CN116068511A - Deep learning-based InSAR large-scale system error correction method - Google Patents

Abstract

The invention discloses an InSAR large-scale system error correction method based on deep learning, which includes the following steps: collecting and preprocessing the data covering the target area, obtaining unwrapped phase data after geocoding, and DEM data after clipping and longitude and latitude data; according to the relationship between relevant data, build a deep neural network model; input the preprocessed data into the deep neural network model, and obtain the simulated large-scale system error unwrapping phase; obtain the corrected unwrapping phase ; Invert the deformation parameters to obtain the corrected deformation results. The invention provides a model capable of simultaneously correcting terrain-related atmospheric phases, orbital residual phases, and partial turbulent phases, and can realize InSAR large-scale system error correction in high and steep mountainous areas, hills, and flat areas.

Description

A method for correcting large-scale system errors of InSAR based on deep learning

Technical Field

The present invention relates to the field of synthetic aperture radar interferometry technology, and in particular to an InSAR large-scale system error correction method based on deep learning.

Background Art

Synthetic aperture radar interferometry is a ground observation measurement technology based on satellite-borne sensors that has developed rapidly in the past three decades. It has the ability to monitor deformation over a large range, all day and all weather. With the continuous improvement of the temporal and spatial resolution of SAR images, the availability of multi-source and multi-temporal SAR images, and the continuous development and progress of InSAR time series algorithms, InSAR technology has been widely used in the field of early identification and monitoring of geological disaster hazards, such as volcano monitoring, ground subsidence monitoring, landslide monitoring, earthquake monitoring, and glacier movement. This technology can effectively make up for the shortcomings of traditional measurement methods. However, due to the existence of large-scale phase errors, such as ionosphere, terrain-related atmospheric phase, and orbital residual phase. These errors will have a serious impact on the accuracy of deformation measurement. For example, when the relative humidity of 20% changes in the atmosphere in time and space, it will cause an error of several decimeters in the deformation results; the orbital phase is similar to the spatial characteristics of large-scale deformation, which will seriously affect the monitoring of large-scale deformation, such as earthquakes and volcanic movements. Therefore, it is of great significance to comprehensively remove large-scale phase errors.

In order to remove large-scale InSAR system errors, many scholars have proposed a large number of methods for different large-scale system error components in the past few decades. For terrain-related atmospheric phase, a correction method based on external data is proposed, such as numerical meteorological products (Weather models), GNSS data, spectral analysis (MODIS, MERIS), or a combination of them; followed by methods based on correction models, such as linear correction method and power-rate correction method. For orbital residual phase, a satellite orbit trajectory compensation model is proposed to accurately estimate the baseline error in InSAR imaging geometry; followed by the characteristics of orbital phase in the spatial domain and frequency domain; then based on the assistance of external data, such as GNSS; and finally a mathematical fitting method, mainly including linear fitting, polynomial fitting and other models.

The correction methods proposed above have been proven to be successful to some extent in removing terrain-related atmospheric phase and orbital residual phase, but these methods are subject to their inherent limitations. Traditional methods based on linear models, polynomial models, etc. cannot fully describe the characteristics of each interferometer pair and orbit, and will be affected by the similar deformation of spatial characteristics and orbital phase and terrain-related atmospheric phase.

Summary of the invention

In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for correcting InSAR large-scale system errors based on deep learning, which provides a model that can simultaneously correct the terrain-related atmospheric phase, orbital residual phase and part of the turbulence phase, solving the problem of limitations of traditional methods.

In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for correcting large-scale system errors of InSAR based on deep learning, characterized in that the method comprises the following steps:

S1: Collect SAR data, DEM data and latitude and longitude data covering the target area, and perform data preprocessing to remove areas with poor coherence, and obtain geocoded unwrapped phase data, cropped DEM data and latitude and longitude data;

S2: Building a deep neural network model based on the relationship between relevant data, wherein the relevant data includes DEM data, longitude and latitude data, terrain-related atmospheric phase, orbital residual phase and turbulence phase;

S3: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model to obtain the simulated large-scale system error unwrapped phase;

S4: Subtract the original geocoded unwrapped phase from the large-scale system error unwrapped phase simulated by the deep neural network model to obtain the corrected unwrapped phase;

S5: The corrected unwrapped phase is de-geocoded to the SAR coordinate system, and the corrected deformation result is obtained by inverting the deformation parameters.

