Unmanned aerial vehicle InSAR phase unwrapping method based on WaveUNet

Abstract

The invention discloses an unmanned aerial vehicle-mounted InSAR phase unwrapping method based on WaveUNet. The method comprises the following steps: collecting InSAR wrapping phase diagram training data; constructing a multi-scale phase unwrapping model, and training the multi-scale phase unwrapping model based on the InSAR wrapped phase diagram training data to obtain a trained multi-scale phase unwrapping model; the multi-scale phase unwrapping model is constructed based on a U-Net network architecture, and a wavelet wiener filtering module is embedded in a feature extraction path of the multi-scale phase unwrapping model; and obtaining an InSAR wrapped phase diagram to be processed, and inputting the InSAR wrapped phase diagram to be processed into the trained multi-scale phase unwrapping model to obtain an absolute phase diagram after the InSAR wrapped phase diagram to be processed is unwrapped. According to the invention, a wavelet wiener filtering module is embedded before each downsampling stage of a U-Net encoder, a multi-scale phase unwrapping model with adaptive denoising is constructed, and the technical problems of low InSAR phase unwrapping precision and poor detail retention in a strong-noise and low-coherence region in a traditional and existing deep learning method are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing information processing technology, and specifically to an unmanned aerial vehicle (UAV) InSAR phase unwrapping method based on WaveUNet. Background Technology

[0002] Interferometric Synthetic Aperture Radar (InSAR) is an important remote sensing technique that extracts surface elevation or deformation information by interferometrically processing multiple SAR images of the same area. Its key lies in extracting information from the data entangled with the surface. The process of recovering a continuous absolute phase from the interference phase within an interval is called phase unwrapping. Because there is a multi-valued mapping relationship between the wrapped phase and the absolute phase, phase unwrapping is essentially an ill-posed inverse problem that requires additional constraints to solve.

[0003] Traditional phase unwrapping methods are mainly divided into two categories: path-following methods and minimum norm methods. Path-following methods (including branching, quality-guided algorithms, and region growing methods) aim to limit unwrapping errors to local areas by constructing reasonable integration paths, but they are prone to path interruption or error propagation in areas with dense noise or abrupt phase changes. Minimum norm methods (including FFT least squares and iterative least squares) transform the unwrapping problem into a global optimization problem. While they have higher computational efficiency, they often lead to the propagation of phase errors throughout the interferogram, reducing the robustness of the unwrapping results. Both types of methods are based on the "phase continuity assumption" (i.e., the phase difference between adjacent pixels does not exceed a certain value). However, in actual SAR images, due to factors such as noise interference, low coherence areas and abrupt terrain changes, this assumption often does not hold true, resulting in a significant decrease in the unwrapping accuracy of traditional methods in complex scenarios.

[0004] In recent years, with the development of deep learning technology, especially the successful application of convolutional neural networks (CNNs) in image processing, researchers have begun to explore data-driven phase unwrapping methods. These methods do not rely on the assumption of phase continuity and can directly learn the unwrapping mapping relationship from the tangled phase, achieving end-to-end phase recovery. Existing deep learning-based unwrapping schemes can be broadly classified into two categories: pure neural network methods and hybrid strategies combining neural networks and traditional methods. However, these methods still have significant limitations: First, most network structures do not embed effective denoising mechanisms at the feature extraction front end, lacking targeted suppression capabilities for common noise such as speckle noise and system noise in interferograms, leading to decreased unwrapping performance and insufficient noise robustness in low signal-to-noise ratio regions; second, traditional CNNs have limitations in capturing multi-scale phase structures and noise patterns, making it difficult to retain important terrain edges and details while denoising; furthermore, existing models mostly use general image processing architectures and do not fully consider the frequency domain characteristics, tangling periodicity, and noise distribution patterns of the interferometric phase, limiting their applicability in complex scenarios such as steep terrain and low coherence regions. Especially in UAV-borne InSAR systems, due to factors such as platform stability, imaging geometry and the complexity of the ground scene, the interferometric phase often exhibits stronger noise interference and more complex fringe patterns, which places higher demands on the denoising capability, detail preservation and multi-scale adaptability of the phase unwrapping method.

