CN113139349A - Method, device and equipment for removing atmospheric noise in InSAR time sequence - Google Patents

Method, device and equipment for removing atmospheric noise in InSAR time sequence Download PDF

Info

Publication number
CN113139349A
CN113139349A CN202110519191.2A CN202110519191A CN113139349A CN 113139349 A CN113139349 A CN 113139349A CN 202110519191 A CN202110519191 A CN 202110519191A CN 113139349 A CN113139349 A CN 113139349A
Authority
CN
China
Prior art keywords
insar
time sequence
deformation
synthesized
trend component
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110519191.2A
Other languages
Chinese (zh)
Other versions
CN113139349B (en
Inventor
林珲
赵倬毅
马培峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Jiangxi Normal University
Original Assignee
Jiangxi Normal University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Jiangxi Normal University filed Critical Jiangxi Normal University
Priority to CN202110519191.2A priority Critical patent/CN113139349B/en
Publication of CN113139349A publication Critical patent/CN113139349A/en
Application granted granted Critical
Publication of CN113139349B publication Critical patent/CN113139349B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/08—Thermal analysis or thermal optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/10—Noise analysis or noise optimisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/12—Timing analysis or timing optimisation
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Geometry (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a method for removing atmospheric noise in an InSAR time sequence, which comprises the following steps: acquiring InSAR deformation time sequence data; inputting the InSAR deformation time sequence data into a deep recursive network denoising model to obtain the InSAR deformation time sequence data for removing atmospheric noise; the depth recursive network denoising model combines seasonal factors as auxiliary input, and an attenuation mechanism is introduced to process the observation value missing problem; the model utilizes simulation InSAR deformation time sequence training optimization. According to the method, the time sequence signal after denoising is obtained by constructing a denoising model based on a deep recursive network, so that atmospheric noise is removed more robustly; the method for simulating the InSAR time sequence data is constructed to synthesize a large amount of simulation time sequence data with abundant samples, and the simulation time sequence data is used for training a deep recursive network model to obtain optimized model parameters; meanwhile, the time sequence data of the synthetic simulation can also help to verify the effectiveness of the denoising model and quantitatively evaluate the performance of the model.

