A method for InSAR deformation field error correction that couples Fourier transform and deep learning
By coupling Fourier transform and deep learning, a deep neural network model was constructed, which solved the problems of accuracy and generalization of InSAR deformation field error correction, achieved high-precision deformation field error correction, and improved the accuracy and robustness of InSAR monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NAT INST OF NATURAL HAZARDS MINISTRY OF EMERGENCY MANAGEMENT OF CHINA
- Filing Date
- 2025-11-19
- Publication Date
- 2026-06-30
AI Technical Summary
In InSAR technology, traditional deformation field error correction methods have limitations in generalization, accuracy, and efficiency when dealing with complex terrain, atmospheric conditions with strong spatiotemporal variations, multi-source error coupling, and low coherence regions. They are difficult to effectively remove major error sources such as atmospheric delay errors, which affects the accuracy and reliability of deformation monitoring.
By employing a method that couples Fourier transform with deep learning, an optimized deep neural network model is constructed. The Fourier transform module is used to fuse frequency and spatial domain features, and the feature encoding and decoding stages are combined to achieve high-precision error correction of InSAR deformation fields.
It achieves high-precision, adaptive, and robust correction of InSAR deformation fields, improves the accuracy of deformation monitoring and the ability to cope with complex scenarios, and overcomes the limitations of traditional methods.
Smart Images
Figure CN121559508B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of remote sensing measurement and geoscientific information processing, and in particular to an InSAR deformation field error correction method that couples Fourier transform and deep learning. Background Technology
[0002] Interferometric Synthetic Aperture Radar (InSAR) is a remote sensing technology that uses spaceborne or airborne synthetic aperture radar systems to acquire high-precision information on surface deformation. The basic principle of InSAR is to extract minute surface deformation information at the centimeter or even millimeter level by comparing the phase difference between two or more SAR images of the same area acquired at different times. It has advantages such as wide coverage, high precision, cloud and fog penetration, and all-weather, all-day monitoring, and is widely used in monitoring land subsidence, geological hazards, crustal movement, volcanic activity, glacial movement, urban safety, and resource surveys.
[0003] Despite its numerous advantages, InSAR technology is susceptible to various errors that significantly impact the accuracy and reliability of the acquired surface deformation fields. The main error sources in InSAR include atmospheric delay error, orbital error, topographic error, phase unwrapping error, and decoherence error. These errors often mask the true, minute deformation signals, leading to misjudgments or insufficient accuracy in deformation monitoring results. Atmospheric delay error originates from the electromagnetic wave propagation delay caused by differences in atmospheric conditions between two SAR images. The phase component caused by atmospheric effects between the two observations is highly complex, exhibiting instability and randomness in magnitude and range. Atmospheric delay errors are generally divided into ionospheric delay and tropospheric delay. Ionospheric delay mainly occurs in low and high latitude regions, having a more severe impact on long-band SAR signals (such as L / P bands); while tropospheric delay occurs in the neutral atmosphere, is independent of radar signal frequency, and its impact is more widespread. Orbital error arises because the precision of satellite orbit determination data is limited, often resulting in residual interference fringes in the interference fringes. Residual topographic errors are topographically related phase errors caused by insufficient accuracy of the external digital elevation model (DEM) or surface variations related to SAR imaging time. Phase unwrapping errors arise from uncertainties in the process of recovering the absolute phase from the wrapped interferometric phase; incorrect unwrapping introduces integer ambiguity, leading to jumps or overall deviations in the deformation field. Temporal decoherence occurs because changes in surface scattering characteristics over time reduce interferometric coherence, making phase information unreliable or lost. Among these errors, atmospheric delay errors have high uncertainty and are currently difficult to simulate and remove using a deterministic method, making them a major error source for InSAR precision deformation measurements.
