CN-122017835-A - InSAR phase unwrapping method based on terrain gradient sensing network

CN122017835ACN 122017835 ACN122017835 ACN 122017835ACN-122017835-A

Abstract

The invention discloses an InSAR phase unwrapping method based on a terrain gradient sensing network, and relates to the field of synthetic aperture radar interferometry. The method comprises the steps of obtaining a phase unwrapping data set, dividing the phase unwrapping data set into a training data set, a verification data set and a test data set, constructing a terrain gradient sensing network, inputting the training data set into the terrain gradient sensing network for iterative training, storing optimal model weights after training is finished, and finally inputting the test data set into the trained terrain gradient sensing network to output a predicted unwrapping phase map. According to the scheme, noise suppression and detail fidelity are simultaneously considered under the conditions of low signal-to-noise ratio and data loss, residual numbers and streak artifacts are suppressed, phase continuity and physical consistency constraint are met, unwrapping precision, stability and scene generalization capability are improved, and a reliable phase unwrapping scheme is provided for synthetic aperture radar interferometry.

Inventors

ZHAO ZHIHONG
TONG AIHUA
WANG HAOYU
Chen Daiyang
ZHANG RAN

Assignees

石家庄铁道大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. An InSAR phase unwrapping method based on a terrain gradient sensing network, the method comprising: S1, acquiring a phase unwrapping data set, and dividing the phase unwrapping data set into a training data set, a verification data set and a test data set; S2, constructing a terrain gradient sensing network, wherein the terrain gradient sensing network comprises a data loading and preprocessing module, a phase diagram terrain gradient change sensing coding module, a gradient guiding dynamic weight multi-scale cavity pooling module, a gradient self-adaptive decoding module and a nerve embedded post-processing module; s3, inputting the training data set into a terrain gradient sensing network for iterative training to generate a trained terrain gradient sensing network, wherein the verification data set is used for monitoring training effects in real time, calculating loss values through a multi-constraint loss function, and storing optimal model weights after training is completed through counter-propagation iterative optimization model parameters; S4, inputting the test data set divided in the phase unwrapping data set into a trained terrain gradient sensing network, and outputting a predicted unwrapping phase map.
2. The InSAR phase unwrapping method based on the terrain gradient sensing network as set forth in claim 1, wherein in step S1: The phase unwrapping dataset includes a training set of 30000 analog phase pairs, a validation set of 1000 analog phase pairs, and a test set of 100 true InSAR phase pairs.
3. The InSAR phase unwrapping method based on the terrain gradient sensing network as claimed in claim 1, wherein: The data loading and preprocessing module is used for carrying out standardization and format conversion operation on an input winding phase diagram, uniformly standardizing the numerical range of the input phase diagram to a [ -pi, pi ] interval, adjusting the size of the phase diagram to 256 multiplied by 256 pixels, ensuring the data dimension consistency in batch training or prediction, adapting to the model input requirement, converting a single-channel phase diagram into a tensor format, and enabling the dimension to be [ B, 1, 256, 256], wherein B is the batch size, carrying out enhancement operation on a training data set, and improving the generalization capability of the model.
4. The InSAR phase unwrapping method based on the terrain gradient sensing network as set forth in claim 1, wherein the phase map terrain gradient change sensing coding module comprises an encoder branch and a terrain gradient sensing module branch, the encoder branch and the terrain gradient sensing module branch are connected in parallel with an input layer and are processed synchronously, the encoder branch sequentially comprises a first residual error unit, a first maximum pooling layer, a second residual error unit, a second maximum pooling layer, a third residual error unit, a third maximum pooling layer, a fourth residual error unit, a fourth maximum pooling layer, a fifth residual error unit, a fifth maximum pooling layer and a sixth residual error unit which are connected in series, each residual error unit comprises two layers of convolution, batch normalization and activation functions and achieves addition of input and convolution output through residual error connection, the terrain gradient sensing module branch comprises a horizontal gradient convolution layer, a vertical gradient convolution layer, a normalization and shaping gradient profile generating unit and a dynamic weight generating unit which are sequentially connected in series, the gradient profile is used for space guidance of a decoding stage, and the dynamic weight is used for multi-scale response modulation.
