Complex terrain SAR object classification method, system, storage medium and electronic equipment based on data compensation

Abstract

The present invention discloses a complex terrain SAR feature classification method based on data compensation. First, the SAR image is preprocessed to obtain backscattered images of different polarization modes. Then, the local incident angle is obtained and its mathematical model is analyzed to determine the overlapping area and the shadow area and compensate for the overlapping area. Then, the compensated descending orbit image is polarized and decomposed to obtain polarization parameters. Then, the relevant polarization features and dual-polarization radar vegetation index of the compensated descending orbit image are extracted, and the polarization feature data and the dual-polarization radar vegetation index are feature fused. Then, a 2D-CNN feature recognition model is constructed and trained and tested using a data set. Finally, the trained model is used to classify the compensated image and the result is output. The method of the present invention can effectively solve the technical problem of image geometric distortion when SAR sensors image complex terrain such as mountains with undulating elevation.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a complex terrain SAR feature classification method, system, storage medium and electronic equipment based on data compensation. Background Art

[0002] Synthetic Aperture Radar (SAR) is an effective remote sensing data source for geographic surveys and topographic mapping. It offers all-day, all-weather observation capabilities and is not susceptible to cloud or fog interference. SAR image object classification is a key research topic within remote sensing image object classification. SAR image object classification is of great significance in areas such as water change detection, soil moisture, and urban transformation, and has garnered widespread attention.

[0003] Due to the side-viewing and range imaging characteristics of SAR sensors, imaging complex terrain, such as mountains with undulating elevations, can result in geometric distortion of the image, such as overlap and shadows. This can cause the scattering characteristics of the image to be inconsistent with the actual scene. Geometrically distorted areas can lead to misclassification of objects due to erroneous scattering characteristics. For example, overlap can cause the backscatter coefficient to be similar to that of built-up areas, making it impossible to perceive the true object type in the overlapped area. Most current classification methods only target flat areas and do not consider the impact of complex terrain. Regarding image feature extraction, features extracted from dual-polarization data essentially only contain polarization features, and in particular, lack features representing vegetation cover. Therefore, how to eliminate the impact of geometrically distorted areas and select effective feature fusion to meet the requirements for object recognition in SAR images with complex terrain and a rich variety of object types is an urgent problem that needs to be solved to achieve complete and accurate object classification in complex terrain areas. Summary of the Invention

[0004] The purpose of the present invention is to provide a complex terrain SAR feature classification method, system, storage medium and electronic equipment based on data compensation, so as to solve the technical problem of image geometric distortion when SAR sensors image complex terrain such as mountains with undulating elevation.

[0005] The technical solutions of the present invention are as follows:

[0006] The present invention discloses a complex terrain SAR object classification method based on data compensation, comprising the following steps:

[0007] Step S101: preprocessing the SAR image to obtain backscattered images of different polarization modes;

[0008] Step S102: Analyze the geometric model of radar satellite imaging to obtain the local incident angle;

[0009] Step S103: Analyze the mathematical model of the local incident angle to determine the overlapped area and the shadow area;

[0010] Step S104: using the ascending track data to compensate for the overlap and shadow areas in the descending track data;

[0011] Step S105: performing polarization matrix conversion on the compensated down-orbit image, and then performing polarization decomposition to obtain polarization parameters;

[0012] Step S106: extracting the relevant polarization features and the dual-polarization radar vegetation index of the compensated descending orbit image, and performing feature fusion on the polarization feature data and the dual-polarization radar vegetation index;

[0013] Step S107: constructing a data set, which is a training set and a validation set;

[0014] Step S108: constructing a ground feature recognition model and using the data set for training and testing;

[0015] Step S109: Use the trained model to classify the compensated image and output the result.

[0016] Preferably, in step S101, the specific method for preprocessing the SAR image is:

[0017] (1) Perform complex data conversion on the SAR ascending and descending images, converting the complex data into amplitude data;

[0018] (2) Multi-view and filtering processing is performed on the amplitude data to optimize the geometric features and achieve the purpose of noise reduction;

[0019] (3) geocoding and terrain correction processing of the image data obtained in step (2);

[0020] (4) Calculate the backscatter value of the ground object based on the image obtained in step (3) to obtain backscatter images of different polarization modes.

