A method to remove atmospheric phase from InSAR

Abstract

The present invention relates to a method for removing the atmospheric phase of InSAR, comprising: obtaining a SAR remote sensing image and generating n unwrapped interferometric phase diagrams based on SBAS technology, wherein each unwrapped interferometric phase diagram includes p permanent scatterer points; removing the atmospheric phase from the n unwrapped interferometric phase diagrams using a quadratic polynomial atmospheric phase correction method, and arranging the atmospheric phase removal results in order according to the removal effect, taking m results with poor removal effect, finding the corresponding unwrapped interferometric phase diagrams without removing the atmospheric phase and classifying them as Class A interferometric phase diagrams; removing the atmospheric phase from each interferometric phase diagram in the Class A interferometric phase diagram using BP neural network training. The method provided by the present invention can remove the atmospheric phase faster and better, thereby improving the accuracy of the deformation rate inversion of the InSAR technology, and providing technical support for better monitoring of surface deformation by the InSAR technology.

Description

Technical Field

[0001] The invention belongs to the field of synthetic aperture radar interferometry (InSAR) technology and relates to a method for removing an atmospheric phase in an InSAR interference graph. Background Art

[0002] InSAR technology is a ground observation technology that has developed rapidly in the past three decades. It has strong cloud and fog penetration capabilities and can obtain large-scale surface settlement information all day and all weather. It has been widely used in monitoring ground settlement caused by landslides, earthquakes, volcanoes, urban water loss, etc. InSAR technology uses the phase difference between two SAR images to obtain the deformation information of the target object, and its theoretical accuracy can reach the millimeter level. During the transmission process, the radar signal will be affected by variable conditions such as atmospheric temperature, humidity, air pressure, and free electron density, resulting in radar signal delay (atmospheric delay). The atmospheric phase caused by atmospheric delay directly affects the accuracy of the deformation result, so the atmospheric phase needs to be accurately removed. At present, the commonly used atmospheric phase removal methods mainly include polynomial fitting, meteorological information model, GPS data and water vapor data correction, etc. However, the above methods are not suitable for atmospheric phase removal of all InSAR interferograms, and the above methods have poor atmospheric phase removal effects on some interferograms. Summary of the invention

[0003] The purpose of the present invention is to provide a method for removing the atmospheric phase of an InSAR interferogram. When the conventional atmospheric phase removal method is not effective, the present invention uses a BP neural network to simulate the atmospheric phase of the InSAR interferogram, and then removes the atmospheric phase simulated by the BP neural network. The present invention can better remove the atmospheric phase of the InSAR interferogram, thereby improving the monitoring accuracy of the InSAR technology.

[0004] In order to achieve the above technical objectives, the present invention adopts the following scheme:

[0005] A method for removing an InSAR atmospheric phase, comprising:

[0006] Acquire remote sensing images and obtain n unwrapped interferometric phase maps based on SBAS technology, each of which includes p permanent scatterer points;

[0007] The atmospheric phase of n unwrapped interferometric phase images is removed by using the quadratic polynomial atmospheric phase correction method, and the atmospheric phase removal effects are arranged in order according to the removal effect. The m results with the poor removal effect are taken, and the corresponding unwrapped interferometric phase images without atmospheric phase removal are found and classified as type A interferometric phase images.

[0008] The longitude and latitude of each interference phase diagram in the A-type interference phase diagram are taken as input, and the corresponding phase value is taken as output, and a BP neural network training model including input layer, hidden layer and output layer is established; the interference phase diagram in the A-type interference phase diagram is subtracted from the atmospheric phase obtained by BP neural network training to obtain the interference phase diagram without the atmospheric phase.

