation on the polarimetric SAR image of the target area, dividing the polarimetric SAR image into a low-scattering region, a strong-scattering region, and other regions;

[0012] Step 2: Based on the three-region segmentation results from Step 1, extract the water-land segmentation map, determine the segmentation boundary coastline, and identify the narrow coastal zone along the coastline.

[0013] Based on the three-region segmentation results of step 1, a strong scattering region segmentation map is extracted to determine the distribution map of strong scattering regions in the coastal narrow zone; the strong scattering sub-regions in the narrow zone are merged according to the contour distance to obtain candidate sub-regions, the strong scattering area in each candidate sub-region is calculated, and the port region of interest is extracted based on the size of the strong scattering area.

[0014] Step 3: Calculate the co-polarization and cross-polarization scattering correlation power of several port regions of interest obtained in Step 2, and perform final identification based on the co-polarization and cross-polarization scattering correlation power distribution of the regions of interest.

[0015] Furthermore, step 1 includes the following steps:

[0016] Step 1.1: Perform a model-based three-component decomposition on the ground object scattering coherence matrix corresponding to the polarimetric SAR image of the target area, and extract the power of the surface scattering, secondary scattering and volume scattering components of the ground objects;

[0017] Step 1.2: Using the surface scattering, secondary scattering, and volume scattering component power of the ground objects obtained in Step 1.1 as segmentation data input, the polarimetric SAR image is segmented into low scattering regions, strong scattering regions, and other regions using the Markov field segmentation method based on multidimensional Gaussian distribution.

[0018] Furthermore, in step 2, based on the three-region segmentation results from step 1, the average power of the three segmented regions is calculated respectively, and the region with the minimum average power is determined to be the water area, thus obtaining the water area-land segmentation map.

[0019] Furthermore, in step 2, circles are drawn point by point along the coastline to determine the narrow coastal zone.

[0020] Furthermore, in step 3, for the port region of interest y l Determine the strongly scattering pixel e located in the strongly scattering region. l If E l If the proportion of pixels with a cross-polarization correlation power higher than Th is greater than ρ, then they are identified as ports, where Th is the set segmentation threshold and ρ is the set proportion threshold.

[0021] Furthermore, the segmentation threshold Th is determined according to the following process:

[0022] The global constant false alarm rate method is adopted. Under the set false alarm rate Pfa, the other regions in the three-region segmentation results of step 1 are used as the background. The histogram distribution of the co-polarization cross-polarization correlation power in the other regions is statistically analyzed. The threshold Th is determined based on the co-polarization cross-polarization correlation power distribution in the other regions.

[0023] Furthermore, the present invention also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement any of the above methods.

[0024] Furthermore, the present invention also proposes a computer system comprising: one or more processors, and the aforementioned computer-readable storage medium for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement any of the aforementioned methods.

[0025] Beneficial effects

[0026] This invention proposes a port detection method based on contextual features, aiming to address the challenge of detecting ports in complex coastal environments where diverse port morphologies make traditional methods based on spatial geometric features ineffective. This invention leverages the contextual features of ports surrounded by water and man-made structures on the land side creating strong scattering. By segmenting ground features, it simultaneously extracts the water area and the strongly scattering built-up area, enabling the extraction of the narrow coastal zone of interest and the region of interest (ROI) of the port. Furthermore, it utilizes the asymmetric scattering characteristics of the man-made structures within the port's context to identify the ROI, thus avoiding the problem of missed detections due to the inability to extract common geometric features of different port morphologies. Ultimately, this achieves high-precision, universal detection of ports of various morphologies in complex scenarios using polarimetric SAR images.

[0027] This invention utilizes the contextual features of ports for port detection. It achieves port detection by extracting port water areas, strong scattering regions, and artificial building areas. It does not rely on the extraction of port geometric features and can simultaneously extract ports of different shapes.

[0028] This invention fully utilizes the scattering characteristics of objects in the port area for port detection. It determines the port's coastal zone of interest by the water-land boundary line, determines the port's region of interest by the strong scattering region of the coastal zone, and identifies the port by the asymmetry of scattering from artificial structures in the port area.

[0029] This invention extracts three components of power polarization parameters for multi-region segmentation of ground features, effectively suppressing strong scattering echo interference in port areas, while simultaneously enabling the extraction of water areas and areas with strong scattering.

[0030] The results show that the present invention enables port detection in complex coastal environments, diverse port forms, and severe noise interference, and enables simultaneous detection of ports of different sizes and types. It effectively reduces missed detections and false alarms, and its detection rate and false alarm rate are far superior to existing methods.

