A multi-channel InSAR elevation inversion method based on deep learning

Abstract

The present invention discloses a multi-channel InSAR elevation inversion method based on deep learning, which includes creating continuous terrain and sudden terrain datasets; placing the created datasets into a multi-channel feature connection and fusion network for training; and placing the multi-channel interferogram to be inverted into a trained multi-channel neural network model to obtain an inverted true elevation image. The method designs a multi-channel feature connection and fusion network, which is used to establish a direct mapping relationship between multiple interferograms and corresponding elevation maps. The network adopts a residual unit that integrates an attention mechanism, and realizes multi-channel feature reuse in a densely connected manner, thereby strengthening the propagation of feature maps and improving the network's ability to fuse information from multiple interferograms. After experiments on interferogram elevation inversion of different terrains, the results show that the method can simultaneously invert elevation maps of continuous terrain and sudden terrain, has good robustness and robustness, and has important application value.

Description

Technical Field

[0001] The present invention belongs to the field of image elevation inversion, relates to image elevation inversion in the application of interferometry technology, and in particular to a multi-channel InSAR elevation inversion method based on deep learning. Background Art

[0002] Elevation inversion is an indispensable step in many interferometry technology applications. The single-channel or multi-channel interferograms obtained in various interferometry technology applications usually have varying degrees of interference phase noise and phase discontinuity problems, which makes the interferogram elevation inversion problem still a very challenging problem.

[0003] Existing Interferometric Synthetic Aperture Radar (InSAR) multi-channel elevation inversion algorithms are roughly divided into the Chinese Remainder Theorem (CRT) algorithm, the Maximum Likelihood (ML) algorithm, the Maximum A Posteriori (MAP) algorithm, and the Graph Cut (GC) algorithm. Among the above algorithms, the Chinese Remainder Theorem algorithm can accurately recover the target elevation information in a pure environment, that is, under the condition of noise-free interference phase. When there is large noise in the interference pattern, the CRT elevation inversion almost fails; the minimum norm method generally focuses on the global situation and finds the optimal solution, but is easily restricted by the Itoh phase continuity assumption; the ML algorithm and the MAP algorithm achieve the optimal estimation parameters by constructing a probability density function, and then reconstruct the elevation. When the interference pattern signal-to-noise ratio is low, the elevation estimation accuracy is often not high; the GC algorithm can effectively obtain the global optimal solution in the convex optimization problem, thereby improving the accuracy of elevation inversion, but the algorithm requires a large amount of memory and is not conducive to the processing of large-scale interference patterns. Summary of the Invention

[0004] In order to solve the above problems, the present invention provides a multi-channel InSAR elevation inversion method based on deep learning with high elevation inversion accuracy and strong robustness.

[0005] The technical solution for achieving the purpose of the present invention is:

[0006] A multi-channel InSAR elevation inversion method based on deep learning includes the following steps:

[0007] S1, creating single-channel elevation maps of continuous terrain and sudden terrain, forming a multi-channel interferogram from the elevation maps of the two terrains, thereby obtaining a data set containing a multi-channel interferogram and a single-channel elevation map;

[0008] S2, puts the data set obtained in S1 into the multi-channel feature connection and fusion network model for training to obtain the trained weights;

[0009] S3, puts the multi-channel interferogram containing continuous terrain and sudden terrain to be inverted into the trained multi-channel feature connection and fusion network model to obtain the inverted elevation image.