[0009] As a preferred implementation, the BP neural network training model is established as follows:

[0010] For each interference phase diagram in the A-type interference phase diagram, the longitude and latitude and phase values ​​of the p permanent scatterer points are converted into a p*3 matrix i; the first column of the matrix i is the longitude of each PS point in the interference phase diagram, the second column is the latitude of each PS point in the interference phase diagram, and the third column is the phase value of each PS point in the interference phase diagram. The first two columns of data in the matrix are normalized to obtain the matrix i2;

[0011] For the first interference phase image in the A-type interference phase image, part of the data of matrix i2 is randomly selected as the training sample of the BP neural network, and the remaining data is used as the test sample. The BP neural network training model is established with the training sample as the input and the corresponding phase value as the output; the trained BP neural network is used to simulate all the data of the matrix i of the first image to obtain the atmospheric phase data of the first interference phase image, and the first interference phase image is subtracted from the obtained atmospheric phase to obtain the interference phase image with the atmospheric phase removed;

[0012] Repeat the steps until all interference phase images in the type A interference phase image are processed, and the InSAR atmospheric phase removal is completed.

[0013] As a preferred implementation, the matrix i1 is obtained by performing equal-interval thinning on the matrix i, and the first two columns of data of the matrix i1 are normalized to [-1, 1] to obtain the matrix i2 for subsequent BP neural network training / testing.

[0014] As a preferred implementation, the preset parameter goal is a performance target for model training, including:

[0015] After the initial training, check whether the training target value is less than goal. If the target value is not less than goal, it is determined that there are abnormal data in the training samples and test samples, and the abnormal data is processed; if the target value is less than goal, use the trained network to simulate the test sample. If the evaluation index between the simulation data simulated by the network and the test sample meets the requirements, the training is completed and the network parameters and structure are saved; otherwise, increase the number of initial hidden layer nodes of the network and retrain using the training samples.

[0016] As a preferred implementation, the abnormal data is processed by: using the current BP neural network to simulate new simulation data using training samples and test data, and calculating the relative error between the simulation data and the data in the third column of the matrix i2;

[0017] When the relative error exceeds 100%, these data are removed as abnormal data points in the sample;

[0018] Increase the number of initial hidden layer nodes of the network, and use the training samples after removing abnormal data for training again. After the training, check whether the training target value is less than the goal. If the target value is less than the goal, use the parameters of this training to simulate the test samples after removing abnormal data. If the evaluation index between the simulation data simulated by the network and the test samples meets the requirements, the training ends and the network parameters and structure are saved; otherwise, increase the number of initial hidden layer nodes, remove abnormal data, and use the training samples after removing abnormal data for training;

[0019] Repeatedly adjust the number of initial hidden layer nodes of the network until the training target value of the BP neural network for the training sample is less than the goal and the evaluation index of the test sample meets the requirements, end the training, and save the network parameters and structure.

[0020] As a preferred implementation, when the number of initial hidden layer nodes of the BP neural network reaches a preset upper limit, and the training target value of the training sample is not less than goal or the evaluation index of the test sample does not meet the requirements, the value of the target value goal is increased, and the number of initial hidden layer nodes of the network is repeatedly adjusted until the training target value of the BP neural network for the training sample is less than the new goal and the evaluation index of the test sample meets the requirements, the training is terminated, and the network parameters and structure are saved.

[0021] As a preferred implementation manner, the evaluation index uses error-related statistical indicators, such as the sum of squared errors, root mean square error, relative error percentage, etc.

[0022] As a preferred implementation, the number of hidden layer nodes of the BP neural network is 2-11.

[0023] As a preferred implementation, the method also includes classifying the remaining nm interference phase maps from which the atmospheric phase has been removed as Class B interference phase maps; after the InSAR atmospheric phase is removed, merging the Class B interference phase map and the Class A interference phase map from which the atmospheric phase has been removed to obtain a new set of interference phase maps, and using the interference phase maps in the new set of interference phase maps to perform deformation rate inversion to obtain a high-precision deformation rate map for the entire study area.

