End-to-end neural network InSAR phase unwrapping method based on mixed attention

Abstract

The invention relates to the technical field of remote sensing image processing, in particular to an end-to-end neural network InSAR (Interferometric Synthetic Aperture Radar) phase unwrapping method based on mixed attention, which takes U-Net as a basic framework, extracts key features and improves resolution through down-sampling and up-sampling operations, thereby effectively recovering detail information. Specifically, a convolutional block attention module (CBAM) and two receptive field modules, namely cavity spatial pyramid pooling (ASPP) and a receptive field module (RFB) are combined to construct an RFAUNet network model, and then simulation data sets generated by two methods of digital elevation inversion and random matrix generation are used for training a neural network until a good unwrapping effect is obtained. And finally, carrying out a phase unwrapping experiment on simulation data and real data to verify the effectiveness and robustness of the RFAUNet network model.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image processing, and in particular to an end-to-end neural network InSAR phase unwrapping method based on hybrid attention. Background Art

[0002] InSAR is an effective radar imaging remote sensing technology. It acquires synthetic aperture radar images of the same area at different observation geometries or time points, and uses the phase difference information to extract surface elevation or monitor surface deformation. It is widely used in fields such as geological disaster monitoring, geodynamic research, and environmental monitoring. However, because two-dimensional phase unwrapping is often an ill-posed problem, meaning that infinite solutions exist, accurately obtaining the correct unwrapped phase is often challenging.

[0003] To address these challenges, traditional phase unwrapping algorithms are primarily categorized into path tracking, global optimization, and Bayesian methods. Path tracking methods, such as branch cutting and mass graph methods, improve unwrapping accuracy by setting a reasonable integration path to prevent global error propagation. Global optimization methods achieve phase unwrapping by minimizing the difference between the wrapped and unwrapped phases. These methods include least squares methods and network flow. Bayesian methods treat phase unwrapping as an optimal state estimation problem and model it using state-space analysis, using a state-space representation. These methods achieve a joint phase unwrapping and phase noise filtering approach. Typical Bayesian methods include the extended Kalman phase unwrapping algorithm, the unscented Kalman phase unwrapping algorithm, and the iterative unscented Kalman filter phase unwrapping method. While these methods address some of the challenges of phase unwrapping, their underlying theoretical frameworks still struggle to overcome the inherent constraints of the phase continuity criterion.

[0004] In recent years, with the development of deep learning, researchers have tried to use deep learning to perform phase unwrapping. For example, regression network phase unwrapping is achieved by directly establishing end-to-end mapping, conditional generative adversarial networks improve the authenticity of unwrapping through adversarial training, and entangled fuzzy number semantic segmentation, gradient three-category semantic segmentation, gradient five-category semantic segmentation, etc. convert unwrapping into pixel-level classification. How to achieve high-precision, strong noise resistance and lightweight phase unwrapping is still a technical gap that needs to be broken through in current research. Summary of the Invention

[0005] The purpose of the present invention is to provide an end-to-end neural network InSAR phase unwrapping method based on hybrid attention, aiming to solve the technical problems that the existing unwrapping methods have poor unwrapping effect in low coherence areas or when the noise influence is large, and it is difficult to perform correct phase unwrapping in relatively dense stripe areas.

[0006] To achieve the above objectives, the present invention provides an end-to-end neural network InSAR phase unwrapping method based on hybrid attention, comprising the following steps:

[0007] Step 1: Construct a training dataset, which includes noisy entangled phase input data and corresponding noise-free true phase labels;

[0008] Step 2: Construct the RFAUNet network model, which is divided into two parts: encoder and decoder, including input layer, convolution layer, tandem hybrid attention layer composed of CBAM module and ASPP module, downsampling layer, RFB layer, upsampling layer, and output layer;

[0009] Step 3: Train the RFAUNet network model based on the training dataset, calculate the difference between the disentangled phase and the true phase through the loss function, update the network parameters and save the trained model weights.

[0010] Optionally, the process of constructing the training dataset in step 1 includes the following steps:

[0011] Step 1.1: Generate the first type of DEM inversion dataset. The first type of DEM inversion dataset is based on the ALOS PALSAR DEM information and synthetic aperture radar phase parameters.