[0005] Therefore, how to achieve high-precision phase unwrapping in noisy environments, especially in low-coherence regions and steep terrain regions, and how to establish an effective multi-scale noise suppression mechanism to fully preserve important terrain feature information while denoising, have become technical problems that urgently need to be solved in this field. Summary of the Invention

[0006] To address the shortcomings of the existing technologies, this invention provides an UAV-borne InSAR phase unwrapping method based on WaveUNet. By embedding a wavelet Wiener filter module before each downsampling stage of the U-Net encoder, an adaptive denoising multi-scale phase unwrapping model is constructed, effectively solving the technical problems of low accuracy and poor detail preservation in InSAR phase unwrapping in strong noise and low coherence regions by traditional and existing deep learning methods.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A method for unmanned aerial vehicle (UAV) InSAR phase unwrapping based on WaveUNet includes the following steps:

[0009] Collect InSAR wrapped phase map training data;

[0010] A multi-scale phase unwrapping model is constructed, and the multi-scale phase unwrapping model is trained based on InSAR wrapped phase map training data to obtain the trained multi-scale phase unwrapping model; the multi-scale phase unwrapping model is constructed based on the U-Net network architecture, and a wavelet Wiener filter module is embedded in its feature extraction path;

[0011] Obtain the InSAR wrapped phase map to be processed, and input the InSAR wrapped phase map to be processed into the trained multi-scale phase unwrapping model to obtain the absolute phase map after unwrapping the InSAR wrapped phase map to be processed.

[0012] As a preferred embodiment, the multi-scale phase unwrapping model includes an encoder, a decoder, and a wavelet Wiener filter module embedded in the encoder. The trained multi-scale phase unwrapping model takes the input InSAR wrapped phase map as the input of the encoder, performs multi-scale adaptive noise suppression step by step through a multi-level wavelet Wiener filter module, and extracts the noise reduction feature maps of each level. The encoder downsamples and extracts deep features from the noise reduction feature maps of each level to obtain deep feature maps. The decoder upsamples the deep feature maps and fuses them with the output feature maps of the corresponding layers of the encoder through skip connections, finally generating the unwrapped absolute phase map as the output of the multi-scale phase unwrapping model.

[0013] As a preferred embodiment, the encoder comprises four sequentially connected coding layers. Each coding layer includes a cascaded wavelet Wiener filter module, a DoubleConv module, and a max-pooling downsampling layer. The DoubleConv module comprises two cascaded convolutional units. The input of the encoder serves as the input to the wavelet Wiener filter module in the first coding layer, the output of the wavelet Wiener filter module in the first coding layer serves as the input to the DoubleConv module in the first coding layer, the output of the DoubleConv module in the first coding layer serves as the input to the max-pooling downsampling module in the second coding layer, the output of each coding layer serves as the input to the next layer connected to it in the encoder, and the output of the fifth coding layer serves as the output of the encoder.

[0014] As a preferred embodiment, each convolutional unit in the DoubleConv module includes a cascaded 3×3 convolutional layer, a BN layer, and a ReLU activation layer;

[0015] The input feature map is first used as the input of the first convolutional unit. It is then processed by a 3×3 convolutional layer for feature extraction, normalized by a BN layer, and non-linearly activated by a ReLU activation layer to obtain primary features. The primary features output by the first convolutional unit are used as the input of the second convolutional unit for depth feature extraction. After performing the same operation as the first convolutional unit, the depth features output by the second convolutional unit are used as the output of the DoubleConv module.

[0016] As a preferred embodiment, the processing procedure of the wavelet Wiener filter module includes the following steps:

[0017] The input feature map is decomposed into a low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH using discrete wavelet transform. Local variance is calculated within the high-frequency sub-bands using a sliding window. The energy distribution of signal and noise is dynamically estimated based on Wiener filtering theory, and an adaptive gain coefficient is calculated. This adaptive gain coefficient is then applied to each high-frequency sub-band for filtering. Finally, the filtered high-frequency sub-bands and the unprocessed low-frequency sub-bands are reconstructed using inverse wavelet transform to obtain the denoised phase feature map.