[0024] Conventional methods for removing atmospheric phases in InSAR interferometric phase maps often have poor effects. The present invention provides a method for removing atmospheric phases in InSAR interferometric phase maps by combining quadratic polynomials with BP neural networks. First, the atmospheric phase of n unwrapped interferometric phase maps is quickly removed by using a quadratic polynomial atmospheric phase correction method. Then, the removal results are arranged in order according to the removal effect. Several results with poor removal effects are taken, and the corresponding unwrapped interferometric phase maps without atmospheric phase removal are searched. The longitude and latitude and phase information of the PS point in the unwrapped interferometric phase map are used to construct a BP neural network, and then the constructed BP neural network is used to simulate the atmospheric phase in the interferometric map, thereby removing the atmospheric phase. The method provided by the present invention can remove the atmospheric phase faster and better, thereby improving the accuracy of the deformation rate inversion of the InSAR technology, and providing technical support for better monitoring of surface deformation by the InSAR technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 a is the geographical location map of Zhinan Village, Yangzhong City; Figure 1 b is the amplitude map of the sentinel image covering Zhinan Village, Yangzhong City after cropping and registration (the image was acquired on August 12, 2016).

[0026] Figure 2 The unwrapped interferometric phase map generated by the Sentinel image acquired on August 12, 2016 and the Sentinel image acquired on September 29, 2016, and the unwrapped interferometric phase map after removing the atmospheric phase using a quadratic polynomial.

[0027] Figure 3 In the figure, a is the final training sample; b is the final test sample.

[0028] Figure 4 In the figure, a is the atmospheric phase simulated using the BP neural network training sample; b is the atmospheric phase simulated using the BP neural network test sample.

[0029] Figure 5 a is the original unwrapped interferogram; b is the unwrapped interferogram after removing the atmospheric phase using a quadratic polynomial; c is the atmospheric phase simulated using a BP neural network using the data of all PS points in the unwrapped interferogram; d is the unwrapped interferogram after removing the atmospheric phase in c.

[0030] Figure 6 Among them, a is the statistical histogram of the original unwrapped interferometric phase; b is the statistical histogram of the unwrapped interferometric phase after removing the atmospheric phase using a quadratic polynomial; c is the statistical histogram of the unwrapped interferometric phase after removing the atmospheric phase using the method provided by the present invention. DETAILED DESCRIPTION

[0031] The specific implementation modes of the present invention are described in detail below with reference to the accompanying drawings.

[0032] 1) The implementation case selection covers Zhinan Village, Yangzhong City, Jiangsu Province as the experimental object, and its specific scope is as follows Figure 1 As shown in a, 42 Sentinel-1A images covering the study area were selected as experimental images. The image acquisition time was from November 22, 2015 to October 30, 2017. The 90m resolution SRTM-3 DEM was used as the DEM for registration and removal of terrain phase. MatlabR2020b and Sentinel-1 image processing open source software ctSent_v2 developed based on matlab were used as the image processing software for this experiment; the ctSent_v2 software automatically selected the image acquired on January 27, 2017 as the main image, and used the ctSent_v2 software to crop and register the 42 Sentinel-1A images. The amplitude map of the image acquired on August 12, 2016 after registration is shown in Figure 2. Figure 1 As shown in Figure 2, each image is sorted according to the image acquisition time. Next, SBAS-InSAR is used to process the 42 registered images. During the SBAS-InSAR time series processing of the ctSent_v2 software, each image only generates an interference pair with the three images behind it, so that 120 image pairs can be generated. The main and auxiliary images corresponding to each image pair are interfered to obtain 120 interference phase images. 4:1 is used as the multi-view ratio to perform multi-view processing on each interference phase image. Gaussian filtering is used to perform filtering and noise reduction on each interference phase image. The coherence threshold is set to 0.35, and the amplitude deviation threshold is set to 0.6. The 120 interference phase images are analyzed in time series to select p PS points. The 120 interference phase images after the PS points are selected are unwrapped to obtain 120 unwrapped interference phase images, and each unwrapped interference phase image includes p PS points.