[0012] Step 1.2: Generate the second type of random matrix data set by randomly generating an N×N matrix and using bicubic interpolation to generate a random matrix data set to simulate the terrain smoothing phase;

[0013] Step 1.3: performing a phase wrapping operation on the first type of DEM inversion dataset and the second type of random matrix dataset, and adding phase noise by adjusting pixel coherence, wherein the coherence range of the phase noise is 0.5-0.9;

[0014] Step 1.4: Take the noisy winding phase as input data and the corresponding noise-free true phase as the label.

[0015] Optionally, the input layer of the RFAUNet network model in step 2 is used to receive single-channel noisy entangled phase data;

[0016] The convolution layer consists of a 3×3 convolution operation, batch normalization, and a LeakReLU activation function to transform the number of feature channels.

[0017] In the serial hybrid attention layer, the CBAM module is used to extract channel attention weights and spatial attention weights; the ASPP module fuses multi-scale context information through a dilated convolution operation;

[0018] The downsampling layer reduces the feature resolution through a 3×3 convolution operation with a stride of 2;

[0019] The RFB layer extracts local features of different receptive fields by combining multi-scale dilated convolution kernels;

[0020] Upsampling layer, which restores feature resolution through 2×2 transposed convolution operation;

[0021] The output layer maps the multi-channel features into a single-channel disentangled phase through a 3×3 convolution operation.

[0022] Optionally, during the training process in step 3, the Adam optimizer is used and the learning rate is set to 10e-4;

[0023] The loss function is the root mean square error function, which updates the parameters by calculating the square root of the average square of the difference between the unwrapped phase and the true phase;

[0024] The ratio of training data to test data is 4:1.

[0025] The present invention provides an end-to-end neural network InSAR phase unwrapping method based on hybrid attention. It uses U-Net as the basic framework, extracts key features and improves resolution through downsampling and upsampling operations, thereby effectively restoring detail information. Specifically, the RFAUNet network model is constructed by combining the convolutional block attention module (CBAM) and two receptive field modules - the atrous spatial pyramid pooling (ASPP) and the receptive field module (RFB). Then, a simulated data set generated by two methods, digital elevation inversion and random matrix generation, is used to train the neural network until a good unwrapping effect is obtained. Finally, the effectiveness and robustness of the RFAUNet network model of the present invention are verified by performing phase unwrapping experiments on simulated data and real data. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0027] Figure 1 This is a step flow diagram of a hybrid attention-based end-to-end neural network InSAR phase unwrapping method of the present invention.

[0028] Figure 2 Schematic diagram of the architecture of the RFAUNet network model of the present invention.

[0029] Figure 3It is a schematic diagram of the serial hybrid attention layer structure of the present invention.

[0030] Figure 4 It is a schematic diagram of the CBAM module structure of the present invention.

[0031] Figure 5 It is a schematic diagram of the ASPP module structure of the present invention.

[0032] Figure 6 It is a schematic diagram of the RFB layer structure of the present invention.

[0033] Figure 7 It is a schematic diagram of the unwrapping results and performance indicators of DEM inversion data in a specific embodiment of the present invention.

[0034] Figure 8 2 is a schematic diagram of the untangling results and performance indicators of random matrix generated data in a specific embodiment of the present invention.

[0035] Figure 9 It is a schematic diagram of the untangling results of real data using different methods in a specific embodiment of the present invention.

[0036] Figure 10 It is a schematic diagram of the re-entanglement results of real data using different methods in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0038] Some of the English term abbreviations used in this article are explained as follows:

[0039] InSAR: Interferometric Synthetic Aperture Radar, synthetic aperture radar interferometry;

[0040] RFAUNet: Receptive Field and Attention U-net, an end-to-end neural network architecture based on hybrid attention;

[0041] CBAM: Convolutional BlockAttention Module, convolutional block attention module;

[0042] ASPP: Atrous Spatial Pyramid Pooling, void spatial pyramid pooling;

[0043] RFB: Receptive Field Block, receptive field module;

[0044] See also Figure 1 The present invention provides an end-to-end neural network InSAR phase unwrapping method based on hybrid attention, comprising the following steps:

[0045] Step 1: Construct a training dataset, which includes noisy entangled phase input data and corresponding noise-free true phase labels;

[0046] Step 2: Construct the RFAUNet network model, which is divided into two parts: encoder and decoder, including input layer, convolution layer, tandem hybrid attention layer composed of CBAM module and ASPP module, downsampling layer, RFB layer, upsampling layer, and output layer;

[0047] Step 3: Train the RFAUNet network model based on the training dataset, calculate the difference between the disentangled phase and the true phase through the loss function, update the network parameters and save the trained model weights.