[0004] Traditional deformation field error correction methods typically rely on specific model assumptions, external auxiliary data, or complex parameter estimation processes. These methods often have limitations when dealing with complex terrain, highly spatiotemporally variable atmospheres, multi-source error coupling, low-coherence areas, or large-scale monitoring. For example, model simplification leads to poor adaptability, external data (such as GNSS and meteorological data) coverage is insufficient or lacks accuracy, parameter estimation processes are complex, and it is difficult to effectively separate coupled error sources. Therefore, deformation field error correction remains a significant challenge for InSAR precision measurements. Summary of the Invention
[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an InSAR deformation field error correction method that couples Fourier transform and deep learning. Through a specific deep learning architecture and data processing strategy, it achieves high-precision, adaptive, and robust correction of the main errors in the InSAR deformation field, overcoming the limitations of existing error correction methods in terms of generalization, accuracy, efficiency, and handling of complex scenarios.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A method for InSAR deformation field error correction that couples Fourier transform and deep learning includes:
[0008] The initial InSAR surface deformation field data and the topographic data of the same region are input into the optimized deep neural network model, and the surface deformation field results after error correction are output.
[0009] The training process of the optimized deep neural network model includes:
[0010] Based on differential interferometric synthetic aperture radar, interferometric data with a short time baseline are acquired, and an interferogram containing real noise characteristics is obtained as error data.
[0011] Acquire terrain data corresponding to the error data and resample it to the same spatial size as the error data;
[0012] Simulated surface deformation field data is generated based on a geophysical model within a preset parameter range and used as training output labels.
[0013] The error data is superimposed with the simulated surface deformation field data to construct a surface deformation field training input containing error signals, and together with the resampled terrain data, they form a training sample, thus completing the division of the training dataset and the test dataset.
[0014] A deep neural network model with Fourier transform modules as basic units is constructed. The model parameters of the deep neural network model are optimized using the training dataset. The optimized deep neural network model is evaluated and saved based on the test dataset.
[0015] Preferably, the Fourier transform module in the optimized deep neural network model is used to fuse frequency domain features and spatial domain features, and the processing flow of the Fourier transform module includes:
[0016] Perform a Fourier transform on the input features to obtain a frequency domain representation;
[0017] Frequency domain features are extracted by sequentially performing 3×3 convolution and 1×1 convolution in the frequency domain.
[0018] Perform an inverse Fourier transform on the frequency domain features to return to the spatial domain;
[0019] The original input is subjected to a 3×3 convolution in the spatial domain to extract spatial features;
[0020] The features returned to the spatial domain are fused with the spatial domain features for output.
[0021] Preferably, the processing flow of the deep neural network model includes a feature encoding stage and a feature decoding stage; in the feature encoding stage, the input features are downsampled step by step and the terrain features are fused; in the feature decoding stage, the features are upsampled step by step and fused with the features at the corresponding scale in the encoding stage through skip connections.
[0022] Preferably, the feature encoding stage includes:
[0023] The initial surface deformation and topography data of W×H×2 are input into the Fourier transform module and the max pooling layer to obtain the W / 2×H / 2×64 feature; where W represents the spatial width of the initial surface deformation field data and topography data on the unified spatial grid; and H represents the spatial height of the initial surface deformation field data and topography data on the unified spatial grid.
[0024] The W / 2×H / 2 terrain data is processed by the terrain feature extraction module to obtain the W / 2×H / 2×64 terrain features, and then fused with the W / 2×H / 2×64 features and processed by the Fourier transform module and max pooling to obtain the W / 4×H / 4×128 features.
[0025] The terrain data of W / 4×H / 4 is processed by the terrain feature extraction module to obtain W / 4×H / 4×128 terrain features, and then fused with the W / 4×H / 4×128 features and processed by the Fourier transform module and max pooling to obtain W / 8×H / 8×256 features.
[0026] The W / 8×H / 8×256 features are input into the Fourier transform module and the multi-scale fusion module, and the W / 8×H / 8×256 features are output.
[0027] Preferably, the feature decoding stage includes:
[0028] The W / 8×H / 8×256 features are upsampled to W / 4×H / 4×256 features, and then fused with the W / 4×H / 4×256 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the features are input into the Fourier transform module to obtain W / 4×H / 4×128 features.
[0029] The W / 4×H / 4×128 features are upsampled to W / 2×H / 2×128 features, and then fused with the W / 2×H / 2×128 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the features are input into the Fourier transform module to obtain W / 2×H / 2×64 features.
[0030] The W / 2×H / 2×64 features are upsampled to W×H×64 features, and then fused with the W×H×64 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the data is input into the Fourier transform module, and the output is W×H×1 error-corrected surface deformation field data.
[0031] Preferably, the terrain feature extraction module consists of a fusion structure of multiple 3×3 convolutions and 1×1 convolutions, used to extract terrain features at different scales and fuse them with features at the corresponding scales.