[0021] Preferably, in step S102, the method for analyzing the geometric model of radar satellite imaging to obtain the local incident angle is:

[0022] The local incident angle θ at that moment is obtained according to the geometric model when the radar collects data. Its expression is:

[0023] θ=β-α

[0024] Among them, β is the side viewing angle of the radar wave; α is the slope angle of the target point.

[0025] Preferably, in step S103, the mathematical model of the local incident angle is analyzed to determine the overlap and shadow areas by:

[0026] When the radar incident angle is less than 0 degrees, the target area is inverted top and bottom, and is an overlapping area; when the local incident angle is greater than 90 degrees, the target area has a signal loss, and is a shadow area.

[0027] Preferably, in step S105, polarization matrix conversion is performed on the compensated down-orbit image, and the generated polarization matrix C2 is expressed as:

[0028]

[0029] Where * represents conjugate transpose, c 11 、c 22 are all real numbers, c 12 is plural;

[0030] The polarization decomposition of the polarization matrix is ​​Cloude polarization decomposition; the obtained polarization parameters include eigenvalues ​​and eigenvectors.

[0031] Preferably, in step S106, the extracted polarization features include scattered power (SPAN), differential intensity (DI), and power ratio (PR); the expressions are as follows:

[0032]

[0033] Among them, |S HH | 2 and |S HV | 2 is the intensity of HH and HV polarizations. dB is the logarithm to the base 10 or log10x of a given number x;

[0034] The expression of the extracted dual-polarization radar vegetation index (DPRVI) is:

[0035] DPRVI=1-DoP*P1

[0036] Where DoP is the degree of polarization and P1 is the normalization parameter.

[0037] Preferably, in step S107, the method for constructing the data set is:

[0038] Google Earth was used to mark the four types of land features in some areas, including farmland, water, woodland and buildings, and the images were divided into training and test sets.

[0039] The present invention also discloses a complex terrain SAR object classification system based on data compensation, comprising:

[0040] A pre-processing unit is configured to perform complex data conversion, multi-view processing, filtering, geocoding and terrain correction on the SAR image, and calculate backscatter values ​​to obtain backscatter images of different polarization modes;

[0041] an overlapping shadow extraction unit configured to extract a local incident angle in the descending track image and determine an overlapping shadow area according to different local incident angle ranges of the overlapping shadow;

[0042] An overlapping shadow compensation unit is configured to compensate the overlapping shadow area in the descending track image using the descending track image, thereby greatly eliminating the geometric distortion phenomenon in the descending track image;

[0043] a feature extraction and fusion unit configured to extract polarization features and dual-polarization features of the compensated image and perform feature fusion;

[0044] Construct a dataset unit, configure it to label the four types of land features in a certain area, namely farmland, water, woodland, and buildings, and divide it into a training image set and a test image set;

[0045] A model training unit is configured to use the fused features to train the model and obtain optimal model parameters;

[0046] The result output unit is configured to use the trained model to classify the compensated image, obtain the classification result, and save the output.

[0047] The present invention also discloses a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes any of the complex terrain SAR feature classification methods based on data compensation.

[0048] The present invention also discloses an electronic device, comprising: a memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the complex terrain SAR feature classification method based on data compensation is implemented.

[0049] The beneficial effects of the present invention are:

[0050] The method of the present invention compensates for the misclassification phenomenon caused by overlapping shadows through ascending and descending orbit compensation. On the basis of dual-orbit polarization characteristics, it incorporates the dual-polarization radar vegetation index to improve the identification ability of farmland and woodland. The method of the present invention can effectively solve the technical problem of image geometric distortion when SAR sensors image complex terrain such as mountains with undulating elevation. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 is a flow chart of the method of the present invention;

[0052] Figure 2 It is the local area captured from the GF-3 input image of Danjiangkou Reservoir;

[0053] Figure 3 For the present invention Figure 2 The results of feature classification in ;

[0054] Figure 4 The complete area of ​​the GF-3 input image for the Danjiangkou Reservoir;