[0033] 2) Use quadratic polynomial to correct the atmospheric phase of 120 unwrapped interferometric phase images, and obtain a total of 120 unwrapped interferometric phase images without atmospheric phase correction. Compare the unwrapped interferometric phase images without atmospheric correction with the unwrapped interferometric phase images with atmospheric correction in turn to find out the interferometric phase images with poor atmospheric phase removal effect using the quadratic polynomial atmospheric phase correction method. The interferometric phase images without atmospheric phase removal corresponding to the interferometric phase images with poor atmospheric phase removal effect using the quadratic polynomial method are classified as type A interferometric phase images. There are m type A interferometric phase images in total. The unwrapped interferometric phase images generated by the images acquired on August 12, 2016 and the images acquired on September 29, 2016 belong to type A interferometric phase images ( Figure 2 ),from Figure 2It can be seen that the unwrapped interferometric phase diagram after atmospheric phase correction using the quadratic polynomial method still includes a large amount of atmospheric phase noise; the interferometric phase diagram that has been de-atmosphericized with a good atmospheric phase correction effect using the quadratic polynomial method is classified as Class B interferometric phase diagram, and there are 120-m Class B interferometric phase diagrams in total.

[0034] 3) Construct a preliminary BP neural network:

[0035] ①Confirm the number of input nodes of BP neural network:

[0036] The input of the BP neural network of the present invention is the latitude and longitude coordinates of the PS point, and the number of neurons in the input layer of the BP neural network of the present invention is 2; for the input layer data, since there are magnitude differences between the data, the present invention normalizes them to between [-1, 1] according to the two input layer neurons, so as to eliminate the magnitude difference problem between the input data.

[0037] ②Confirm the number of output nodes of the network:

[0038] The corresponding output of the BP neural network of the present invention is the phase value of the PS point, so the number of neurons in the output layer of the network is 1.

[0039] ③Confirm the number of hidden layer nodes in the network:

[0040] The empirical formula for selecting the number of hidden layers is as follows:

[0041]

[0042] in is the number of neurons in the input layer, is the number of hidden layer nodes, is the number of neurons in the output layer, It is usually a positive integer less than or equal to 10.

[0043] ④ Construct a BP neural network. According to the empirical formula of the number of hidden layers, the number of neurons in the input layer in the present invention is b =2, the number of neurons in the output layer c =1, number of hidden layer nodes The value is between [2, 11], and the number of nodes in the initial hidden layer of the network is b 1 =2; Use the newff function in Matlab to build a BP neural network, set the hidden layer parameter of the neural network to 2, the learning rate of the BP neural network to 0.001, the number of training steps to 1000, and the initial training target value of the network to goal=0.001. At this time, a preliminary BP neural network has been built.

[0044] 4) Convert the longitude and latitude and phase values ​​of p PS points in the unwrapped interferometric phase map generated by the images acquired on August 12, 2016 and the images acquired on September 29, 2016 into a p*3 matrix i, the first column of which is the longitude of each PS point in the interferometric phase map, the second column is the latitude of each PS point in the interferometric phase map, and the third column is the phase value of each PS point in the interferometric phase map. Perform thinning of the matrix i with an equal interval of 10 to obtain the matrix i1, normalize the first two columns of the matrix i1 to [-1, 1] to obtain the matrix i2, randomly select 70% of the data of the matrix i2 as the training samples of the BP neural network, and use the remaining 30% of the data of the matrix i2 as the test samples of the BP neural network. Use the constructed preliminary BP neural network to train the training samples of the matrix i2, and then perform the following steps:

[0045] ① After the initial training, check whether the training target value is less than goal. If the target value is not less than goal, proceed to step ②. If the target value is less than goal, use the trained network to simulate 30% of the test samples of matrix i2. If the evaluation index between the simulation data simulated by the test samples and the test samples meets the requirements, the training ends and the network parameters and structure are saved. If the evaluation index between the simulation data simulated by the test samples and the test samples does not meet the requirements, the number of initial hidden layer nodes of the network is taken. b 1 =3, use the training samples for training again, and repeatedly adjust the number of initial hidden layer nodes of the network until the training target value of the BP neural network for the training samples is less than goal and the evaluation index of the test samples meets the requirements, then end the training and save the network parameters and structure.