Description

Method, device and equipment for removing atmospheric noise in InSAR time sequence
Technical Field
The invention relates to the technical field of remote sensing science and synthetic aperture radar interferometric data processing, in particular to a method, a device and equipment for removing atmospheric noise in an InSAR time sequence.
Background
As interferometric synthetic aperture radar (InSAR) technology has evolved in recent years, it is increasingly being used to monitor surface deformations. However, the radar signal has a time delay when passing through the atmosphere, the uncertain time delay is caused by variable meteorological conditions, and the atmosphere delay introduces a great deal of noise to InSAR observation data.
Atmospheric delays, which are typically created by the refraction of radar signals through the troposphere, can be divided into vertically stratified portions and turbulently mixed portions. Wherein the vertically layered portion is dependent on air pressure and temperature, and changes more slowly with time; the portion of turbulent mixing depends primarily on the distribution of water vapor, which changes rapidly with time. Therefore, the atmospheric retardation of most of the vertically layered portion will be cancelled out in the interference image in which the effect of the turbulent mixing portion is larger than that of the vertically layered portion. In the practical application of the time-series InSAR technology, when there are more data close in time and turbulent mixing noise is dominant (an area with small elevation difference), atmospheric noise can be regarded as random noise irrelevant in a time domain. Since atmospheric noise usually appears as a high-frequency random signal in the time domain, time-domain low-pass filtering is often used to estimate and remove the atmospheric noise.
Common time-domain filtering methods (e.g., gaussian filtering methods) play an important role in removing atmospheric noise, but their performance is greatly influenced by parameter settings (e.g., filter weights and filter window size). Ideally, radar data is available from each orbital cycle of a given satellite, however, for operational reasons, the problem of data loss can often arise. The conventional time domain filtering method needs to complement the missing data first, so that an undesirable complementary result has a large influence on the filtering result.
Disclosure of Invention
The embodiment of the invention provides a method, a device and equipment for removing atmospheric noise in an InSAR time sequence, which are used for solving the problem that the atmospheric noise removed by traditional time-domain filtering is influenced by parameter setting in the background technology.
The embodiment of the invention provides a method for removing atmospheric noise in an InSAR time sequence, which comprises the following steps:
acquiring InSAR deformation time sequence data;
inputting the InSAR deformation time sequence data into a deep recursive network denoising model to obtain the InSAR deformation time sequence data for removing atmospheric noise;
wherein, the deep recursive network denoising model is as follows: a deep recurrent neural network model which is trained by adopting the synthesized simulation InSAR deformation time sequence data; the synthesized simulated InSAR deformation time sequence data comprises: a synthesized surface deformation trend component.
Further, the synthesized surface deformation trend component includes: a linear trend component, a deceleration trend component, an acceleration trend component and a stability trend component;
the linear trend component is expressed by a linear equation as:
Tl(t)＝-A·t
wherein A controls the deformation speed, and the stable trend component is a special case of a linear trend component when A is zero;
the deceleration trend component, settling as an approximately logarithmic function of time, is expressed as:
Td(t)＝-Bd·log(t+td)+Cd
wherein, BdControlling the speed of change of the speed of deceleration deformation, tdDepending on the start time of the observation, CdIs a constant, when T is 0, Td(t)＝0；
The acceleration tendency component is expressed by a reversed logarithmic function as:
Ta(t)＝Ba·log(tf-t)+Ca
wherein, tfTime to failure for landslide or landslide, tfT is the expected life, BaControlling the speed of change of the acceleration deformation speed, CaIs a constant, when T is 0, Ta(t)＝0。
Further, the synthesized simulated InSAR deformation time series data further includes: a synthetic seasonal component;
taking the thermal expansion and the cold shrinkage caused by the change of the air temperature as an example, the expression is as follows:
S(t)＝As·SF(t)
Figure BDA0003063250640000031
Figure BDA0003063250640000032
where SF (t) represents the normalized seasonal factor, the unit variance and the temperature variation with mean zero,
Figure BDA0003063250640000033
representing the amplitude of seasonal oscillations, H0Representing the initial height of the earth's surface, gamma represents the linear thermal expansion coefficient, lambda determines the amplitude of the oscillation, and omega controls the period of the oscillation, here fixed at a period of 1 year, with the starting phase phi depending on the starting observation point in the year.
Further, the synthesized simulated InSAR deformation time series data further includes: the introduced simulated atmospheric noise and observation missing mode;
atmospheric noise simulated by additive white Gaussian noise is adopted, after a simulation deformation time sequence is generated, part of data points are randomly discarded to simulate an observation value missing mode, and finally the synthesized InSAR simulation deformation time sequence is expressed as follows:
D(t)＝T(t)+S(t)
X(t)＝(D(t)+N(t))·M(t)
wherein, n (t) represents noise, m (t) represents missing mask, d (t) represents synthesized surface deformation, t (t) represents deformation trend component, and s (t) represents seasonal component.
Further, the deep recursive network denoising model includes:
the input layer takes the synthesized InSAR simulation deformation time sequence as input;
the GRU-D network layer processes the missing data and outputs a non-missing intermediate layer multi-dimensional characteristic vector sequence;
stacking GRU network layers and outputting a multi-dimensional feature vector sequence of a top layer;
and the full-link output layer converts the characteristic vector sequence output by the recursive network layer into a final deformation time sequence for removing atmospheric noise.
Further, the GRU-D network layer is of a bidirectional structure, and a bidirectional attenuation mechanism is realized by using the values before and after.
An embodiment of the present invention further provides a device for removing atmospheric noise in an InSAR timing sequence, including:
the deformation time sequence acquisition module is used for acquiring InSAR deformation time sequence data;
the deformation time sequence denoising module is used for inputting the InSAR deformation time sequence data into the deep recursive network denoising model to obtain the InSAR deformation time sequence data for removing atmospheric noise;