[0032] Preferably, the multi-scale fusion module extracts multi-scale features using dilated convolutions with dilation rates of 2, 4, and 6, and outputs W / 8×H / 8×256 features after stepwise fusion.
[0033] Preferably, the attention module includes channel attention and spatial attention, which are used to weight the importance of channels and spatial locations during the skip connection fusion process with the corresponding scale features in the feature encoding stage.
[0034] Preferably, the short time baseline used in the training process is 6 days or 12 days.
[0035] The present invention discloses the following technical effects:
[0036] This invention establishes a relationship model between the SAR satellite-observed surface deformation field and the actual surface deformation field, which can effectively and accurately separate the actual surface deformation from the observed mixed signals. It overcomes the limitations of traditional deformation error correction methods in terms of generalization, accuracy, efficiency, and handling of complex scenarios, and achieves high-precision error correction of the InSAR deformation field, thereby improving the accuracy of InSAR monitoring of surface deformation. Attached Figure Description
[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 A flowchart of the method provided in an embodiment of the present invention;
[0039] Figure 2 A flowchart illustrating the technical route provided in the embodiments of the present invention;
[0040] Figure 3 This is a structural diagram of the Fourier transform module provided in an embodiment of the present invention;
[0041] Figure 4 The diagram shows the structure of the deep neural network model FFT-NET with coupled Fourier transform module provided in this embodiment of the invention. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0044] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides an InSAR deformation field error correction method coupled with Fourier transform and deep learning, comprising:
[0045] Step 100: Input the initial InSAR surface deformation field data and the topographic data of the same area into the optimized deep neural network model, and output the surface deformation field results after error correction;
[0046] The training process of the optimized deep neural network model includes:
[0047] Step 101: Obtain interferometric data with a short time baseline based on differential interferometric synthetic aperture radar, and obtain an interferogram containing real noise characteristics as error data;
[0048] Step 102: Obtain the terrain data corresponding to the error data and resample it to the same spatial size as the error data;
[0049] Step 103: Generate simulated surface deformation field data within a preset parameter range based on the geophysical model, and use it as training output labels;
[0050] Step 104: Overlay the error data with the simulated surface deformation field data to construct a surface deformation field training input containing the error signal, and combine it with the resampled terrain data to form a training sample, and complete the division of the training dataset and the test dataset.
[0051] Step 105: Construct a deep neural network model with Fourier transform modules as the basic units, optimize the model parameters of the deep neural network model using the training dataset, and evaluate and save the optimized deep neural network model based on the test dataset.
[0052] In another embodiment of the invention, such as Figure 2 As shown, the technical approach of this embodiment includes the following steps:
[0053] Step 1: Using differential interferometric synthetic aperture radar (D-InSAR) technology, short-time baseline interferometry (6 days or 12 days) is used to obtain interferograms containing real noise features as error data;
[0054] Step 2: Based on the geographical location of the error data, obtain the corresponding terrain data and resample it to the same size as the error data;
[0055] Step 3: Based on the geophysical model, within the preset parameter range, generate simulated surface deformation data as real surface deformation field data, and use it as the output label for deep learning training;
[0056] Step 4: Overlay the error data from Step 1 and the surface deformation field data from Step 3 to construct a surface deformation field dataset containing errors. Combine this dataset with the terrain data from Step 2 and use it as input samples for deep learning training.
[0057] Step 5: Divide the input and output data into training and test datasets according to a certain ratio;
[0058] Step 6: Design the Fourier transform module as the basic module of the deep learning model, build the deep neural network model FFT-NET with the Fourier transform module coupled. The model input is the surface deformation field data and terrain data containing errors in Step 4, and the model output is the real surface deformation field data in Step 3.
[0059] Step 7: Train the deep neural network model FFT-NET built in Step 6 using the training dataset and optimize the model parameters;
[0060] Step 8: Evaluate the performance of the trained model using the test dataset and save the optimized deep neural network model;
[0061] Step 9: Use D-InSAR or other InSAR techniques to perform interferometric processing on the SAR image to obtain initial surface deformation field data;
[0062] Step 10: Input the initial surface deformation data and corresponding terrain data obtained in Step 9 into the deep neural network model FFT-NET, and output the error-corrected surface deformation field results.
