CN115963498A - Glacier zone InSAR large-gradient phase unwrapping method - Google Patents

Abstract

The present invention provides a large-gradient phase unwrapping method for InSAR in glacier areas, comprising the following steps: generating training samples based on the interferogram simulation technology of DEM differential values in glacier areas, and constructing a symmetrical phase unwrapping network mixed with a convolutional layer and a Transformer After the model trains the training samples, the predicted phase discontinuity information is input into the maximum flow\min cut algorithm as a probability mass map to estimate the phase winding count and complete the phase unwrapping. The invention utilizes the powerful learning summarization ability and data mining ability of deep learning to get rid of the dependence of phase unwrapping on the assumption of phase continuity.

Description

A large gradient phase unwrapping method for InSAR in glacier regions

technical field

The invention belongs to the technical field of space-borne synthetic aperture radar interferometry, and in particular relates to an InSAR large-gradient phase unwrapping method in a glacier area.

Background technique

Under the background of global warming, glaciers are in a state of rapid material loss and gradual instability. Due to the huge volume and mass of glaciers, large-scale sudden separation of glaciers (also called glacier collapse) can cause serious mountain disasters. Several recent studies have found that glaciers often show obvious signs of pulsation before a large-scale sudden separation occurs. In addition, the glacier jump itself can also cause some mountain disasters, such as the melting water generated by the rapid melting of the jumping ice body, causing mudslides, the jumping ice body blocking the canyon and causing the flood of the barrier lake, and the jumping body destroying the glacier lake and causing the glacier lake outburst flood. Therefore, monitoring glacier surge has important guiding significance for glacier disaster perception. Since glacier jumping can cause significant changes in glacier surface elevation, the monitoring of ice surface elevation change is one of the important means to study glacier jumping. Spaceborne Synthetic Aperture Radar Interferometry (InSAR) has all-weather observation capabilities and can provide a key data source for continuous, timely and comprehensive monitoring of glacier surface elevation changes. When generating the ice surface DEM, the single-shot and double-receive SAR images are not affected by temporal decoherence of the glacier surface, glacier movement, and atmospheric changes. Therefore, after the launch of German TanDEM-X single-shot and double-receive image data in 2011, InSAR technology has been widely used in the field of glacier surface elevation change monitoring.

Phase unwrapping is the core step of InSAR technology to obtain surface elevation. Since the original interferometric phase fringes in mountainous areas generated by general SAR images are very dense, the success rate of direct unwrapping is extremely low. It is necessary to make a difference between the original phase and the terrain phase simulated by the external DEM, and then unwrap the differential phase. A new DEM can be obtained by converting the unfolded differential phase into an elevation difference and adding an external DEM. Traditional phase unwrapping methods are almost based on the assumption of phase continuity, that is, the phase difference between adjacent pixels does not exceed π. However, when the glacier jumps or collapses, the surface elevation change may cause the phase difference between adjacent pixels on the InSAR differential phase map to exceed π. Unwrapping with the traditional method will cause obvious jumps in the unwrapping phase, which eventually leads to gross errors in the key target areas in the generated InSAR DEM, which cannot be used to estimate the elevation change of the glacier surface.

The traditional InSAR phase unwrapping methods are almost all based on the assumption of phase continuity, that is, the phase difference between adjacent pixels does not exceed . Traditional phase unwrapping methods cannot obtain accurate unwrapped phases when the assumption of phase continuity is violated by dramatic phase changes or noise caused by significant elevation changes in glacier regions. In addition, even if the assumed condition of phase continuity can be satisfied, the path tracing method is easy to form a large number of closed areas that cannot be unwrapped under the condition of low signal-to-noise ratio, resulting in the "island" phenomenon; The error is propagated to the whole unwrapping area; the efficiency of processing interference data based on statistical methods is low, and the requirements for computer hardware equipment are high.