wherein, the deep recursive network denoising model is as follows: a deep recurrent neural network model which is trained by adopting the synthesized simulation InSAR deformation time sequence data; the synthesized simulated InSAR deformation time sequence data comprises: a synthesized surface deformation trend component;
the synthesized surface deformation trend component comprises: a linear trend component, a deceleration trend component, an acceleration trend component and a stability trend component;
the linear trend component is expressed by a linear equation as:
Tl(t)＝-A·t
wherein A controls the deformation speed, and the stable trend component is a special case of a linear trend component when A is zero;
the deceleration trend component, settling as an approximately logarithmic function of time, is expressed as:
Td(t)＝-Bd·log(t+td)+Cd
wherein, BdControlling the speed of change of the speed of deceleration deformation, tdDepending on the start time of the observation, CdIs a constant, when T is 0, Td(t)＝0；
The acceleration tendency component is expressed by a reversed logarithmic function as:
Ta(t)＝Ba·log(tf-t)+Ca
wherein, tfTime to failure for landslide or landslide, tfT is the expected life, BaControlling the speed of change of the acceleration deformation speed, CaIs a constant, when T is 0, Ta(t)＝0。
Further, the synthesized simulated InSAR deformation time series data further includes:
taking the thermal expansion and the cold shrinkage caused by the change of the air temperature as an example, the expression is as follows:
S(t)＝As·SF(t)
Figure BDA0003063250640000041
Figure BDA0003063250640000042
wherein 5F (t) represents normalized seasonal factors, unit variance and temperature variation with mean zero,
Figure BDA0003063250640000043
representing the amplitude of seasonal oscillations, H0Representing the initial height of the earth's surface, gamma the linear thermal expansion coefficient, lambda determines the amplitude of the oscillation, omega controls the period of the oscillation, here fixed at a period of 1 year, the starting phase phi depends on the starting view of the yearAnd (6) inspecting points.
Further, the synthesized simulated InSAR deformation time series data further includes: the introduced simulated atmospheric noise and observation missing mode;
atmospheric noise simulated by additive white Gaussian noise is adopted, after a simulation deformation time sequence is generated, part of data points are randomly discarded to simulate an observation value missing mode, and finally the synthesized InSAR simulation deformation time sequence is expressed as follows:
D(t)＝T(t)+S(t)
X(t)＝(D(t)+N(t))·M(t)
wherein, n (t) represents noise, m (t) represents missing mask, d (t) represents synthesized surface deformation, t (t) represents deformation trend component, and s (t) represents seasonal component.
The embodiment of the present invention further provides a computer device, which includes a memory and a processor, where the memory stores a computer program, and is characterized in that the processor implements the steps of the method when executing the computer program.
The embodiment of the invention provides a method, a device and equipment for removing atmospheric noise in an InSAR time sequence, and compared with the prior art, the method, the device and the equipment have the following beneficial effects:
the invention provides a data-driven atmospheric noise removal method, which is suitable for correcting errors caused by atmospheric delay in an InSAR time sequence, and is characterized in that a model based on a deep recursive network is constructed, a deep recursive network denoising model is combined with seasonal factors as auxiliary input, and an attenuation mechanism is introduced to process the problem of observation value loss; the model utilizes simulation InSAR deformation time sequence training optimization, inputs an original time sequence signal, directly outputs a time sequence signal after denoising, and more robustly removes noise, specifically:
(1) the invention constructs a method for simulating InSAR time sequence data, so as to synthesize a large amount of simulation time sequence data with abundant samples, train a deep recursion network and obtain optimized model parameters; and the time sequence data of the synthetic simulation can also help to verify the effectiveness of the denoising model and quantitatively evaluate the performance of the model.
(2) In the case of observation value missing in an actual InSAR deformation time sequence, a traditional method for processing data missing firstly performs value complementing on missing data, for example, forward-file and linear interpolation, and the value complementing and subsequent time sequence processing are often divided into two steps by the method, so that global optimization cannot be performed. Moreover, since the original InSAR timing signal contains high noise, a large bias is introduced by simple forward file or linear interpolation. The invention introduces a GRU-D network structure, and the network structure introduces an attenuation mechanism to model missing data, so that an end-to-end model can be constructed to carry out global optimization.
(3) Seasonal fluctuations in the InSAR timing signals are primarily related to thermal expansion and contraction due to temperature changes, seasonal precipitation, or seasonal groundwater draws. These underlying environmental changes are therefore defined as seasonal factors SF (seasonal factor), which are introduced by the present invention as an auxiliary input to the deep recursive model in order to help the model to better understand the seasonal oscillation patterns that may exist in the InSAR timing signals.
Drawings
FIG. 1 is a flow chart of simulation timing data synthesis provided by an embodiment of the present invention;
fig. 2 is a diagram of a deep recursive network model structure according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Referring to fig. 1 to 2, an embodiment of the present invention provides a method for removing atmospheric noise in an InSAR timing sequence, where the method specifically includes:
the first step is as follows: synthesizing simulated deformation time series data
Table subsidence or displacement, influenced by geological conditions, may exhibit a long-term, monotonic tendency to deform and may exhibit some seasonal oscillation under the influence of weather and hydrological conditions. Thus, the surface deformation D (t) can be generally decomposed into a trend component and a seasonal component, formulated as:
D(t)＝T(t)+S(t)
where T (t) represents a trend component and S (t) represents a seasonal component. Based on the decomposition method, the process of synthesizing the simulation deformation time sequence is divided into three parts of synthesizing trend components, synthesizing seasonal components and introducing noise and missing modes.
1) Component of synthetic trend
The trend component of surface deformation may exhibit linear, acceleration and deceleration behavior, and additionally, the four different trends shown in fig. 1 are simulated taking into account stable ground points (no apparent deformation).
The linear trend, which is most often observed in practice and may occur in the early stages of ground consolidation, shows a constant deformation speed and is expressed by the linear equation:
Tl(t)＝-A·t
where a controls the deformation speed, the steady trend can be seen as a special case of a linear trend when a is zero.
The tendency to slow down usually occurs later in the consolidation of the ground, where the surface tends to stabilize and settle as an approximately logarithmic function of time, which can be expressed as:
Td(t)＝-Bd·log(t+td)+Cd
wherein, BdControlling the speed of change of the deformation speed tdDepending on the start time of the observation, CdIs a constant, when T is 0, Td(t)＝0。
The acceleration tendency releases the danger signal that landslide or landslide may occur, and the reasons for the acceleration deformation are many, and here, a reversed logarithmic function is expressed as:
Ta(t)＝Ba·log(tf-t)+Ca
wherein, tfIndicating the time to failure, t, of a landslide or landslidef-t represents the expected lifetime, BaControlling the speed of change of the deformation speed, CaIs a constant, when T is 0, Ta(t)＝0。
2) Synthesizing seasonal components
The deformation of the earth surface often shows a certain seasonality under the influence of air temperature and hydrological conditions, the interference of seasonal fluctuation also increases the difficulty in removing atmospheric noise, and the traditional time domain filtering method is easily influenced by the seasonality. Based on this, the present invention analyzes the synthesis seasonal component, specifically as follows:
seasonal oscillations in ground deformation are mainly caused by seasonal air temperature, precipitation, or groundwater pumping changes, collectively referred to herein as Seasonal Factors (SF). Taking the thermal expansion and contraction caused by the change of air temperature as an example, the seasonal oscillation depends on:
S(t)＝γ·H0·ΔΘ(t)
wherein H0Representing the initial height of the earth's surface and gamma represents the linear thermal expansion coefficient. Δ Θ (t) represents the surface temperature variation, which can be approximated by a sine function, and can be expressed as:
ΔΘ(t)＝λ·sm(ωt+φ)
where λ determines the amplitude of the oscillation, ω controls the period of the oscillation (here fixed to a period of 1 year), and the starting phase is dependent on the starting observation point in the year.
After finishing, the following expression is adopted:
Figure BDA0003063250640000081
Figure BDA0003063250640000082
S(t)＝As·SF(t)
where SF (t) represents the normalized seasonal factor (temperature variation with unit variance and mean of zero),
Figure BDA0003063250640000083
representing the amplitude of the seasonal oscillation. Similarly, seasonal variations caused by other seasonal factors may also be expressed in the above form.
3) Introducing noise and loss patterns
Based on the previous discussion of atmospheric noise, which can be viewed as time-domain uncorrelated random noise, it is modeled with additive white gaussian noise. In order to simulate the data missing problem frequently occurring in InSAR time sequence, part of data points are randomly discarded after simulation time sequence data are generated. The final generated simulation deformation time sequence can be expressed as:
X(t)＝(D(t)+N(t))·M(t)
where n (t) represents noise and m (t) represents a missing mask.
By combining the three parts of time sequence synthesis, a large amount of varied simulation deformation time sequences can be synthesized. The complete simulation timing synthesis flow can refer to fig. 1.
The second step is that: construction of deep recursive network denoising model
Recording the deformation time sequence of InSAR as X ═ X1，x2，…，xn)∈RnWherein x istRepresents the t-th observation value, and the interval between two adjacent observation values is the same. To mark missing data, a missing mask vector M ═ M (M) is introduced1，m2，...，mn)∈{0，1}n，mt0 represents xtIs missing. The corresponding normalized seasonal factor signal is noted as SF ═ s1，s2，...，sn)∈RnSF will serve as an auxiliary input sequence for the denoising model. Recording the output deformation time sequence Y for removing the atmospheric noise as (Y)1，y2，...，yn)∈Rn. Based on the above notation, the denoising model can be expressed as:
Y＝f(X，M，SF；θ)
where θ represents a model parameter.
The constructed deep recursive network can be divided into four parts, namely an input layer, a GRU-D layer, a stacked GRU layer and a full link output layer. The overall network structure can be seen in fig. 2, where all recursive network layers employ a bi-directional recursive pattern.
When the input layer combines the deformation time sequence X and the seasonal factor SF into a binary
Figure BDA0003063250640000091
Wherein
Figure BDA0003063250640000092
To introduce a data miss pattern, the corresponding value in timing X is ignored according to the miss mask vector M. To have the same binary sequence
Figure BDA00030632506400000914
The match, miss mask vector is also expanded to a binary vector
Figure BDA0003063250640000093
Wherein
Figure BDA0003063250640000094
Since the seasonal factor SF is usually easier to obtain, it is set here
Figure BDA0003063250640000095
Is all 1. For the sake of brevity of expression, in the following formulas, it will be omitted
Figure BDA0003063250640000096
And
Figure BDA0003063250640000097
the bar symbol of (2).
The GRU-D layer is designed to deal with the data loss problem, GRU-D is a variant of GRU that introduces trainable attenuation parameters by introducing an attenuation rate γ in the input sequence and implicit states of the recursive networktTo achieve gammatIs defined as:
γt＝exp{-max(0，Wγδt+bγ)}∈(0，1]
wherein, deltatRepresenting the interval from the last observed data, which can be obtained by missing the mask vector M, WγAnd bγAre parameters that can be learned. Based on this formula, γtWill follow deltatIs increased to approach 0. Gamma raytThe decay mechanism will be introduced into the input sequence and implicit state by the following form:
Figure BDA0003063250640000098
Figure BDA0003063250640000099
wherein x istIs the last observed value of the signal that was,
Figure BDA00030632506400000915
is an empirical mean or default value of the sequence, ht-1Is an implicit state of the previous iteration step,
Figure BDA00030632506400000910
and
Figure BDA00030632506400000911
acting as a decay mechanism on the input sequence and the implicit state, respectively, indicates a point-to-point multiplication. With deltatIncrease of, via attenuation
Figure BDA00030632506400000912
Will tend toward the empirical mean
Figure BDA00030632506400000913
Attenuated
Figure BDA0003063250640000101
Will tend towards 0. However, there is no empirical mean or default for InSAR time series that exhibit a trend of deformation. Therefore, the invention will
Figure BDA00030632506400001014
The term is removed and the GRU-D layer is extended to a bi-directional structure to utilize both the previous and the subsequent values, the bi-directional attenuation mechanism being as follows:
Figure BDA0003063250640000102
Figure BDA0003063250640000103
Figure BDA0003063250640000104
Figure BDA0003063250640000105
wherein the content of the first and second substances,
Figure BDA0003063250640000106
and
Figure BDA0003063250640000107
the last observed values for the forward and reverse iterations, respectively (t '< t, t' > t),
Figure BDA0003063250640000108
and
Figure BDA0003063250640000109