[0010] Furthermore, the S1 includes the following steps:

[0011] S1-1, construct a 3×3 to 18×18 random initial matrix through a random function, and use the double-triple interpolation method to interpolate and enlarge the initial matrix to 256 pixels × 256 pixels to obtain a simulated continuous terrain elevation map I;

[0012] S1-2, obtain ASTER GDEMV2 elevation image data with an accuracy of 30m from the geospatial data cloud, select the real elevation map of Sichuan Province, China, where the terrain is relatively continuous and smooth, and intercept the elevation matrix with a size of 256 pixels × 256 pixels to obtain the real continuous terrain elevation map II;

[0013] S1-3, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size of which does not exceed the size of the image itself, and then by setting the local areas to zero, a sudden change terrain elevation map III is obtained, the size of which is 256 pixels × 256 pixels;

[0014] S1-4, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size of which does not exceed the size of the image itself, and then re-assigning values to the local areas to obtain a sudden change terrain elevation map IV, the size of which is 256 pixels × 256 pixels;

[0015] S1-5, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size does not exceed the size of the image itself, and then the sudden terrain elevation is obtained by setting the local areas to zero and re-assigning values. Figure V , size is 256 pixels × 256 pixels;

[0016] S1-6, simulated continuous terrain elevation map I, real continuous terrain elevation map II, sudden change terrain elevation map III, sudden change terrain elevation map IV and sudden change terrain elevation map obtained from S1-1-S1-5 Figure V As the entire elevation map, three true phase maps are generated, and then the three true phase maps are re-wrapped and noise is added to obtain three noisy wrapped phase maps, namely multi-channel interferograms;

[0017] S1-7, simulated continuous terrain elevation map I, real continuous terrain elevation map II, sudden change terrain elevation map III, sudden change terrain elevation map IV, sudden change terrain elevation map Figure V The multi-channel interferograms generated by S1-6 constitute a data set containing multi-channel interferograms and single-channel elevation maps.

[0018] Furthermore, the multi-channel feature connection and fusion network model described in S2 has a main structure divided into three horizontal channels and four vertical network blocks.

[0019] The three lateral channels are symmetrically structured, receiving three interferograms and forming their own interferogram processing channels. From a horizontal perspective, the three channels are parallel network layers with the same network structure, and can extract useful information from the three-channel interferograms separately.

[0020] Four vertical network blocks connect different interference map channels and integrate them to finally output an elevation map. Vertically, the first three network blocks serve as the main three-channel interference map processing modules. In one network block, a convolutional layer, a residual network layer embedded with an attention mechanism, and a convolutional layer are used as processing units for a channel. All three channels use this unit to process different interference map information, and multi-channel feature fusion is formed by merging the channels. The three network blocks realize multi-channel cascade and dense connection through serial and skip connections, which enhances the reuse of feature maps and the connection between different interference map channels. The fourth network block is three consecutive convolutional layers, which integrate and process the feature maps and output them as a single-channel elevation image.

[0021] The S2 comprises the following steps:

[0022] S2-1, the multi-channel interference map obtained in S1 enters the input layer of the multi-channel feature connection and fusion network model, adjusts the number of image channels through the multi-channel convolution layer, passes through the residual network layer embedded with the attention mechanism, and then outputs through the convolution layer to obtain a multi-channel feature map;

[0023] S2-2, the multi-channel feature map obtained by S2-1 is channel-merged, and its output is used as the input of the next multi-channel convolutional layer and jump-connected to the next channel-merging layer to form a dense connection;

[0024] S2-3, serialize S2-1 and S2-2 as a network block, repeat this network block to form three main multi-channel interference image processing network blocks, and the multi-channel feature map passes through these network blocks in turn and is output;

[0025] S2-4: The feature map obtained in S2-3 is passed through three ordinary convolutional layers to reduce the number of channels to 1, and a single-channel elevation map is obtained as the network output. The error between the feature map and the true elevation is calculated using the mean squared error (MSE) loss function, and the network parameters are updated after backpropagation.

[0026] S2-5, train the entire data set. After 18-20 hours of training, the trained network weights are obtained.

[0027] Furthermore, in S3, the multi-channel interference map to be inverted containing continuous terrain and sudden terrain enters the first vertical network block from the input layer, wherein the three horizontal parallel channel layers process three entangled phase maps respectively. Each channel uses the convolution layer, the residual network layer embedded with the attention mechanism and the convolution layer as processing units to process the corresponding entangled phase map. Finally, the channel merging layer performs multi-channel merging, and its output is passed to the next vertical network block and jump-connected to the next channel merging layer to form a dense connection. After the three main vertical network blocks process the multi-channel interference map, it passes through the fourth vertical network block, i.e., the last three serial convolution layers to output the elevation image.

