GB-InSAR atmospheric phase compensation method based on LSTM

Abstract

This invention relates to the field of atmospheric phase compensation technology and discloses a GB-InSAR atmospheric phase compensation method based on LSTM. Because the atmospheric delay coefficients of electromagnetic wave propagation may vary in each azimuth, the atmosphere has a spatially variable effect on radar data monitoring. Therefore, conventional atmospheric phase compensation methods are no longer applicable. This paper proposes an LSTM-based atmospheric phase compensation method. First, the correlation coefficient and phase average value of the PS points are calculated to construct an evaluation factor. The evaluation factor is selected and input into the LSTM to obtain the temporal phase variation characteristics of the PS points. A classification model is constructed and all PS points are classified using the K-means method. The class with the smallest standard deviation is selected as the stable point to verify the reliability of the model and obtain a suitable classification model. The classification model is then used to classify all PS points and select stable PS points. Finally, Lowess interpolation fitting is performed based on the stable PS points to compensate for the atmospheric phase of the PS points. This method is suitable for scenarios with complex weather conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of atmospheric phase compensation, and in particular to a GB-InSAR atmospheric phase compensation method based on LSTM. Background Art

[0002] The propagation coefficient of electromagnetic waves is affected by atmospheric conditions. When atmospheric conditions change, the electromagnetic wave coefficient will also change. Therefore, the electromagnetic wave coefficient is different at different times. The deformation variables calculated by GB-InSAR through two differential interferometry calculations contain meteorological disturbance errors. In order to obtain the deformation variables more accurately, it is necessary to perform atmospheric phase compensation on the pixels in the interferometric phase map.

[0003] Currently, two common atmospheric phase compensation methods are used in GB-InSAR measurements. The first is to use meteorological data to calculate the atmospheric phase based on the atmospheric refractive index model. The second is to establish an atmospheric parameter model based on the PS point. Under good weather conditions, the atmosphere can be considered to be spatially homogeneous. Based on the atmospheric homogeneity, a linear model that varies with slant range is established to estimate the atmospheric phase at the PS point. However, both methods have drawbacks. In actual monitoring scenarios, meteorological conditions change over time, and the atmosphere is heterogeneous in time and space. Especially under complex weather conditions, the atmospheric changes in space are more complex. Using a multi-parameter model to estimate the atmospheric phase will result in large errors. Therefore, it is necessary to study compensation methods for spatiotemporal atmospheric phase variation. Summary of the Invention

[0004] The present invention aims to provide a GB-InSAR atmospheric phase compensation method based on LSTM to solve the problem of large atmospheric phase estimation error in existing methods under complex weather conditions.

[0005] In order to achieve the above object, the basic scheme of the present invention is as follows: a GB-InSAR atmospheric phase compensation method based on LSTM comprises the following steps:

[0006] Step 1: Calculate the correlation coefficient and phase average of PS points and construct the evaluation factor.

[0007] Step 2: Select the evaluation factor and input it into LSTM to obtain the phase change characteristics of the PS point time series and build a classification model.

[0008] Step 3: Use the K-means method to classify all PS points, select the category with the smallest standard deviation as the stable point to verify the reliability of the model and obtain a suitable classification model.

[0009] Step 4: Use the classification model to classify all PS points and select stable PS points.

[0010] Step 5: Perform Lowess interpolation fitting based on the stable PS point to compensate for the atmospheric phase of the PS point.

[0011] The principle and beneficial effects of the present invention are as follows: by considering the correlation of atmospheric phase in time series, the present method is proposed, so that it can be applied to scenes with more complex weather environments.

[0012] Furthermore, LSTM is a long-short time series network. The long-short time series network adds a forget gate and a self-loop connection point on the basis of the output gate and input gate of the recurrent neural network. In the training process of the long-short time series network, the tanh function is used as the core function and the Simgomd function is used as the activation function.

[0013] Beneficial effects: The long-short time series network can record or delete information through the forget gate to achieve forgetting and memory functions, and mainly achieves forgetting and memory functions through the input gate and output gate. The input gate can choose to forget information, and the output gate determines the output value of the memory storage cell and whether it will be used as the input neuron of the next period.