[0018] As a preferred embodiment, the formula for calculating the adaptive gain coefficient is:

[0019] ;

[0020] In the formula, For adaptive gain coefficients; The signal variance of the current sub-band; Noise variance estimated from the highest frequency subband HH;

[0021] Filtering using adaptive gain coefficients is expressed as follows:

[0022] ;

[0023] In the formula, These are the filtered wavelet coefficients; These are the original wavelet coefficients.

[0024] As a preferred embodiment, the decoder includes four cascaded decoding layers, each decoding layer including a cascaded DoubleConv module and a transposed convolutional upsampling layer; wherein, the input of the decoding layer serves as the input of the DoubleConv module, the output of the DoubleConv module serves as the input of the transposed convolutional upsampling layer, and the output of the transposed convolutional upsampling layer serves as the output of the decoding layer.

[0025] As a preferred embodiment, the multi-scale phase unwrapping model further includes a bottleneck layer, which comprises a cascaded wavelet Wiener filter module and a DoubleConv module;

[0026] The entangled phase map input to the multi-scale phase unwrapping model serves as the input to the first encoding layer. The output of each encoding layer serves as the input for connecting to the next layer in the multi-scale phase unwrapping model. The output of the fourth encoding layer serves as the input to the bottleneck layer. The output of the bottleneck layer is upsampled by transposed convolution and then skip-connected with the output of the fourth DoubleConv module. After alignment using bilinear interpolation, it serves as the input to the fourth decoding layer. The input to each decoding layer is the feature map fused with the output of the DoubleConv module in its corresponding encoding layer through skip connections and bilinear interpolation with the output of the previous layer in the multi-scale phase unwrapping model. The output of the first decoding layer is convolved by 3×3 and serves as the overall output of the multi-scale phase unwrapping model.

[0027] As a preferred embodiment, the multi-scale phase unwrapping model is trained in the following manner:

[0028] The unwrapped phase map that has been unwrapped in advance is used as a training sample to form a training sample set, which is then input into the multi-scale phase unwrapping model. The learning rate and hyperparameters of the multi-scale phase unwrapping model are set, and a mean squared error loss function is constructed to evaluate the unwrapping performance. The parameters of the multi-scale phase unwrapping model are optimized and updated with the goal of minimizing the mean squared error loss function, and then the multi-scale phase unwrapping model is trained.

[0029] As a preferred embodiment, the mean squared error loss function is:

[0030]

[0031] In the formula, For the sample size, Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0032] Compared with the prior art, the present invention has the following technical effects:

[0033] 1. This invention achieves multi-scale, adaptive noise suppression of the input phase map by embedding a wavelet Wiener filtering module before each downsampling stage of the encoder. This module uses discrete wavelet transform to decompose the phase information and performs Wiener filtering based on local variance estimation in the high-frequency subband. While effectively removing speckle noise and system noise, it can significantly preserve phase edge and detail information, thereby greatly improving the phase unwrapping accuracy and robustness of the model under strong noise interference such as low coherence region and steep terrain. Compared with traditional phase unwrapping methods, it can better handle the phase unwrapping problem under low signal-to-noise ratio conditions.

[0034] 2. This invention integrates a wavelet Wiener filter module before each downsampling stage of the U-Net encoder, forming a multi-level, multi-scale noise suppression architecture. This enables the model to identify and suppress noise at different feature scales, enhancing its adaptability to complex noise patterns and avoiding the imbalance between global and local noise suppression in traditional single-scale denoising methods. This results in a more comprehensive and stable improvement in phase signal-to-noise ratio.

[0035] 3. This invention employs a U-Net symmetric encoding / decoding structure with skip connections and introduces a DoubleConv module for feature enhancement, enabling the network to simultaneously fuse deep semantic information and shallow detail features. This not only accelerates the feature extraction and reconstruction process but also effectively alleviates the gradient vanishing problem through skip connections, improving training stability and convergence speed, and achieving efficient end-to-end phase mapping. Furthermore, during the network's forward propagation, bilinear interpolation is used to calibrate the dimensions of the transposed upsampled feature map and the skip connection feature map in real time, avoiding information loss or training instability caused by resolution mismatch, thus improving the method's practicality and generalization performance.