[0028] Advantages or beneficial effects of the present invention:

[0029] The present invention uses a multi-channel feature connection and fusion network as a framework to construct a multi-channel dataset suitable for continuous and sudden terrain, realizing a direct mapping relationship from multiple interference maps to elevation maps. The multi-channel feature connection and fusion network adopts a residual unit integrated with an attention mechanism, which improves the attention to important channels while protecting the integrity of the interference map information. It uses a dense connection method to realize multi-channel feature reuse, strengthens the propagation of feature maps, and improves the network's ability to fuse information from multiple interference maps. The trained network can quickly and efficiently perform elevation inversion on multiple interference maps to obtain their predicted elevation values. Experiments on interference map elevation inversion for different types of continuous and sudden terrains show that this method has good robustness and high inversion accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1a Schematic diagram of multi-channel feature connection and fusion network model training in an embodiment of the present invention;

[0031] Figure 1b Schematic diagram of inputting multiple interferograms to be inverted into a trained network model to obtain inversion results in an embodiment of the present invention;

[0032] Figure 2aThe initial matrix is expanded and interpolated to 256 pixels × 256 pixels using the bicubic interpolation method in the embodiment of the present invention to obtain a continuous terrain elevation map and three noisy interferograms; the first image on the left is the continuous terrain elevation map, and the second, third, and fourth images are noisy interferograms;

[0033] Figure 2b In the embodiment of the present invention, a continuous terrain elevation matrix of 256 pixels × 256 pixels and three noisy interferograms are extracted from a real elevation map of part of Sichuan Province, China. The first image on the left is the continuous terrain elevation matrix, and the second, third, and fourth images are noisy interferograms.

[0034] Figure 2c In the embodiment of the present invention, Figure 2a and Figure 2b The sudden change terrain elevation map and three noise interference maps after processing of the randomly selected continuous terrain elevation map; the first one on the left is the sudden change terrain elevation map, and the second, third and fourth ones are noise interference maps;

[0035] Figure 2d In the embodiment of the present invention, Figure 2a and Figure 2b The sudden change terrain elevation map and three noise interference maps after processing of the randomly selected continuous terrain elevation map; the first one on the left is the sudden change terrain elevation map, and the second, third and fourth ones are noise interference maps;

[0036] Figure 2e In the embodiment of the present invention, Figure 2a and Figure 2b The sudden change terrain elevation map and three noise interference maps after processing of the randomly selected continuous terrain elevation map; the first one on the left is the sudden change terrain elevation map, and the second, third and fourth ones are noise interference maps;

[0037] Figure 2a to Figure 2e The horizontal axis represents the horizontal size of the image, the vertical axis represents the vertical size of the image, and the chromaticity bar on the right represents the value range of the image;

[0038] Figure 3 This is a diagram of the multi-channel feature connection and fusion network elevation inversion model in an embodiment of the present invention, where Conv is the convolution layer + batch normalization + activation function ReLU, SE-Res is the residual unit embedded in the attention mechanism, concat is channel merging, Block is the vertical network block; Channel is the channel, and DEM is the elevation;

[0039] Figure 4ais the residual unit model diagram, where 1×1 and 3×3 are the convolution kernel sizes, BN is batch normalization, ReLU is the activation function, and add is the "add" operation; Feature Map in represents the input feature map, and Feature Map out represents the output feature map;

[0040] Figure 4b is the attention mechanism model diagram; ReLU is the activation function, Sigmoid is the activation function;

[0041] Figure 4c This is an example diagram of a residual unit model embedded with an attention mechanism, where 1×1 and 3×3 are convolution kernel sizes, BN is batch normalization, ReLU is the activation function, SE is the attention mechanism module, and add is the "add" operation;

[0042] Figure 5 It is a dense connection method diagram, where Dense connection is dense connection and channel-wise connection is channel splicing. Indicates channel merging. DETAILED DESCRIPTION

[0043] The present invention is further described below with reference to the embodiments and accompanying drawings. It is apparent that the embodiments described are only a portion of the embodiments of the present invention, not all of them. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort are intended to fall within the scope of protection of the present invention.