The beneficial effect of the above scheme is: through the above technical scheme, the orbital phase and the terrain-related atmospheric phase are comprehensively considered to obtain their functional relationship with the unwrapping phase. While eliminating the terrain-related atmospheric phase and the orbital phase, the spatial characteristics of the turbulent phase are considered, and part of the turbulent phase can be eliminated. At the same time, the terrain-related atmospheric phase, the orbital residual phase and part of the turbulent phase are corrected, which solves the problems limited by the traditional method and can identify deformation more accurately.

Furthermore, in S1, the SAR data is preprocessed into InSAR data, including the following steps:

S1-1: Obtain two SLC images of the same study area in time order and select one of them as the main image;

S1-2: The influence of speckle noise is reduced by setting the multi-view ratio of 10:2 in the distance and azimuth of the image, and the corresponding filtering parameters are set according to the size of the noise, and the filtered coherence data is obtained by filtering;

S1-3: Setting a disentanglement threshold based on the data distribution of the filtered coherence data to ensure sufficient training data and prevent the influence of disentanglement error data;

S1-4: According to the unwrapping threshold, the minimum cost flow method is selected to perform phase unwrapping on the wrapped phase calculated from the selected image interferogram;

S1-5: The unwrapped phase data in the SAR coordinate system is transferred to the same geographic coordinates as the cropped DEM data and longitude and latitude data through geocoding to obtain the geocoded unwrapped phase data.

The beneficial effect of the above further scheme is: through the above technical scheme, the collected data is preprocessed to obtain the data required for training the deep neural network.

Furthermore, the deep neural network model construction in S2 includes the following steps:

S2-1: The fully connected layer network of the deep neural network model is used as a feature module to extract the input DEM data and longitude and latitude data information. Combined with the relationship between DEM data, longitude and latitude data and terrain-related atmospheric phase, orbital residual phase and turbulence phase, the channel attention mechanism module of the deep neural network model is selected as the weighted feature extraction module of the data;

S2-2: Perform feature compression operation through weighted feature extraction module, compress the spatial dimension, and then generate corresponding weights for each relevant data feature dimension through restoration operation, including the following formula:

The eigenvalue compression formula is as follows:

为输入的数据，

和

分别为输入数据的长和宽，

为压缩操作，

为原始数据压缩后第

个二维矩阵，

和

均为图像像素坐标；in,

For the input data,

and

are the length and width of the input data respectively,

For compression operation,

The original data is compressed

A two-dimensional matrix,

and

All are image pixel coordinates;

The weight calculation formula is as follows:

为还原操作，

为RELU激活函数，

为RELU激活函数参数的降维层，

为RELU激活函数参数的升维层，

为特征通道权重，

为原始数据压缩后的结果，

为全局平均池化；in,

To restore the operation,

is the RELU activation function,

is the dimension reduction layer of the RELU activation function parameters,

is the dimension-raising layer of the RELU activation function parameters,

is the feature channel weight,

is the result after the original data is compressed.

is global average pooling;

S2-3: Apply the weight to each original relevant data feature channel. The formula is as follows:

为特征映射，

为尺度因子，

为最终得到的图像数据，

为特征映射与尺度因子在通道上的卷积。

is the feature map,

is the scale factor,

is the final image data,

is the convolution of the feature map and the scale factor on the channel.

The beneficial effect of the above further scheme is: through the above technical scheme, taking into account the data used, a deep neural network model is built to facilitate the subsequent acquisition of the simulated large-scale system error unwrapped phase, and at the same time, weights are applied to each of the original relevant data feature channels, which is conducive to learning the importance of different data channels.

Furthermore, the deep neural network model selects MSE as the loss function for parameter evaluation and the Adam optimization algorithm as the parameter optimization algorithm.

The beneficial effect of the above further scheme is: taking into account that the relationship between the terrain-related atmospheric phase and the terrain function varies in space, that is, the functional relationship between the terrain and the atmospheric phase in different regions is different, through the above technical scheme, the model can fit a better global parameter.