The existing deep learning phase unwrapping methods still have the following problems: (1) Most of the existing methods are directly tested on simulated data, which makes the model insufficiently consider the spatial distribution characteristics of the real InSAR interferometric phase, and phase classification is prone to occur in complex terrain (2) The network designed by the existing method is relatively shallow, and it is difficult to fit the phase data with more residuals, and the deep learning models used almost all use the convolutional layer as the basic unit, and the performance of the model is affected by the convolutional layer. It is not sensitive to the global position of the feature, and it is difficult to track the long-distance dependencies in the image. (3) Since the phase winding count estimation is based on the minimum norm, it is easy to propagate the wrongly predicted phase gradient information to the entire region.

Therefore, a large-gradient phase unwrapping method for InSAR in glacier regions is needed in this field.

Contents of the invention

The purpose of the present invention is to provide a large-gradient InSAR phase unwrapping method in glacier regions to solve the problem that the traditional InSAR phase unwrapping methods proposed in the background are almost all based on the assumption of phase continuity and have different shortcomings.

The technical solution of the present invention is a method for large-gradient phase unwrapping of InSAR in glacier regions, comprising the following steps:

Step 1. Generate training samples based on the interferogram simulation technology of DEM differential values in the glacier area,

Step 2. Train the training samples by constructing a symmetrical phase unwrapping network model mixed with a convolutional layer and a Transformer,

Step 3. Input the predicted phase discontinuity information as a probability mass map into the maximum flow\min cut algorithm to estimate the phase winding count and complete the phase unwrapping.

In a specific embodiment, in the step 1, the differential value of the SRTM (ShuttleRadar Topography Mission) DEM data and the COP-DEM data of the glacier-covered mountainous area is selected for training sample simulation.

In a specific implementation, the specific process of step 1 includes: unifying the two-period DEM elevation reference, registering the two-period DEM, difference two-period DEM, simulating the topographic difference phase map based on the DEM difference map, and then adding the simulated Atmospheric turbulent noise, Gaussian noise, large-amplitude noise of random shape, and large-amplitude noise are used to simulate severe decorrelation caused by water bodies and mountain shadows, estimate the coherence of the simulated interference phase after adding noise, and finally wrap the simulated absolute phase to get Simulates winding phases.

In a specific implementation, in step 2, the residual map is calculated based on the simulated winding phase map based on the "phase continuity assumption", the glacier label map is clipped according to the position of the simulated phase map, and finally the above two factors and The simulated winding phase map is used as the feature of the three-channel input of the deep learning model; the phase discontinuity point is calculated from the simulated absolute phase map, and the phase discontinuity point map is used as the output of the deep model.

In a specific implementation, in the step 2, a CNN and Transformer hybrid symmetric phase unwrapping architecture GTPU-Net is constructed, including a skip connection (skipconnection) of an encoder, a decoder and an attention mechanism; The basic unit of the model encoder is the Vision Transformer.

In a specific implementation, the step 3 specifically includes: defining an energy function in a Markov random field, using the phase discontinuity information as the prior knowledge of the phase change used in the estimation of the winding count, and constructing a Markov For directed graphs, the weights of each side of the directed graph are calculated according to the energy function, and the winding count k that minimizes the energy function is solved based on the maximum flow\minimum cut algorithm to complete the phase unwrapping.

The beneficial effects of the present invention include:

1. The present invention utilizes the powerful learning summary ability and data mining ability of deep learning to get rid of the dependence of phase unwrapping on the assumption of phase continuity.

2. The present invention uses the DEM differential value of the glacier area to simulate the dual-station InSAR differential phase, and adds a variety of noises according to the phase characteristics of the glacier area, so that the obtained InSAR phase training samples are more in line with the scene characteristics of the glacier surface elevation change monitoring scene.

3. The present invention builds a symmetric phase unwrapping network model with a mixture of convolutional layers and Transformers, and fuses the local information extracted by the convolution module and the global information extracted by the Transformer through skip connections to achieve high-precision segmentation of interference images and solve the problem of convolution Layers are not sensitive to the global location of image features and are difficult to track long-range dependencies in images.