[0044] Example:

[0045] The basic process of the multi-channel InSAR elevation inversion method based on deep learning proposed in the present invention is described below with reference to the accompanying drawings.

[0046] The multi-channel InSAR elevation inversion method model based on deep learning is as follows: Figure 1a and Figure 1b As shown, Figure 1a The diagram of network training is shown in Figure 2. The mapping relationship between multi-channel interferogram and single-channel elevation map is established through the training data set, and the trained network model is obtained. The test map containing different types to be inverted is input into the trained network model to obtain the elevation inversion result, such as Figure 1b shown.

[0047] The multi-channel InSAR elevation inversion method based on deep learning includes the following steps:

[0048] S1) Create single-channel elevation maps of continuous terrain and sudden terrain, and form multi-channel interferograms from the elevation maps of the two terrains, thereby obtaining a total of 24,000 data sets containing multi-channel interferograms and single-channel elevation maps;

[0049] S1) comprises the following steps:

[0050] S1-1), constructing a 3×3 to 18×18 random initial matrix through a random function, and using the double-triple interpolation method to expand the initial matrix to 256 pixels × 256 pixels to obtain a simulated continuous terrain elevation map I;

[0051] S1-2) 30m-accurate ASTER GDEMV2 elevation image data were obtained from the geospatial data cloud. A real elevation map of Sichuan Province, China, with relatively continuous and smooth terrain, was selected from the image data. An elevation matrix with a size of 256 pixels × 256 pixels was then intercepted to obtain the real continuous terrain elevation map II.

[0052] S1-3) , by selecting multiple local areas from the two continuous terrain elevation maps of S1-1) and S1-2), the number of the areas is generally 5-6, and the total size does not exceed the size of the image itself, and then the local areas are set to zero to obtain a sudden change terrain elevation map III, the size of which is 256 pixels × 256 pixels;

[0053] S1-4), by selecting multiple local areas from the two continuous terrain elevation maps of S1-1) and S1-2), the number of the areas is generally 5-6, and the total size does not exceed the size of the image itself, and then re-assigning the values of the local areas to obtain a sudden change terrain elevation map IV, the size of which is 256 pixels × 256 pixels;

[0054] S1-5) selects multiple local areas from the two continuous terrain elevation maps of S1-1) and S1-2), the number of areas is generally 5-6, and the total size does not exceed the size of the image itself, and then obtains the sudden terrain elevation by setting the local areas to zero and re-assigning values. Figure V , size is 256 pixels × 256 pixels;

[0055] S1-6), simulated continuous terrain elevation map I, real continuous terrain elevation map II, sudden change terrain elevation map III, sudden change terrain elevation map IV and sudden change terrain elevation map obtained by S1-1)-S1-5). Figure V As the entire elevation map, three true phase maps are generated, and then the three true phase maps are re-wrapped and noise is added to obtain three noisy wrapped phase maps, namely multi-channel interferograms;

[0056] S1-7), simulated continuous terrain elevation map I, real continuous terrain elevation map II, sudden change terrain elevation map III, sudden change terrain elevation map IV, sudden change terrain elevation map Figure V The multi-channel interferograms generated by (S1-6) and (S1-6) constitute a dataset containing multi-channel interferograms and single-channel elevation maps.