Furthermore, S3 includes the following steps:

S3-1: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model, the spatial features of the DEM data and longitude and latitude data are extracted through the fully connected module of the deep neural network model, and the weights of the DEM data, longitude and latitude data and unwrapped phase data are obtained through the channel attention mechanism module;

S3-2: Obtain a functional relationship between the DEM data, the longitude and latitude data, and the unwrapped phase data according to the weights of the DEM data, the longitude and latitude data, and the unwrapped phase data to obtain a pre-trained weight;

S3-3: Based on the pre-trained weights, the final prediction result is obtained through the deep neural network model to obtain the simulated large-scale system error disentanglement phase.

The beneficial effect of the above further scheme is: through the above technical scheme, after the data is input into the deep neural network model for processing, the unwrapped phase is obtained after removing the large-scale system error.

Furthermore, S5 includes the following steps:

S5-1: Re-encode the corrected unwrapped phase to SAR coordinates, perform DInSAR inversion deformation operation, refine and re-level the orbit phase by selecting and refining stable ground control points, eliminate the orbit error phase in differential interferometry, and then invert the deformation parameters;

S5-2: Geocode the inverted deformation results to obtain the deformation results after correcting the systematic errors.

The beneficial effect of the above further scheme is: the obtained corrected unwrapped phase is re-encoded to the SAR coordinates, and the corresponding DInSAR inversion deformation operation is continued to obtain the deformation result map after the system error is corrected. After comparison, the advantages of the present invention can be clearly seen.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 is a flow chart of a method for correcting large-scale system errors of InSAR based on deep learning.

Figure 2 is a technical flow chart of a deep learning-based InSAR large-scale system error correction method.

Figure 3 shows some of the InSAR large-scale error correction results using FC-Net.

Figure 4 is a comparison of the deformation results before and after FC-Net correction.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

As shown in FIG1 , a method for correcting large-scale system errors of InSAR based on deep learning is characterized in that the method comprises the following steps:

S1: Collect SAR data, DEM data and latitude and longitude data covering the target area, and perform data preprocessing to remove areas with poor coherence, and obtain geocoded unwrapped phase data, cropped DEM data and latitude and longitude data;

S2: Building a deep neural network model based on the relationship between relevant data, wherein the relevant data includes DEM data, longitude and latitude data, terrain-related atmospheric phase, orbital residual phase and turbulence phase;

S3: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model to obtain the simulated large-scale system error unwrapped phase;

S4: Subtract the original geocoded unwrapped phase from the large-scale system error unwrapped phase simulated by the deep neural network model to obtain the corrected unwrapped phase;

S5: The corrected unwrapped phase is de-geocoded to the SAR coordinate system, and the corrected deformation result is obtained by inverting the deformation parameters.

In addition, S1 performs InSAR data preprocessing on the SAR data, including the following steps:

S1-1: Obtain two SLC images of the same study area in time order and select one of them as the main image;

S1-2: The influence of speckle noise is reduced by setting the multi-view ratio of 10:2 in the distance and azimuth of the image, and the corresponding filtering parameters are set according to the size of the noise, and the filtered coherence data is obtained by filtering;

S1-3: Setting a disentanglement threshold based on the data distribution of the filtered coherence data to ensure sufficient training data and prevent the influence of disentanglement error data;

S1-4: According to the unwrapping threshold, the minimum cost flow method is selected to perform phase unwrapping on the wrapped phase calculated from the selected image interferogram;

S1-5: The unwrapped phase data in the SAR coordinate system is transferred to the same geographic coordinates as the cropped DEM data and longitude and latitude data through geocoding to obtain the geocoded unwrapped phase data.

The deep neural network model construction described in S2 includes the following steps:

S2-1: The fully connected layer network of the deep neural network model is used as a feature module to extract the input DEM data and longitude and latitude data information. Combined with the relationship between DEM data, longitude and latitude data and terrain-related atmospheric phase, orbital residual phase and turbulence phase, the channel attention mechanism module of the deep neural network model is selected as the weighted feature extraction module of the data;

S2-2: Perform feature compression operation through weighted feature extraction module, compress the spatial dimension, and then generate corresponding weights for each relevant data feature dimension through restoration operation, including the following formula:

The eigenvalue compression formula is as follows:

为输入的数据，

和

分别为输入数据的长和宽，

为压缩操作，

为原始数据压缩后第

个二维矩阵，

和

均为图像像素坐标；in,

For the input data,

and

are the length and width of the input data, respectively.