[0036] 4. This invention uses the mean squared error loss function combined with an adaptive optimization algorithm to train the model in an end-to-end manner, enabling the network to directly learn the complex mapping relationship of phase unwrapping from the data. This eliminates the dependence on the traditional phase continuity assumption. Compared with the traditional unwrapping algorithm based on iterative optimization, this method not only improves the unwrapping efficiency, but also significantly improves the unwrapping accuracy and model generalization ability in complex scenarios through data-driven optimization. Attached Figure Description

[0037] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0038] Figure 1 This invention discloses the overall process of an UAV-borne InSAR phase unwrapping method based on WaveUNet.

[0039] Figure 2 This is a diagram of the multi-scale phase unwrapping model architecture used in this invention;

[0040] Figure 3 This refers to the downsampling stage in this embodiment of the invention;

[0041] Figure 4 This refers to the upsampling stage in this embodiment of the invention;

[0042] Figure 5 This is a schematic diagram of the phase unwrapping result and local region slice curve in an embodiment of the present invention;

[0043] Figure 6This is a schematic diagram of the traditional UNet phase unwrapping result and slice curve in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but only to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0045] The present invention will now be described in further detail with reference to the accompanying drawings.

[0046] With the widespread application of Interferometric Synthetic Aperture Radar (InSAR) technology in surface deformation monitoring and elevation modeling, phase unwrapping, as a crucial step in recovering the absolute phase from the wrapped phase, directly impacts the reliability of the final monitoring results. Traditional phase unwrapping methods heavily rely on the assumption of "phase continuity," which is prone to failure in complex real-world scenarios such as low coherence regions, strong noise interference, or abrupt terrain changes, leading to error propagation and decreased accuracy. Although deep learning technology provides a new data-driven approach for phase unwrapping, existing network architectures generally lack effective mechanisms for processing the noise characteristics of interferometric phases. Particularly in noisy environments, they exhibit insufficient noise suppression, weak detail preservation, and inadequate multi-scale feature fusion, limiting their practicality and robustness in lightweight, high-dynamic platforms such as UAV-borne InSAR.

[0047] To overcome the above-mentioned shortcomings, this invention proposes an UAV-borne InSAR phase unwrapping method based on WaveUNet. This method constructs an adaptive denoising multi-scale phase unwrapping model by embedding a wavelet Wiener filter module before each downsampling stage of the U-Net encoder. This effectively solves the technical problems of low accuracy and poor detail preservation in InSAR phase unwrapping in strong noise and low coherence regions by traditional and existing deep learning methods.

[0048] Specifically, the present invention proposes an unmanned aerial vehicle (UAV) InSAR phase unwrapping method based on WaveUNet, such as... Figure 1As shown, the process includes the following steps: acquiring an InSAR wrapped phase map; constructing a multi-scale phase unwrapping model, training the multi-scale phase unwrapping model based on the InSAR wrapped phase map to obtain a trained multi-scale phase unwrapping model; the multi-scale phase unwrapping model is based on a U-Net network architecture and embeds a wavelet Wiener filter module in its feature extraction path; acquiring the InSAR wrapped phase map to be processed, inputting the InSAR wrapped phase map to be processed into the trained multi-scale phase unwrapping model to obtain an absolute phase map of the unwrapped InSAR wrapped phase map.

[0049] The UAV-borne InSAR phase unwrapping method based on WaveUNet of the present invention will be described in more detail below.