are input values that are passed through the attenuation mechanism in two directions respectively,
Figure BDA00030632506400001010
and
Figure BDA00030632506400001011
are the implicit states of the previous step of the forward and reverse iterations respectively,
Figure BDA00030632506400001012
and
Figure BDA00030632506400001013
the implicit states of the previous step through the forward and reverse iterations of the decay mechanism, respectively. Based on the above formula, the input and implicit states through the decay mechanism are affected by their preceding and following observations, with closer proximity having greater effect. In addition to the decay mechanism, the missing mask vector M is also directly input to the updated iteration function of the recursive network based on the original GRU-D model.
Above the GRU-D layer is a stacked GRU layer to increase the complexity and expressiveness of the model. And finally, the output layer adopts a fully-linked network structure and is used for converting the output characteristic vector of the GRU layer into a unary noise-removed output sequence.
The third step: training network models using synthetic data
And training a recursive network model by utilizing the synthesized simulation time sequence data, and then applying the recursive network model to the real InSAR deformation time sequence data. In order to make the model better suited to the real data, the synthesis parameters of the simulation data are set according to the distribution characteristics of the real data. Specifically, the maximum deformation amount and the maximum seasonal oscillation amplitude are set to cover the deformation and oscillation conditions of the real data, and the deformation amount and the oscillation amplitude of the corresponding simulation data are randomly generated in a uniform distribution within a set range. The parameter ω controlling the period of the seasonal oscillation will be fixed and will make the simulated data have the same period as the real data (typically 1 year period). The STD (standard deviation) and the data missing ratio of additive white gaussian noise need to be set to a level similar to the real data. A model based on a recursive network can theoretically accept sequences of arbitrary length as input, but generating simulation timings of similar length to the real InSAR timing data will help to obtain a better model performance. In order for the model to learn the seasonal characteristics of the time sequence better, the synthetic time sequence should not be shorter than a full seasonal oscillation period. The remaining synthesis parameters will be generated randomly.
The synthesized simulation timing data will be divided into training data, verification data and test data for training and selecting the optimal model parameters. The synthesized simulation time sequence X (t) containing noise and missing data is used as the input of the model, the corresponding simulation time sequence D (t) containing no noise and missing data is used as the output supervision data, and the corresponding seasonal factor SF (t) used for synthesizing the seasonal component of the time sequence is used as the auxiliary input of the model. The Dropout mechanism is introduced in the training process, so that overfitting can be avoided, and the optimal parameters are selected by adopting an early stopping method (early stopping). The simulation time sequence of the input model is globally and linearly normalized to be in the range of [ -1, 1], and the original value is used for the output supervision data.
The fourth step: denoising real InSAR deformation time sequence input network model
Before the real InSAR time sequence data is input into the network model, the same linear normalization operation as that of the synthetic simulation data is needed, and corresponding seasonal factor signals, such as air temperature, are also normalized (the unit variance and the mean value are zero). And combining the normalized InSAR deformation time sequence and the seasonal factor signal into a binary sequence input model to obtain the InSAR deformation time sequence with atmospheric noise removed.
Principles of embodiments of the invention
An InSAR deformation time sequence simulation method based on a physical mechanism comprises the following steps:
ground deformation can generally be broken down into two components, a long-term trend component and a periodic seasonal component. According to such a decomposition method, the flow of the synthesis sequence includes three parts: synthesizing trend components, synthesizing seasonal components, and introducing atmospheric noise and data loss modes. According to different change trend behaviors possibly presented by the surface deformation, three different mathematical models are used for respectively simulating linear, acceleration and deceleration trend components. The seasonal component of the periodic fluctuation is simulated with a sinusoidal function. Since the invention primarily considers turbulent mixing noise, which can be considered a random noise, it is modeled with white gaussian noise. After the time series data are synthesized, a part of the data are randomly removed to simulate the condition of possible missing data.
Constructing a denoising model of the deep recursive network:
the network model provided by the invention takes GRU as a recursive network unit and mainly comprises an input layer, a GRU-D network layer, a stacked GRU network layer and a full link output layer. The input layer combines an InSAR deformation time sequence and a seasonal factor SF time sequence to form a bivariate time sequence, and simultaneously combines a mask vector (marking data missing information) as input; the GRU-D network layer is used for processing the missing data and outputting a multi-dimensional characteristic vector sequence of the non-missing intermediate layer; stacking GRU network layers to further improve the model expression capability and output a multi-dimensional feature vector sequence of the top layer; and the full-link output layer converts the characteristic vector sequence output by the recursive network layer into a final deformation time sequence for removing atmospheric noise. The technical scheme takes a GRU recursive network model as an example, the method is also suitable for other recursive models, such as simple RNN, LSTM and the like, and the model can be a bidirectional recursive network or a unidirectional recursive network.
Model training:
the invention adopts a supervised learning method to train a network model, and training data are synthetic simulation InSAR deformation time sequence data. According to the characteristics of real scene data in specific application, parameters (such as deformation size, acceleration and deceleration amplitude, periodic oscillation amplitude and the like) in a data synthesis algorithm are correspondingly adjusted to generate a large amount of simulation data matched with the real data.
Although the embodiments of the present invention have been disclosed in the foregoing for illustrative purposes, those skilled in the art will appreciate that various modifications, additions and substitutions are possible, without departing from the scope and spirit of the invention as disclosed in the accompanying drawings.