For compression operation,

The original data is compressed

A two-dimensional matrix,

and

All are image pixel coordinates;

The weight calculation formula is as follows:

为还原操作，

为RELU激活函数，

为RELU激活函数参数的降维层，

为RELU激活函数参数的升维层，

为特征通道权重，

为原始数据压缩后的结果，

为全局平均池化；in,

To restore the operation,

is the RELU activation function,

is the dimension reduction layer of the RELU activation function parameters,

is the dimension-raising layer of the RELU activation function parameters,

is the feature channel weight,

is the result after the original data is compressed.

is global average pooling;

S2-3: Apply the weight to each original relevant data feature channel. The formula is as follows:

为特征映射，

为尺度因子，

为最终得到的图像数据，

为特征映射与尺度因子在通道上的卷积。

is the feature map,

is the scale factor,

is the final image data,

is the convolution of the feature map and the scale factor on the channel.

The deep neural network model uses MSE as the loss function for parameter evaluation and the Adam optimization algorithm as the parameter optimization algorithm.

S3 includes the following steps:

S3-1: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model, the spatial features of the DEM data and longitude and latitude data are extracted through the fully connected module of the deep neural network model, and the weights of the DEM data, longitude and latitude data and unwrapped phase data are obtained through the channel attention mechanism module;

S3-2: Obtain a functional relationship between the DEM data, the longitude and latitude data, and the unwrapped phase data according to the weights of the DEM data, the longitude and latitude data, and the unwrapped phase data to obtain a pre-trained weight;

S3-3: Based on the pre-trained weights, the final prediction result is obtained through the deep neural network model to obtain the simulated large-scale system error disentanglement phase.

S5 includes the following steps:

S5-1: Re-encode the corrected unwrapped phase to SAR coordinates, perform DInSAR inversion deformation operation, refine and re-level the orbit phase by selecting and refining stable ground control points, eliminate the orbit error phase in differential interferometry, and then invert the deformation parameters;

S5-2: Geocode the inverted deformation results to obtain the deformation results after correcting the systematic errors.

In one embodiment of the present invention, as shown in FIG2 , SAR image data and precise orbit data covering the target area are collected, SLC data is obtained through registration processing to form an original interference pair, and external elevation data is used to remove the flat ground phase and terrain phase, thereby forming a differential interference pair, and the entangled phase is unwrapped and geocoded by the minimum cost flow unwrapping algorithm, and the pre-processed geocoded unwrapped phase, elevation and longitude and latitude data are input into the deep neural network to obtain a simulated unwrapped phase, including a simulated terrain-related atmospheric phase and a simulated orbital residual phase, and then the geocoded unwrapped phase is subtracted from the unwrapped phase simulated by the network model to obtain a corrected unwrapped phase, and the corrected unwrapped phase is inversely geocoded to the SAR coordinate system, and the deformation parameters are inverted to obtain the corrected deformation result. In this embodiment, a fully connected channel attention network is used as a deep neural network.

The present invention provides a method for correcting large-scale InSAR system errors in steep mountainous areas, hills and plain areas, and proposes a model that can simultaneously correct terrain-related atmospheric phases, orbital residual phases and part of the turbulent phases. The model is of great significance for InSAR to obtain the actual deformation rate and correct interpretation. Figure 3 provides the results of error correction, and gives a comparison chart of the interference pair after filtering, the original unwrapped phase, the simulated unwrapped phase and the corrected unwrapped phase results. It can be seen from Figures 3 and 4 (a and c in Figure 4 are the deformation results before correction, and b and d are the deformation results after correction) that when using DInSAR for deformation monitoring, this scheme can greatly reduce the influence of terrain-related atmospheric phases, orbital residual phases and part of the small-scale turbulent phases by using this method, so that deformation can be identified from areas affected by the atmosphere and orbital phase.

Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the invention.