[0046] ② The target value of the initial training is not less than goal. Since the training and test samples used in the present invention contain certain abnormal data, the present invention handles the abnormal data as follows during implementation:

[0047] Use the current network to simulate new simulation data using training samples and test data , simulation data The relative error with the third column of sample data is expressed as follows:

[0048]

[0049] in is the relative error, It is the simulation data simulated by BP neural network. The third column of data is the training and testing samples.

[0050] When the relative error between the simulated data of the BP neural network and the third column data of the sample exceeds 100%, these data are regarded as abnormal data points in the sample, these abnormal data are removed, and the number of initial hidden layer nodes of the network is taken b 1 =3, use the training samples after removing abnormal data for training again, check whether the training target value is less than goal after the training, if the target value is less than goal, use the parameters of this training to simulate the test samples after removing abnormal data, if the evaluation index between the simulation data simulated by the test samples and the test samples meets the requirements, the training ends and the network parameters and structure are saved; if the target value of the training sample is not less than goal, the number of initial hidden layer nodes of the network is taken b 1 =4, remove abnormal data, use the training samples after removing abnormal data for training, and repeatedly adjust the number of initial hidden layer nodes of the network until the training target value of the BP neural network for the training sample is less than goal and the evaluation index of the test sample meets the requirements, end the training, and save the network parameters and structure.

[0051] ③ When the number of initial hidden layer nodes of the network is b 1 =11, if neither the training target value of the training sample nor the evaluation index of the test sample meets the requirements, the target value goal is increased, and the number of initial hidden layer nodes of the network is repeatedly adjusted until the training target value of the BP neural network for the training sample is less than the new goal and the evaluation index of the test sample meets the requirements, the training is terminated, and the network parameters and structure are saved. Figure 3 a is the training sample finally used for training after removing abnormal data points, Figure 3 b is the test sample finally used for testing after removing abnormal data points, Figure 4 a is to use BP neural network Figure 3 The atmospheric phase simulated by the training sample in a, Figure 4 b is the use of BP neural network Figure 3 Simulated atmospheric phase for the training samples in b.