5. The InSAR phase unwrapping method based on the terrain gradient sensing network as claimed in claim 4, wherein: The phase diagram terrain gradient change perception coding module receives the preprocessed data, extracts multi-scale phase characteristics and generates terrain gradient guiding information, two parallel branches are formed by an encoder consisting of 6 coding layers and 1 terrain gradient perception module, wherein the number of characteristic channels of the encoder branch is increased from 1 layer by layer to 256, the spatial resolution is halved from 256 to 8 x 8 layer by layer, each coding layer of the encoder takes two 3x 3 convolutions as cores, the step size is 1, the batch normalization BN and ReLU activation is carried out to accelerate training convergence and enhance nonlinear expression, and then the input x and the characteristics after convolution transformation are added through residual connection to form Wherein: ; And (3) with Respectively a first 3 x 3 convolution, a second 3 x 3 convolution, Normalizing the batch; The parallel terrain gradient sensing module calculates horizontal gradients and vertical gradients on the input phase map and normalizes the horizontal gradients and the vertical gradients to generate a gradient distribution map and dynamic weights; The terrain gradient sensing module is deployed in parallel with 6 coding layers, synchronously receives the preprocessed winding phase diagram, and has dimensions of [ B,1,256,256], wherein B is the batch size; The processing method of the terrain gradient sensing module comprises four steps: 1) Gradient calculation, namely calculating first partial derivatives of an input phase diagram along an x axis and a y axis respectively by adopting a Sobel operator adapting to periodic characteristics of the phase diagram to obtain a horizontal gradient and a vertical gradient, wherein the calculation formula is as follows: ; Wherein the method comprises the steps of Is a horizontal gradient which is formed by a plurality of vertical gradients, Is a vertical gradient which is formed by a vertical gradient, And Is a 3×3 adaptive Sobel convolution kernel; Calculating the gradient amplitude of each pixel representing the steepness of the terrain and the trend of the terrain representing the gradient direction, wherein the calculation formula is as follows: ; Wherein the method comprises the steps of Is a horizontal gradient which is formed by a plurality of vertical gradients, Is a vertical gradient, and outputs a gradient amplitude diagram with the dimensions of [ B,1,256,256] And gradient pattern ; 2) And carrying out gradient normalization, namely calculating the maximum value and the minimum value of the gradient amplitude map according to batches, and carrying out normalization operation on each pixel, wherein the calculation formula is as follows: ; Wherein, the Is the maximum value of the gradient magnitude map, Outputting normalized gradient amplitude map as minimum value of gradient amplitude map And a gradient pattern preserving the original values ; 3) Providing basis for gradient guidance requirement of a decoding layer, firstly performing 1×1 lightweight convolution operation on a normalized gradient amplitude map to keep the channel number as 1, generating a pixel-by-pixel attention weight map, performing adaptive scaling on characteristic responses of different spatial positions through the attention weight map, enabling high gradient and structure abrupt change regions to obtain higher weight to enhance edge and detail expression, enabling low gradient and flat regions to obtain lower weight to inhibit noise disturbance, and finally generating a gradient distribution map with the dimension of [ B,1,256,256], wherein the calculation formula is as follows: , Wherein the method comprises the steps of Representing a1 x 1 lightweight convolution operation, Representing a gradient-guided spatial weighting operation on the normalized gradient magnitude map, The gradient distribution map is directly transmitted into each up-sampling branch of the decoding layer, and provides basis for gradient guiding and enhancing links in the decoding stage; 4) For dynamic expansion rate weight generation, normalized gradient amplitude diagram is firstly generated Performing global average pooling operation to obtain an average gradient value of a batch level, distributing adaptive weights for 4 branches with different expansion rates of a cavity space pyramid pooling ASPP module based on the average gradient value, packaging the weights into a learnable parameter, and finally outputting a dynamic expansion rate weight vector with the dimension of 4, wherein the calculation formula is as follows: ; Wherein the method comprises the steps of Representing a global average pooling operation, Representing a branch weight allocation operation based on the average gradient value, For the dynamic expansion rate weight vector, the weight vector can be directly transmitted into a cavity space pyramid pooling ASPP module and is used for weighting and fusing the characteristics of branches with different expansion rates, so that the dynamic adaptation of the multi-scale characteristics to the terrain gradient is realized.
6. The InSAR phase unwrapping method based on the terrain gradient sensing network of claim 1, wherein the gradient guiding dynamic weight multi-scale hole pooling module comprises a first hole convolution branch with expansion rate of 1, a second hole convolution branch with expansion rate of 3, a third hole convolution branch with expansion rate of 6, a fourth hole convolution branch with expansion rate of 12 and a global context branch, outputs of the five branches are fused with a batch normalization and activation function through a layer of 1×1 convolution after channel dimension splicing to output bottleneck enhancement features, wherein the dynamic weight is associated with weights of the four hole convolution branches, and adaptive modulation is performed on each branch response.