Claims (6)

1. A method for correcting large-scale InSAR system errors based on deep learning, characterized in that the method comprises the following steps: S1: Collect SAR data, DEM data and latitude and longitude data covering the target area, and perform data preprocessing to remove areas with poor coherence, and obtain geocoded unwrapped phase data, cropped DEM data and latitude and longitude data; S2: Building a deep neural network model based on the relationship between relevant data, wherein the relevant data includes DEM data, longitude and latitude data, terrain-related atmospheric phase, orbital residual phase and turbulence phase; S3: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model to obtain the simulated large-scale system error unwrapped phase; S4: Subtract the original geocoded unwrapped phase from the large-scale system error unwrapped phase simulated by the deep neural network model to obtain the corrected unwrapped phase; S5: The corrected unwrapped phase is de-geocoded to the SAR coordinate system, and the corrected deformation result is obtained by inverting the deformation parameters. 2. The InSAR large-scale systematic error correction method based on deep learning according to claim 1 is characterized in that the InSAR data preprocessing in S1 comprises the following steps: S1-1: Obtain two SLC images of the same study area in time order and select one of them as the main image; S1-2: The influence of speckle noise is reduced by setting the multi-view ratio of 10:2 in the distance and azimuth of the image, and the corresponding filtering parameters are set according to the size of the noise, and the filtered coherence data is obtained by filtering; S1-3: Setting a disentanglement threshold based on the data distribution of the filtered coherence data to ensure sufficient training data and prevent the influence of disentanglement error data; S1-4: According to the unwrapping threshold, the minimum cost flow method is selected to perform phase unwrapping on the wrapped phase calculated from the selected image interferogram; S1-5: The unwrapped phase data in the SAR coordinate system is transferred to the same geographic coordinates as the cropped DEM data and longitude and latitude data through geocoding to obtain the geocoded unwrapped phase data. 3. The InSAR large-scale system error correction method based on deep learning according to claim 1 is characterized in that the deep neural network model building in S2 comprises the following steps: S2-1: The fully connected layer network of the deep neural network model is used as a feature module to extract the input DEM data and longitude and latitude data information. Combined with the relationship between DEM data, longitude and latitude data and terrain-related atmospheric phase, orbital residual phase and turbulence phase, the channel attention mechanism module of the deep neural network model is selected as the weighted feature extraction module of the data; S2-2: Perform feature compression operation through weighted feature extraction module, compress the spatial dimension, and then generate corresponding weights for each relevant data feature dimension through restoration operation, including the following formula: The eigenvalue compression formula is as follows:

in,

For the input data,

and

are the length and width of the input data, respectively.

For compression operation,

The original data is compressed

A two-dimensional matrix,

and

All are image pixel coordinates; The weight calculation formula is as follows:

in,

To restore the operation,

is the RELU activation function,

is the dimension reduction layer of the RELU activation function parameters,

is the dimension-raising layer of the RELU activation function parameters,

is the feature channel weight,

is the result after the original data is compressed.

is global average pooling; S2-3: Apply the weight to each original relevant data feature channel. The formula is as follows:

is the feature map,

is the scale factor,

is the final image data,

is the convolution of the feature map and the scale factor on the channel. 4. According to the InSAR large-scale system error correction method based on deep learning in claim 3, it is characterized in that the deep neural network model uses MSE as the loss function for parameter evaluation and selects the Adam optimization algorithm as the parameter optimization algorithm. 5. The InSAR large-scale systematic error correction method based on deep learning according to claim 1, characterized in that S3 comprises the following steps: S3-1: The geocoded unwrapped phase data, cropped DEM data and longitude and latitude data obtained after preprocessing are input into the deep neural network model, the spatial features of the DEM data and longitude and latitude data are extracted through the fully connected module of the deep neural network model, and the weights of the DEM data, longitude and latitude data and unwrapped phase data are obtained through the channel attention mechanism module; S3-2: Obtain a functional relationship between the DEM data, the longitude and latitude data, and the unwrapped phase data according to the weights of the DEM data, the longitude and latitude data, and the unwrapped phase data to obtain a pre-trained weight; S3-3: Based on the pre-trained weights, the final prediction result is obtained through the deep neural network model to obtain the simulated large-scale system error disentanglement phase. 6. The InSAR large-scale system error correction method based on deep learning according to claim 1, characterized in that S5 comprises the following steps: S5-1: Re-encode the corrected unwrapped phase to SAR coordinates, perform DInSAR inversion deformation operation, refine and re-level the orbit phase by selecting and refining stable ground control points, eliminate the orbit error phase in differential interferometry, and then invert the deformation parameters; S5-2: Geocode the inverted deformation results to obtain the deformation results after correcting the systematic errors.

Images