[0050] 1. Multi-scale phase unwrapping model based on U-Net network architecture

[0051] This embodiment employs the classic U-Net encoder-decoder symmetric structure and makes corresponding adjustments to the U-Net network for the phase unwrapping task. Specifically, as follows... Figure 2 As shown, the multi-scale phase unwrapping model (WaveUNet) includes an encoder, a decoder, and a wavelet Wiener filter module embedded in the encoder. The trained multi-scale phase unwrapping model takes the input InSAR wrapped phase map as the input of the encoder, performs multi-scale adaptive noise suppression step by step through a multi-level wavelet Wiener filter module, and extracts the noise reduction feature maps of each level. The encoder downsamples the noise reduction feature maps of each level and extracts deep features to obtain deep feature maps. The decoder upsamples the deep feature maps and fuses them with the output feature maps of the corresponding layers of the encoder through skip connections. Finally, the unwrapped absolute phase map is generated as the output of the multi-scale phase unwrapping model.

[0052] The following sections will introduce them separately.

[0053] 1.1 Encoder

[0054] In this embodiment, the encoder includes four sequentially connected coding layers. Each coding layer includes a cascaded wavelet Wiener filter module, a DoubleConv module, and a max-pooling downsampling layer. The DoubleConv module includes two cascaded convolutional units. The input of the encoder serves as the input of the wavelet Wiener filter module in the first coding layer, the output of the wavelet Wiener filter module in the first coding layer serves as the input of the DoubleConv module in the first coding layer, the output of the DoubleConv module in the first coding layer serves as the input of the max-pooling downsampling module in the second coding layer, the output of each coding layer serves as the input of the next layer connected to it in the encoder, and the output of the fifth coding layer serves as the output of the encoder.

[0055] In specific implementation, such as Figure 3 As shown, the encoder in this embodiment includes an input stage and four downsampling stages. Each stage first performs wavelet filtering to denoise the image, and a batch normalization (BN) layer is added between each convolutional layer and ReLU activation layer to accelerate the training and convergence speed of U-Net and prevent gradient vanishing, gradient exploding, and overfitting. The number of channels in the feature map doubles with each downsampling pass, increasing to 1024 channels after four downsampling passes.

[0056] 1.2 DoubleConv module

[0057] In this embodiment, each convolutional unit in the DoubleConv module includes a cascaded 3×3 convolutional layer, a BN layer, and a ReLU activation layer;

[0058] In the DoubleConv module, the input feature map is first used as the input of the first convolutional unit. Features are extracted through a 3×3 convolutional layer, normalized through a BN layer, and then non-linearly activated through a ReLU activation layer to obtain primary features. The primary features output by the first convolutional unit are used as the input of the second convolutional unit for depth feature extraction. After performing the same operation as the first convolutional unit, the depth features output by the second convolutional unit are used as the output of the DoubleConv module.

[0059] 1.3 Wavelet Wiener Filter Module

[0060] In this embodiment, a wavelet Wiener filtering module is used to preprocess the input features before feature extraction, thereby effectively suppressing noise interference. Specifically, the input feature map is decomposed into a low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH using discrete wavelet transform (DWT). Within the high-frequency sub-bands, a sliding window is used to calculate the local variance. Based on Wiener filtering theory, the energy distribution of the signal and noise is dynamically estimated, and an adaptive gain coefficient is calculated. This adaptive gain coefficient is applied to each high-frequency sub-band for filtering. The filtered high-frequency sub-bands and the unprocessed low-frequency sub-bands are then reconstructed using inverse wavelet transform (IDWT) to obtain the denoised phase feature map.

[0061] The formula for calculating the adaptive gain coefficient is as follows:

[0062] ;

[0063] In the formula, The adaptive gain coefficient, ranging from [0,1]; The signal variance of the current sub-band; Noise variance estimated from the highest frequency subband HH;

[0064] Filtering using adaptive gain coefficients is expressed as follows:

[0065] ;

[0066] In the formula, These are the filtered wavelet coefficients; These are the original wavelet coefficients.

[0067] 1.4 Decoder

[0068] In this embodiment, the decoder includes four cascaded decoding layers, each decoding layer including a cascaded DoubleConv module and a transposed convolutional upsampling layer; wherein, the input of the decoding layer serves as the input of the DoubleConv module, the output of the DoubleConv module serves as the input of the transposed convolutional upsampling layer, and the output of the transposed convolutional upsampling layer serves as the output of the decoding layer.