Claims (10)

1. A method for removing atmospheric noise in an InSAR time sequence is characterized by comprising the following steps:
acquiring InSAR deformation time sequence data;
inputting the InSAR deformation time sequence data into a deep recursive network denoising model to obtain the InSAR deformation time sequence data for removing atmospheric noise;
wherein, the deep recursive network denoising model is as follows: a deep recurrent neural network model which is trained by adopting the synthesized simulation InSAR deformation time sequence data; the synthesized simulated InSAR deformation time sequence data comprises: a synthesized surface deformation trend component.
2. The method of removing atmospheric noise in an InSAR sar time sequence as claimed in claim 1, wherein said synthesized terrain deformation trend component comprises: a linear trend component, a deceleration trend component, an acceleration trend component and a stability trend component;
the linear trend component is expressed by a linear equation as:
Tl(t)＝-A·t
wherein A controls the deformation speed, and the stable trend component is a special case of a linear trend component when A is zero;
the deceleration trend component, settling as an approximately logarithmic function of time, is expressed as:
Td(t)＝-Bd·log(t+td)+Cd
wherein, BdControlling the speed of change of the speed of deceleration deformation, tdDepending on the start time of the observation, CdIs a constant, when T is 0, Td(t)＝0；
The acceleration tendency component is expressed by a reversed logarithmic function as:
Ta(t)＝Ba·log(tf-t)+Ca
wherein, tfTime to failure for landslide or landslide, tfT is the expected life, BaControlling the speed of change of the acceleration deformation speed, CaIs a constant, when T is 0, Ta(t)＝0。
3. The method for removing atmospheric noise in an InSAR timing sequence according to claim 1, wherein the synthesized simulated InSAR morphed timing sequence data further includes: a synthetic seasonal component;
taking the thermal expansion and the cold shrinkage caused by the change of the air temperature as an example, the expression is as follows:
S(t)＝As·SF(t)
Figure FDA0003063250630000021
Figure FDA0003063250630000022
where SF (t) represents the normalized seasonal factor, the unit variance and the temperature variation with mean zero,
Figure FDA0003063250630000023
representing the amplitude of seasonal oscillations, H0Representing the initial height of the earth's surface, gamma represents the linear thermal expansion coefficient, lambda determines the amplitude of the oscillation, and omega controls the period of the oscillation, here fixed at a period of 1 year, with the starting phase phi depending on the starting observation point in the year.
4. The method for removing atmospheric noise in an InSAR timing sequence according to claim 1, wherein the synthesized simulated InSAR morphed timing sequence data further includes: the introduced simulated atmospheric noise and observation missing mode;
atmospheric noise simulated by additive white Gaussian noise is adopted, after a simulation deformation time sequence is generated, part of data points are randomly discarded to simulate an observation value missing mode, and finally the synthesized InSAR simulation deformation time sequence is expressed as follows:
D(t)＝T(t)+S(t)
X(t)＝(D(t)+N(t))·M(t)
wherein, n (t) represents noise, m (t) represents missing mask, d (t) represents synthesized surface deformation, t (t) represents deformation trend component, and s (t) represents seasonal component.
5. The method for removing atmospheric noise in an InSAR time sequence as claimed in claim 1, wherein the deep recursive network denoising model comprises:
the input layer takes the synthesized InSAR simulation deformation time sequence as input;
the GRU-D network layer processes the missing data and outputs a non-missing intermediate layer multi-dimensional characteristic vector sequence;
stacking GRU network layers and outputting a multi-dimensional feature vector sequence of a top layer;
and the full-link output layer converts the characteristic vector sequence output by the recursive network layer into a final deformation time sequence for removing atmospheric noise.
6. The method for removing atmospheric noise in an InSAR time sequence as claimed in claim 5, wherein the GRU-D network layer is a bidirectional structure, and the bidirectional attenuation mechanism is implemented by using the values before and after.
7. An apparatus for removing atmospheric noise in an InSAR time sequence, comprising:
the deformation time sequence acquisition module is used for acquiring InSAR deformation time sequence data;
the deformation time sequence denoising module is used for inputting the InSAR deformation time sequence data into the deep recursive network denoising model to obtain the InSAR deformation time sequence data for removing atmospheric noise;
wherein, the deep recursive network denoising model is as follows: a deep recurrent neural network model which is trained by adopting the synthesized simulation InSAR deformation time sequence data; the synthesized simulated InSAR deformation time sequence data comprises: a synthesized surface deformation trend component;
the synthesized surface deformation trend component comprises: a linear trend component, a deceleration trend component, an acceleration trend component and a stability trend component;
the linear trend component is expressed by a linear equation as:
Tl(t)＝-A·t
wherein A controls the deformation speed, and the stable trend component is a special case of a linear trend component when A is zero;
the deceleration trend component, settling as an approximately logarithmic function of time, is expressed as:
Td(t)＝-Bd·log(t+td)+Cd
wherein, BdControlling the speed of change of the speed of deceleration deformation, tdDepending on the start time of the observation, CdIs a constant, when T is 0, Td(t)＝0；
The acceleration tendency component is expressed by a reversed logarithmic function as:
Ta(t)＝Ba·log(tf-t)+Ca
wherein, tfTime to failure for landslide or landslide, tfT is the expected life, BaControlling the speed of change of the acceleration deformation speed, CaIs a constant, when T is 0, Ta(t)＝0。
8. The apparatus for removing atmospheric noise in an InSAR sequence as claimed in claim 7, wherein the synthesized simulated InSAR morphed sequence data further comprises:
taking the thermal expansion and the cold shrinkage caused by the change of the air temperature as an example, the expression is as follows:
S(t)＝As·SF(t)
Figure FDA0003063250630000031
Figure FDA0003063250630000032
where SF (t) represents the normalized seasonal factor, the unit variance and the temperature variation with mean zero,
Figure FDA0003063250630000041
representing the amplitude of seasonal oscillations, H0Representing the initial height of the earth's surface, gamma representing the linear thermal expansion coefficient, and lambda determining the oscillationThe amplitude, ω, controls the period of oscillation, fixed here as a period of 1 year, with the starting phase φ depending on the starting observation point in the year.
9. The apparatus for removing atmospheric noise in an InSAR sequence as claimed in claim 7, wherein the synthesized simulated InSAR morphed sequence data further comprises: the introduced simulated atmospheric noise and observation missing mode;
atmospheric noise simulated by additive white Gaussian noise is adopted, after a simulation deformation time sequence is generated, part of data points are randomly discarded to simulate an observation value missing mode, and finally the synthesized InSAR simulation deformation time sequence is expressed as follows:
D(t)＝T(t)+S(t)
X(t)＝(D(t)+N(t))·M(t)
wherein, n (t) represents noise, m (t) represents missing mask, d (t) represents synthesized surface deformation, t (t) represents deformation trend component, and s (t) represents seasonal component.
10. A computer device comprising a memory and a processor, the memory storing a computer program, wherein the processor when executing the computer program implements the steps of the method of any of claims 1 to 6.
CN202110519191.2A 2021-05-12 2021-05-12 Method, device and equipment for removing atmospheric noise in InSAR time sequence Active CN113139349B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110519191.2A CN113139349B (en) 2021-05-12 2021-05-12 Method, device and equipment for removing atmospheric noise in InSAR time sequence