[0052] 5) After the above process is completed, the optimal number of hidden layer nodes of the BP neural network of the present invention is 6, and the optimal training target value is goal=0.018. At this point, a 2-6-1 BP neural network is established. The BP neural network is used to simulate all data of matrix i to obtain the atmospheric phase data of the first interference phase diagram in the A-type interference phase diagram ( Figure 5 c) Subtract the atmospheric phase obtained by the BP neural network from the first interference phase diagram in the A-type interference phase diagram to obtain an interference phase diagram with a better atmospheric phase removal effect ( Figure 5d). The statistical results of the phase mean and standard deviation of the three interference phases are shown in Table 1. Figure 5 a and Figure 5 By comparing with b, it can be found that the effect of removing the atmospheric phase using the method provided by the present invention is significantly better than the quadratic polynomial method. Figure 5 The mean value of the original phase in a is 3.274 rad, and the standard deviation of the phase is 1.277 rad; Figure 5 The mean value of the interference phase in b using the quadratic polynomial to remove the atmospheric phase is almost 0 rad, and the phase standard deviation is 1.145 rad. From the perspective of the phase standard deviation, the quadratic polynomial can hardly remove the atmospheric phase from the original phase; Figure 5 The interference phase mean of the atmospheric phase removed by the method provided by the present invention in d is 0.015 rad, and the phase standard deviation is 0.725 rad. From the perspective of the phase mean and standard deviation, the method provided by the present invention can remove a large amount of atmospheric phase noise, and the removal effect is significant. Figure 5 As shown in the black box, after the atmospheric phase is removed by the quadratic polynomial, there is still a lot of atmospheric phase noise there, but after the atmospheric phase is removed by the method of the present invention, the atmospheric phase in the interference phase is almost completely removed, and the effect is extremely significant. To further verify the reliability of the method provided by the present invention, the original interference phase in the black box, the interference phase after the atmospheric phase is removed by the quadratic polynomial, and the interference phase after the atmospheric phase is removed by the method of the present invention are statistically analyzed in histograms. The statistical results are shown in Figure 1. Figure 6 As shown, it can be seen from the figure that: the original interference phase is mainly distributed between [2, 9] rad, the phase mean is 5.782 rad, and the phase standard deviation is 1.669 rad, indicating that there is a large amount of atmospheric phase noise here; the interference phase removed by the quadratic polynomial is mainly distributed between [-2, 4] rad, the phase mean is 1.229 rad, and the phase standard deviation is 1.641 rad, indicating that there is still a large amount of atmospheric phase noise here; and the interference phase after removing the atmospheric phase using the method provided by the present invention is mainly evenly distributed between [-2, 2] rad, the phase mean is 0.050 rad, and the phase standard deviation is 0.799 rad, indicating that there is only a very small amount of atmospheric phase noise here; from the perspective of the phase standard deviation, the phase standard deviation of the quadratic polynomial method is almost the same as the standard deviation of the original phase, while the phase standard deviation of the method provided by the present invention is 2 times smaller than that of the quadratic polynomial method, which shows that the effect of removing the atmospheric phase by the method provided by the present invention is 2 times better than that of the quadratic polynomial method. Figure 6 The statistical results once again prove that the method provided by the present invention is more effective in removing the atmospheric phase.

[0053] Table 1 Comparison of statistical results of three interference phases

[0054] Phase mean / rad Phase standard deviation / rad Raw interferometric phase 3.274 1.277 Second-order polynomial to remove the interference phase of atmospheric phase 0 1.145 The method of the present invention removes the interference phase of the atmospheric phase 0.015 0.725 Original interference phase (in the black box) 5.782 1.669 The quadratic polynomial removes the interference phase of the atmospheric phase (in the black box) 1.229 1.641 The method of the present invention removes the interference phase of the atmospheric phase (in the black box) 0.050 0.799

[0055] 6) Perform steps 3) to 5) on the remaining interference phase maps of category A in turn to obtain m-1 corresponding interference phase maps with the atmospheric phase removed. Combine the m interference phase maps with the atmospheric phase removed using the BP neural network and the 120-m interference phase maps of category B to obtain a new interference phase map set C. Use the interference phase maps in set C to perform deformation rate inversion to obtain a high-precision deformation rate map of the entire study area.

Claims

1. A method for removing the atmospheric phase of InSAR, characterized in that: include: Acquire remote sensing images and obtain n unwrapped interferometric phase maps based on SBAS technology, each of which includes p permanent scatterer points; The atmospheric phase of n unwrapped interferometric phase images is removed by using the quadratic polynomial atmospheric phase correction method, and the atmospheric phase removal results are arranged in order according to the removal effect. The m results with the poor removal effect are taken, and the corresponding unwrapped interferometric phase images without atmospheric phase removal are found and classified as type A interferometric phase images. The longitude and latitude of each interference phase diagram in the A-type interference phase diagram are taken as input, and the corresponding phase value is taken as output, and a BP neural network training model including input layer, hidden layer and output layer is established; the interference phase diagram in the A-type interference phase diagram is subtracted from the atmospheric phase obtained by BP neural network training to obtain the interference phase diagram without the atmospheric phase.