[0014] Furthermore, the training process of the long-short time series network consists of the following parts:

[0015] The first part is the forget gate, which determines whether information is recorded or forgotten. The formula is as follows:

[0016]

[0017] Where, f t Represents the output of the forget gate. When it is 1, it means that the current information will be input to the next period. When it is 0, the information will be deleted. The size between [0,1] indicates how much information will be recorded in the next time. W f represents the weight, b f represents the bias, σ represents the activation function, h t-1 Indicates the information recorded in the previous period, x t represents the input of this period,

[0018] The second part is that after the forget gate makes a judgment, the input gate will determine whether the input of this period will be recorded. The formula is as follows:

[0019]

[0020]

[0021] Among them, i t represents the output of the input gate, W i The weight matrix represented by b i represents the bias term, The value representing the new cell state, WC represents the weight of the current input unit state, b C Indicates the bias of the current input cell state,

[0022] The third part is the update, which completes the update of the long and short time series network based on the output of the activation function and the information recorded in the previous period. The formula is as follows:

[0023]

[0024] The fourth part is the long and short time series network for input and output. When outputting, the long and short time series network will judge the output and decide whether to output the result based on the current unit state. It mainly depends on the activation function and the tanh function. The formula is as follows:

[0025]

[0026] h t =o t *tanh(C t ) (17)

[0027] Among them t represents the output of the output gate, W0 represents the weight matrix, b0 represents the bias term, h t Indicates the output result.

[0028] Beneficial effects: The Simgomd function can normalize the input samples and convert the input values ​​to [0,1]. This can reduce the sample gap caused by large numerical deviations between samples, which is very helpful for training. However, the Simgomd function has defects and is prone to slow convergence and gradient disappearance problems. Its formula is as follows:

[0029]

[0030] Tanh has advantages over the Simgmod function. It avoids the defects of Simgmod. Therefore, it is often used as the core function of long and short time series networks.

[0031] Furthermore, PS points are divided into two categories, namely stable PS points and unstable PS points. Each PS point is described by two observation variables: phase average value and complex correlation coefficient.

[0032] Furthermore, the phase average is calculated as the phase average of the kth PS point Expressed as in Represents the phase value of PS point k after the interference unwrapping of the i-th image and the first image.

[0033] Furthermore, the calculation method of the complex correlation coefficient is to calculate the correlation coefficient of each PS point with the complex phase of all PS points and average it to get And the formula is,

[0034]

[0035] in, represents the information of PS point k in the i-th image, and W represents the information of all PS points.

[0036] Furthermore, since the dimensions and value ranges of the phase average and complex correlation coefficient are different, a dimensionless normalization process is required first. The formula is as follows:

[0037]

[0038] Then add the two elements to get the evaluation factor And select x% before and after the evaluation criteria to train with LSTM, and then classify the remaining PS points, the first x% as stable points, and the last x% as unstable points.

[0039] Furthermore, in order to verify the classification model constructed by LSTM, the K-means method is used to cluster all PS points. Then, by calculating the standard deviation of all cluster categories, the category with the minimum standard deviation is selected as the stable point category to verify the classification model. The principle of the K-means algorithm is to divide the given sample set into clusters according to distance, and the K-means algorithm can set the center point of the cluster and divide the points with similar adjacent distances into the same category according to the distance. Suppose the sample set x is divided into K clusters (C1, C1, ..., C k ), the sum of squared errors E between clusters is expressed as:

[0040]

[0041] Among them, ||·|| represents the second-order norm, that is, the modulus of the vector, μ i It is cluster C i The means of are adjacent, and the formula is expressed as:

[0042]

[0043] Here, |·| represents the first-order norm, that is, the number of points in the cluster,

[0044] Cluster all PS points, the cluster category is k, and calculate the standard deviation of each cluster. The calculation formula is expressed as:

[0045]

[0046] in, represents the phase value of the i-th point in the k-th cluster, represents the phase mean of the k-th cluster.

