A method for extracting buildings from InSAR terrain products

Abstract

The present invention provides a building extraction method for InSAR terrain products, comprising: data preparation and preprocessing; screening samples as training sets, feature calculation, feature discrimination calculation, and feature fusion using principal component analysis; inputting the feature-fused samples into a neural network for training and optimization; performing feature calculation and feature fusion, and using the trained neural network to perform feature extraction; obtaining preliminary building extraction results; identifying shadows and overlapping areas with similar textures to buildings in a study area; removing the identified shadows and overlapping results from the preliminary building extraction results, screening out some misidentified areas, and performing classification and post-processing to obtain a final building extraction result, thereby realizing the combination of SAR image texture features extracted by the neural network and external terrain factors to extract buildings, thereby improving the accuracy of building extraction using SAR images in areas with large terrain fluctuations, such as mountainous areas.

Description

Technical Field

[0001] The present invention relates to the technical field of Synthetic Aperture Radar (SAR) image interpretation, and in particular to a building extraction method for InSAR terrain products. Background Art

[0002] With the development of the global economy and the accelerating pace of urbanization, artificial structures are undergoing rapid renewal. As a primary target for social development, disaster relief, and earthquake monitoring, studying regional changes in buildings is crucial for understanding a region's development from a macro perspective, providing critical support for social development. Furthermore, it can precisely identify rescue areas during inevitable natural disasters, thereby protecting people's lives and property. Synthetic Aperture Radar (SAR) offers significant advantages, including all-weather capability and the ability to penetrate clouds and rain, enabling it to obtain surface information in virtually any environment. Therefore, using SAR imagery to extract buildings has become a hot topic of research. However, SAR imaging differs from optical remote sensing, and its unique speckle noise poses significant challenges to SAR image interpretation. Furthermore, due to its slant-range projection imaging mode, the acquired SAR images exhibit distinct characteristics depending on the varying terrain of the study area. Therefore, developing a comprehensive building extraction method for SAR imagery has been a major challenge.

[0003] Neural networks have been a hot topic in SAR image building extraction in recent years. The main idea is to allow the network to acquire the ability to automatically extract buildings from SAR images by learning sample data. The advantage of this method is that it can automatically extract buildings from new data using existing samples, greatly improving the efficiency of the extraction method while effectively ensuring accuracy. However, due to the differences in terrain in different regions, the imaging mechanism of SAR images causes areas with large terrain undulations, such as mountainous areas, and mountainous areas with large slopes to exhibit texture features similar to buildings. Using neural networks, it is impossible to perform high-precision building extraction from SAR images in areas with large terrain undulations. Therefore, for SAR image building extraction in mountainous terrain, traditional texture feature calculation methods and neural network deep learning cannot achieve good results. Research on SAR image building extraction methods suitable for mountainous terrain has a guiding role in economic development in mountainous areas and responding to natural disasters. Summary of the Invention

[0004] The purpose of the present invention is to provide a building extraction method for InSAR terrain products that combines neural networks and external terrain factors, starting from the specific interpretability of terrain factors on SAR image building extraction, and to address the problem that SAR image buildings are easily confused with mountains when the terrain is undulating, thereby affecting the accuracy of building extraction.

[0005] In order to achieve the above object, the present invention provides a method for extracting buildings from InSAR terrain products, comprising the following steps:

[0006] S1, data preparation and preprocessing, the data are original data and the terrain information of the corresponding study area;

[0007] S2, select samples as training sets, calculate features of samples, calculate feature discrimination through Bhattacharyya distance, and apply principal component analysis for feature fusion;

[0008] S3: Input the feature-fused samples into the neural network to extract and train the pooled convolution features, select the loss function and optimizer, and modify the neural network structure and parameters according to the training results until the neural network loss function reaches convergence;

[0009] S4, performs feature calculation and feature fusion, and uses the trained neural network to perform feature extraction;

[0010] S5, stitching the small area images of the building extraction result set to obtain the preliminary building extraction results under the real terrain conditions;

[0011] S6, obtain satellite imaging parameters and reconstruct the topographic information of the study area; calculate the slope and aspect corresponding to each pixel, analyze the relationship and difference between imaging angle and slope and aspect, and identify shadows and overlapping areas with similar textures to buildings in the study area;

[0012] S7, remove the identified shadows and overlapping results from the preliminary building extraction results, and filter out some misidentified areas by setting the slope threshold, and obtain the final building extraction results after classification and processing.

[0013] Furthermore, in S1, the original data is multi-viewed to obtain a SAR intensity image; the SAR intensity image is geocoded and converted to the real terrain; the SAR intensity image is filtered; and the slope and aspect of the corresponding study area terrain information is calculated using external DEM data.

[0014] Furthermore, the training samples in S2 are all single-polarization SAR image data. The features of the SAR intensity images are calculated, and the Bhattacharyya distance is used to evaluate the discrimination ability of the calculated multi-category features. The three categories of features with the strongest discrimination ability are selected, and the principal component analysis method is used to eliminate the coherence between the features; the final processed features are fused to form a new feature image as the input of the neural network.

[0015] Furthermore, the feature calculation in S2 uses the gray-level co-occurrence matrix to count the joint probability distribution of the two pixel positions; let S be the set of pixel pairs with specific spatial connections in the target area R, then the co-occurrence matrix P is defined as:

[0016] P(i,j)=[p(i,j,d,θ)] (1)

[0017] That is, the frequency of simultaneous occurrence of pixel pairs with a step size of d, a direction of θ, and grayscales i and j (i, j = 0, 1, 2, 3, ..., N) on the image; where θ is the angle of the line connecting the two pixels;

[0018] Texture features are further extracted from it, and variance, contrast, and mean are used as features. The calculation formula is:

[0019]

[0020]

[0021]

[0022] Furthermore, the neural network selected in S3 is the U-Net network, and the optimizer type used for training is the stochastic gradient descent optimizer; the training samples input into the neural network include SAR intensity image feature fusion images and building label images.

[0023] Furthermore, in S4, the SAR intensity image of the study area is filtered, the features of the SAR intensity image are calculated, and the features selected by S2 are used for feature fusion. At the same time, the large scene image after feature fusion is cropped into small scene images for neural network recognition, and the segmented small scene image data is input into the neural network to meet the network input parameter size requirements. The trained neural network is used to extract features from batch images to obtain preliminary building extraction results s.

[0024] Furthermore, in S6, the acquired terrain information of the corresponding study area is preprocessed to keep its resolution and pixel size consistent with the SAR intensity image, and the external DEM data is resampled.

[0025] Furthermore, S6 uses external DEM data and satellite imaging parameters to establish a functional relationship between DEM and geometric distortion to identify geometric distortion areas. Slope, aspect, and elevation are calculated pixel by pixel, and based on the relationship between these and the satellite incidence angle and azimuth parameters, it is determined whether the area is a shadow overlap area. This is also determined based on the satellite flight mode.

[0026] For satellites in orbit raising mode:

[0027] Judgment 1:

[0028]

[0029] Where aspect(i,j) is the aspect angle of pixel point (i,j), which is the angle rotated clockwise with the north direction as the starting axis; slope(i,j) is the slope angle of pixel point (i,j); azi_angle(i,j) is the flight azimuth of the SAR satellite at the pixel point, which is the angle rotated clockwise with the north direction as the starting axis; inc_angle(i,j) is the incident angle of the SAR satellite at the pixel point;

[0030] For this judgment, if the relationship between its parameters satisfies all the equations in (5) at the same time, the pixel is considered to belong to the distortion area;

[0031] Judgment 2:

[0032]

[0033] slope(i,j)>inc_angle(i,j) (7)

[0034] For this judgment, if the relationship between its parameters satisfies any one or two of the equations (6) and (7), the pixel is considered to belong to the distortion area.

[0035] For satellites in descending orbit mode:

[0036] Judgment 1:

[0037]

[0038] For this judgment, if the relationship between its parameters satisfies all the equations in (8) at the same time, the pixel is considered to belong to the distortion area;

[0039] Judgment 2:

[0040]

[0041] slope(i,j)>(90-inc angle(i,j) ) (10)

[0042] For this judgment, if the relationship between its parameters satisfies any one or two of the equations (9) and (10), the pixel is considered to belong to the distorted area.

[0043] The above solution of the present invention has the following beneficial effects:

[0044] The method provided by the present invention is suitable for extracting buildings in difficult areas such as mountainous areas. Based on the texture features of SAR images, it is assisted by external terrain information to jointly extract buildings. The accuracy of the resulting building extraction results is greatly improved. This method fully utilizes the image information and terrain information in the study area to identify built-up areas. It not only shows good results in difficult areas such as mountainous areas, but also has high accuracy in flat areas.

[0045] The present invention does not rely on the original SAR data, which significantly improves the universality of the method and can be used to extract buildings from any type of SAR data. It mainly uses SAR images and supplements them with terrain information, combining the two to extract buildings.

[0046] The neural network used in the present invention is more flexible in selection. The neural network can better overcome the subtle differences between different data for building extraction and achieve a good balance between efficiency and accuracy. The neural network selected in the scheme is the U-Net network. In addition, the deeplabV3+ and VGG networks can also achieve good results in this invention.

[0047] Other beneficial effects of the present invention will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the present invention;

[0049] Figure 2 Part of the data used for network training of the present invention is shown (where (a) is the original intensity image, (b) is the image after feature fusion, and (c) is the corresponding building label result);

[0050] Figure 3 Schematic diagram of data used in an embodiment of the present invention (where (a) is an intensity image and (b) is an optical image of the corresponding study area);

[0051] Figure 4 Schematic diagrams of shadows and overlapping areas in an embodiment of the present invention (where (a) is a schematic diagram of a shadow area, and (b) is a schematic diagram of an overlapping area);

[0052] Figure 5: Comparison diagram before and after the removal of shadows and overlapping areas in an embodiment of the present invention (where (a) is before the removal of shadows and overlapping areas, and (b) is after the removal of shadows and overlapping areas);

[0053] Figure 6 1 is a diagram showing the building extraction effect in an embodiment of the present invention (where (a) is the standard building mask result, and (b) is the building extraction result of the present invention). DETAILED DESCRIPTION

[0054] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0055] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on this disclosure, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects described herein can be used to implement an apparatus and / or practice a method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this apparatus and / or practice this method.

[0056] It should also be noted that the diagrams provided in the following embodiments are merely schematic illustrations of the basic concepts of the present disclosure. The diagrams only show components relevant to the present disclosure and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the configuration, quantity, and proportion of each component may be varied at will, and the component layout may be more complex. Furthermore, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, those skilled in the art will appreciate that the described aspects may be practiced without these specific details.

[0057] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting buildings from an InSAR terrain product, comprising the following steps:

[0058] S1, data preparation and preprocessing.

[0059] In this embodiment, the data is raw data and terrain information corresponding to the study area. The raw data is multi-viewed to obtain a SAR intensity image; the SAR intensity image is geocoded to convert it to the real terrain; and the SAR intensity image is filtered.

[0060] It should be noted that the acquired SAR data resolution ranges from 2 to 20 meters, which can be adjusted and input based on actual conditions and mission requirements. Data preprocessing includes multi-look processing of the original SLC (single-look complex) imagery to reduce speckle noise. Appropriate filtering can be selected based on the actual speckle noise level of the SAR imagery. Topographic information for the study area, such as slope and aspect, is calculated using external DEM data to provide data support for subsequent processing.

[0061] S2, select appropriate samples as training sets, calculate features of the samples, calculate feature discriminability through Bhattacharyya distance, and apply principal component analysis for feature fusion.

[0062] In this embodiment, the training samples are all single-polarization SAR image data. Feature calculation is performed on the SAR intensity image, and the Bhattacharyya distance is used to evaluate the discrimination ability of the calculated multiple features. The three features with the strongest discrimination ability are selected, and the principal component analysis method is used to eliminate the coherence between the features. The final processed features are subjected to feature fusion to form a new feature image as the input of the neural network.

[0063] It should be noted that the selection of training samples should be appropriately changed according to the neural network used for training. In order to ensure the training effect of the neural network and the accuracy of preliminary building extraction, the number of samples should preferably be greater than 500.

[0064] In this embodiment, the feature calculation uses the Gray Level Co-occurrence Matrix (GLCM) to calculate the joint probability distribution of the positions of two pixels. Let S be the set of pixel pairs with specific spatial connections in the target area R, then the co-occurrence matrix P can be defined as:

[0065] P(i,j)=[p(i,j,d,θ)] (1)

[0066] This formula can be described as: the frequency of simultaneous occurrence of pixel pairs with grayscales i and j (i, j = 0, 1, 2, 3, ..., N) in the image, satisfying the conditions of step size d, direction θ. θ is the angle of the vector connecting the two pixels, and texture features are usually obtained along four angles: 0°, 45°, 90°, and 135°.

[0067] Further texture features are extracted from it, and variance, contrast, and mean are used as features. The calculation formulas are as follows:

[0068]

[0069]

[0070]

[0071] In S3, the samples after feature fusion are input into the neural network for continuous extraction and training of pooled convolution features, and the appropriate loss function and optimizer are selected. The neural network structure and neural network parameters are modified according to the training effect until the neural network loss function reaches convergence.

[0072] In this embodiment, the selected neural network is the U-Net network, and the type of optimizer used for training is the stochastic gradient descent optimizer (SGD). A large number of studies have shown that the SGD optimizer can maintain high stability when applying large data sets.

[0073] It should be noted that the training samples ultimately fed into the neural network consist of a fused SAR intensity image feature and a building label image. The images used for training are segmented and cropped small scene images, with some overlap. The neural network used for training is a U-Net VGG series network, whose backbone feature extraction consists of convolution and max pooling. By continuously sampling and reading features from multiple dimensions of SAR images, it achieves multi-level feature recognition, effectively improving the accuracy of automated building extraction. By selecting an appropriate loss function and learning optimizer for evaluation, and adjusting the neural network structure and parameters, excellent training results were achieved.

[0074] S4, performs feature calculation and feature fusion on the SAR intensity image, and uses the trained neural network to extract features.

[0075] Specifically, this step filters the SAR intensity image of the study area, calculates its features, and fuses them using the features selected by S2. Simultaneously, Matlab is used to crop the large scene image after feature fusion into small scenes for neural network recognition. The segmented small scene image data is then fed into the neural network to meet the network input parameter size requirements. The trained neural network is then used to extract features from the batch of images, yielding preliminary building extraction results s.

[0076] It should be noted that the SAR satellite imaging mode and polarization mode selected in this step should be consistent with the training samples as much as possible.

[0077] S5, stitching the small area images of the building extraction result set s to obtain the preliminary building extraction result S under the real terrain conditions.

[0078] It should be noted that in this embodiment, the neural network input and output are both small scene images, which cannot represent the building extraction under real terrain conditions, let alone perform accuracy assessment. Therefore, it is necessary to splice the neural network output results and fuse them at the edges to obtain the preliminary building extraction result S under real terrain conditions. Specifically, Matlab is used to splice the small scene image results to restore them to the extraction results under real terrain conditions.

[0079] S6, obtain satellite imaging parameters, use the terrain information of the corresponding study area, reconstruct the resolution of the DEM data to make it the same as the resolution of the SAR intensity image; calculate the slope and aspect corresponding to each pixel, analyze the relationship and difference between the imaging angle and the slope and aspect, and identify shadows and overlapping areas with textures similar to buildings in the study area.

[0080] In this step, the terrain information of the corresponding study area should be preprocessed first so that its resolution and pixel size are completely consistent with the SAR image. The resampling operation of the DEM data will also affect the subsequent building extraction accuracy to a certain extent. Therefore, selecting a DEM with higher accuracy and performing appropriate resampling operations are the key in this step.

[0081] It's important to note that because SAR satellites use side-on imaging, their line of sight is at an angle to the vertical. Uneven terrain can prevent certain parts of the surface from being illuminated by the satellite, reducing the interpretability of the SAR image. Furthermore, mountainous areas with a certain slope facing the satellite appear as bright areas in SAR images, exhibiting texture characteristics similar to buildings. These areas cannot be distinguished using SAR imagery alone, so external DEM data is needed to identify and remove them.

[0082] The main principle of this step is to establish a functional relationship between DEM and geometric distortion through external reference DEM data and satellite imaging parameters to identify geometric distortion areas. First, the slope, aspect and elevation of each pixel are calculated. Based on the relationship between them and the satellite incidence angle and azimuth parameters, it is determined whether it is a shadow overlap area. The judgment of shadow overlap areas also needs to be determined based on the satellite flight mode.

[0083] For satellites in orbit raising mode:

[0084] Judgment 1:

[0085]

[0086] Where aspect(i,j) is the aspect angle at pixel (i,j), measured clockwise from true north; slope(i,j) is the slope angle at pixel (i,j); azi_angle(i,j) is the SAR satellite's flight direction at the pixel, measured clockwise from true north; and inc_angle(i,j) is the SAR satellite's angle of incidence at the pixel. It should be noted that the angle of incidence and azimuth are identical for each pixel in an image, as specified in the satellite imaging parameter file.

[0087] For this judgment, if the mutual relationships satisfy the above three equations at the same time, it can be considered that the pixel point belongs to the distorted area.

[0088] Judgment 2:

[0089]

[0090] slope(i,j)>inc_angle(i,j) (7)

[0091] For the second judgment, if the relationship between its parameters satisfies any one or two of the equations (6) and (7), the pixel point can be considered to belong to the distorted area.

[0092] For satellites in descending orbit mode:

[0093] Judgment 1:

[0094]

[0095] For this judgment, if the above three equations are satisfied at the same time, the pixel point can be considered to belong to the distorted area.

[0096] Judgment 2:

[0097]

[0098] slope(i,j)>(90-inc angle(i,j) ) (10)

[0099] For the second judgment, if the relationship between its parameters satisfies any one or two of the equations (9) and (10), the pixel point can be considered to belong to the distorted area.

[0100] S7, remove the shadows and overlapping results identified in S6 from the preliminary building extraction result S, and filter out some misidentified areas by setting the slope threshold, and perform appropriate classification post-processing to obtain the final building extraction result T.

[0101] Therefore, S7 combines S5 and S6 to filter out the final building extraction results, effectively improving the building extraction accuracy in the mountainous area.

[0102] The method provided in this embodiment combines neural networks and external terrain factors. Based on SAR intensity images, it considers the advantages of traditional feature calculation methods in representing SAR texture features. Based on the efficient recognition ability of neural networks for multi-level and multi-directional features, it introduces external reference DEM data to assist in eliminating shadows and overlapping areas with similar building textures. By combining image texture features and considering external terrain factors to identify and extract buildings, it can effectively improve the accuracy of building extraction and make the building extraction process basically automated.

[0103] from Figure 1 As can be seen in the figure, the introduction of external reference DEM data, the identification of shadows and overlapping areas, and the screening of slopes all fully consider the impact of external terrain on building distribution. Combining feature calculation methods with neural network feature extraction capabilities, it is possible to extract buildings in similar areas such as mountainous areas, making full use of prior information and effectively improving the accuracy of building extraction in these areas. This opens the possibility of further achieving high-precision building extraction across all terrains.

[0104] The following is a specific case to further illustrate and demonstrate this method.

[0105] The training sample set is selected. The training sample is the HH band SAR image of full polarization data taken by the Gaofen-3 satellite (GF-3). The data resolution is 10m. The size of each training sample is 256×256 pixels, and there are 25,000 SAR image data in total. The training data set and the corresponding building label part are shown in the figure below. Figure 2 As shown, (a) is the original intensity image, (b) is the image after feature fusion, and (c) is the building label.

[0106] The data for the study area comes from the German Aerospace Center's TerraSAR-X / Tandem-X dual-satellite system. This satellite system has an absolute elevation accuracy of better than 10 meters. The data selected is from Macau, China, covering the area from 22°03′26″N to 22°25′51″N and 113°20′32″E to 113°32′59″E. The landforms are composed of low hills and flat land, with high terrain in the south and low terrain in the north. The data is shown in the figure below. Figure 3 As shown, (a) is the intensity image and (b) is the corresponding optical image.

[0107] Figure 4 The figure shows the reflection of SAR echo signal in the face of very steep terrain. In this case, Figure 4In (a), the terrain segment EFG cannot receive the signal sent by the radar, resulting in point EG not being imaged. This is the reason for the existence of the shadow area. The slope EF is defined as "active shadow", while the part of FG that is located on the slope EF and cannot reach the radar beam because it is blocked by the slope EF is called "passive shadow". Figure 4 As can be seen in (b), the imaging order of slope BC is reversed and superimposed on the images of slices AB and CD. This is the overlap region, where slope BC is also called "active overlap," while the AB and CD segments at either end of slope BC are "near overlap" and "far overlap," respectively.

[0108] Figure 5 The comparison pictures before and after the shadow and overlapping area removal are shown, where (a) is before the shadow and overlapping area removal, and (b) is after the shadow and overlapping area removal.

[0109] In this embodiment, the building accuracy assessment uses internal and external accuracy to evaluate. The internal accuracy is calculated by statistically analyzing the standard deviation within the class, which can be used to determine the difference within the class. The smaller the difference within the class, the higher the classification accuracy. The specific calculation formula is as follows:

[0110]

[0111]

[0112] Among them, x i is the gray value of the i-th pixel belonging to a certain category, is the average grayscale value of pixels belonging to this category, and σ is the internal accuracy.

[0113] The external accuracy is measured using overall accuracy and the Kappa coefficient. For the building category, the 2019 global built-up area mask file, generated from Sentinel data and published in 2022 by the German Space Agency (DLR), was used as the actual ground truth for comparison, as shown in Table 1.

[0114] Table 1 Classification results

[0115]

[0116] Strictly speaking, the overall accuracy OA is defined as the ratio of the number of correctly extracted objects to the total number of objects. The specific calculation formula is as follows:

[0117]

[0118] Among them, TP, TN, FP, and FN are the numbers of pixels that meet the classification requirements. The Kappa coefficient K is a ratio that represents the reduction in errors between classification and completely random classification. Its value usually falls between 0 and 1. Its numerical range is as follows: 0.0-0.20 very low consistency; 0.21-0.40 general consistency; 0.41-0.60 moderate consistency; 0.61-0.80 high consistency; 0.81-1.0 almost perfect consistency. The specific calculation formula is as follows:

[0119]

[0120]

[0121] Figure 6 The following table shows the building extraction results, where (a) is the standard building mask result and (b) is the experimental result of the present invention. The accuracy of building extraction using the above scheme is evaluated as shown in the following table. The accuracy of the building classification within the building category is 63.576, and the accuracy of the building classification outside the building category is 0.92 and 0.72 respectively. The accuracy evaluation shows that the extracted building category is highly consistent with the true result, as shown in Table 2.

[0122] Table 2 Accuracy assessment

[0123]

[0124] In summary, this embodiment provides a building extraction method suitable for difficult areas such as mountains. Based on the texture features of SAR images, buildings are extracted with the assistance of external terrain information. The accuracy of the obtained building extraction results is greatly improved. This method fully utilizes the image information and terrain information in the study area to identify built-up areas. It not only shows good results in difficult areas such as mountains, but also has high accuracy in flat areas. This embodiment proposes a special terrain building extraction method suitable for all SAR data. This method does not rely on the original SAR data, which greatly improves the universality of this method and can be used to extract buildings from any type of SAR data. SAR images are used as the main method, supplemented by terrain information, and the two are combined to extract buildings.

[0125] Based on the same inventive concept, this embodiment further provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the aforementioned method for extracting buildings from InSAR terrain products is implemented.

[0126] Computer-readable media include, but are not limited to, any type of disk (including floppy disks, hard disks, optical disks, CD-ROMs, and magneto-optical disks), ROMs, RAMs, EPROMs (Erasable Programmable Read-Only Memory), EEPROMs, flash memories, magnetic cards, or optical cards. In other words, computer-readable media include any medium that can store or transmit information in a form that can be read by a device (e.g., a computer).

[0127] The computer-readable storage medium provided in this embodiment has the same inventive concept and the same beneficial effects as the aforementioned method, and will not be described in detail here.

[0128] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for extracting buildings from InSAR terrain products, characterized in that: The steps include: S1, data preparation and preprocessing, the data are original data and the terrain information of the corresponding study area; S2, select samples as training sets, calculate features of samples, calculate feature discrimination through Bhattacharyya distance, and apply principal component analysis for feature fusion; S3: Input the feature-fused samples into the neural network to extract and train the pooled convolution features, select the loss function and optimizer, and modify the neural network structure and parameters according to the training results until the neural network loss function reaches convergence; S4, performs feature calculation and feature fusion, and uses the trained neural network to perform feature extraction; S5, stitching the small area images of the building extraction result set to obtain the preliminary building extraction results under the real terrain conditions; S6, obtain satellite imaging parameters and reconstruct the topographic information of the study area; Calculate the slope and aspect corresponding to each pixel, analyze the relationship and difference between imaging angle and slope and aspect, and identify shadows and overlapping areas with similar textures to buildings in the study area; S7, removing the identified shadows and overlapping results from the preliminary building extraction results, and filtering out some misidentified areas by setting a slope threshold, and obtaining the final building extraction results after classification and processing; In S1, the original data is multi-viewed to obtain a SAR intensity image; the SAR intensity image is geocoded and converted to the real terrain; the SAR intensity image is filtered; the slope and aspect of the corresponding study area terrain information is calculated using external DEM data; The training samples in S2 are all single-polarization SAR image data. The features of the SAR intensity image are calculated, and the Bhattacharyya distance is used to evaluate the discrimination ability of the calculated multi-category features. The three categories of features with the strongest discrimination ability are selected, and the principal component analysis method is used to eliminate the coherence between the features. The final processed features are fused to form a new feature image as the input of the neural network.

2. The method for extracting buildings from InSAR terrain products according to claim 1, characterized in that: The feature calculation in S2 uses the gray level co-occurrence matrix to count the joint probability distribution of the two pixel positions; let S be the set of pixel pairs with specific spatial connections in the target area R, then the co-occurrence matrix P is defined as: P(i,j)=[p(i,j,d,θ)] (1) That is, the frequency of simultaneous occurrence of pixel pairs with a step size of d, a direction of θ, and grayscales i and j (i, j = 0, 1, 2, 3, ..., N) on the image; where θ is the angle of the line connecting the two pixels; Texture features are further extracted from it, and variance, contrast, and mean are used as features. The calculation formula is:

3. The method for extracting buildings from InSAR terrain products according to claim 2, characterized in that: The neural network selected in S3 is the U-Net network, and the optimizer used for training is the stochastic gradient descent optimizer; the training samples input into the neural network include SAR intensity image feature fusion images and building label images.

4. The method for extracting buildings from InSAR terrain products according to claim 1, characterized in that: In S4, the SAR intensity image of the study area is filtered, the features of the SAR intensity image are calculated, and the features selected by S2 are used for feature fusion. At the same time, the large scene image after feature fusion is cropped into small scene images for neural network recognition, and the segmented small scene image data is input into the neural network to meet the network input parameter size requirements. The trained neural network is used to extract features from batch images to obtain preliminary building extraction results s.

5. The method for extracting buildings from InSAR terrain products according to claim 4, characterized in that: In S6, the acquired terrain information of the corresponding study area is preprocessed to make its resolution and pixel size consistent with the SAR intensity image, and the external DEM data is resampled.

6. The method for extracting buildings from InSAR terrain products according to claim 5, characterized in that: In S6, a functional relationship between DEM and geometric distortion is established using external DEM data and satellite imaging parameters to identify geometric distortion areas. Slope, aspect, and elevation are calculated pixel by pixel, and based on their relationship with satellite incidence angle and azimuth parameters, shadow overlap is determined. This is also determined based on satellite flight patterns. For satellites in orbit raising mode: Judgment 1: Where aspect(i,j) is the aspect angle of pixel point (i,j), which is the angle rotated clockwise with the north direction as the starting axis; slope(i,j) is the slope angle of pixel point (i,j); azi_angle(i,j) is the flight azimuth of the SAR satellite at the pixel point, which is the angle rotated clockwise with the north direction as the starting axis; inc_angle(i,j) is the incident angle of the SAR satellite at the pixel point; For this judgment, if the relationship between its parameters satisfies all the equations in (5) at the same time, the pixel is considered to belong to the distortion area; Judgment 2: slope(i,j)>inc_angle(i,j) (7) For this judgment, if the relationship between its parameters satisfies any one or two of the equations (6) and (7), the pixel is considered to belong to the distortion area. For satellites in descending orbit mode: Judgment 1: For this judgment, if the relationship between its parameters satisfies all the equations in (8) at the same time, the pixel is considered to belong to the distortion area; Judgment 2: slope(i,j)>(90-inc angle(i,j) ) (10) For this judgment, if the relationship between its parameters satisfies any one or two of the equations (9) and (10), the pixel is considered to belong to the distorted area.

Images