7. The InSAR phase unwrapping method based on the terrain gradient aware network of claim 6, wherein the processing method of the gradient guiding dynamic weight multi-scale hole pooling module comprises the following steps: each branch of the parallel basic branches comprises a basic expansion rate corresponding to 3 multiplied by 3 convolution, a BN batch normalization layer and a ReLU activation function, and features with specific scales are extracted; Based on dynamic expansion rate weight, the output characteristics of each basic branch are differentially weighted, the high gradient region enhances the branch response with large expansion rate, and the gentle region enhances the branch response with small expansion rate, and the calculation formula is as follows: ; Wherein, the Is the first The original output of the individual base branches is, For the dynamic enhanced feature, 0.5 is an enhancement factor, and the contribution of each branch is balanced; Secondly, the global average pooling branch comprises an adaptive average pooling output size, 1 multiplied by 1 convolution, a BN layer and a ReLU activation, and is responsible for extracting global context information; Up-sampling the output characteristics of the global average pooling branch to 8 x 8 size, and keeping the dimension consistent with the parallel basic branch characteristics; Finally, 4 dynamic enhancement branches and global features are spliced in the channel dimension, and are fused into 256 channel features through 1X 1 convolution, wherein the calculation formula is as follows: ; Wherein, the Is a fusion feature map output by the hole space pyramid pooling module, and is the feature output by the coding residual error block After 4 void convolutions with different expansion rates, dynamic weights generated by a terrain gradient sensing unit And (5) obtaining weighted summation.
8. The InSAR phase unwrapping method based on the terrain gradient sensing network as claimed in claim 1, wherein: The gradient self-adaptive decoding module comprises a first gradient self-adaptive upsampling unit, a second gradient self-adaptive upsampling unit, a third gradient self-adaptive upsampling unit, a fourth gradient self-adaptive upsampling unit and a fifth gradient self-adaptive upsampling unit which are connected in series and are respectively connected with a first connecting layer to a fifth connecting layer, wherein each gradient self-adaptive upsampling unit comprises a transposed convolution upsampling layer, the convolution kernel size is 2 multiplied by 2, the step size is 2, a gradient weight generating layer receives a gradient distribution diagram matched with the current scale and generates a spatial attention weight through 1 multiplied by 1 convolution and Sigmoid, the enhancement convolution comprises two layers of 3 multiplied by 3 convolution and batch normalization, an activation function and a residual fusion layer, the enhancement branch is added with an upsampling main branch after being weighted according to attention, the upsampling output of each stage and the output of the corresponding connecting layer are spliced in a channel dimension, the upsampling output of each stage is fed into a next upsampling unit after being refined by the convolution residual fusion layer, the resolution is restored to 256 multiplied by 256 in sequence, and the key detail restoration is executed on a high gradient area in the whole process.
9. The InSAR phase unwrapping method based on the terrain gradient sensing network of claim 8, wherein in the gradient adaptive decoding module: the gradient self-adaptive decoding module firstly performs standard up-sampling on input features, realizes doubling of spatial resolution through 2×2 transpose convolution, and has a step length of 2, and a calculation formula is as follows: ; Wherein, the The features are input for the module and, As a feature after standard up-sampling, For the convolution kernel size, Step length, then outputting gradient distribution map of topography gradient sensing unit Upsampling to AND Same size, generating spatial attention weights The pertinence of the high gradient edge area is enhanced, and the calculation formula is as follows: ; Wherein, the For the up-sampled gradient profile, For the purpose of the channel splicing, Outputting the enhanced features, and entering an up-sampling residual block for further refinement; in the up-sampling residual block and skip connection fusion stage, the output characteristics and the skip connection characteristics of the corresponding scale of the encoder are combined Splicing in channel dimension, extracting fusion feature by two layers of 3X 3 convolution, stabilizing gradient propagation and feature reuse in residual mode, and combining the spatial attention weight obtained in the previous step Slightly adjusting splicing characteristics before fusion to strengthen information transmission of high gradient edges, wherein a calculation formula is as follows: ; Wherein, the For the characteristics of the corresponding level of the encoder, For the channel-splicing operation, Is the final output of the decoding layer and serves as input to the next decoding stage.
10. The InSAR phase unwrapping method based on the terrain gradient sensing network as set forth in claim 1, wherein the neural embedded post-processing module comprises a residue correlation correction part and a continuity optimization part, wherein the residue correlation correction part is based on adjacent differences of phases, continuous residue indication features are constructed through sine mapping, residue attention map is generated through a shallow convolution network, and is predicted to correct by another convolution network after being spliced with a phase map, the continuity optimization part calculates discontinuous responses through a fixed 3X 3 Laplace operator, accordingly generates adaptive smooth weights, meanwhile, small convolution network predicts smooth correction, and the phase is weighted and refined in combination, namely global continuity is enhanced in a flat area, real transition details are reserved in a mutation area, and finally, unwrapped phase maps are output through 1X 1 convolution.