[0047] Beneficial effect: The PS in the cluster with the smallest phase standard deviation is classified as a stable PS point, which is used to test the accuracy of the classification model obtained before and after x% of the evaluation criteria. When the model meets the expected criteria, the current model is used to classify all PS points and select stable PS points. Finally, the stable PS is used to compensate for the atmospheric phase. The compensation method uses local weighted regression fitting.

[0048] Furthermore, Lowess is a local weighted regression, which is a non-parametric statistical learning method. Local weighted learning can fit the global model through local simplification and simple local model assumptions.

[0049] Beneficial effects: In practical applications, the distribution of data often cannot provide sufficient distribution. Local weighted learning gives different weights to training samples according to the distance between the prediction point and the direct point of the training sample. The closer the training sample is to the predicted point, the greater the weight. On the contrary, the farther the training sample is from the predicted point, the smaller the weight. This allows local weighted learning to differentially learn local information of the data.

[0050] Furthermore, for the stable PS point, first take a screenshot of the window, and the intercepted distance is selected as the set of points with the minimum distance to the specified point. The expression is as follows:

[0051]

[0052] According to the set window d span , select the set of points with the smallest distance to the specified point within the wide range, the expression is as follows,

[0053]

[0054] Each point at a distance from a specified point has its own weight, which is:

[0055]

[0056] Using these points for local weighted regression, the loss function is:

[0057]

[0058] A fitting surface f(x,y) is constructed using local weighted regression, and the obtained f(x,y) is the estimated atmospheric interferometry phase value. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of a traditional LSTM model of the LSTM-based GB-InSAR atmospheric phase compensation method in an embodiment of the present invention.

[0060] Figure 2 These are on-site pictures and radar pictures of the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention.

[0061] Figure 3 Radar image and differential interferometry phase map of the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention.

[0062] Figure 4 This is a selection distribution diagram of the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention.

[0063] Figure 5 This is the phase change curve when 3% is selected for the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention.

[0064] Figure 6 This is a result diagram of selecting different evaluation factor percentages for the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention.

[0065] Figure 7 This is a point clustering result and minimum standard deviation category diagram of the LSTM-based GB-InSAR atmospheric phase compensation method in an embodiment of the present invention.

[0066] Figure 8 This is a classification result diagram of the LSTM-based GB-InSAR atmospheric phase compensation method in an embodiment of the present invention.

[0067] Figure 9 This is a comparison chart of the atmospheric compensation results based on LSTM of the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention and the atmospheric compensation of the conventional method.

[0068] Figure 10 This is a flow chart of the GB-InSAR atmospheric phase compensation method based on LSTM in an embodiment of the present invention. DETAILED DESCRIPTION

[0069] The following is further described in detail through specific implementation methods:

[0070] The embodiment is basically as shown in the attached Figure 1 and Figure 10 As shown in the figure: The GB-InSAR atmospheric phase compensation method based on LSTM includes the following steps:

[0071] Step 1: Calculate the correlation coefficient and phase average of PS points and construct the evaluation factor.

[0072] The calculation method of the phase average is the phase average of the kth PS point Expressed as in Represents the phase value of PS point k after interference unwrapping between the i-th image and the first image. The calculation method of the complex correlation coefficient is to calculate the correlation coefficient of each PS point with all PS points and average it. And the formula is,

[0073]

[0074] in, Represents the information of PS point k in the i-th image, and W represents the information of all PS points

[0075] Since the dimensions and value ranges of the phase average and complex correlation coefficient are different, they need to be normalized to eliminate the dimensions first. The formula is as follows:

[0076]

[0077] Then add the two elements to get the evaluation factor

[0078] Step 2: Select the evaluation factor and input it into LSTM to obtain the phase change characteristics of the PS point time series and build a classification model.