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110519191.2A CN113139349B (en) 2021-05-12 2021-05-12 Method, device and equipment for removing atmospheric noise in InSAR time sequence

Publications (2)

Publication Number Publication Date
CN113139349A true CN113139349A (en) 2021-07-20
CN113139349B CN113139349B (en) 2022-11-29

Family

ID=76817866

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110519191.2A Active CN113139349B (en) 2021-05-12 2021-05-12 Method, device and equipment for removing atmospheric noise in InSAR time sequence

Country Status (1)

Country Link
CN (1) CN113139349B (en)

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6477468B1 (en) * 2000-08-28 2002-11-05 The United States Of America As Represented By The Secretary Of The Air Force Atmospheric density modeling process using data from multiple satellites
US20090005997A1 (en) * 2006-01-30 2009-01-01 Willen Dennis E Method for Spatial Filtering of Electromagnetic Survey Data
CN104020495A (en) * 2014-06-24 2014-09-03 中国矿业大学(北京) Automatic underground pipeline parameter recognizing method on basis of ground penetrating radar
CN104123464A (en) * 2014-07-23 2014-10-29 中国国土资源航空物探遥感中心 Method for inversion of ground feature high elevation and number of land subsidence through high resolution InSAR timing sequence analysis
CN108446516A (en) * 2018-04-04 2018-08-24 长安大学 A kind of construction fracture and surface subsidence deformation decomposition method based on inversion of the force
CN110174651A (en) * 2019-05-16 2019-08-27 河海大学 Weather radar wind power plant clutter suppression method based on low-rank Hankel matrix completion
CN110333494A (en) * 2019-04-10 2019-10-15 马培峰 A kind of InSAR timing deformation prediction method, system and relevant apparatus
CN110781169A (en) * 2019-10-24 2020-02-11 首都师范大学 Self-adaptive multi-source InSAR monitoring ground settlement time sequence data splicing method and system
CN111337923A (en) * 2020-04-10 2020-06-26 中国水利水电第四工程局有限公司 Method for establishing landslide deformation time course model through time sequence InSAR data
CN211375602U (en) * 2019-12-18 2020-08-28 国网湖北省电力有限公司电力科学研究院 Mobile intelligent noise live-action cloud picture modeling system for transformer substation
CN111722706A (en) * 2019-03-21 2020-09-29 英飞凌科技股份有限公司 Method and system for recognizing aerial written characters based on radar network