[0048] The following is a further explanation based on the specific implementation steps:

[0049] The training dataset in step 1 is divided into two parts. The first is a DEM inversion dataset, which generates true phases by combining ALOS PALSAR DEM information with the phase parameters of its synthetic aperture radar. The second is a random matrix dataset, which randomly generates an N×N matrix and uses bicubic interpolation to generate smooth true phases. These two parts of data are mixed and twisted to generate twisted phases. Phase noise is added by adjusting the coherence of the twisted phases. Finally, the noisy twisted phases are used as features and the noise-free true phases as labels to create the training dataset.

[0050] The RFAUNet network model structure built in step 2 is as follows Figure 2 As shown, including:

[0051] Input layer: used to input the noisy entangled phase into the neural network and convert the single-channel image into multi-channel features for phase unwrapping of the neural network.

[0052] Convolutional layer: The convolutional layer consists of 3×3 convolution, batch normalization, and activation function. 3×3 convolution is used to extract channel features. Batch normalization helps improve the generalization ability of the model by standardizing each batch of data. The activation function uses the LeakReLU function to promote more efficient gradient propagation and network training while preventing neuron death.

[0053] Concatenating mixed attention layers such as Figure 3 As shown, it includes CBAM module and ASPP module. The CBAM module structure is as follows Figure 4 As shown, the ASPP module structure is as follows Figure 5 As shown in Figure 3, the concatenated mixed attention layers improve the feature extraction of global and local information while expanding the receptive field.

[0054] Downsampling layer: The downsampling layer uses convolution with a kernel size of 3×3 and a stride of 2 instead of pooling to better capture image details and contextual relationships while retaining more local information details.

[0055] RFB layer: RFB structure is as follows Figure 6 As shown in Figure 2, the RFB layer enhances the receptive field of the network by introducing multi-scale dilated convolutions and convolution kernels of different sizes, enabling it to capture richer local features at different scales.

[0056] Upsampling layer: 2×2 transposed convolution is used for upsampling to restore the image resolution while retaining more details and spatial structure.

[0057] Output layer: A convolution with a kernel size of 3×3 is used to integrate the multi-channel features passing through the network to obtain the noise-free true phase.

[0058] In step 3, the RFAUNet network model was trained using the training dataset, with noisy entangled phases as features and noise-free true phases as labels. Adam was used as the optimizer with a learning rate of 10e-4. The mean squared error (MSE) was used as the loss function, and model performance was evaluated by calculating the square root of the mean square of the difference between the unentangled and true phases. Each training round used a batch size of 8, consisting of 256×256 entangled phases, for a total of 140 rounds. The dataset consisted of 10,000 training data sets, 8,000 of which were training data sets and 2,000 of which were testing data sets.

[0059] Furthermore, the present invention also tests RFAUNet using simulated data and real data:

[0060] RFAUNet was experimentally compared with the quality graph method (QG), the unscented Kalman filter (UKF), the least squares method (MCF), the gradient three-classification UNet, and the regression disentanglement network ResUNet. Using simulated data, the root mean square error (RMSE) and structural similarity index (SSIM) between the disentangled phase and the true phase were used as evaluation metrics. RMSE quantifies the error between the disentangled phase and the true phase, while SSIM assesses the similarity between the disentangled phase and the true phase.

[0061] In the simulated data, Figure 7Tables 1 and 2 show the unwrapping results and performance indicators of DEM inversion data. Figure 8 Tables 3 and 4 show the unwrapping results and performance indicators of random matrix generated data. It can be seen that under different coherences, RFAUNet has a good unwrapping effect. While its RMSE is the smallest, it has the best SSIM, which is better than other unwrapping methods, and the obtained unwrapped image is more continuous and smooth.