[0079] Select x% before and after the evaluation criteria to use LSTM training, and then classify the remaining PS points. The first x% are stable points, and the last x% are unstable points. The LSTM training process consists of the following parts:

[0080] The first part is the forget gate, which determines whether information is recorded or forgotten. The formula is as follows:

[0081]

[0082] Where, f t Represents the output of the forget gate. When it is 1, it means that the current information will be input to the next period. When it is 0, the information will be deleted. The size between [0,1] indicates how much information will be recorded in the next time. W f represents the weight, b f represents the bias, σ represents the activation function, h t-1 Indicates the information recorded in the previous period, x t represents the input of this period,

[0083] The second part is that after the forget gate makes a judgment, the input gate will determine whether the input of this period will be recorded. The formula is as follows:

[0084]

[0085] Among them, i t represents the output of the input gate, W i The weight matrix represented by b i represents the bias term, The value representing the new cell state, W C represents the weight of the current input unit state, b C Indicates the bias of the current input cell state,

[0086] The third part is the update, which completes the update of the long and short time series network based on the output of the activation function and the information recorded in the previous period. The formula is as follows:

[0087]

[0088] The fourth part is the long and short time series network for input and output. When outputting, the long and short time series network will judge the output and decide whether to output the result based on the current unit state. It mainly depends on the activation function and the tanh function. The formula is as follows:

[0089]

[0090] h t =o t *tanh(C t ) (17)

[0091] Among them t represents the output of the output gate, W0 represents the weight matrix, b0 represents the bias term, h t Indicates the output result.

[0092] Step 3: Use the K-means method to classify all PS points, select the category with the smallest standard deviation as the stable point to verify the reliability of the model and obtain a suitable classification model.

[0093] In order to verify the classification model constructed by LSTM, the K-means method is used to cluster all PS points. Then, by calculating the standard deviation of all cluster categories, the category with the minimum standard deviation is selected as the stable point category to verify the classification model. The principle of the K-means algorithm is to divide the given sample set into clusters according to distance, and the K-means algorithm can set the center point of the cluster and divide the points with similar adjacent distances into the same category according to the distance. Suppose the sample set x is divided into K clusters (C1, C1, ..., C k ), the sum of squared errors E between clusters is expressed as:

[0094]

[0095] Among them, ||·|| represents the second-order norm, that is, the modulus of the vector, μ i It is cluster C i The means of are adjacent and can be expressed as:

[0096]

[0097] Here, |·| represents the first-order norm, that is, the number of points in the cluster.

[0098] Step 4: Use the classification model to classify all PS points and select stable PS points.

[0099] Cluster all PS points, the cluster category is k, and calculate the standard deviation of each cluster. The calculation formula can be expressed as:

[0100]

[0101] in, represents the phase value of the i-th point in the k-th cluster, represents the phase mean of the k-th cluster,

[0102] The PS in the cluster with the smallest phase standard deviation is classified as a stable PS point to test the accuracy of the classification model obtained before and after x% of the evaluation criteria. When the model meets the expected criteria, the current model is used to classify all PS points and select the stable PS point.

[0103] Step 5: Perform Lowess interpolation fitting based on the stable PS point to compensate for the atmospheric phase of the PS point. Lowess is a local weighted regression, which is a non-parametric statistical learning method. Local weighted learning can fit the global model through local simplification and simple local model assumptions. Then, for the stable PS point, first screenshot the window, and the intercepted distance is selected as the set of points with the minimum distance to the specified point. The expression is as follows:

[0104]

[0105] According to the set window d span , select the set of points with the smallest distance to the specified point within the wide range, the expression is as follows,

[0106]

[0107] Each point at a distance from a specified point has its own weight, which is:

[0108]

[0109] Using these points for local weighted regression, the loss function is:

[0110]

[0111] A fitting surface f(x,y) is constructed using local weighted regression, and the obtained f(x,y) is the estimated atmospheric interferometry phase value.

[0112] Experimental information:

[0113] like Figure 2 Figure (a) shows a photograph of the landslide scene. The landslide is located in the Zhongtai Formation of Liujing Village, Tudi Township, Wulong District, at 29°28'53"N, 107°55'43"E. It is approximately 25 kilometers from Wulong City, with National Highway 319 providing direct access, making transportation relatively convenient. The slope is covered with dense vegetation, and the radar image shows a relatively stable scattering effect from the exposed slope. The slope detection radar, developed by Beijing Institute of Technology's Raytheon Company, operates in the 1.5- ...