Description

InSAR phase unwrapping method based on terrain gradient sensing network Technical Field The invention relates to the field of synthetic aperture radar interferometry, in particular to an InSAR phase unwrapping method based on a terrain gradient sensing network. Background The synthetic aperture radar interferometry technology has become a core means of earth surface deformation monitoring and elevation inversion by virtue of the advantages of all-day, all-weather and high spatial resolution, and phase unwrapping is used as a key link of data processing of the technology, and the accuracy and scene suitability of the technology directly determine the reliability of final earth surface data. The core goal of phase unwrapping is to recover continuous phase information reflecting the real surface morphology and deformation from the wrapping phase in the range of [ -pi, pi ], but in practical application, the problems of complex terrain gradient change, noise interference, dense phase jump, uneven distribution of phase loss and interference areas and the like are solved, and the unwrapping precision and robustness are always restricted. The current phase unwrapping technology is mainly divided into a traditional method and a deep learning method. The traditional phase unwrapping method has the inherent defects that a path tracking method, a minimum norm method, a network flow method and the like all depend on a manually designed physical model or mathematical rules, and the self-adaptive adjustment capability is lacking. In a scene with low signal-to-noise ratio and dense residual points, the unwrapping error accumulation is easily caused by noise and measurement error interference, when facing a steep terrain and a strong destructive interference area, the unwrapping phase fracture, excessive smoothness or local error global propagation are difficult to balance noise suppression and detail reservation, and the problems of low automation degree, calculation efficiency and scene suitability are difficult to consider. The existing deep learning method has obvious optimization bottleneck. First, complex terrain is not sufficiently adaptable. The mainstream model adopts the multiscale module of fixed knot structure more, lacks topography gradient perception mechanism, can't adjust the receptive field according to topography abrupt degree developments, leads to the flat regional detail to be too smooth, and steep regional context is caught inadequately. Secondly, edge detail restoration lacks pertinence. The phase gradient information is not effectively utilized in the up-sampling process of the decoder, the phase jump at the abrupt change of the terrain is positioned inaccurately, and the problems of blurred edges and discontinuous stripe breakage connection are easy to occur. Thirdly, the physical rationality and the processing flow are disjointed. The partial method relies on the traditional unwrapping frame or off-line post-processing, does not embed special physical logic for phase unwrapping, has single constraint of a loss function, is easy to cause accurate numerical value of unwrapping results but physical distortion, and the partial model also depends on additional auxiliary data or large-scale labeling data, so that generalization capability is limited. Disclosure of Invention The invention aims to solve the technical problem of providing an InSAR phase unwrapping method which can realize the unification of noise suppression and detail reservation and improve the precision, robustness and scene suitability of phase unwrapping. In order to solve the technical problems, the invention adopts the technical scheme that the InSAR phase unwrapping method based on the terrain gradient sensing network comprises the following steps: S1, acquiring a phase unwrapping data set, and dividing the phase unwrapping data set into a training data set, a verification data set and a test data set; S2, constructing a terrain gradient sensing network, wherein the terrain gradient sensing network comprises a data loading and preprocessing module, a phase diagram terrain gradient change sensing coding module, a gradient guiding dynamic weight multi-scale cavity pooling module, a gradient self-adaptive decoding module and a nerve embedded post-processing module; S3, inputting the training data set into the terrain gradient sensing network for iterative training to generate a trained terrain gradient sensing network, wherein the verification data set is used for monitoring training effects in real time, calculating loss values through a multi-constraint loss function, and storing optimal model weights after training is completed through counter-propagation iterative optimization model parameters; And S4, inputting the test data set into a trained terrain gradient sensing network, and outputting a predicted unwrapping phase map. The further technical scheme is that the phase diagram terrain gradient change sensing coding module compr