2. The method according to claim 1, characterized in that The BP neural network training model is established in the following way: For each interference phase diagram in the A-type interference phase diagram, the longitude and latitude and phase values ​​of the p permanent scatterer points are converted into a p*3 matrix i; the first column of the matrix i is the longitude of each PS point in the interference phase diagram, the second column is the latitude of each PS point in the interference phase diagram, and the third column is the phase value of each PS point in the interference phase diagram. The first two columns of data in the matrix are normalized to obtain the matrix i2; For the first interference phase image in the A-type interference phase image, part of the data of matrix i2 is randomly selected as the training sample of the BP neural network, and the remaining data is used as the test sample. The BP neural network training model is established with the training sample as the input and the corresponding phase value as the output; the trained BP neural network is used to simulate all the data of the matrix i of the first image to obtain the atmospheric phase data of the first interference phase image, and the first interference phase image is subtracted from the simulated atmospheric phase to obtain the interference phase image with the atmospheric phase removed; Repeat the steps until all interference phase images in the type A interference phase image are processed and the InSAR atmospheric phase removal is completed.

3. The method according to claim 2, characterized in that Matrix i is thinned out with equal spacing to obtain matrix i1, and the first two columns of data in matrix i1 are normalized to [-1, 1] to obtain matrix i2 for subsequent BP neural network training / testing.

4. The method according to claim 2, characterized in that: The preset parameter goal is the performance target for model training, including: After the initial training, check whether the training target value is less than goal. If the target value is not less than goal, it is determined that there are abnormal data in the training samples and test samples, and the abnormal data is processed; if the target value is less than goal, use the trained network to simulate the test sample. If the evaluation index between the simulation data simulated by the network and the test sample meets the requirements, the training is completed and the network parameters and structure are saved; otherwise, increase the number of initial hidden layer nodes of the network and retrain using the training samples.

5. The method according to claim 4, characterized in that The method of processing abnormal data is as follows: using the current BP neural network to simulate new simulation data using training samples and test data, and calculating the relative error between the simulation data and the data in the third column of matrix i2; When the relative error exceeds 100%, these data are removed as abnormal data points in the sample; Increase the number of initial hidden layer nodes of the network, and use the training samples after removing abnormal data for training again. After the training, check whether the training target value is less than the goal. If the target value is less than the goal, use the trained network to simulate the test samples after removing abnormal data. If the evaluation index between the simulation data simulated by the network and the test samples meets the requirements, the training ends and the network parameters and structure are saved. Otherwise, increase the number of initial hidden layer nodes, remove abnormal data, and use the training samples after removing abnormal data for training; Repeatedly adjust the number of initial hidden layer nodes of the network until the training target value of the BP neural network for the training sample is less than the goal and the evaluation index of the test sample meets the requirements, end the training, and save the network parameters and structure.

6. The method according to claim 4 or 5, characterized in that: When the number of initial hidden layer nodes of the BP neural network reaches the preset upper limit, and the training target values ​​of the training samples and the test samples are not less than goal, the target value goal is increased, and the number of initial hidden layer nodes of the network is repeatedly adjusted until the training target value of the BP neural network for the training samples is less than the new goal and the evaluation index of the test samples meets the requirements, the training is terminated, and the network parameters and structure are saved.

7. The method according to any one of claims 4 to 6, characterized in that: The evaluation index uses a statistical index related to the error.

8. The method according to claim 1, characterized in that The number of hidden layer nodes of BP neural network is 2~11.

9. The method according to claim 1, characterized in that: It also includes classifying the remaining nm interference phase maps from which the atmospheric phase has been removed as Type B interference phase maps; after the InSAR atmospheric phase is removed, merging the Type B interference phase map and the Type A interference phase map from which the atmospheric phase has been removed to obtain a new interference phase map set, and using the interference phase maps in the new interference phase map set to perform deformation rate inversion to obtain a high-precision deformation rate map for the entire study area.

Images