[0055] Figure 5 For the present invention Figure 4 The results of feature classification in ;

[0056] Figure 6 Schematic diagram of complex terrain SAR object classification system based on data compensation. DETAILED DESCRIPTION

[0057] like Figure 1 As shown, a complex terrain SAR object classification method based on data compensation of the present invention includes the following steps:

[0058] Step S101: preprocessing the SAR image to obtain backscattered images of different polarization modes; specifically, the following steps are included:

[0059] (1) Perform complex data conversion on the SAR ascending and descending images, converting the complex data into amplitude data;

[0060] (2) Multi-view and filtering processing is performed on the amplitude data to optimize the geometric features and achieve the purpose of noise reduction;

[0061] Multi-view processing and filtering can optimize geometric features and reduce noise to a certain extent, making it easier for subsequent processing.

[0062] (3) geocoding and terrain correction processing of the image data obtained in step (2);

[0063] Image correction is the premise of further processing. Geocoding uses a geometric correction method based on the RD positioning model and uses DEM data for geocoding terrain correction (GTC).

[0064] (4) Calculate the backscatter value of the ground object based on the image obtained in step (3) to obtain backscatter images of different polarization modes.

[0065] Step S102: Analyze the geometric model of radar satellite imaging to obtain the local incident angle;

[0066] The local incident angle θ at that moment is obtained according to the geometric model when the radar collects data. Its expression is:

[0067] θ=β-α (1)

[0068] Among them, β is the side viewing angle of the radar wave; α is the slope angle of the target point.

[0069] Step S103: Analyze the mathematical model of the local incident angle to determine the overlapped area and the shadow area; the specific method is:

[0070] When the radar incident angle is less than 0 degrees, the target area is inverted top and bottom, and is an overlapping area; when the local incident angle is greater than 90 degrees, the target area has a signal loss, and is a shadow area.

[0071] Step S104: using the ascending track data to compensate for the overlap and shadow areas in the descending track data;

[0072] Step S105: performing polarization matrix conversion on the compensated down-orbit image, and then performing polarization decomposition to obtain polarization parameters;

[0073] The polarization matrix C2 is generated by performing polarization matrix conversion on the compensated down-orbit image:

[0074]

[0075] Where * represents conjugate transpose, c 11 、c 22 are all real numbers, c 12 is a complex number; the polarization decomposition of the polarization matrix is ​​Cloude polarization decomposition; the analysis is mainly based on the eigenvalues ​​and eigenvectors of C2. The following physical quantities related to eigenvalues ​​and eigenvectors are defined:

[0076] Scattering entropy H (Entropy) indicates the degree of randomness of target scattering:

[0077]

[0078] Among them, P i is the eigenvalue normalization parameter; i is the i-th eigenvalue of the polarization matrix C2.

[0079] The average scattering angle α (Alpha) represents the average scattering mechanism of the target scattering mechanism from surface scattering to volume scattering to dihedral scattering:

[0080] α=P1α1+P2α2 (5)

[0081] Among them, P1 and P2 are eigenvalue normalization parameters; α1 and α2 are the internal degrees of freedom of the scatterer, ranging from 0° to 90°.

[0082] Anti-entropy A (Anisotropy), which indicates the degree of anisotropy of target scattering:

[0083]

[0084] Among them, λ1 is the first eigenvalue; λ2 is the second eigenvalue.

[0085] Step S106: extracting the relevant polarization features and the dual-polarization radar vegetation index of the compensated descending orbit image, and performing feature fusion on the polarization feature data and the dual-polarization radar vegetation index;

[0086] The polarization features extracted in this step include scattered power (SPAN), differential intensity (DI), and power ratio (PR); the expressions are as follows:

[0087]

[0088] Among them, |S HH | 2 and |S HV | 2 is the intensity of HH and HV polarizations. dB (decibel) is the base 10 logarithm of a given number x; the expression for the extracted Dual Polarization Radar Vegetation Index (DPRVI) is:

[0089] DPRVI=1-DoP*P1

[0090] Where DoP is the degree of polarization and P1 is the normalization parameter.

[0091] Step S107: construct a data set, which includes a training set and a validation set. The method for constructing the data set is:

[0092] Use Google Earth to mark four types of land features in some areas: farmland, water, woodland and buildings, and divide the training image set and the test image set into an appropriate ratio of 8:2.

[0093] Step S108: constructing a ground feature recognition model and using the data set for training and testing;

[0094] The 2D-CNN model can be used. The model mainly consists of four parts: input layer, convolution layer, maximum pooling normalization layer and batch normalization layer, and fully connected layer. The input layer inputs the data set. The convolution layer is used to extract 12 layers of convolution kernels to extract input layer features. Parameters can be selected by applying strides, pooling, and filters. Convolution is performed on the image and the ReLU activation function is applied to the matrix. The maximum pooling normalization layer and batch normalization layer filter the features extracted by the convolution layer. The pooling layer is followed by the fully connected layer, which converts all the feature matrices of the pooling layer into a large -dimensional feature vector. The fully connected layer is generally placed at the end of the network structure and is used to classify the training set and validation set.

[0095] Set bathch_size to 32 and the maximum number of iterations to 300. Then input the training dataset and label set into the constructed network model to train the network. Every 60 generations, use the validation dataset to test the network performance and save the model parameters obtained from the current training.

[0096] Step S109: Use the trained model to classify the compensated image and output the result.

[0097] The local area and the complete study area intercepted from the GF-3 input image of Danjiangkou Reservoir are shown in Figure 2. Figure 2 and Figure 4 As shown, the results of the object classification of the local area and the whole area of ​​the GF-3 input image of Danjiangkou Reservoir using the method of the present invention are respectively as follows: Figure 3 and Figure 5 As shown by Figure 3 and Figure 5 It can be seen that the present invention has a good effect on classifying objects in complex scenes with high accuracy, and can largely avoid the influence of complex terrain; and because the present invention has high classification accuracy and high processing efficiency, it also has good practical value.

[0098] like Figure 6 As shown, the present invention provides a complex terrain SAR object classification system based on data compensation, comprising:

[0099] A pre-processing unit is configured to perform complex data conversion, multi-view processing, filtering, geocoding and terrain correction on the SAR image, and calculate backscatter values ​​to obtain backscatter images of different polarization modes;

[0100] an overlapping shadow extraction unit configured to extract a local incident angle in the descending track image and determine an overlapping shadow area according to different local incident angle ranges of the overlapping shadow;

[0101] An overlapping shadow compensation unit is configured to compensate the overlapping shadow area in the descending track image using the descending track image, thereby greatly eliminating the geometric distortion phenomenon in the descending track image;

[0102] a feature extraction and fusion unit configured to extract polarization features and dual-polarization features of the compensated image and perform feature fusion;

[0103] Construct a dataset unit, configure it to label the four types of land features in a certain area, namely farmland, water, woodland, and buildings, and divide it into a training image set and a test image set;

[0104] A model training unit is configured to use the fused features to train the model and obtain optimal model parameters;

[0105] The result output unit is configured to use the trained model to classify the compensated image, obtain the classification result, and save the output.

[0106] A computer-readable storage medium of the present invention stores a computer program. When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the complex terrain SAR feature classification method based on data compensation of the present invention.

[0107] An electronic device of the present invention comprises: a memory and a processor, wherein the memory stores a program that can be run on the processor, and the processor implements the complex terrain SAR feature classification method based on data compensation of the present invention when executing the program.

[0108] Finally, it should be noted that the above description is merely a preferred embodiment of the present application and is not intended to limit the present application. Those skilled in the art will readily appreciate that various modifications and variations are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A complex terrain SAR feature classification method based on data compensation, characterized in that: The following steps are involved: Step S101: preprocessing the SAR image to obtain backscattered images of different polarization modes; Step S102: Analyze the geometric model of radar satellite imaging to obtain the local incident angle; Step S103: Analyze the mathematical model of the local incident angle to determine the overlapped area and the shadow area; Step S104: using the ascending track data to compensate for the overlap and shadow areas in the descending track data; Step S105: performing polarization matrix conversion on the compensated down-orbit image, and then performing polarization decomposition to obtain polarization parameters; Step S106: extracting the relevant polarization features and the dual-polarization radar vegetation index of the compensated descending orbit image, and performing feature fusion on the polarization feature data and the dual-polarization radar vegetation index; Step S107: constructing a data set, which is a training set and a validation set; Step S108: constructing a ground feature recognition model and using the data set for training and testing; Step S109: Use the trained model to classify the compensated image and output the result.