[0031] 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

[0032] 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:

[0033] Figure 1 : Flowchart of a port detection method based on contextual features in polarimetric SAR images;

[0034] Figure 2 : Schematic diagram of narrow coastal zone extraction;

[0035] Figure 3 The following are the results of RADARSAT-2 port inspections in the Singapore area as illustrated in the example:

[0036] (a) Pauli pseudocolor map. (b) Three-region segmentation results. (c) Land-water segmentation map. (d) Narrow coastal zone. (e) Distribution map of regions of interest with strong scattering in the narrow zone. (f) Port region of interest, where the region of interest is marked with a rectangle. (h) Region corresponding to the port region of interest in the Pauli pseudocolor map, where the port region of interest is marked with a rectangle without a number, and the rectangle with a number is the detected false alarm region. (i) Horizontal co-polarization cross-polarization correlation power map. (j) Final port detection results. Detailed Implementation

[0037] 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.

[0038] like Figure 3 As shown in (a), a port is a large-scale man-made structure located at the junction of land and sea. Due to differences in coastal topography and marine environment, the morphology of ports is highly complex. However, the contextual environment of ports shares commonalities, such as one side being surrounded by water and the other side being a land area containing numerous man-made structures. Both the water area and man-made structures related to the port's context have strong scattering characteristics. This invention addresses the problem that traditional methods based on spatial geometric features are insufficient for detecting ports in complex coastal environments due to their diverse morphologies, and proposes a port detection method based on contextual features.

[0039] Figure 1 The basic method flow of this invention is presented, mainly consisting of three steps: three-region segmentation of land features, extraction of port regions of interest, and port identification. Details are as follows:

[0040] Step 1: Feature segmentation steps:

[0041] Acquire a polarimetric SAR image of the target region; perform three-region segmentation on the polarimetric SAR image of the target region, dividing the polarimetric SAR image into a low-scattering region, a strong-scattering region, and other regions;

[0042] The specific process is as follows:

[0043] Step 1.1: Perform a model-based three-component decomposition on the ground object scattering coherence matrix corresponding to the polarimetric SAR image of the target area, and extract the power of the surface scattering, secondary scattering and volume scattering components of the ground objects;

[0044] Step 1.2: Using the surface scattering, secondary scattering, and volume scattering component power of the ground objects obtained in Step 1.1 as segmentation data input, the polarimetric SAR image is segmented into low scattering regions, strong scattering regions, and other regions using the Markov field segmentation method based on multidimensional Gaussian distribution.

[0045] Statistical polarimetric SAR images of ports reveal a common characteristic: they are located on coastlines in uneven regions with strong scattering properties. Port detection requires segmentation of both water and land areas, as well as segmentation of strong and weak scattering regions within the land. Therefore, the image must be segmented into at least low-scattering regions (water areas), strong-scattering built-up areas, and other areas (primarily vegetation). A problem with directly using segmentation methods based on regional statistical characteristics is that port water and vegetation areas are often affected by sidelobes from strong-scattering building echoes along the coast, leading to incorrect segmentation and resulting in missed detections and false alarms. Therefore, a model-based three-component decomposition of the ground object scattering coherence matrix, extracting surface scattering, secondary scattering, and volume scattering components for segmentation, can effectively reduce these erroneous segmentations. The specific principle and method are as follows:

[0046] If the scattering coherence matrix of a single pixel of the target is T, a surface scattering, secondary scattering and volume scattering model of ground object scattering is established. The three-component decomposition decomposes T into the sum of three scattering components: surface scattering, secondary scattering and volume scattering.

[0047] T = P s T surface +P d T double +P v T volume (1)

[0048] Where T surface T double T volume 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.

[0049] To avoid the negative power problem, the modified three-component decomposition method proposed in the literature (W An, Y. Cui, and J. Yang, "Three-Component Model-Based Decomposition for Polarimetric SAR data," IEEE Trans. Geosci. RemoteSens., vol. 48, no. 6, pp. 2732-2739, June 2010.) is used to decompose the dedirected coherence matrix T′, and the volume scattering model T′ is decomposed. volume The model is a 3×3 unit diagonal matrix. After corrected ternary decomposition, the three-dimensional feature vector {(P} of the image can be obtained. s (x, y), P d (x, y), P vLet I be an image plane consisting of three regions {Ω}. k The segmentation is composed of {Q = q(x, y)}, where k ∈ {1, 2, 3}. The segmentation result is represented as {Q = q(x, y)}, where q(x, y) ∈ {1, 2, 3}. The optimal segmentation model is then expressed as:

[0050] Q = argmax Q P(Q|I) (2)

[0051] In the formula, P(Q|I)∝P(I|Q)P(Q), P(I|Q) is the data likelihood probability, and P(Q) is the prior probability of the segmentation.

[0052] Likelihood probability is related to the distribution of pixel data in each region. Under the Markov assumption, likelihood probability is expressed as the product of the probabilities of each pixel given the class, i.e., P(I|Q)=∏ (x，y) P(I|q(x,y)). The probability distribution of eigenvectors is difficult to derive; therefore, it is assumed that the probability distribution of eigenvectors in homogeneous regions follows a three-dimensional Gaussian distribution.

[0053]

[0054] in ∑ k is the k-mean covariance matrix of the region. This represents the average value of k in the region.

[0055] The MRF segmentation model models the categories of neighboring pixels as a two-dimensional Markov field, and the prior probabilities are described using a Gibbs distribution:

[0056]

[0057] In the formula, Z is the normalization factor, and for pixel i, U(q) i The class of pixel i is determined by the class of its neighboring pixels. If the 4- or 8-neighborhood of pixel i is represented as N... i ,but Where V d (q i )=β(1-δ(q i q d )), β>0, δ(·,·) is the impulse function, and the energy is 0 if the categories are the same.

[0058] From equations (2), (3), and (4), the objective function of the segmentation model can be obtained:

[0059]

[0060] The three-region MRF segmentation model can be solved using an iterative conditional model. The segmentation initialization is obtained by using the three-region K-means clustering method on the three component power eigenvectors.

[0061] Step 2: Extracting the region of interest at the port:

[0062] The port is located on the land-sea boundary, therefore the region of interest (ROI) of the port lies within the coastal zone. Extracting the coastal zone requires separating the water and land from the segmented image. If the result of the MRF three-region segmentation is Q, the average power of each of the three segmented regions is calculated. The region with the lowest average power is identified as the water area, thus obtaining the land-sea segmentation binary image A, where the water area corresponds to a value of 1 and the land area to a value of 0. The coastal zone can be determined by drawing circles point-by-point along the land-sea segmentation boundary, such as... Figure 2 As shown in the diagram, for the land-sea separation map, the land-sea separation boundary line B is first determined by the binary map boundary tracing algorithm, then the radius Rad of the narrow zone is determined, and the coastal narrow zone C is determined by drawing circles centered on each point on B.

[0063] Man-made structures in port areas typically exhibit strong scattering; therefore, the region of interest (ROI) for the port can be extracted by extracting strong scattering regions within the narrow coastal region. First, strong scattering regions are identified from the segmentation result Q. These strong scattering regions correspond to the region D with the highest average power. Thus, the strong scattering ROI E = C∧D for all ports within the narrow coastal region can be obtained.

[0064] Port area artificial structures are typically large and may comprise multiple structures; therefore, strong scattering regions within a port usually have a certain area. Utilizing this contextual characteristic, the region of interest (ROI) can be extracted by identifying the largest region within a narrow-band strong scattering region E. However, due to the complexity of the distribution of port artificial structures, structures within a narrow-band region may be truncated into multiple segments. Therefore, for all connected sub-regions in E, region merging is required within a set distance threshold. If the merged result is F, the minimum bounding rectangle of the merged region is determined, thus obtaining the candidate ROI. If the merged candidate set is U = {u...} k}(k∈{1,...,M}), the corresponding minimum bounding rectangle set is V={v k}(k∈{1,...,M}), for candidate target u k Determine its minimum bounding rectangle v k And the candidate target is located in the strongly scattering pixel E in E. k If E k Area larger than the strong scattering area threshold of the narrowband region t Then it is determined to be the port region of interest Y = {y k}(k∈{1,...,n y}).

[0065] In summary, the specific steps for extracting the region of interest for a port are as follows:

[0066] Step 2.1: Utilizing the contextual features of the port being surrounded by water, and based on the three-region segmentation results from Step 1, calculate the average power of the three regions respectively. The region with the lowest average power is determined to be the water area, thus obtaining the water-land segmentation map. Determine the segmentation boundary coastline, and draw circles point by point along the coastline to determine the narrow coastal zone.

[0067] Step 2.2: Utilizing the contextual features of a large number of strong scattering artificial structures distributed on the land side of the port area, and based on the three-region segmentation results of Step 1, extract the strong scattering region segmentation map to determine the distribution map of strong scattering regions in the narrow coastal area.

[0068] Step 2.3: Taking advantage of the large area of ​​artificial buildings with strong scattering in the port area, the strong scattering sub-regions in the narrow band area are merged according to the contour distance to obtain candidate sub-regions. The strong scattering area in each candidate sub-region is calculated, and the region of interest in the port is extracted based on the size of the strong scattering area.