[0063] In another embodiment of the invention, such as Figure 3 As shown, the Fourier transform module in this embodiment aims to fuse frequency domain features with spatial domain features. A detailed structural description is as follows:
[0064] Step 1: First, the input data is transformed to the frequency domain using a Fourier transform.
[0065] Step 2: Input the frequency domain data into the convolutional layers (3×3 convolution and 1×1 convolution) for frequency domain feature extraction;
[0066] Step 3: Perform inverse Fourier transform on the frequency domain features;
[0067] Step 4: Input data into a convolutional layer (3×3 convolution) for feature extraction to obtain spatial features;
[0068] Step 5: Fuse the spatial features with the features after inverse Fourier transform, and output the fused feature map.
[0069] In another embodiment of the invention, such as Figure 4 As shown, the structure of the deep neural network model FFT-NET with coupled Fourier transform modules is described below. The FFT-NET model mainly consists of two parts: feature encoding and feature decoding.
[0070] The feature encoding stage includes the following steps:
[0071] Step 1: Input the initial surface deformation and topographic data of W×H×2 into the Fourier transform module and the max pooling layer. After the max pooling layer, the output size is W / 2×H / 2×64.
[0072] Step 2: Input the terrain data of W / 2×H / 2 into the terrain feature extraction module to extract the terrain feature W / 2×H / 2×64, and fuse it with the output of Step 1 to obtain the feature W / 2×H / 2×128. Then input it into the Fourier transform module and the max pooling layer. After the max pooling layer, the output size is W / 4×H / 4×128.
[0073] Step 3: Input the terrain data of W / 4×H / 4 into the terrain feature extraction module to extract the terrain feature W / 4×H / 4×128, and fuse it with the output of Step 2 to obtain the feature W / 4×H / 4×256. Then input it into the Fourier transform module and the max pooling layer. After the max pooling layer, the output size is W / 8×H / 8×256.
[0074] Step 4: Input the output of Step 3 into the Fourier transform module and the multi-scale fusion module. The output size is W / 8×H / 8×256.
[0075] The terrain feature extraction module consists of multiple 3×3 and 1×1 convolutional layers fused together to extract terrain features; the multi-scale fusion module extracts features at different scales by dilated convolutions with dilation rates of 2, 4 and 6 respectively, and then fuses them step by step to extract multi-scale features.
[0076] In another embodiment of the present invention, the feature decoding stage includes the following steps:
[0077] Step 1: Input the feature map W / 8×H / 8×256 into the upsampling layer to obtain feature W / 4×H / 4×256. Then, input it together with the feature map W / 4×H / 4×256 from the feature encoding stage (corresponding to step 3) into the attention module to obtain feature W / 4×H / 4×256. Finally, after fusing the features, input them into the Fourier transform module, and the output size is W / 4×H / 4×128.
[0078] Step 2: Input the feature map W / 4×H / 4×128 into the upsampling layer to obtain feature W / 2×H / 2×128. Then, input it together with the feature map W / 2×H / 2×128 from the feature encoding stage (corresponding to Step 2) into the attention module to obtain feature W / 2×H / 2×128. Finally, after fusing the features, input them into the Fourier transform module, and the output size is W / 2×H / 2×64.
[0079] Step 3: Input the W / 2×H / 2×64 feature map into the upsampling layer to obtain the feature W×H×64. Then, input it together with the feature map W×H×64 from the feature encoding stage (corresponding to Step 1) into the attention module to obtain the feature W×H×64. Finally, after fusing the features, input them into the Fourier transform module. The output size is W×H×1, which is the corrected surface deformation data.
[0080] The attention module consists of channel attention and spatial attention. The channel attention mechanism focuses on the importance of different channels in the feature map, while the spatial attention mechanism focuses on the importance of different spatial locations in the feature map.
[0081] The beneficial effects of this invention are as follows:
[0082] This invention establishes an InSAR deformation field error correction model FFT-NET that couples Fourier transform and deep learning. This model utilizes the frequency domain feature analysis capability of Fourier transform and the spatial feature learning advantage of deep learning model to achieve accurate separation of deformation signal and error noise, significantly improving the measurement accuracy and robustness of InSAR deformation field.