[0057] Specifically, in the multi-channel InSAR interferometry application, the true height h, the true interferometric phase φ and the actual observed phase (i.e. winding phase) has the following relationship:

[0058] h=κ i φ i ,i=1,2,…n (1)

[0059]

[0060] Wherein, κ is the height sensitivity, which characterizes the relationship between the true elevation and true phase of the observed target and is determined by the antenna system parameters. In the method of the present invention, the antenna system parameters are: slant range 840 km, incident angle 19°, wavelength 0.057 m, long baseline length 415 m, medium baseline length 272 m, short baseline length 151 m; j represents a complex number, jφ represents a complex phase, That is, the winding phase; i represents the number of channels. In this embodiment, three channels are selected, that is, i=3 is selected in formulas (1) and (2); when constructing the data set required for the network model, the original elevation map is first generated, and three groups of different height sensitivities are obtained using the antenna system parameters and formula (1), thereby obtaining three groups of true interference phase maps, and then formula (2) is used to obtain three winding phase maps corresponding to the original elevation map, and then Gaussian noise with a standard deviation of 0.5 is added to these three winding phase maps to obtain noisy winding maps. Five types of single-channel elevation maps and multi-channel interference maps are shown as follows: Figure 2a 、 Figure 2b 、 Figure 2c 、 Figure 2d and Figure 2e As shown, it is generated as follows:

[0061] like Figure 2a As shown, a random initial matrix of 3×3 to 18×18 is constructed through a random function, and the initial matrix is expanded and enlarged to 256 pixels × 256 pixels using a double-triple interpolation method. The phase range of the label image is 0 to 200 radians;

[0062] like Figure 2b As shown, a matrix of 256 pixels × 256 pixels was intercepted from the real elevation map of some areas in Sichuan Province, China, and the phase range of the label image is 0 to 200 radians;

[0063] like Figure 2c As shown, through Figure 2a and Figure 2b The two continuous terrain elevation maps are used to select multiple local areas, usually 5-6, with a total size not exceeding the size of the image itself. Then, the sudden terrain elevation map III is obtained by setting the local areas to zero. The size is 256 pixels × 256 pixels, and the phase range of the label image is 0 to 200 radians.

[0064] like Figure 2d As shown, through Figure 2a and Figure 2b The two continuous terrain elevation maps are used to select multiple local areas, usually 5-6, with a total size not exceeding the size of the image itself. Then, the sudden terrain elevation map IV is obtained by re-assigning values to the local areas. The size is 256 pixels × 256 pixels, and the phase range of the label image is 0 to 200 radians.

[0065] like Figure 2e As shown, through Figure 2a and Figure 2b The two continuous terrain elevation maps select multiple local areas, generally 5-6 areas, with a total size not exceeding the size of the image itself, and then obtain the sudden terrain elevation by setting the local area to zero and re-assigning the value Figure V , the size is 256 pixels × 256 pixels, and the phase range of the label image is 0 to 200 radians;

[0066] Then, three interference patterns corresponding to the elevation map are generated by formula (1) and formula (2) and noise is added. The image size in the dataset is 256 pixels × 256 pixels, and the phase range of the label image is 0 to 200 radians.

[0067] The above dataset contains 20,000 training datasets and 4,000 validation datasets; Figure 2a-2e They are respectively the data sets of five data generation methods according to S1), among which Figure 2a The first image on the left is the elevation map generated according to the S1-1) method, and the three images on the right are the corresponding multi-channel interferograms generated according to the S1-6) method; Figure 2b The first image on the left is the elevation map generated according to the S1-2) method, and the three images on the right are the corresponding multi-channel interferograms generated according to the S1-6) method; Figure 2c The first image on the left is the elevation map generated according to the S1-3) method, and the three images on the right are the corresponding multi-channel interferograms generated according to the S1-6) method; Figure 2d The first image on the left is the elevation map generated according to the S1-4) method, and the three images on the right are the corresponding multi-channel interferograms generated according to the S1-6) method; Figure 2eThe first image on the left is the elevation map generated according to the S1-5) method, and the three images on the right are the corresponding multi-channel interferograms generated according to the S1-6) method; Figure 2a-2e All single-channel elevation maps and multi-channel interferograms are all dataset types in S1-7);