4. The present invention performs winding count estimation based on Markov random energy field and maximum flow/minimum cut algorithm, solves the problem of inconsistency between the unwrapped phase and the original interferometric phase after regression prediction based on deep learning, and suppresses the low coherence area at the same time The propagation of mispredicted phase gradient information makes the large-gradient phase unwrapping results of InSAR in glacier regions more reliable.

In addition to the objects, features and advantages described above, the present invention has other objects, features and advantages. Hereinafter, the present invention will be described in further detail with reference to the drawings.

Description of drawings

The accompanying drawings constituting a part of this application are used to provide further understanding of the present invention, and the schematic embodiments and descriptions of the present invention are used to explain the present invention, and do not constitute an improper limitation of the present invention. In the attached picture:

Fig. 1 is a comparison diagram of sample data simulation; where: (a) is the unfolded phase map; (b) is the interferogram; (c) is the glacier boundary map; (d) is the discontinuity point calculated based on the assumption of phase continuity; ( e) is the coherence map; (f) is the discontinuity calculated based on the unwrapped phase.

Figure 2 is a comparison diagram of the interferogram of a certain glacier in Xinluhai, the unfolded phase generated based on the GAMMA software minimum cost flow method, and the unfolded phase generated by the present invention; the area marked 1 is the ice tongue area.

Figure 3 is a comparison diagram of the interferogram of another glacier in Xinluhai, the unfolded phase generated based on the GAMMA software minimum cost flow method, and the unfolded phase generated by the present invention; the marked 2 area is the ice tongue area.

Fig. 4 shows the phase continuity of the InSAR interferogram of the present invention.

Detailed ways

The embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings. The specific embodiments described here are only used to explain the present invention, and are not intended to limit the present invention.

Example 1

In order to make the phase sample information more consistent with the glacier differential interferogram scenario and ensure the effectiveness of model training, the present invention, based on the multi-period real pulsating glacier DEM differential map, carries out InSAR interferometry that takes into account the significant elevation changes and multi-source noise on the glacier surface phase simulation. Its principle and process are as follows:

Considering that the effective information used by the model is phase gradient information, and the spatial resolution of the interferogram after multi-view processing is about 30m, the present invention selects the differential value of SRTM DEM data and COP-DEM data in glacier-covered mountainous areas for training samples simulation. Using the orbital parameters of the TanDEM-X satellite, the difference between the SRTM DEM data and the COP-DEM data is transformed into the terrain phase in the radar coordinate system:

where ψ(m,n) is the topographic phase at point (m,n), B _⊥ is the vertical baseline length, λ is the wavelength, R is the satellite orbital altitude, θ is the satellite incident angle, h(m,n) reverse geographic The encoded DEM differential map. Add Perlin noise to the terrain phase with a probability of 30% to simulate the influence of atmospheric turbulence on the interferometric phase, and add a large gradient deformation phase simulated by a two-dimensional Gaussian function with a probability of 20% to obtain the initial simulated phase.

Random complex Gaussian noise is added to the initial simulated phase map, and the coherence of the interferogram is estimated according to the added noise, which is used as a phase quality index for traditional algorithm unwrapping. At the same time, considering the complex terrain of the glacier area, factors such as water bodies and mountain shadows may cause severe decorrelation, random-shaped large-amplitude noise is added to the simulated phase map with a probability of 10%, and finally the simulated phase winding is obtained through phase winding .

The present invention regards the phase unwrapping problem as a semantic segmentation problem, but it is different from the existing deep learning phase unwrapping method that divides images into multiple categories based on the winding count. The present invention defines phase discontinuity points as adjacent For the point where the phase difference exceeds, the phase map is classified only according to whether the pixel is a phase discontinuity point, which reduces the stringent requirements of the model performance. The interferogram, residual map and glacier boundary map are used as input data, and the phase Discontinuity point prediction map as output data. At the same time, in order to solve the problem that the convolutional layer is not sensitive to the global position of the feature and it is difficult to track the long-distance dependencies in the image, and ensure the performance and stability of the model, the present invention adopts a CNN and Transformer hybrid symmetric phase unwrapping architecture GTPU- Net, which consists of an encoder, a decoder, and a skip connection that adds an attention mechanism.

