CN-116879896-B - Method, system and device for reconstructing multi-channel interferogram elevation

CN116879896BCN 116879896 BCN116879896 BCN 116879896BCN-116879896-B

Abstract

The invention discloses a multi-channel interferogram elevation reconstruction method based on a dual-channel network, which comprises the steps of S1, constructing the dual-channel network consisting of an elevation reconstruction flow and a boundary detection flow, S2, constructing a data set with different topographic features, defining an optimal loss function and training the dual-channel network, S3, inputting the multi-channel interferogram into the trained dual-channel network, and establishing a dual-mapping relation among the multi-channel interferogram, the elevation graph and the boundary graph to finish the multi-channel interferogram elevation reconstruction. The invention can realize the multi-channel interference pattern elevation reconstruction based on the double-flow network.

Inventors

CAI CHANGQING
GAO MENGTIAN
XU HAOWEN
LIANG ZENGXIAN
WU JIAZHEN
YU XIAOWEN
LI DANHONG
LIANG MAOLIN
Lu Siya

Assignees

广州大学

Dates

Publication Date: 20260505
Application Date: 20230711

Claims (10)

1. A method for multi-channel interferogram elevation reconstruction based on a dual-flow network, comprising: s1, constructing a double-flow network consisting of an elevation reconstruction flow and a boundary detection flow; S2, constructing a data set with different topographic features, defining an optimal loss function and training a double-flow network; S3, inputting the multichannel interferograms into a trained double-flow network, establishing a double-mapping relation between the multichannel interferograms and the elevation and boundary diagrams, and completing elevation reconstruction of the multichannel interferograms; The boundary detection flow mainly consists of residual blocks, and the residual blocks detect boundary information of the observed topography and combine the boundary information into the elevation reconstruction flow.
2. The method of claim 1, wherein S1 comprises constructing a dual stream network of an elevation reconstruction stream and a boundary detection stream, the dual stream network comprising an encoder and a decoder.
3. The method according to claim 2, wherein the encoder comprises a deep convolutional neural network, a shrunken spatial pyramid pool, a spatial attention module and a convolution layer which are sequentially connected, the deep convolutional neural network is used for extracting different scale feature maps of high-level semantic information in the multi-channel interference map, the shrunken spatial pyramid pool is used for fusing the feature maps of different scales, the spatial attention module is used for enhancing the attention to the spatial information of the fused feature map output by the shrunken spatial pyramid pool, the convolution layer of a 1×1 convolution+batch normalization+rectification linear unit is used for adjusting the channel number of the output feature map, and the encoder outputs 256 channel feature maps with high semantic information with a resolution of 16×16.
4. A method according to claim 3, characterized in that the decoder comprises in particular: In the decoding stage, the deep convolutional neural network outputs a low-level characteristic diagram, the low-level characteristic diagram is adjusted through 1×1 convolution +BN +ReLU operation, 128 channel characteristic diagrams with the size of 64×64 are obtained after adjustment, 256 channel characteristic diagrams with the size of 64×64 are obtained after 4+1×1 convolution up-sampling operation is carried out on the characteristic diagrams with high semantic information output by an encoder, the channel characteristic diagrams are further processed by a space attention module and a 1×1Conv+BN +ReLU unit after being cascaded with the output characteristic diagrams, 256 channel characteristic diagrams with the size of 256×256 are obtained through a 4+1×1 convolution up-sampling unit after being processed, then 32 channel characteristic diagrams with phase details and edge information output by a residual block unit in a boundary detection stream are cascaded with the residual block characteristic diagrams, and the elevation diagram of an observation topography corresponding to a wrapped phase image input from the network is further processed by a 3×3Conv+BN +ReLU unit, and reconstruction is completed.
5. A system for multi-channel interferogram elevation reconstruction based on a dual-flow network, comprising: The double-flow network module is used for constructing a double-flow network consisting of an elevation reconstruction flow and a boundary detection flow; the training module is used for constructing a data set with different topographic features, defining an optimal loss function and training a double-flow network; The reconstruction module is used for inputting the multichannel interferogram into a trained double-flow network, establishing a double-mapping relation between the multichannel interferogram and the elevation map and between the multichannel interferogram and the boundary map, and completing elevation reconstruction of the multichannel interferogram; The boundary detection flow mainly consists of residual blocks, and the residual blocks detect boundary information of the observed topography and combine the boundary information into the elevation reconstruction flow.
6. The system of claim 5, wherein the dual-stream network module is specifically configured to construct a dual-stream network comprised of an elevation reconstruction stream and a boundary detection stream, the dual-stream network comprising an encoder and a decoder.
7. The system of claim 6, wherein the encoder comprises a deep convolutional neural network, a collapsed spatial pyramid pool, a spatial attention module and a convolution layer, which are sequentially connected, wherein the deep convolutional neural network is used for extracting different scale feature maps of high-level semantic information in the multi-channel interferograms, the collapsed spatial pyramid pool is used for fusing the feature maps of different scales, the spatial attention module is used for enhancing the attention to the spatial information of the fused feature maps output by the collapsed spatial pyramid pool, the convolution layer of a 1X 1 convolution+batch normalization+rectification linear unit is used for adjusting the channel number of the output feature maps, and the encoder outputs 256 channel feature maps with high semantic information with a resolution of 16X 16.
8. The system of claim 7, wherein the decoder comprises: In the decoding stage, the deep convolutional neural network outputs a low-level characteristic diagram, the low-level characteristic diagram is adjusted through 1×1 convolution +BN +ReLU operation, 128 channel characteristic diagrams with the size of 64×64 are obtained after adjustment, 256 channel characteristic diagrams with the size of 64×64 are obtained after 4+1×1 convolution up-sampling operation is carried out on the characteristic diagrams with high semantic information output by an encoder, the channel characteristic diagrams are further processed by a space attention module and a 1×1Conv+BN +ReLU unit after being cascaded with the output characteristic diagrams, 256 channel characteristic diagrams with the size of 256×256 are obtained through a 4+1×1 convolution up-sampling unit after being processed, then 32 channel characteristic diagrams with phase details and edge information output by a residual block unit in a boundary detection stream are cascaded with the residual block characteristic diagrams, and the elevation diagram of an observation topography corresponding to a wrapped phase image input from the network is further processed by a 3×3Conv+BN +ReLU unit, and reconstruction is completed.
9. An apparatus for multi-channel interferogram elevation reconstruction based on a dual-flow network, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, which when executed by the processor, performs the steps of the method for multi-channel interferogram elevation reconstruction based on a dual-flow network as in any one of claims 1 to 4.
10. A computer readable storage medium, characterized in that it has stored thereon a program for realizing information transfer, which when executed by a processor, realizes the steps of the method for multi-channel interferogram elevation reconstruction based on a dual-flow network according to any one of claims 1 to 4.