[0068] S2) Put the data set obtained in S1) into the multi-channel feature connection and fusion network model for training to obtain the trained weights;

[0069] S2) comprises the following steps:

[0070] S2-1), the multi-channel interference map obtained in S1) enters the input layer of the multi-channel feature connection and fusion network model, passes through the multi-channel convolution layer, passes through the residual network layer embedded with the attention mechanism, and then passes through the convolution layer output to obtain a multi-channel feature map; S2-2), the multi-channel feature map obtained in S2-1) is channel-merged, and its output serves as the input of the next multi-channel convolution layer and jump-connected to the next channel merging layer to form a dense connection;

[0071] S2-3), serialize S2-1) and S2-2) as a network block, repeat this network block to form three main multi-channel interference image processing network blocks, and the multi-channel feature map passes through these network blocks in turn and is output;

[0072] In step S2-4), the feature map obtained in step S2-3) is passed through three ordinary convolutional layers to reduce the number of channels to 1, thereby obtaining a single-channel elevation map. In step S2-5), the entire dataset is trained. After 19.2 hours of training, the trained network weights are obtained.

[0073] The residual module and attention mechanism are introduced as follows:

[0074] As a long-lasting network model in neural networks, the residual network is easy to optimize and has been applied to many complex scenarios due to its characteristics of increasing depth to improve accuracy. Figure 4a This is a bottleneck residual network, where conv represents a standard convolutional layer, 1×1 and 3×3 represent the convolution kernel sizes, BN stands for batch normalization, and ReLU is the activation function, also known as the linear rectification function. The feature map first passes through a convolutional module with kernels of 1×1, 3×3, and 3×3, and then is combined with the initial input through the activation function to produce the output. The residual module prevents network degradation and improves network performance.

[0075] Attention mechanism such as Figure 4bAs shown, it can allocate computing resources to more important tasks, focusing on the information that is more critical to elevation inversion among the numerous interferogram information, greatly filtering out irrelevant information, and improving the efficiency and accuracy of elevation inversion. It can be divided into three modules based on computational requirements. The compression part is a global pooling layer that compresses the H×W×C feature map into a 1×1×C real number sequence feature according to the spatial dimension. The real number weight of each feature channel is adjusted to increase the focus on important channels. The excitation part utilizes the channel information aggregated in the previous step to learn the correlation between different interferogram channels. The two fully connected layers can limit model complexity and assist generalization. The restoration operation applies channel weights to the original input feature map, restoring the feature map size while strengthening the output of important feature maps.

[0076] The residual module embedded with the attention mechanism is as follows Figure 4c As shown in the figure, overall, the module maintains the basic structure of the bottleneck residual unit, and the attention mechanism module is directly connected to the bottleneck residual unit, which can adjust the network's ability to pay attention to different interference map channels and enhance the network's interference map feature extraction ability.

[0077] Densely connected methods such as Figure 5 As shown in Figure 1, its characteristic is that it achieves feature reuse through the continuous cascading of feature maps, closely connecting the forward and backward channels, and further realizing a mechanism for aggregating interference fringe features and elevation detail information. These characteristics enable the network to significantly reduce the number of parameters while maintaining model accuracy and stability, alleviate the gradient vanishing problem, and enhance the transfer of feature maps.

[0078] Furthermore, the multi-channel feature connection and fusion network elevation inversion model described in S2) is as follows: Figure 3 As shown in the figure, the structure is simple and effective. Each horizontal channel processes an interference pattern, and each vertical network block connects and integrates the detailed information of different interference pattern channels. This structure can collaboratively and efficiently process the information of three interference patterns and ultimately accurately restore the true elevation.