In order to solve the problem of unbalanced binary classification samples, the present invention adopts a fusion strategy of two loss functions (Loss) of binary classification cross-entropy loss function (BCELoss) and focal loss function (Focal Loss). At the same time, model training is performed twice for the horizontal and vertical discontinuities of the pixels, and the predicted phase discontinuity information is input into the unwrapping algorithm as a probability mass map to estimate the phase winding count.

The invention estimates the phase winding count and obtains the unfolded phase based on the maximum flow/minimum cut theory and phase discontinuity information. Figure 4 shows the continuity of pixel point (i, j) and its first-order neighborhood in the interferogram. where (i,j)∈G ₀ , h _ij and v _ij denote horizontal and vertical discontinuities, respectively.

Define the energy function in a Markov random field:

和

分别表示像素水平和垂直差异，同时有：where k is the winding count, ψ is the interference phase, V(.) is the clique potential,

and

represent the pixel horizontal and vertical differences, respectively, and have:

The goal of this method is to find the winding count k that minimizes the energy function. The h _ij and v _ij are replaced by the phase continuity probabilities in the horizontal and vertical directions predicted by the GTPU-Net model as the phase change prior knowledge used in the entanglement count estimation. For the convex potential V, the minimization of E(k|ψ) can be achieved through a series of binary optimizations, each binary problem is mapped to a binary image, and can be obtained by computing the maximum flow/minimum cut on the image Binary minimization of an objective function.

The interferogram simulation method in step 1 of the present invention is not the common numerical simulation and DEM back calculation, but the interferometric phase simulation based on the DEM difference map of the multi-period real pulsating glacier area and the fusion of multi-source noise, taking into account the elevation change characteristics of the glacier surface method.

The present invention designs a symmetric phase unwrapping architecture that combines a convolutional layer and a Transformer, and adds an attention mechanism to the skip connection at the same time, which can solve the problem that the convolutional layer is not sensitive to the global position of image features and it is difficult to track long-distance dependencies in the image question. At the same time, in order to enable the model to better learn and summarize the phase discontinuity information in the scene of mountain glacier jumping, the semantic information closely related to the distribution of phase discontinuity points is selected as the model input, and the phase information is completed on the basis of phase information encoding and decoding. Prediction of discontinuity information, thus ensuring the robustness of the model.

The present invention does not use the minimum ^L1 norm to estimate the winding count, but builds a Markov directed graph based on the interferogram, and inputs the phase discontinuity information predicted by the deep learning model as the prior information of the phase change to the maximum In the flow \ minimum cut model, this method can effectively suppress the propagation of the phase gradient information incorrectly predicted by the deep learning model in the phase winding count estimation method. Compared with the winding count estimation method based on the L ¹ norm, this method is more robust sex.

The invention considers the InSAR phase training sample simulation technology of the glacier surface elevation variation characteristics. Most of the existing interferogram simulation methods use simple numerical simulation or DEM inverse calculation, which do not take into account the spatial distribution characteristics of real InSAR differential interferometric phases. The present invention proposes an interferogram simulation method based on the DEM differential value of the glacier area, which uses multi-period DEM differential values to simulate the large gradient phase caused by the significant elevation change in the glacier area, and then fuses multi-source noise to fully restore the InSAR in the real glacier pulsation scene The spatial distribution feature of differential interferometric phase improves the quality of training samples and ensures the effectiveness of model training.

The invention considers the network model construction technology of the global characteristics of the InSAR interferogram. The performance of network models adopted by existing deep learning phase unwrapping methods is limited by the insensitivity of convolutional layers to the global position of features and the difficulty of tracking long-range dependencies in images. The invention constructs a symmetric phase unwrapping network model mixed with a convolution layer and a Transformer, fuses the local information extracted by the convolution module and the global information extracted by the Transformer through skip connections, and realizes high-precision segmentation of interference images.