[0069] In specific implementation, such as Figure 4 As shown, the decoder in this embodiment includes four upsampling stages and an output stage. Upsampling is performed through transposed convolutions to reconstruct feature information and restore the features extracted by the encoder to their original size. Skip connections are then established between the decoder and the features of the corresponding layer in the encoder. Each convolutional layer is followed by a ReLU activation layer. The encoder and decoder are connected through skip connections, thus U-Net can simultaneously preserve both deep and shallow semantic information.

[0070] In addition, the multi-scale phase unwrapping model also includes a bottleneck layer, which includes a cascaded wavelet Wiener filter module and a DoubleConv module;

[0071] The entangled phase map input to the multi-scale phase unwrapping model serves as the input to the first encoding layer. The output of each encoding layer serves as the input for connecting to the next layer in the multi-scale phase unwrapping model. The output of the fourth encoding layer serves as the input to the bottleneck layer. The output of the bottleneck layer is upsampled by transposed convolution and then skip-connected with the output of the fourth DoubleConv module. After alignment using bilinear interpolation, it serves as the input to the fourth decoding layer. The input to each decoding layer is the feature map fused with the output of the DoubleConv module in its corresponding encoding layer through skip connections and bilinear interpolation with the output of the previous layer in the multi-scale phase unwrapping model. The output of the first decoding layer is convolved by 3×3 and serves as the overall output of the multi-scale phase unwrapping model.

[0072] 2. Training of the multi-scale phase unwrapping model

[0073] In this embodiment, the training process of the multi-scale phase unwrapping model is as follows:

[0074] Pre-unwrapped phase maps, already unwrapped, are used as training samples to form a training sample set, which is then input into the multi-scale phase unwrapping model. The initial learning rate of the multi-scale phase unwrapping model is set to 1e-4, decreasing to 0.1 times the original rate every 30 epochs. A mean squared error loss function is constructed to evaluate unwrapping performance. The parameters of the multi-scale phase unwrapping model are optimized and updated with the goal of minimizing the mean squared error loss function, thereby training the multi-scale phase unwrapping model. Gradient clipping is employed during training to prevent gradient explosion.

[0075] The mean squared error loss function is:

[0076]

[0077] In the formula, For the sample size, Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0078] 3. Examples

[0079] To better illustrate the effectiveness of the UAV-borne InSAR phase unwrapping method based on WaveUNet in this embodiment, a specific embodiment will be described below.

[0080] In the phase unwrapping task, in order to quantitatively evaluate the performance of the proposed method, this embodiment conducts comparative experiments on the same dataset using the method of this embodiment and the traditional UNet network, and analyzes the results from both visual and quantitative perspectives.

[0081] Figure 5 The phase unwrapping results and evaluation metrics of the method in this embodiment are shown, wherein... Figure 5 (b) is the entanglement phase diagram. Figure 5 (c) is the unwrapping phase diagram. Figure 5 (d) is the true phase diagram. Figure 5 (a) is a slice curve of the unwrapped phase map and the true phase map along the dotted line. As can be seen from the curve, the unwrapping result in this embodiment is highly consistent with the true phase in terms of trend. Figure 6 This demonstrates the untangling results and evaluation metrics of the traditional UNet under the same conditions. Among them, Figure 6 (b) is the entanglement phase diagram. Figure 6 (c) is the unwrapping phase diagram. Figure 6 (d) is the true phase diagram. Figure 6 (a) is a slice curve of the unwound phase map and the true phase map along the dotted line.

[0082] To further quantify the comparison, we calculated the root mean square error (RMSE) and peak signal-to-noise ratio (PSNR) of the two methods. The method in this embodiment achieved an RMSE of 0.0014 and a PSNR of 61.08 dB; while the traditional UNet achieved an RMSE of 0.0049 and a PSNR of 46.68 dB. Numerically, it can be seen that the method in this embodiment significantly outperforms the traditional method in both key metrics.

[0083] Based on the advantages of the visual fit and quantitative indicators of the slice curves mentioned above, it can be concluded that the method in this embodiment has higher accuracy and reliability in phase unwrapping tasks, and effectively improves the quality of phase reconstruction.