[0114] like Figure 3 As shown, 30 radar images collected continuously at 11:00 on July 28, 2020, with an average acquisition time of two minutes, are now analyzed. Figure (a) shows a radar image of the area, with the pixel amplitudes processed by dB. Using the first image as the primary image and the last image as the auxiliary image, differential interferometry is performed, resulting in the interferometric phase diagram (Figure (b). The phases of many pixels in Figure (b) change slowly with slant range and azimuth, indicating that the interferometric phase exhibits strong spatial variability. The area indicated by the red circle is the deformation region, and the interferometric phase is significantly different from that of other areas.

[0115] The figure below shows the result of selecting PS points using the amplitude deviation method. The amplitude deviation threshold is set to 0.25 and the amplitude threshold is set to -25dB. The total number of PS points after selection is 18171. Figure 4 shown.

[0116] Experimental results:

[0117] Get the classification model:

[0118] For the 18,171 high-quality pixels selected in the previous section, we first obtain their interference phases in the 29 differential interferometer phase images. We then calculate the correlation coefficient between each high-quality pixel and all other high-quality pixels and average the correlation coefficients to obtain the corresponding correlation coefficients for each high-quality pixel. We then calculate the average phase value of each high-quality pixel in the 29 differential interferometer phase images. Finally, we normalize the correlation coefficients and the average phase values ​​and use them as reference elements to construct the evaluation factor.

[0119] Different percentages before and after are selected as high-quality sets and low-quality sets, and their differences are shown using time-series interferometry phase diagrams, as shown below. Figure 5 As shown in (a) and (b), Figure 5 (a) represents the phase change curve of the first 3% points, and (b) represents the phase change curve of the last 3%. When the evaluation factors are selected before and after 3%, the selected high-quality set PS points show obvious differences in the 29 differential interferometer phase images. The first 3% of the evaluation factors fluctuate significantly in the 29 differential interferometer phase images, with the highest reaching 3rad and the lowest at -4rad. The last 3% fluctuates smoothly in the 29 differential interferometer phase images, and the phase fluctuation does not exceed 2rad.

[0120] An LSTM was used to extract the time-series phase characteristics of the pre- and post-evaluation factors in 29 differential interferometry phase images. A classification model was then constructed. All PS points were then classified to identify stable and unstable PS points. Table 1 below shows the number of stable and unstable points selected based on different percentages of the evaluation factors. As can be seen, the number of stable points decreases as the percentage of the selected pre- and post-evaluation factors increases. However, the interpolation fitting process requires a sufficient number of stable points. Therefore, when selecting the evaluation factor percentage, it is important to avoid excessively low numbers of stable points after classification, hindering subsequent interpolation fitting to compensate for atmospheric phases. To avoid inconsistent classification results each time the classification model is constructed, the same percentage of the evaluation factor was selected and the LSTM classification model was constructed multiple times. A PS point was classified as stable if its frequency among stable points exceeded 90% across multiple classification results.

[0121] Table 1 Number of stable and unstable points selected according to different evaluation factor percentages

[0122]

[0123] After selecting the stable and unstable points, the last radar image and the first radar image are used to perform differential interference to display the stable PS points and unstable PS points. The display results are shown in the figure below. Figure 6 (a), (c), and (e) are the unstable PS point distributions obtained by selecting 1%, 2%, and 3% of the evaluation factors before and after input into LSTM to build the classification model. Figure 6 (b), (d), and (f) are the distribution diagrams of the obtained stable PS points. It can be seen from the figures that when the front and rear 1% are selected, relatively more stable points are obtained, which is conducive to the subsequent interpolation fitting to compensate for the atmospheric phase.

[0124] In order to verify the accuracy of the model, K-means was used to cluster all PS points. The clustering results are as follows Figure 7 (a) shows. Calculate the standard deviation of all categories and select the category with the smallest standard deviation. The result is as follows Figure 7 As shown in (b), this category is considered a stable point category. Classification models constructed with different evaluation factor percentages are then incorporated to verify the accuracy of the classification models. The resulting accuracy is shown in the table below. To mitigate errors, the accuracy is averaged over multiple classifications. As can be seen from the table, the accuracy jumps when the evaluation factor percentage increases from 1% to 2%. This suggests that the model with a 1% evaluation factor percentage is more reliable.