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6477468B1 (en) * 2000-08-28 2002-11-05 The United States Of America As Represented By The Secretary Of The Air Force Atmospheric density modeling process using data from multiple satellites
US20090005997A1 (en) * 2006-01-30 2009-01-01 Willen Dennis E Method for Spatial Filtering of Electromagnetic Survey Data
CN104020495A (en) * 2014-06-24 2014-09-03 中国矿业大学(北京) Automatic underground pipeline parameter recognizing method on basis of ground penetrating radar
CN104123464A (en) * 2014-07-23 2014-10-29 中国国土资源航空物探遥感中心 Method for inversion of ground feature high elevation and number of land subsidence through high resolution InSAR timing sequence analysis
CN108446516A (en) * 2018-04-04 2018-08-24 长安大学 A kind of construction fracture and surface subsidence deformation decomposition method based on inversion of the force
CN111722706A (en) * 2019-03-21 2020-09-29 英飞凌科技股份有限公司 Method and system for recognizing aerial written characters based on radar network
CN110333494A (en) * 2019-04-10 2019-10-15 马培峰 A kind of InSAR timing deformation prediction method, system and relevant apparatus
CN110174651A (en) * 2019-05-16 2019-08-27 河海大学 Weather radar wind power plant clutter suppression method based on low-rank Hankel matrix completion
CN110781169A (en) * 2019-10-24 2020-02-11 首都师范大学 Self-adaptive multi-source InSAR monitoring ground settlement time sequence data splicing method and system
CN211375602U (en) * 2019-12-18 2020-08-28 国网湖北省电力有限公司电力科学研究院 Mobile intelligent noise live-action cloud picture modeling system for transformer substation
CN111337923A (en) * 2020-04-10 2020-06-26 中国水利水电第四工程局有限公司 Method for establishing landslide deformation time course model through time sequence InSAR data

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
WEI SONG ET AL.: "Automatic Sea-Ice Classification of SAR Images Based on Spatial and Temporal Features Learning", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 *
ZHUOYI ZHAO ET AL.: "Recurrent neural networks for atmospheric noise removal from InSAR time series with missing values", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 *
林珲 等: "监测城市基础设施健康的星载MT-InSAR方法介绍", 《测绘学报》 *
田馨: "InSAR技术形变监测中的干涉条件研究", 《中国博士学位论文全文数据库 基础科学辑》 *

Also Published As

Publication number Publication date
CN113139349B (en) 2022-11-29

Similar Documents

Publication Publication Date Title
Fablet et al. Bilinear residual neural network for the identification and forecasting of geophysical dynamics
CN111814875B (en) Ship sample expansion method in infrared image based on pattern generation countermeasure network
Jung et al. Towards global empirical upscaling of FLUXNET eddy covariance observations: validation of a model tree ensemble approach using a biosphere model
Rank Application of Bayesian trained RBF networks to nonlinear time-series modeling
CN111597751B (en) Crude oil film absolute thickness inversion method based on self-expanding depth confidence network
Berry et al. Forecasting turbulent modes with nonparametric diffusion models: Learning from noisy data
CN113189561A (en) Sea clutter parameter estimation method, system, equipment and storage medium
Yu et al. Learning chaotic dynamics using tensor recurrent neural networks
Yang et al. Making invisible visible: Data-driven seismic inversion with spatio-temporally constrained data augmentation
Wang et al. Sample entropy‐based adaptive wavelet de‐noising approach for meteorologic and hydrologic time series
Chen et al. An efficient and statistically accurate Lagrangian data assimilation algorithm with applications to discrete element sea ice models
CN113139349B (en) Method, device and equipment for removing atmospheric noise in InSAR time sequence
WO2020178936A1 (en) Method for generating neural network model, and control device using neural network model
Sangireddy et al. Multiresolution analysis of characteristic length scales with high‐resolution topographic data
CN112199887A (en) Semi-supervised resistivity inversion method and system based on countermeasure generation network and pseudo-labeling
Saha et al. Downscaling extreme rainfall using physical-statistical generative adversarial learning
CN116842472A (en) Land evapotranspiration remote sensing estimation method based on depth space-time coding and decoding network
WO2019098154A1 (en) Time-sequential data generation device and program
Marchuk et al. Numerical study of the global ocean equilibrium circulation
Ouala et al. Learning ocean dynamical priors from noisy data using assimilation-derived neural nets
Giannitrapani et al. Additive models for correlated data with applications to air pollution monitoring
Yano Downscaling, parameterization, decomposition, compression: a perspective from the multiresolution analysis
Nikitin et al. Statistics-based models of flood-causing cyclones for the Baltic Sea region
Shi et al. Universal Functional Regression with Neural Operator Flows
Mateo‐García et al. Generative Adversarial Networks in the Geosciences

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant