A method for characterizing ground object scattering properties and detecting buildings in PolInSAR images

By combining polarimetric interferometry coherence, target decomposition, and iterative processing of self-organizing mapping networks, the problem of volume scattering overestimation in rotating building regions is solved, and accurate detection of building regions is achieved.

CN118762161BActive Publication Date: 2026-11-17HARBIN INST OF TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410736815.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-06-07
Publication Date
2026-11-17
Estimated Expiration
2044-06-07

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively distinguish the polarization responses of forests and rotating buildings, leading to an overestimation of volume scattering in rotating building areas and inaccurate building detection.

Method used

By combining polarimetric interferometry coherence, target decomposition, and self-organizing mapping network, and through iterative processing, volume scattering of rotating building regions is corrected, building region segmentation is refined, and accurate building detection is achieved.

Benefits of technology

The volume scattering of the rotating building area was effectively corrected, the scattering ambiguity between the forest and the rotating building was removed, and the building area was accurately extracted.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118762161B_ABST
    Figure CN118762161B_ABST
Patent Text Reader

Abstract

This invention discloses a method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images. The method includes a preprocessing step, a target decomposition step, a building detection step, a region segmentation refinement step, and an iterative update step. The preprocessing step extracts features from the input data. The target decomposition step decomposes the polarimetric data using polarimetric interferometry. The building detection step combines the target decomposition results with a self-organizing map network to cluster and identify building categories for building detection. The region segmentation refinement step further clusters the building detection results to refine the region segmentation. The iterative update step feeds back the refined segmentation results to the target decomposition step and repeats the aforementioned steps until a termination condition is met. This invention can obtain reasonable ground object scattering mechanisms and their proportions, correct volume scattering components, and simultaneously obtain accurate building detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing technology, and relates to a method for characterizing the scattering characteristics of ground objects and detecting buildings in polarimetric interferometric SAR (PolInSAR) images. Background Technology

[0002] Polarimetric Synthetic Aperture Radar (PolSAR) can emit electromagnetic waves with different polarization states to acquire information related to ground material, dielectric constant, and orientation. Interferometric Synthetic Aperture Radar (InSAR) can obtain information related to target height and its degree of variation through interferometric processing. PolInSAR combines the two, acquiring even more information and has wide applications in remote sensing information processing and interpretation.

[0003] Ground object scattering mechanism analysis and scattering characteristic characterization are fundamental to subsequent interpretation and applications, and are an important part of PolInSAR research. Model-based target decomposition is an important implementation method. In target decomposition research, Freeman and Durden proposed a three-component target decomposition method in 1998, including surface scattering, dihedral scattering, and volume scattering, which can characterize the scattering characteristics of ground objects. However, it also faces the problem of overestimation of volume scattering in rotating building areas. Since then, most research on target decomposition has focused on solving this problem, developing two methods: one is to use polarimetric azimuth to correct polarimetric data, and the other is to increase the types of scattering mechanisms and make full use of the elements in the polarimetric data. However, despite a lot of effort invested in these two aspects, the problem of overestimation of volume scattering in rotating building areas has not been effectively solved, or the target decomposition results cannot balance the reasonable interpretation of rotating building areas with other areas, and cannot effectively distinguish between forests and rotating buildings. Fundamentally, this is because forests and rotating buildings are very similar in polarimetric response, which is also known as ambiguity. The target decomposition based on the model is not orthogonal, and the scattering modeling also lacks rigorous calculation and derivation. Therefore, it is difficult to solve this problem from the perspective of polarization response, which also makes the subsequent building detection inaccurate. Summary of the Invention

[0004] The purpose of this invention is to provide a method for characterizing the scattering characteristics of ground objects and detecting buildings in PolInSAR images. This method corrects the overestimated volume scattering component in rotated building areas, provides a reasonable characterization of ground object scattering characteristics, and achieves accurate building detection. This invention can obtain a reasonable ground object scattering mechanism and its proportion, correct the volume scattering component, and simultaneously obtain accurate building detection results.

[0005] The objective of this invention is achieved through the following technical solution:

[0006] A method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images includes the following steps:

[0007] Step 1: Preprocessing: The polarimetric interferometric input data is registered, filtered, subjected to interferometric processing, and optimized for polarimetric interferometric coherence to generate building indicators, orientation maps, and phase differences based on polarimetric interferometric coherence. Figure 3 One feature;

[0008] Step 2, Target Decomposition: Based on building indicators, perform region segmentation or directly use the refined region segmentation results from the iteration process to assist in the target decomposition of polarization data, so as to generate the energy and proportion of each scattering component;

[0009] Step 3, Building Detection: Using the target decomposition results and the three features obtained in Step 1 as input, a self-organizing map network is used for clustering, and the building categories are identified to achieve building detection.

[0010] Step 4, Refinement of Region Segmentation: Using the optimal polarization interference coherence as input, the self-organizing map network is used again to perform fine clustering on the building regions generated in Step 3, and the building categories are divided into two types according to the coherence: regions with higher coherence are non-rotating building regions, and regions with lower coherence are rotating building regions, thus refining the segmentation of building regions.

[0011] Step 5, Iterative Update: Determine if the termination condition is met. If not, feed back the region segmentation results obtained in Step 3 and Step 4 to Step 2, and execute Step 2 to Step 4 again. If the condition is met, stop the iteration and output the target decomposition results and building detection results.

[0012] Compared with the prior art, the present invention has the following advantages:

[0013] 1. The scattering of isolated rotating building areas is defined as rotating dihedral scattering, and the scattering of densely rotating building areas is defined as coherent volume scattering, providing a reasonable scattering mechanism for ground feature interpretation;

[0014] 2. By combining polarimetric interferometry coherence, target decomposition, and self-organizing mapping network, and with iterative loop processing, the differences between natural regions, densely rotating building regions, and non-densely rotating building regions are deeply explored, effectively correcting the volume scattering of rotating building regions and removing the scattering ambiguity between forests and rotating buildings.

[0015] 3. Based on polarimetric interferometry coherence and target decomposition results, accurate extraction of building areas is achieved for scenes with both densely rotating buildings and forests. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the algorithm for characterizing the scattering characteristics of ground objects and detecting buildings;

[0017] Figure 2 This is the SAR amplitude map of the HH channel of the UAVSAR data used in the embodiment. In the map, patch_NA and the green circled area are the regions of interest in natural areas, patch_non_DRBA and the red circled area are the regions of interest in non-dense rotating building areas, and patch_DRBA and the yellow circled area are the regions of interest in dense building areas.

[0018] Figure 3 The image shows the building detection results of the UAVSAR data used in the example. White represents the detected areas, and black represents the undetected areas. Detailed Implementation

[0019] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0020] This invention provides a method for characterizing the scattering characteristics of ground objects and detecting buildings in PolInSAR images. The method includes a preprocessing step, a target decomposition step, a building detection step, a region segmentation refinement step, and an iterative update step. The preprocessing step extracts features from the input data. The target decomposition step decomposes the polarimetric data using polarimetric interferometry. The building detection step combines the target decomposition results with a self-organizing map network to cluster and identify building categories, thereby achieving building detection. The region segmentation refinement step further clusters the building detection results to achieve refined region segmentation. The iterative update step feeds back the refined segmentation results to the target decomposition step and repeats the aforementioned steps until a termination condition is met. Figure 1 As shown, the specific steps are as follows:

[0021] Step 1: Preprocessing: The polarimetric interferometric input data is registered, filtered, subjected to interferometric processing, and optimized for polarimetric interferometric coherence to generate building indicators, orientation maps, and phase differences based on polarimetric interferometric coherence. Figure 3 The specific steps are as follows: (1 feature)

[0022] Step 1: Register, filter, perform interference processing on the polarization interference input data, and optimize the polarization interference coherence to obtain the coherence coefficients for each polarization state and the interference fringe patterns γ under the HH, HV, and VV polarization channels. HH γ HV and γVV And the optimal, second-optimal and third-optimal polarization interference coherence coefficients γ opt1 γ opt2 and γ opt3 .

[0023] Steps 1 and 2: Calculate the building indicator P based on polarization interferometry coherence. bd :

[0024]

[0025] Where, γ i The coefficient represents the coherence coefficient under state i, where i represents different polarization states or optimal and suboptimal states. HH, HV, and VV represent the HH polarization state, HV polarization state, and VV polarization state, respectively. opt1 and opt2 represent optimal and suboptimal coherence.

[0026] Step 13: Interference fringe pattern γ based on HH polarization state HH Calculate the orientation diagram θ and the phase difference diagram Δφ:

[0027]

[0028] Where θ∈[0,180°), the step is 5°, L is the length, set to 10, and (i,j) is the coordinate of the current pixel;

[0029]

[0030] Where θ is the calculated orientation map.

[0031] Step 2, Target Decomposition: Based on building indicators, perform region segmentation or directly use the refined region segmentation results from the iteration process to assist in the target decomposition of polarization data, in order to generate the energy and proportion of each scattering component. The specific steps are as follows:

[0032] Step 2.1: In the initial stage, based on the building indicator P obtained in Step 1... bd The natural and building areas in the figure are segmented using the segmentation algorithm shown in equation (4):

[0033]

[0034] Where Th is the threshold, expressed as:

[0035] Th = max(P) bd ), when (1+(γ) HH +γ VV ) / P bd )>M (5);

[0036] The parameter M is related to the length of the time baseline; the longer the time baseline, the larger the parameter M. When the time baseline is short and the average coherence of the building area is around 0.8, M is set to 6; when the time baseline is long and the coherence of the building area is around 0.6, M is set to 16.5. For other coherence conditions, the parameter M is adjusted according to the above characteristics.

[0037] Based on the segmentation results, the target decomposition is performed as shown in equation (6):

[0038]

[0039] Where [T] represents the polarization coherence matrix, f represents the scattering energy, and the subscripts s, d, v, and rd represent surface scattering, dihedral scattering, volume scattering, and rotating dihedral scattering, respectively, and the coherence matrices of each scattering are as follows:

[0040]

[0041] The energy expression for the volume scattering or rotational dihedral scattering component is:

[0042] P v or P rd =4T 33 (10);

[0043] Define x 11 =T 11 -2T 33 x 22 =T 22 -T 33 T 11 T 12 T 22 and T 33 The elements are in the polarization coherence matrix [T], and the subscripts represent the row and column they are in.

[0044] When x 11 >x 22 When α = 0, β = T 12 / x 11 ,at this time:

[0045] P s =x 11 +|T 12 | 2 / x 11 (11);

[0046] P d =x 22 -|T 12 | 2 / x 11 (12);

[0047] When x 11 ≤x 22 When β = 0, α = T 12 / x 22 ,at this time:

[0048] P s =x 22 +|T 12 | 2 / x 22 (13);

[0049] P d =x 11 -|T 12 | 2 / x 22 (14);

[0050] Thus, the surface scattering component P is obtained. s Dihedral scattering component P d Volume scattering component P v and the dihedral scattering component P of rotation rd .

[0051] Step 22: During the iteration, the region segmentation result is provided by the region segmentation refinement step in the previous iteration. The region in the figure is segmented into natural regions, non-dense rotating building regions, and dense rotating building regions. Based on this, the target decomposition as shown in Equation (15) is performed:

[0052]

[0053] Where [T] represents the polarization coherence matrix, f represents the scattering energy, and the subscripts s, d, v, rd and cv represent surface scattering, dihedral scattering, volume scattering, rotating dihedral scattering and coherent volume scattering, respectively.

[0054] The polarization coherence matrix of coherent volume scattering is expressed as:

[0055]

[0056] The energy expression for coherent volume scattering is:

[0057] P cv =4T 33 (17);

[0058] The energy solutions for other scattering components are the same as those for equations (11)-(14).

[0059] Thus, the surface scattering component P is obtained. s Dihedral scattering component P d Volume scattering component P vThe rotating dihedral scattering component P rd and coherent volume scattering component P cv .

[0060] In this step, the scattering generated by an isolated rotating dihedral structure is rotating dihedral scattering, while the scattering generated by a dense and rotating building area is coherent volume scattering. Here, rotation is defined as the angle between the target's orientation and the radar azimuth.

[0061] Step 3, Building Detection: Using the target decomposition results and the three features obtained in Step 1 as input, a self-organizing map network is used for clustering, and the building categories are identified to achieve building detection. The specific steps are as follows:

[0062] Step 3.1. Let the input of the self-organizing map (SOM) be X = [x l The coherence, building indicators, orientation map, and phase difference map obtained in step one, along with the scattering components obtained from target decomposition, are used as the SOM data, i.e.:

[0063]

[0064] SOM output layer Defined as a two-dimensional grid, 3×3 in size, each cell of the grid... Each neuron, referred to as a neuron, represents a cluster category. p1 and p2 represent the row and column positions of that neuron in the grid. Each SOM neuron is connected to the input features via weights. connect.

[0065] Step 3.2: SOM achieves clustering through competitive learning. The specific steps are as follows:

[0066] Step 321: Calculate the input features X = [x l and the connection weights of each neuron Euclidean distance between As shown in equation (19):

[0067]

[0068] Among them, L input L represents the dimension of the input feature vector in the initial stage. input =12, during the iteration process L input =13.

[0069] Step 322: Classify the current pixel into the category with the smallest Euclidean distance. This category is called the winner neuron, denoted as . As shown in equation (20):

[0070]

[0071] Where P is the dimension of the output neuron grid, P = 3; (p1, p2) are the coordinates of the category in the output neuron grid; (win1, win2) are the coordinates of the winner neuron in the output neuron grid.

[0072] Step 3: In order to achieve self-organization, the weights of the winner neuron and its neighboring neurons are updated according to equations (21)-(24);

[0073]

[0074] L(t) = L0e -t / λ (twenty two);

[0075]

[0076] σ(t)=σ0·exp(-t / λ) (24);

[0077] Where t represents the number of iterations; L(t), θ(t), and σ(t) represent the learning rate function, neighborhood function, and radius of the neighborhood function, respectively; L0 and σ0 represent the initial learning rate and initial radius, respectively; and λ represents the time constant, used to decay the learning rate and radius.

[0078] Steps 3 and 4: Repeat the update process in step 3 3 200 to 1000 times to finally obtain the clustering results of the P×P pattern.

[0079] Step 35: After obtaining the clustering results, identify the category to which the building area belongs in order to perform building detection. The specific steps are as follows:

[0080] Step 3.51: The initial segmentation result during the iteration process or the segmentation result from the previous iteration process, after removing the surface scattering-dominant region, can be used as a reference region to obtain the main building categories: Select the scattering component with the highest proportion within the reference region as the main building category, and represent its coordinates in the SOM output neural network as (p B1 ,p B2 ).

[0081] Step 352: Due to the diversity of building area scattering, there is usually more than one building category. When the main building category accounts for more than 70% of the reference area, a distance-based method is used to determine the remaining building categories, i.e.:

[0082] |(p b1 ,p b2 )-(p B1 ,pB2 )| 2 ≤2 (25);

[0083] Among them, (p B1 ,p B2 (p) represents the coordinates of the main building category in the 2D clustering output layer grid. b1 ,p b2 The coordinates of the building category in the grid of the two-dimensional clustering output layer are represented by ). Otherwise, a statistical method is used to determine the remaining building categories, that is, the categories in the reference area are sorted from largest to smallest according to their proportion and the cumulative proportion is calculated. The clustering category with a cumulative proportion of more than 90% is retained as the building category.

[0084] Step 353: At this point, all building categories (p) have been obtained. b1,b2 The building area refers to the area belonging to the building category, that is:

[0085] I build :(y win1,win2 )∈(p b1,b2 (26).

[0086] Step 4: Refinement of Building Region Segmentation: Using optimal polarimetric interferometric coherence as input, a self-organizing map network is used again to perform fine clustering on the building regions generated in Step 3. Buildings are then divided into two categories based on their coherence: regions with higher coherence are non-rotating building regions, and regions with lower coherence are rotating building regions. This refines the building region segmentation. The specific steps are as follows:

[0087] Step 41: Using optimal polarization interference coherence γ opt1 Using SOM as input, clustering is performed again, and the clustering process is the same as that in equations (19)-(24), to obtain the clustering output results.

[0088] Step 4.2: Calculate Building Area I build Average optimal polarization interference coherence and the average optimal polarization interferometry coherence of the regions to which each category belongs. Similar categories are merged according to equations (27) and (28) to obtain a refined building area segmentation result: Dense Rotating Building Area I 密集旋转建筑 Non-dense rotating building area I 非密集旋转建筑 The remaining area is natural region I. 自然 :

[0089]

[0090] Step 5, Iterative Update: Determine if the termination condition is met. If not, feed the region segmentation results obtained in Steps 3 and 4 back to Step 2, and execute Steps 2 through 4 again. If the condition is met, stop the iteration and output the target decomposition result. The specific steps are as follows:

[0091] Step 51: Set the evaluation metric R for determining the termination of the iteration:

[0092] R = 100 × (|A) new -A old |) / A old (%) (29);

[0093] Among them, A new and A old These represent the building detection results in the current and previous iterations, respectively, which are the number of pixels belonging to the building area.

[0094] Step 5.2: When R gradually converges with iteration and R≤1%, stop the iteration; otherwise, feed back the refined region segmentation result to Step 2, update the region segmentation result, and execute Steps 2 to 4 again until the iteration termination condition is met, and output the target decomposition result and building detection result.

[0095] Example:

[0096] This embodiment uses a set of UAVSAR polarimetric interferometric data to verify the proposed method for characterizing ground object scattering characteristics and detecting buildings. The SAR amplitude map of the HH channel is shown below. Figure 2 As shown in Table 1, the parameter settings for UAVSAR images are as follows.

[0097] Table 1 UAVSAR Image Parameters

[0098]

[0099] The target is decomposed using the method proposed in this invention, and based on... Figure 2 The proportions of each scattering component in the regions of interest (ROIs) defined in the paper are statistically analyzed, as shown in Table 2. The table shows that natural areas (mostly forest areas) maintain volume scattering dominance, while in areas without dense rotating buildings (mostly buildings with orientations parallel to the azimuth), dihedral scattering is dominant. However, in areas with dense rotating buildings, volume scattering is the lowest. Furthermore, scattering from man-made targets, including dihedral scattering, rotating dihedral scattering, and coherent volume scattering, is absolutely dominant. This indicates that the method proposed in this invention has a significant volume scattering correction effect in rotating building areas and can reasonably characterize the scattering characteristics of ground features.

[0100] Table 2 Target decomposition results of UAVSAR

[0101]

[0102] At the same time, building inspection results can also be obtained, such as Figure 3 As shown, it can be seen that buildings, whether oriented parallel to the direction or densely rotating, can be correctly detected. Moreover, there are very few false alarms in the forest area. Further calculations show that the detection rate is 92.5% and the false alarm rate is 4.7%, indicating high detection accuracy.

Claims

1. A method for characterizing the scattering characteristics of ground objects and detecting buildings in PolInSAR images, characterized in that... The method includes the following steps: Step 1, Preprocessing: The polarimetric interferometric input data is registered, filtered, interferometrically processed, and polarimetric interferometric coherence optimized to generate three features, including building indicators, orientation maps, and phase difference maps based on polarimetric interferometric coherence. Step 2, Target Decomposition: Based on building indicator, perform region segmentation or directly use the refined region segmentation results during the iteration process to assist in the target decomposition of polarization data, so as to generate the energy and proportion of each scattering component. Among them, the scattering of isolated rotating building regions is defined as rotating dihedral scattering, and the scattering of dense rotating building regions is defined as coherent volume scattering. The rotation is defined as the angle between the target's orientation and the radar azimuth. Step 3, Building Detection: Using the target decomposition results and the three features obtained in Step 1 as input, a self-organizing map network is used for clustering, and the building categories are identified to achieve building detection. Step 4, Refinement of Region Segmentation: Using the optimal polarization interference coherence as input, the self-organizing map network is used again to perform fine clustering on the building regions generated in Step 3, and the building categories are divided into two types according to the coherence: regions with higher coherence are non-rotating building regions, and regions with lower coherence are rotating building regions, thus refining the segmentation of building regions. Step 5, Iterative Update: Determine if the termination condition is met. If not, feed back the region segmentation results obtained in Step 3 and Step 4 to Step 2, and execute Step 2 to Step 4 again. If the condition is met, stop the iteration and output the target decomposition results and building detection results.

2. The method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images according to claim 1, characterized in that... The specific steps of step one are as follows: Step 1: Register, filter, perform interference processing on the polarization interference input data, and optimize the polarization interference coherence to obtain the coherence coefficients for each polarization state and the interference patterns in the HH, HV, and VV polarization channels. , and And the optimal, second-optimal and third-optimal polarization interferometric coherence coefficients , and ; Steps 1 and 2: Calculating building indicator signals based on polarization interferometry coherence. : (1); in, Representing state The coherence coefficient under the following standard is indicated by the following standard: Representing different polarization states, or optimal and suboptimal states, HH, HV, and VV represent the HH, HV, and VV polarization states, respectively. and This represents optimal and suboptimal coherence; Step 13: Interference fringe pattern based on HH polarization state Calculate orientation map Phase difference diagram : (2); in, Step for , The length is set to 10. The coordinates of the current pixel; (3); in, This is the calculated orientation map.

3. The method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images according to claim 2, characterized in that... The specific steps of step two are as follows: Step 2.1: In the initial stage, based on the building indicators obtained in Step 1... The natural and building areas in the figure are segmented using the segmentation algorithm shown in equation (4): (4); in, The threshold is expressed as: (5); Among them, parameters It is a parameter related to the length of the time baseline; Based on the segmentation results, the target decomposition is performed as shown in equation (6): (6); in, Represents the polarization coherence matrix. Indicates scattered energy, subscript , , and Let f(x) represent surface scattering, dihedral scattering, volume scattering, and rotational dihedral scattering, respectively, with the coherence matrices for each type of scattering as follows: (7); (8); (9); The energy expression for the volume scattering or rotational dihedral scattering component is: (10); definition , ,in , , and Polarization coherence matrix The elements in the array have subscripts that indicate their row and column. when hour, , ,at this time: (11); (12); when hour, , ,at this time: (13); (14); Thus, the surface scattering component is obtained. Dihedral scattering components Volume scattering component and the dihedral scattering component of rotation ; Step 22: When performing the iteration, the region segmentation result is provided by the region segmentation refinement step in the previous iteration. The region in the figure is segmented into natural regions, non-dense rotating building regions, and dense rotating building regions. Based on this, the target decomposition as shown in Equation (15) is performed: (15); in, Represents the polarization coherence matrix. Indicates scattered energy, subscript , , , and These represent surface scattering, dihedral scattering, volume scattering, rotating dihedral scattering, and coherent volume scattering, respectively. The polarization coherence matrix of coherent volume scattering is expressed as: (16); The energy expression for coherent volume scattering is: (17); The energy solutions for other scattering components are the same as those for equations (11)-(14); Thus, the surface scattering component is obtained. Dihedral scattering components Volume scattering component Rotational dihedral scattering components and coherent volume scattering components .

4. The method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images according to claim 3, characterized in that... The specific steps of step three are as follows: Step 3.

1. Let the input of the self-organizing map network (SOM) be... The coherence, building indicators, orientation map, and phase difference map obtained in step one, along with the scattering components obtained from target decomposition, are used as the SOM data, i.e.: (18); SOM output layer Defined as a two-dimensional grid, 3×3 in size, each cell of the grid... These are called neurons, representing a cluster category. and This represents the row and column position of the neuron in the grid. Each SOM neuron is connected to the input feature via weights. connect; Step 3.2: SOM achieves clustering through competitive learning. The specific steps are as follows: Step 321: Calculate input features and the connection weights of each neuron Euclidean distance between As shown in equation (19): (19); in, The dimension of the input feature vector in the initial stage During the iteration process ; Step 322: Classify the current pixel into the category with the smallest Euclidean distance. This category is called the winner neuron, denoted as . As shown in equation (20): (20); in, To output the dimensions of the neuron grid, ; The coordinates of the category in the output neuron grid; These are the coordinates of the winner neuron in the output neuron grid. Step 3: In order to achieve self-organization, the weights of the winner neuron and its neighboring neurons are updated according to equations (21)-(24); (21); (22); (23); (24); in, Represents the number of iterations; , and These represent the learning rate function, the neighborhood function, and the radius of the neighborhood function, respectively. and These represent the initial learning rate and the initial radius, respectively. Represents the time constant, used to decay the learning rate and radius; Steps 3 and 4: Repeat step 3.3 of the update process 200 to 1000 times to finally obtain... Clustering results of the patterns; Step 35: After obtaining the clustering results, identify the category to which the building area belongs in order to perform building detection. The specific steps are as follows: Step 3.51: The initial segmentation result during the iteration process or the segmentation result from the previous iteration process, after removing the surface scattering-dominant region, is used as the reference region. The scattering component with the highest proportion within the reference region is selected as the main building category, and its coordinates in the SOM output neural mesh are represented as follows: ; Step 352: When the proportion of the main building category in the reference area exceeds 70%, the remaining building categories are determined using a distance-based method, i.e.: (25); in, The coordinates of the main building categories in the output layer grid of the two-dimensional clustering. The coordinates of the building category in the grid of the two-dimensional clustering output layer are used; otherwise, a statistical method is used to determine the remaining building categories, that is, the categories in the reference area are sorted from largest to smallest according to their proportions and the cumulative proportions are calculated. The clustering categories with a cumulative proportion of more than 90% are retained as building categories. Step 353: At this point, all categories of buildings have been obtained. The building area refers to the area that belongs to the building category, that is: (26)。 5. The method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images according to claim 4, characterized in that... The specific steps of step four are as follows: Step 41: Using optimal polarization interference coherence Using SOM as input, clustering is performed again, and the clustering process is the same as that in equations (19)-(24), to obtain the clustering output results. ; Step 42: Calculate the building area Average optimal polarization interferometry coherence and the average optimal polarization interferometric coherence of each category's region Similar categories are merged according to equations (27) and (28) to obtain a refined building area segmentation result: dense rotating building area Non-dense rotating building areas The remaining area is a natural area. : (27); (28)。 6. The method for characterizing ground object scattering characteristics and detecting buildings in PolInSAR images according to claim 5, characterized in that... The specific steps of step five are as follows: Step 51: Set the evaluation criteria for determining the termination of the iteration. : (29); in, and These represent the building detection results in the current and previous iterations, respectively, i.e., the number of pixels belonging to the building area; Step 52, when As the iterations gradually converge and If the iteration stops, the refined region segmentation result is fed back to step two, the region segmentation result is updated, and steps two through four are executed again until the iteration termination condition is met, and the target decomposition result and building detection result are output.

Citation Information

Patent Citations

  • Synthetic aperture radar (SAR) image segmentation method based on deconvolution network and mapping inference network

    CN105389798A

  • Radar remote sensing image scene feature description and classification method

    CN110032974A