The invention is based on the Markov random energy field and the InSAR phase winding number estimation technology of the maximum flow/minimum cut algorithm. Existing phase winding number estimation methods tend to propagate mispredicted winding number gradient information to the entire target area. The invention designs an energy function considering the prior distribution information of the phase discontinuity based on the Markov random field, transforms the interference graph into a directed graph including source points and sink points, and uses the energy function to assign weights to the edges of the directed graph , to optimize the winding number estimation by a max-flow/min-cut algorithm, which effectively suppresses the propagation of mispredicted phase gradient information.

Figure 1 shows the simulated data generated by the InSAR phase training sample simulation method that takes into account the characteristics of glacier surface elevation changes proposed by the present invention, wherein the interferogram, glacier boundary and residual map are the input data of the deep learning model, and the phase discontinuity points are used to predict The graph is used as the output data, and then the discontinuity point prediction graph is input into the max-flow/min-cut algorithm as an InSAR quality graph for the estimation of the winding count, and finally the phase unwrapping is completed.

Figure 2 and Figure 3 show the case of glacier monitoring in Xinluhai. The ice tongue area is the most active area of glaciation and is also the ablation area of the glacier. In the figure, the ice tongue area 1 is marked in the unwrapping map generated by GAMMA software , 2 were not correctly unwrapped, but using the unwrapping method of the present invention, combined with the shape of the real glacier of Xinluhai Glacier, it can be seen that the glacier part has been correctly unwrapped.

The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be assumed that the specific implementation of the present invention is limited to these descriptions. For those of ordinary skill in the technical field of the present invention, without departing from the concept of the present invention, some simple deductions and substitutions can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims (6)

1. An InSAR large gradient phase unwrapping method in a glacier region is characterized by comprising the following steps:

step 1, generating a training sample based on an interferogram simulation technology of the DEM differential value of the glacier region,

step 2, training a training sample through the constructed symmetrical phase unwrapping network model of the convolutional layer and the Transformer,

and 3, inputting the predicted phase discontinuity information serving as a probability mass chart into a maximum flow/minimum cut algorithm, and performing phase winding counting estimation to finish phase unwrapping.

2. The InSAR large-gradient phase unwrapping method as recited in claim 1, wherein in the step 1, a difference value between SRTM (shuttleRadarTopographiyMissision) DEM data and COP-DEM data of the mountain area covered by the glacier is selected for training sample simulation.

3. The glacier zone InSAR large gradient phase unwrapping method as set forth in claim 2, wherein the specific process of the step 1 comprises: unifying two-stage DEM elevation datum, registering the two-stage DEM, differentiating the two-stage DEM, simulating a terrain differential phase diagram based on a DEM differential diagram, then adding simulated atmospheric turbulence noise, gaussian noise and random-shaped large-amplitude noise, wherein the large-amplitude noise is used for simulating serious decorrelation caused by water and mountain shadow, estimating the coherence of interference phases after the noise is added, and finally winding the simulated absolute phases to obtain simulated winding phases.

4. The InSAR large-gradient phase unwrapping method in the glacier region according to claim 1, wherein in the step 2, a residual error map is calculated based on a phase continuity assumption according to a simulated wrapped phase map, a glacier label map is cut according to the position of the simulated phase map, and finally the two factors and the simulated wrapped phase map are used as features input by three channels of a deep learning model; and calculating phase discontinuity points according to the simulated absolute phase diagram, and outputting the phase discontinuity point diagram as a depth model.

5. The Islamic district InSAR large gradient phase unwrapping method as recited in claim 1, wherein in the step 2, a CNN and Transformer mixed symmetric phase unwrapping structure GTPU-Net is constructed, which comprises an encoder, a decoder and a skip connection (skip connection) with attention mechanism; the basic unit of the model encoder is VisionTransformer.

6. The glacier zone InSAR large gradient phase unwrapping method of claim 1, wherein the step 3 specifically comprises: defining an energy function in a Markov random field, taking phase discontinuity information as phase change prior knowledge used in winding count estimation, constructing a Markov directed graph, calculating the weight of each side of the directed graph according to the energy function, solving a winding count k for minimizing the energy function based on a maximum flow \ minimum cut algorithm, and finishing phase unwrapping.

Images