[0079] The three horizontal channels, channel 1, channel 2, and channel 3, are symmetrically structured, receiving three interferograms and forming their own interferogram processing channels. The four vertical network blocks, Block 1, Block 2, Block 3, and Block 4, connect and integrate the different interferogram channels to ultimately output an elevation map. Horizontally, channel 1, channel 2, and channel 3 are parallel network layers, each with the same parallel network block structure, capable of extracting useful information from the three-channel interferograms. Vertically, Block 1, Block 2, and Block 3 serve as the primary three-channel interferogram processing module. Taking Block 1 as an example, the convolutional layer (convolution + batch normalization + ReLU activation function), the residual unit (SE-Res) with an embedded attention mechanism, and the convolutional layer (convolutional layer) serve as a processing unit for a channel. The three channels in Block 1 process different interferogram information using this unit. Channel merging (concat) forms a channel-by-channel merge. The output of the concat layer serves as the input to Block 2 and forms dense connections with the concat layer of Block 2. The three different blocks form a multi-channel cascade, which strengthens the reuse of feature maps and the connection between different interference map channels. Block 4 is a series of three Conv layers that integrate and process the feature maps output by Block 3 and output the final elevation map;

[0080] S3) The multi-channel interferogram to be inverted, which contains continuous terrain and sudden terrain, is placed into the trained multi-channel feature connection and fusion network model to obtain the inverted elevation image.

[0081] In S3), the multi-channel interference map containing continuous terrain and sudden terrain to be inverted enters the first vertical network block from the input layer, where the three horizontal parallel channel layers process three entangled phase maps respectively. Each channel uses the convolution layer, the residual network layer embedded with the attention mechanism and the convolution layer as processing units to process the corresponding entangled phase map. Finally, the channel merging layer performs multi-channel merging, and its output is passed to the next vertical network block and jump-connected to the next channel merging layer to form a dense connection. After the three main vertical network blocks process the multi-channel interference map, it passes through the fourth vertical network block, that is, the last three serial convolution layers to output the elevation image.

[0082] The dataset for this example was generated using MATLAB 2018b simulation software. The model development platform was Python 3.7, using the Tensorflow 1.13.0 framework and Keras 2.3.0. The computer parameters used for network model training and experimental testing were as follows: an NVIDIA GeForce RTX 2080Ti GPU, a Xeon-W2145 CPU, and 64GB of RAM. Inverting a single multi-channel interferogram took 0.03 seconds.

[0083] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A multi-channel InSAR elevation inversion method based on deep learning, characterized in that: The following steps are involved: S1, creating single-channel elevation maps of continuous terrain and sudden terrain, forming a multi-channel interferogram from the elevation maps of the two terrains, thereby obtaining a data set containing a multi-channel interferogram and a single-channel elevation map; S2, puts the data set obtained in S1 into the multi-channel feature connection and fusion network model for training to obtain the trained weights; The multi-channel feature connection and fusion network model described in S2 has a main structure divided into three horizontal channels and four vertical network blocks. The three lateral channels are symmetrically structured, receiving three interferograms and forming their own interferogram processing channels. From a horizontal perspective, the three channels are parallel network layers with the same network structure, extracting useful information from the three-channel interferograms respectively. Four vertical network blocks connect and integrate different interference map channels to ultimately output an elevation map. Vertically, the first three of the four network blocks serve as the main three-channel interference map processing modules. Within each network block, a convolutional layer, a residual network layer embedded with an attention mechanism, and a convolutional layer serve as a processing unit for each channel. All three channels process different interference map information using this unit, and the channels are merged to form multi-channel feature fusion. Serial and skip connections are used between the three network blocks to achieve multi-channel cascading and dense connections, enhancing the reuse of feature maps and the connection between different interference map channels. The fourth network block consists of three consecutive convolutional layers that integrate and process the feature maps, outputting a single-channel elevation image. S3, puts the multi-channel interferogram containing continuous terrain and sudden terrain to be inverted into the trained multi-channel feature connection and fusion network model to obtain the inverted elevation image.