[0084] 4. Overview

[0085] In summary, the UAV-borne InSAR phase unwrapping method based on WaveUNet proposed in this embodiment has the following technical advantages compared with the prior art:

[0086] 1. This invention achieves multi-scale, adaptive noise suppression of the input phase map by embedding a wavelet Wiener filtering module before each downsampling stage of the encoder. This module uses discrete wavelet transform to decompose the phase information and performs Wiener filtering based on local variance estimation in the high-frequency subband. While effectively removing speckle noise and system noise, it can significantly preserve phase edge and detail information, thereby greatly improving the phase unwrapping accuracy and robustness of the model under strong noise interference such as low coherence region and steep terrain. Compared with traditional phase unwrapping methods, it can better handle the phase unwrapping problem under low signal-to-noise ratio conditions.

[0087] 2. This invention integrates a wavelet Wiener filter module before each downsampling stage of the U-Net encoder, forming a multi-level, multi-scale noise suppression architecture. This enables the model to identify and suppress noise at different feature scales, enhancing its adaptability to complex noise patterns and avoiding the imbalance between global and local noise suppression in traditional single-scale denoising methods. This results in a more comprehensive and stable improvement in phase signal-to-noise ratio.

[0088] 3. This invention employs a U-Net symmetric encoding / decoding structure with skip connections and introduces a DoubleConv module for feature enhancement, enabling the network to simultaneously fuse deep semantic information and shallow detail features. This not only accelerates the feature extraction and reconstruction process but also effectively alleviates the gradient vanishing problem through skip connections, improving training stability and convergence speed, and achieving efficient end-to-end phase mapping. Furthermore, during the network's forward propagation, bilinear interpolation is used to calibrate the dimensions of the transposed upsampled feature map and the skip connection feature map in real time, avoiding information loss or training instability caused by resolution mismatch, thus improving the method's practicality and generalization performance.

[0089] 4. This invention uses the mean squared error loss function combined with an adaptive optimization algorithm to train the model in an end-to-end manner, enabling the network to directly learn the complex mapping relationship of phase unwrapping from the data. This eliminates the dependence on the traditional phase continuity assumption. Compared with the traditional unwrapping algorithm based on iterative optimization, this method not only improves the unwrapping efficiency, but also significantly improves the unwrapping accuracy and model generalization ability in complex scenarios through data-driven optimization.

[0090] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for unmanned aerial vehicle (UAV) InSAR phase unwrapping based on WaveUNet, characterized in that, Includes the following steps: Collect InSAR wrapped phase map training data; A multi-scale phase unwrapping model is constructed, and the multi-scale phase unwrapping model is trained based on InSAR wrapped phase map training data to obtain the trained multi-scale phase unwrapping model; the multi-scale phase unwrapping model is constructed based on the U-Net network architecture, and a wavelet Wiener filter module is embedded in its feature extraction path; Obtain the InSAR wrapped phase map to be processed, and input the InSAR wrapped phase map to be processed into the trained multi-scale phase unwrapping model to obtain the absolute phase map after unwrapping the InSAR wrapped phase map to be processed.

2. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 1, characterized in that, The multi-scale phase unwrapping model includes an encoder, a decoder, and a wavelet Wiener filter module embedded in the encoder. The trained multi-scale phase unwrapping model takes the input InSAR wrapped phase map as the input of the encoder, performs multi-scale adaptive noise suppression step by step through a multi-level wavelet Wiener filter module, and extracts the noise reduction feature maps of each level. The encoder downsamples and extracts deep features from the noise reduction feature maps of each level to obtain deep feature maps. The decoder upsamples the deep feature maps and fuses them with the output feature maps of the corresponding layers of the encoder through skip connections. Finally, the unwrapped absolute phase map is generated as the output of the multi-scale phase unwrapping model.

3. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 2, characterized in that, The encoder comprises four sequentially connected coding layers. Each coding layer includes a cascaded wavelet Wiener filter module, a DoubleConv module, and a max-pooling downsampling layer. The DoubleConv module comprises two cascaded convolutional units. The input of the encoder serves as the input of the wavelet Wiener filter module in the first coding layer, the output of the wavelet Wiener filter module in the first coding layer serves as the input of the DoubleConv module in the first coding layer, the output of the DoubleConv module in the first coding layer serves as the input of the max-pooling downsampling module in the second coding layer, the output of each coding layer serves as the input of the next layer connected to it in the encoder, and the output of the fifth coding layer serves as the output of the encoder.

4. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 3, characterized in that, Each convolutional unit in the DoubleConv dual convolutional module includes a cascaded 3×3 convolutional layer, a BN layer, and a ReLU activation layer; The input feature map is first used as the input of the first convolutional unit. Features are extracted through a 3×3 convolutional layer, normalized through a BN layer, and then non-linearly activated through a ReLU activation layer to obtain the primary features. The primary features output by the first convolutional unit are used as input to the second convolutional unit for deep feature extraction. After performing the same operation as the first convolutional unit, the deep features output by the second convolutional unit are used as the output of the DoubleConv module.

5. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 3, characterized in that, The processing procedure of the wavelet Wiener filter module includes the following steps: The input feature map is decomposed into a low-frequency sub-band LL and three high-frequency sub-bands LH, HL, and HH using discrete wavelet transform. Local variance is calculated within the high-frequency sub-bands using a sliding window. The energy distribution of signal and noise is dynamically estimated based on Wiener filtering theory, and an adaptive gain coefficient is calculated. This adaptive gain coefficient is then applied to each high-frequency sub-band for filtering. The filtered high-frequency subband and the unprocessed low-frequency subband are reconstructed using inverse wavelet transform to obtain the denoised phase feature map.

6. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 5, characterized in that, The formula for calculating the adaptive gain coefficient is as follows: ； In the formula, For adaptive gain coefficients; The signal variance of the current sub-band; Noise variance estimated from the highest frequency subband HH; Filtering using adaptive gain coefficients is expressed as follows: ； In the formula, These are the filtered wavelet coefficients; These are the original wavelet coefficients.

7. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 6, characterized in that, The decoder comprises four cascaded decoding layers, each of which includes a cascaded DoubleConv module and a transposed convolutional upsampling layer. The input of the decoding layer serves as the input of the DoubleConv module, the output of the DoubleConv module serves as the input of the transposed convolutional upsampling layer, and the output of the transposed convolutional upsampling layer serves as the output of the decoding layer.

8. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 7, characterized in that, The multi-scale phase unwrapping model also includes a bottleneck layer, which includes a cascaded wavelet Wiener filter module and a DoubleConv module. The entangled phase map input to the multi-scale phase unwrapping model serves as the input to the first encoding layer. The output of each encoding layer serves as the input for connecting to the next layer in the multi-scale phase unwrapping model. The output of the fourth encoding layer serves as the input to the bottleneck layer. The output of the bottleneck layer is upsampled by transposed convolution and then skip-connected with the output of the fourth DoubleConv module. After alignment using bilinear interpolation, it serves as the input to the fourth decoding layer. The input to each decoding layer is the feature map fused with the output of the DoubleConv module in its corresponding encoding layer through skip connections and bilinear interpolation with the output of the previous layer in the multi-scale phase unwrapping model. The output of the first decoding layer is convolved by 3×3 and serves as the overall output of the multi-scale phase unwrapping model.

9. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 1, characterized in that, The multi-scale phase unwrapping model is trained in the following manner: The unwrapped phase map that has been unwrapped in advance is used as a training sample to form a training sample set, which is then input into the multi-scale phase unwrapping model. The learning rate and hyperparameters of the multi-scale phase unwrapping model are set, and a mean squared error loss function is constructed to evaluate the unwrapping performance. The parameters of the multi-scale phase unwrapping model are optimized and updated with the goal of minimizing the mean squared error loss function, and then the multi-scale phase unwrapping model is trained.

10. The UAV-borne InSAR phase unwrapping method based on WaveUNet according to claim 9, characterized in that, The mean squared error loss function is: ； In the formula, For the sample size, Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

Images