[0125] Table 1 Classification model accuracy under different evaluation factors

[0126]

[0127] Atmospheric phase compensation:

[0128] From the above, we can see that when selecting 1% before and after the evaluation factor, the classification results are more reliable than when selecting 2% before and after the evaluation factor to build a classification model. Therefore, it is necessary to select before 1% to build a classification model. When selecting 2% and 3%, fewer stable points are obtained compared to 1%, which is not conducive to Lowess interpolation fitting to compensate for the atmospheric phase. Here, the time series phase of the 1% PS points before and after the evaluation factor is selected as the input to the LSTM to build a classification model. After obtaining the classification model, all remaining PS points are classified. The classification results are as follows Figure 8 As shown in (a) and (b), Figure 8 (a) is the classification result of unstable points, Figure 8 (b) is the classification result of stable points.

[0129] After obtaining the stable PS point, the atmospheric phase compensation is performed on all PS points by interpolation fitting. The interpolation fitting compensation method here uses the LOWESS method. After compensation, the 29 compensated differential interferometer phase images are accumulated, and the results are as follows Figure 9 (b) is shown. To verify the effectiveness of this method, the results are compared with the conventional compensation method. As can be seen from the figure, the phase after the method used in this chapter compensates for the atmospheric phase is Figure 9 (b) Phase compensation relative to conventional methods Figure 9 (a), the phases of most PS points are around 0rad.

[0130] In order to more intuitively compare the difference between the atmospheric phase after compensation based on the conventional method and the LSTM method proposed in this chapter, three PS points with the smallest amplitude deviation are selected from different slant distances as reference points. Figure 9 (c) Figure 9 As shown in (d), Figure 9 (c) is the reference point phase curve under conventional compensation, and (d) is the reference point phase curve under LSTM compensation. Figure 9As can be seen in (c), based on the LSTM atmospheric phase compensation method, the phase curve of the reference point after compensation is smoother, with the highest phase being 0.2 rad, while Figure 9 In (c), the phase of the compensation result using the conventional method reaches a maximum of 1 rad, and the phase of the reference point accumulates over time, which is caused by the residual atmospheric phase. This shows that the atmospheric phase compensation method in this chapter improves the compensation results compared to the conventional compensation method.

[0131] The above is only an embodiment of the present invention, and common knowledge such as the specific structure and / or characteristics of the scheme is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. The LSTM-based GB-InSAR atmospheric phase compensation method is characterized by: The following steps are included: Step 1: Calculate the complex correlation coefficient and phase average of the PS point to construct the evaluation factor; The calculation method of the complex correlation coefficient is to calculate the correlation coefficient of each PS point with the complex phase of all PS points and calculate the average to obtain And the formula is, in, represents the complex phase of PS point k in the i-th image, W represents the complex phase of all PS points, Step 2: Select the evaluation factor and input it into LSTM to obtain the phase change characteristics of the PS point time series and build a classification model. Step 3: Use the K-means method to classify all PS points, select the category with the smallest standard deviation as the stable point to verify the reliability of the model and obtain a suitable classification model. Step 4: Use the classification model to classify all PS points and select stable PS points. Step 5: Perform Lowess interpolation fitting based on the stable PS point to compensate for the atmospheric phase of the PS point.

2. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 1, wherein The LSTM is a long-short time series network. The long-short time series network adds a forget gate and a self-loop connection point on the basis of the output gate and input gate of the recurrent neural network. In the training process of the long-short time series network, the tanh function is used as the core function and the Simgomd function is used as the activation function.

3. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 2, characterized in that The training process of the long-short time series network consists of the following parts: The first part is the forget gate, which determines whether information is recorded or forgotten. The formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f ) (12) Where, f t Represents the output of the forget gate. When it is 1, it means that the current information will be input to the next period. When it is 0, the information will be deleted. The size between [0,1] indicates how much information will be recorded in the next time. W f represents the weight, b f represents the bias, σ represents the activation function, h t-1 Indicates the information recorded in the previous period, x t represents the input of this period, The second part is that after the forget gate makes a judgment, the input gate will determine whether the input of this period will be recorded. The formula is as follows: i t =σ(W i ·[h t-1 ,x t ]+b i ) (13) Among them, i t represents the output of the input gate, W i The weight matrix represented by b i represents the bias term, is the candidate value of the unit state of the long short-term memory network at the current moment, W c represents the weight of the current input unit state, b C Indicates the bias of the current input cell state, The third part is the update, which completes the update of the long and short time series network based on the output of the activation function and the information recorded in the previous period. The formula is expressed as follows: Among them, C t : is the unit state value of the long short-term memory network at the current moment; C t-1 : is the unit state value of the long short-term memory network at the previous moment; is the candidate value of the unit state of the long short-term memory network at the current moment, The fourth part is the output of the long and short time series network. When outputting, the long and short time series network will judge the output and decide whether to output the result based on the current unit state. It mainly depends on the activation function and the tanh function. The formula is as follows: about t =σ(W0[h t-1 ,x t ]+b0) (16) h t =o t *tanh(C t ) (17) Among them t represents the output of the output gate, W0 represents the weight matrix, b0 represents the bias term, h t Indicates the output result, C t : is the unit state of the long short-term memory network at the current moment.

4. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 3, characterized in that The PS points are divided into two categories, namely stable PS points and unstable PS points. Each PS point is described by two observation variables: phase average value and complex correlation coefficient.

5. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 4, characterized in that The phase average value is calculated as the phase average of the kth PS point Expressed as in Represents the phase value of PS point k after the interference unwrapping of the i-th image and the first image.

6. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 5, characterized in that Since the dimensions and value ranges of the phase average value and complex correlation coefficient are different, a dimensionless normalization process must be performed first. The formula is as follows: in, is the average multiple correlation coefficient of PS points; is the minimum value of the multiple correlation coefficient; is the maximum value of the multiple correlation coefficient, Then add the two elements to get the evaluation factor And select x% before and after the evaluation criteria to train with LSTM, and then classify the remaining PS points, the first x% as stable points, and the last x% as unstable points.

7. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 6, characterized in that In order to verify the classification model constructed by LSTM, the K-means method is used to cluster all PS points. Then, by calculating the standard deviation of all cluster categories, the category with the minimum standard deviation is selected as the stable point category to verify the classification model. The principle of the K-means algorithm is to divide the given sample set into clusters according to distance, and the K-means algorithm can set the center point of the cluster and divide the points with similar adjacent distances into the same category according to the distance. Suppose the sample set x is divided into K clusters (C1, C1, ..., C k ), the sum of squared errors E between clusters is expressed as: Among them, ||·|| represents the second-order norm, that is, the modulus of the vector, μ i It is cluster C i The means of are adjacent, and the formula is expressed as: Here, |·| represents the first-order norm, that is, the number of points in the cluster, Cluster all PS points, the cluster category is k, and calculate the standard deviation of each cluster. The calculation formula is expressed as: in, represents the phase value of the i-th point in the k-th cluster, represents the phase mean of the k-th cluster, and n is the number of PS points in the k-th cluster after PS points are clustered.

8. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 7, characterized in that The Lowess method is a local weighted regression method, which belongs to a non-parametric statistical learning method. Local weighted learning can fit a global model through local simplification and simple local model assumptions.

9. The GB-InSAR atmospheric phase compensation method based on LSTM as claimed in claim 8, characterized in that For the selected stable PS point, first intercept its image sub-window, and the intercepted distance is selected as the set of points with the minimum distance to the specified point, which is expressed as follows: According to the set window d span , select the point with the smallest distance to the specified point to form a point set. The expression is as follows: Each point at a distance from a specified point has its own weight, which is: Using these points for local weighted regression, the loss function is: A fitting surface f(x, y) is constructed using local weighted regression, and the obtained f(x, y) is the estimated atmospheric interferometry phase value.

Images