Description

Method, system and device for reconstructing multi-channel interferogram elevation Technical Field The invention relates to the field of interference pattern reconstruction, in particular to a method, a system and a device for reconstructing a multichannel interference pattern elevation. Background Interferometric synthetic aperture radar (InSAR) is capable of obtaining elevation information of a terrain surface by applying a known relationship between a phase difference obtained from two SAR images and a terrain elevation value, independent of weather conditions and sunlight, and is widely used in various fields including terrain mapping, surface deformation monitoring, and detection of special targets. The phase difference obtained from the two SAR images is also commonly referred to as the unwrapped phase, which is mapped to the dominant value interval (-pi, pi) by modulo 2 pi mapping due to the inverse trigonometric operation of the complex InSAR signal. Phase Unwrapping (PU) is one of the key technologies for interferometric SAR data processing, playing a very important role in the application of InSAR technology. Single-base line PU methods, represented by Goldstein branch cut methods, quality-oriented methods, and network-flow methods, generally work well with strict adherence to the phase-continuity assumption (i.e., the absolute value of the phase difference between a pixel and its neighbors is less than pi). However, the phase continuity assumption required for a single baseline PU is not always satisfied due to the presence of discontinuities and/or interference noise, such as in steep mountains, cliffs, and urban areas, in which case the PU measuring the wrapped phase becomes an uncomfortable problem. The multi-baseline PU eliminates the limitation of phase continuity assumption which a single-baseline PU must follow, and can effectively solve the problem of elevation reconstruction of the terrain with larger gradient change. In the last two decades, multi-channel InSAR elevation reconstruction algorithms based on different strategies (i.e., multi-baseline PU algorithm in the present study) have been sequentially proposed, including Chinese Remainder Theorem (CRT) method [8], least Squares (LS), maximum likelihood Method (ML), maximum A Posteriori (MAP), cluster Analysis (CA) method, subspace Projection (SPJ) method, two-stage planning method (TSPA), and nonlinear filter based state estimation method, etc. The CRT method firstly constructs a congruence equation of the phase related to the length of the interference base line, and then solves the equation by utilizing the Chinese remainder theorem to obtain the phase corresponding to the observed terrain height value. CRT methods can generally accurately reconstruct the elevation of an observed terrain provided that the interference baseline length strictly meets the relevant conditions required for the method, but their performance is severely limited by the noise present in the interferogram. In order to improve the anti-noise performance of the CRT method, an improved CRT method is reported therein. The multi-baseline LS method can be seen as an extension of the traditional single-baseline LS method, by averaging the phases weighted by the respective baseline lengths under the LS criterion to obtain a global solution of the terrain elevation. The anti-noise performance of the multi-baseline LS method is generally better than that of the CRT method, however, when the elevation gradient is extracted from the plurality of interferograms corresponding to the interference baselines, the elevation reconstruction accuracy of the observed terrain obtained by the method will be severely reduced, and there will be a great deviation from the true elevation gradient of the observed terrain. The ML method regards the unwrapped phase/or terrain height as a parameter in a statistical distribution framework and establishes a probability density function of the unwrapped phase/or terrain height, and then achieves optimal estimation of the parameter by maximizing the probability density function constructed according to the ML criteria. The ML approach requires a large number of observation sources to obtain reliable results, which can greatly increase the cost of implementing a satisfactory multichannel InSAR system. Like the ML method, the MAP method is a markov statistical method based on a bayesian estimator, in which a priori model of the phase/altitude of the terrain is built using a markov random field. The MAP method reduces the dependence on the number of interferograms required to some extent compared to the ML method, and sometimes even if the number of interferograms is smaller than that required by the ML method, these methods can obtain reliable elevation estimation of the observed scene, but require higher computational complexity and more time consumption. The CA-based method clusters all pixels in the interferograms into different groups accord