[0083] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0084] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for correcting InSAR deformation field errors by coupling Fourier transform and deep learning, characterized in that, include: The initial InSAR surface deformation field data and the topographic data of the same region are input into the optimized deep neural network model, and the surface deformation field results after error correction are output. The training process of the optimized deep neural network model includes: Based on differential interferometric synthetic aperture radar, interferometric data with a short time baseline are acquired, and an interferogram containing real noise characteristics is obtained as error data. Acquire terrain data corresponding to the error data and resample it to the same spatial size as the error data; Simulated surface deformation field data is generated based on a geophysical model within a preset parameter range and used as training output labels. The error data is superimposed with the simulated surface deformation field data to construct a surface deformation field training input containing error signals, and together with the resampled terrain data, they form a training sample, thus completing the division of the training dataset and the test dataset. A deep neural network model with Fourier transform modules as basic units is constructed. The model parameters of the deep neural network model are optimized using the training dataset. The optimized deep neural network model is evaluated and saved based on the test dataset.
2. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 1, characterized in that, The Fourier transform module in the optimized deep neural network model is used to fuse frequency domain features and spatial domain features. The processing flow of the Fourier transform module includes: Perform a Fourier transform on the input features to obtain a frequency domain representation; Frequency domain features are extracted by sequentially performing 3×3 convolution and 1×1 convolution in the frequency domain. Perform an inverse Fourier transform on the frequency domain features to return to the spatial domain; The original input is subjected to a 3×3 convolution in the spatial domain to extract spatial features; The features returned to the spatial domain are fused with the spatial domain features for output.
3. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 1, characterized in that, The processing flow of the deep neural network model includes a feature encoding stage and a feature decoding stage. In the feature encoding stage, the input features are downsampled step by step and the terrain features are fused together. In the feature decoding stage, the features are upsampled step by step and fused with the features at the corresponding scale in the encoding stage through skip connections.
4. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 3, characterized in that, The feature encoding stage includes: The initial surface deformation and topography data of W×H×2 are input into the Fourier transform module and the max pooling layer to obtain the W / 2×H / 2×64 feature; where W represents the spatial width of the initial surface deformation field data and topography data on the unified spatial grid; and H represents the spatial height of the initial surface deformation field data and topography data on the unified spatial grid. The W / 2×H / 2 terrain data is processed by the terrain feature extraction module to obtain the W / 2×H / 2×64 terrain features, and then fused with the W / 2×H / 2×64 features and processed by the Fourier transform module and max pooling to obtain the W / 4×H / 4×128 features. The terrain data of W / 4×H / 4 is processed by the terrain feature extraction module to obtain W / 4×H / 4×128 terrain features, and then fused with the W / 4×H / 4×128 features and processed by the Fourier transform module and max pooling to obtain W / 8×H / 8×256 features. The W / 8×H / 8×256 features are input into the Fourier transform module and the multi-scale fusion module, and the W / 8×H / 8×256 features are output.
5. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 4, characterized in that, The feature decoding stage includes: The W / 8×H / 8×256 features are upsampled to W / 4×H / 4×256 features, and then fused with the W / 4×H / 4×256 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the features are input into the Fourier transform module to obtain W / 4×H / 4×128 features. The W / 4×H / 4×128 features are upsampled to W / 2×H / 2×128 features, and then fused with the W / 2×H / 2×128 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the features are input into the Fourier transform module to obtain W / 2×H / 2×64 features. The W / 2×H / 2×64 features are upsampled to W×H×64 features, and then fused with the W×H×64 features of the corresponding scale in the encoding stage through skip connections. After being processed by the attention module, the data is input into the Fourier transform module, and the output is W×H×1 error-corrected surface deformation field data.
6. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 4, characterized in that, The terrain feature extraction module consists of a fusion structure of multiple 3×3 convolutions and 1×1 convolutions, used to extract terrain features at different scales and fuse them with features at the corresponding scales.
7. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 4, characterized in that, The multi-scale fusion module extracts multi-scale features using dilated convolutions with dilation rates of 2, 4, and 6, and outputs W / 8×H / 8×256 features after stepwise fusion.
8. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 5, characterized in that, The attention module includes channel attention and spatial attention, which are used to weight the importance of channels and spatial locations during the skip connection fusion process with the corresponding scale features in the feature encoding stage.
9. The InSAR deformation field error correction method coupled with Fourier transform and deep learning according to claim 1, characterized in that, The short time baseline used in the training process is 6 days or 12 days.