2. The multi-channel InSAR elevation inversion method based on deep learning according to claim 1, characterized in that: S1 includes the following steps: S1-1, construct a 3×3~18×18 random initial matrix through random function, and use the double triple interpolation method to interpolate and enlarge the initial matrix to 256 pixels × 256 pixels to obtain the simulated continuous terrain elevation map I; S1-2, obtain ASTER GDEMV2 elevation image data with an accuracy of 30m from the geospatial data cloud, select the real elevation map of Sichuan Province, China, where the terrain is relatively continuous and smooth, and intercept the elevation matrix with a size of 256 pixels × 256 pixels to obtain the real continuous terrain elevation map II; S1-3, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size of which does not exceed the size of the image itself, and then by setting the local areas to zero, a sudden change terrain elevation map III is obtained, the size of which is 256 pixels × 256 pixels; S1-4, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size of which does not exceed the size of the image itself, and then re-assigning the values of the local areas to obtain the sudden change terrain elevation map IV, the size of which is 256 pixels × 256 pixels; S1-5, by selecting multiple local areas from the two continuous terrain elevation maps S1-1 and S1-2, the total size of which does not exceed the size of the image itself, and then by setting the local areas to zero and re-assigning values to obtain a sudden change terrain elevation map V, the size of which is 256 pixels × 256 pixels; S1-6, using the simulated continuous terrain elevation map I, the real continuous terrain elevation map II, the sudden change terrain elevation map III, the sudden change terrain elevation map IV and the sudden change terrain elevation map V obtained from S1-1 to S1-5 as all elevation maps, generating three real phase maps, and then re-entering the three real phase maps and adding noise to obtain three noisy entangled phase maps, i.e., multi-channel interferograms; S1-7, a data set consisting of the simulated continuous terrain elevation map I, the real continuous terrain elevation map II, the sudden change terrain elevation map III, the sudden change terrain elevation map IV, the sudden change terrain elevation map V generated by S1-1-S1-5 and the multi-channel interferogram generated by S1-6, contains multi-channel interferograms and single-channel elevation maps.

3. The multi-channel InSAR elevation inversion method based on deep learning according to claim 1, characterized in that: S2 includes the following steps: S2-1, the multi-channel interference map obtained in S1 enters the input layer of the multi-channel feature connection and fusion network model, passes through the multi-channel convolution layer, and then passes through the residual network layer embedded with the attention mechanism, and then outputs through the convolution layer to obtain a multi-channel feature map; S2-2, the multi-channel feature map obtained by S2-1 is channel-merged, and its output is used as the input of the next multi-channel convolutional layer and jump-connected to the next channel-merging layer to form a dense connection; S2-3, serialize S2-1 and S2-2 as a network block, repeat this network block to form three main multi-channel interference image processing network blocks, and the multi-channel feature map passes through these network blocks in turn and is output; S2-4, the feature map obtained in S2-3 is passed through three ordinary convolutional layers, the number of channels is reduced to 1, and a single-channel elevation map is obtained as the network output. The error between the map and the true elevation is calculated through the mean square error of the loss function, and the network parameters are updated after back propagation; S2-5, train the entire data set. After 18-20 hours of training, the trained network weights are obtained.

4. The multi-channel InSAR elevation inversion method based on deep learning according to claim 1, characterized in that: In S3, the multi-channel interference map containing continuous terrain and sudden terrain to be inverted enters the first vertical network block from the input layer, where the three horizontal parallel channel layers process three entangled phase maps respectively. Each channel uses the convolution layer, the residual network layer embedded with the attention mechanism and the convolution layer as processing units to process the corresponding entangled phase map. Finally, the channel merging layer performs multi-channel merging, and its output is passed to the next vertical network block and jump-connected to the next channel merging layer to form a dense connection. After the three main vertical network blocks process the multi-channel interference map, it passes through the fourth vertical network block, that is, the last three serial convolution layers to output the elevation image.

Images