[0069] Step 3: Port Identification:

[0070] Utilizing the asymmetric scattering characteristics of artificial structures in the port area, the co-polarization and cross-polarization scattering correlation power of several regions of interest in the port obtained in step 2 is calculated, and the final identification is performed based on the distribution of co-polarization and cross-polarization scattering correlation power in the regions of interest.

[0071] The strong scattering region segmented by MRF is a segmentation result based on the statistical properties of the coherence matrix. Non-artificial building areas exist within the segmented strong scattering regions. Therefore, some false alarms exist in the extracted region of interest, requiring identification using the unique polarization scattering characteristics of artificial buildings. Artificial buildings exhibit asymmetric reflection relative to natural features, and under horizontal and vertical orthogonal polarization bases, the co-polarization and cross-polarization correlation power of symmetrical target scattering is almost zero. (S HH Horizontal co-polarization scattering power, S VV For vertical co-polarization scattering power, S HV (For horizontal and vertical cross-polarization power), therefore, port identification can be performed using reflection symmetry.

[0072] The strength of co-polarization and cross-polarization correlation power can distinguish between asymmetric man-made structures and symmetric features, but it faces the problem of difficulty in determining the segmentation threshold. Considering that MRF segmentation obtains three-region segmentation results of the image, using the other regions (vegetation regions) in the MRF segmentation results as background, the histogram distribution of co-polarization and cross-polarization correlation power in the other regions is statistically analyzed, and a global constant false alarm rate (CFAR) method is used to determine the threshold Th. Under a set false alarm rate Pfa, the segmentation threshold Th is determined based on the co-polarization and cross-polarization correlation power distribution in the other regions. Then, the extracted ports' regions of interest are identified sequentially. For the port region of interest y... lDetermine its minimum bounding rectangle region Rect l and the strongly scattering pixel E located in E l If E l If the proportion of pixels whose cross-polarization correlation power is higher than the segmentation threshold Th is greater than the proportion threshold ρ, then the pixel is identified as a port. The threshold ρ is between 0.2 and 0.6, and is 0.5 in this example.

[0073] The following is a specific example of port inspection using RADARSAT-2 fully polarimetric SAR data for a portion of Singapore:

[0074] The RADARSAT-2 fully polarimetric SAR data for parts of Singapore has a resolution of approximately 5m × 5m and a data size of 1000 × 1000. Figure 3 (a) is a Pauli pseudocolor map of the region's data. Observation reveals numerous ports of varying shapes along the coast. Port breakwaters are relatively small, with indistinct geometric features. Some large breakwaters are blended seamlessly with coastal man-made structures, while some smaller breakwaters have irregular shapes. Directly using port detection methods based on geometric features would result in numerous missed detections. These ports with varying geometric features share a common characteristic: they are all surrounded by water, and their landside is composed of numerous man-made structures with strong scattering properties.

[0075] When performing MRF segmentation, the Gibbs distribution parameter β is set to 0.5. Assume the image resolution is R. x ×R y If the width of the dam is 100m, then the number of pixels w for the dam width is... When extracting the region of interest (ROI) for a port, the radius (rad) is set to w / 2 when extracting the narrow coastal region. It is assumed that the minimum bounding rectangle in the narrow region has dimensions at least w, representing strong scattering. Therefore, the strong scattering area threshold (Area) for the narrow region is determined. t Take w 2 When identifying ports, the false alarm rate Pfa is set to 0.01, and the area ratio ρ is determined to be 0.5.

[0076] Figure 3 (b)-(j) represent the detection results of the proposed method:

[0077] in Figure 3 (b) shows the MRF three-region segmentation results for comparison. Figure 3 As can be seen from (a) and (b), both the water area and the strong scattering structures were well segmented. By extracting the three-component segmentation power component for segmentation, the disturbed water area and the area disturbed by the strong scattering artificial structures in the port area were correctly segmented. Figure 3 (c) is the land-water segmentation map extracted from the three-region segmentation results. The waters related to the port were correctly extracted.Figure 3 (d) is a narrow coastal zone defined by the land-water boundary line. Figure 3 (e) shows the distribution of strong scattering over a narrow coastal zone. (Comparison) Figure 3 (a) and Figure 3 (e) It can be found that the strong scattering man-made structures along the coast of each port, including breakwaters, moored ships, bridges, etc., are all contained within the extracted narrow coastal area. Figure 3 (f) represents the defined port area of ​​interest. Figure 3 (h) shows the region of interest (ROI) for the port in the Pauli pseudocolor image, where the ROI is marked with a red rectangle, and the marked white rectangles represent detected false alarm regions. (Comparison) Figure 3 (a) and Figure 3 (h) It can be observed that all eight ports with different shapes were correctly detected as ports of interest. Since the large cross-sea bridges along the coast are also large areas of strong scattering, the detected regions of interest contain some false alarms, such as... Figure 3 (h) contains regions 1-3. Figure 3 (i) is the calculated horizontal co-polarization cross-polarization correlation power distribution map. It can be observed that the correlation power of asymmetric targets such as artificial buildings in the port area is significantly higher than that of asymmetric areas such as vegetation and bare land. Figure 3 (j) represents the final port inspection result. (Comparison) Figure 3 (h) and Figure 3 (j) can be found that, Figure 3 Figure 3 All three false alarms in (h) were correctly suppressed, and the proposed method achieved accurate port detection.

[0078] Table 1 compares the proposed method with a port detection method based on multi-directional breakwater scanning, showing the results. Pd and Pf represent the detection rate and false alarm rate, respectively. The table shows that the proposed method detected all eight ports with a 100% detection rate, while the comparative method only detected four ports with a 50% detection rate. The comparative method detected three false alarms in each port, resulting in a false alarm rate of 42.8%, while the proposed method detected no false alarms. Therefore, the proposed method's detection rate and false alarm rate are significantly superior to the comparative method.

[0079] Table 1 compares the performance and false alarm rates of the proposed method with those of the multi-directional breakwater scanning method.

[0080]

[0081] 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 port detection method based on contextual features, characterized in that: Includes the following steps: Step 1: Acquire the polarimetric SAR image of the target area; perform three-region segmentation on the polarimetric SAR image of the target area, dividing the polarimetric SAR image into a low-scattering region, a strong-scattering region, and other regions; Step 2: Based on the three-region segmentation results from Step 1, extract the water-land segmentation map, determine the segmentation boundary coastline, and identify the narrow coastal zone along the coastline. Based on the three-region segmentation results of step 1, the strong scattering region segmentation map is extracted to determine the distribution map of strong scattering regions in the narrow coastal zone. Strong scattering sub-regions in the narrow band region are merged according to the contour distance to obtain candidate sub-regions. The strong scattering area in each candidate sub-region is calculated, and the port region of interest is extracted based on the size of the strong scattering area. Step 3: Calculate the co-polarization and cross-polarization scattering correlation power of several port regions of interest obtained in Step 2, and perform final identification based on the co-polarization and cross-polarization scattering correlation power distribution of the regions of interest: The final identification in step 3 includes: for the port area of ​​interest Identify the strongly scattering pixels located in the strongly scattering region. ,like The percentage of pixels with cross-polarization correlation power higher than Th is greater than Then, it is identified as a port, where Th is the set segmentation threshold. The set percentage threshold ranges from 0.2 to 0.

6. The segmentation threshold Th is determined using a global constant false alarm rate method. The specific process is as follows: [The text abruptly ends here, so the translation stops as well.] Next, taking the other regions in the three-region segmentation results of step 1 as the background, the distribution of the common polarization cross-polarization related power histogram of the other regions is statistically analyzed, and the threshold Th is determined based on the distribution of the common polarization cross-polarization related power of the other regions.

2. The port detection method based on contextual features according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1: Perform a model-based three-component decomposition on the ground object scattering coherence matrix corresponding to the polarimetric SAR image of the target area, and extract the power of the surface scattering, secondary scattering and volume scattering components of the ground objects; Step 1.2: Using the surface scattering, secondary scattering, and volume scattering component power of the ground objects obtained in Step 1.1 as segmentation data input, the polarimetric SAR image is segmented into low scattering regions, strong scattering regions, and other regions using the Markov field segmentation method based on multidimensional Gaussian distribution.

3. The port detection method based on contextual features according to claim 1, characterized in that: In step 2, based on the three-region segmentation results from step 1, the average power of the three segmented regions is calculated respectively. The region with the minimum average power is determined to be the water area, thus obtaining the water area-land segmentation map.

4. The port detection method based on contextual features according to claim 1, characterized in that: In step 2, circles are drawn point by point along the coastline to determine the narrow coastal zone.

5. A computer-readable storage medium storing computer-executable instructions, characterized in that: When executed, the instructions are used to implement the method of any one of claims 1 to 4.

6. A computer system, comprising: One or more processors, the computer-readable storage medium of claim 5, for storing one or more programs, characterized in that: when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1 to 4.

Images