2. The complex terrain SAR feature classification method based on data compensation according to claim 1 is characterized in that: In step S101, the specific method for preprocessing the SAR image is: (1) Perform complex data conversion on the SAR ascending and descending images, converting the complex data into amplitude data; (2) Multi-view and filtering processing is performed on the amplitude data to optimize the geometric features and achieve the purpose of noise reduction; (3) geocoding and terrain correction processing of the image data obtained in step (2); (4) Calculate the backscatter value of the ground object using the image obtained in step (3) to obtain backscatter images of different polarization modes.

3. The complex terrain SAR feature classification method based on data compensation according to claim 1, characterized in that: In step S102, the geometric model of radar satellite imaging is analyzed to obtain the local angle of incidence by: obtaining the local angle of incidence θ at that moment based on the geometric model when the radar collects data. The expression is: θ=β-α Among them, β is the side viewing angle of the radar wave; α is the slope angle of the target point.

4. The complex terrain SAR feature classification method based on data compensation according to claim 1, characterized in that: In step S103, a mathematical model of the local incident angle is analyzed to determine the overlap and shadow areas. When the radar incident angle is less than 0 degrees, the target area is inverted top and bottom, and is an overlapping area; when the local incident angle is greater than 90 degrees, the target area has a signal loss, and is a shadow area.

5. The complex terrain SAR feature classification method based on data compensation according to claim 1, characterized in that: In step S105, polarization matrix conversion is performed on the compensated down-orbit image, and the generated polarization matrix C2 is expressed as: Where * represents conjugate transpose, c 11 、c 22 are all real numbers, c 12 is a complex number; the polarization decomposition of the polarization matrix is ​​Cloude polarization decomposition; the obtained polarization parameters include eigenvalues ​​and eigenvectors.

6. The complex terrain SAR feature classification method based on data compensation according to claim 1, characterized in that: In step S106, the extracted polarization features include scattered power (SPAN), differential intensity (DI), and power ratio (PR); the expressions are as follows: Among them, |S HH | 2 and |S HV | 2 is the intensity of HH and HV polarization in dB; the expression of the extracted dual-polarization radar vegetation index (DPRVI) is: DPRVI=1-DoP*P1 Where DoP is the degree of polarization and P1 is the normalization parameter.

7. The complex terrain SAR feature classification method based on data compensation according to claim 1, characterized in that: In step S107, the method for constructing the data set is: Google Earth was used to mark the four types of land features in some areas, including farmland, water, woodland and buildings, and the images were divided into training and test sets.

8. A complex terrain SAR object classification system based on data compensation, characterized by: include: A pre-processing unit is configured to perform complex data conversion, multi-view processing, filtering, geocoding and terrain correction on the SAR image, and calculate backscatter values ​​to obtain backscatter images of different polarization modes; an overlapping shadow extraction unit configured to extract a local incident angle in the descending track image and determine an overlapping shadow area according to different local incident angle ranges of the overlapping shadow; An overlapping shadow compensation unit is configured to compensate the overlapping shadow area in the descending track image using the descending track image, thereby greatly eliminating the geometric distortion phenomenon in the descending track image; a feature extraction and fusion unit configured to extract polarization features and dual-polarization features of the compensated image and perform feature fusion; Construct a dataset unit, configure it to label the four types of land features in a certain area, namely farmland, water, woodland, and buildings, and divide it into a training image set and a test image set; A model training unit is configured to use the fused features to train the model and obtain optimal model parameters; The result output unit is configured to use the trained model to classify the compensated image, obtain the classification result, and save the output.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the device where the computer-readable storage medium is located executes the complex terrain SAR feature classification method based on data compensation according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a program that can be run on the processor, and when the processor executes the program, the complex terrain SAR feature classification method based on data compensation according to any one of claims 1 to 7 is implemented.

Images