[0062] Table 1 RMSE of unwrapping of DEM inversion simulation data

[0063]

[0064] Table 2 Unwrapped SSIM of DEM inversion simulation data

[0065]

[0066] Table 3 RMSE of disentanglement of simulated data generated by random matrices

[0067]

[0068] Table 4 SSIM of disentangled data generated by random matrix

[0069]

[0070]

[0071] In real data, Figure 9 The disentanglement results of different methods on real data are: Figure 10 Figure 1 shows the re-entanglement results of different methods on real data. For highly coherent real data, such as those from the Three Gorges region and the Antofagasta region of Chile, RFAUNet and other methods demonstrated a certain level of unwrapping capability. However, in less coherent regions, such as Bayingolin Mongol Autonomous Prefecture in Xinjiang and Haibei Tibetan Autonomous Prefecture in Qinghai, other methods exhibited varying degrees of unwrapping errors, while RFAUNet achieved relatively good unwrapping results. In summary, RFAUNet demonstrated excellent unwrapping capabilities across a wide range of real data, producing smooth and clear unwrapped phases.

[0072] The above results show that the end-to-end neural network InSAR phase unwrapping method based on hybrid attention has good unwrapping ability and robustness, and also has certain unwrapping performance under low coherence conditions.

[0073] The above disclosure is merely one or more preferred embodiments of the present invention, and certainly cannot be used to limit the scope of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made in accordance with the claims of the present invention still fall within the scope of the invention.

Claims

1. A hybrid attention-based end-to-end neural network InSAR phase unwrapping method, characterized by: The following steps are involved: Step 1: Construct a training dataset, which includes noisy entangled phase input data and corresponding noise-free true phase labels; Step 2: Construct the RFAUNet network model, which is divided into two parts: encoder and decoder, including input layer, convolution layer, tandem hybrid attention layer composed of CBAM module and ASPP module, downsampling layer, RFB layer, upsampling layer, and output layer; Step 3: Train the RFAUNet network model based on the training dataset, calculate the difference between the disentangled phase and the true phase through the loss function, update the network parameters and save the trained model weights.

2. The hybrid attention-based end-to-end neural network InSAR phase unwrapping method according to claim 1, characterized in that The process of constructing the training dataset in step 1 includes the following steps: Step 1.1: Generate the first type of DEM inversion dataset. The first type of DEM inversion dataset is based on the ALOS PALSAR DEM information and synthetic aperture radar phase parameters. Step 1.2: Generate the second type of random matrix data set by randomly generating an N×N matrix and using bicubic interpolation to generate a random matrix data set to simulate the terrain smoothing phase; Step 1.3: performing a phase wrapping operation on the first type of DEM inversion dataset and the second type of random matrix dataset, and adding phase noise by adjusting pixel coherence, wherein the coherence range of the phase noise is 0.5-0.9; Step 1.4: Take the noisy winding phase as input data and the corresponding noise-free true phase as the label.

3. The hybrid attention-based end-to-end neural network InSAR phase unwrapping method according to claim 1, characterized in that The input layer of the RFAUNet network model in step 2 is used to receive single-channel noisy entangled phase data; The convolution layer consists of a 3×3 convolution operation, batch normalization, and a LeakReLU activation function to transform the number of feature channels. In the cascaded mixed attention layer, the CBAM module is used to extract channel attention weights and spatial attention weights; The ASPP module fuses multi-scale context information through a dilated convolution operation; The downsampling layer reduces the feature resolution through a 3×3 convolution operation with a stride of 2; The RFB layer extracts local features of different receptive fields by combining multi-scale dilated convolution kernels; Upsampling layer, which restores feature resolution through 2×2 transposed convolution operation; The output layer maps the multi-channel features into a single-channel disentangled phase through a 3×3 convolution operation.

4. The hybrid attention-based end-to-end neural network InSAR phase unwrapping method according to claim 1, characterized in that During the training process in step 3, the Adam optimizer is used and the learning rate is set to 10e-4; The loss function is the root mean square error function, which updates the parameters by calculating the square root of the average square of the difference between the unwrapped phase and the true phase; The ratio of training data to test data is 4:1.

Images