CN115293198A

CN115293198A - Method for improving GNSS-R height finding inversion accuracy based on multi-hidden-layer neural network

Info

Publication number: CN115293198A
Application number: CN202210761676.7A
Authority: CN
Inventors: 郑伟; 吴凡; 王强
Original assignee: China Academy of Space Technology CAST
Current assignee: China Academy of Space Technology CAST
Priority date: 2022-06-29
Filing date: 2022-06-29
Publication date: 2022-11-04

Abstract

The invention discloses a method for improving GNSS-R height finding inversion accuracy based on a multi-hidden-layer neural network, which comprises the following steps: acquiring multi-source data; extracting the characteristics of the multi-source data to obtain a characteristic parameter set; constructing a sea surface height prediction model constructed based on a multi-hidden-layer neural network; and taking the characteristic parameter set as the input of the sea surface height prediction model, and outputting the corresponding sea surface height through the sea surface height prediction model. By the method, the GNSS-R height measurement inversion accuracy is improved, and the reliability of the novel MHL-NN model is verified by using the global average sea surface DTU18 subjected to tide correction. In addition, in order to obtain a feature set more suitable for a sea surface height inversion model, 14 sets of feature sets with different information details are used for respectively training a novel MHL-NN model, and the sensitivity of the sea surface height inversion performance to different input parameters is analyzed.

Description

Method for improving GNSS-R height finding inversion accuracy based on multi-hidden-layer neural network

Technical Field

The invention belongs to the technical field of intersection such as satellite altimetry and marine surveying and mapping, and particularly relates to a method for improving GNSS-R (global navigation satellite system-reflection) altimetry inversion accuracy based on a multi-hidden-layer neural network.

Background

Sea surface height is an important parameter in ocean science research, and plays an important role in establishing a global tide model, observing large-scale ocean circulation, monitoring global climate change and the like. A Global Navigation Satellite system reflectometer (GNSS-R) is a novel bistatic Satellite remote sensing technology with a transmitting and receiving dual-position, which takes Navigation Satellite signals as signal sources and carries out reflection surface physical quantity inversion by measuring the difference between direct signals and reflected signals reflected by the sea surface. Since the GNSS-R technology is applied to the field of sea surface height measurement by Martin-Neira in 1993 for the first time, the technology is successfully verified on platforms such as shore-based platforms, air-based platforms, space-based platforms and the like. Compared with the traditional radar altimeter, the GNSS-R altimeter has the advantages of low cost, multiple signal sources, high space-time resolution and the like, is beneficial to obtaining the mesoscale ocean altitude signal and keeps continuous observation.

Successful transmission of TechDemosat-1 (TDS-1) in the UK, CYGNSS in the USA and A/B double-star wind-catching number in China successfully marks that the GNSS-R technology is stepped into a new stage of detecting global surface parameters. Clarizia et al initially uses TDS-1 data to preliminarily study satellite-borne GNSS-R sea height inversion, and verifies in the North Pacific ocean and the south Atlantic ocean respectively by adopting a DTU10 average sea model, and the results show that the GNSS-R data can successfully capture large-scale changes of the sea height. Xu et al utilized TDS-1 satellite data to treat 351 lakes (area greater than 500 km) around the world ² And the elevation of each lake is lower than 3000 m), and the result shows that the lake surface inversion result of the TDS-1 satellite has strong correlation with the inversion results of the CryoSat-2 satellite, the Jason satellite and the Envisat satellite, but has larger error. Mashburn et al further utilize TDS-1 data to perform sea surface height inversion, and establish various error models involved in the satellite-borne GNSS-R height inversion process, and analysis shows that the satellite-borne GNSS-R can accurately invert the global sea surface height, and the residual error compared with the DTU10 average sea surface height is about 6.4m. But adds Delay-Doppler Map (DDM) informationCompression into a single scalar does not completely reflect sea height information. Li and the like use CYGNSS data to compare and analyze sea surface height inversion accuracy of three re-tracking methods of HALF, DER and Z-V models, and the result shows that the Z-V model re-tracking method has higher inversion accuracy. Mashburn et al utilizes CYGNSS satellite-borne GNSS-R data of Indonesia to invert the sea height by adopting a VZ18 model retracing method to obtain about 6m inversion deviation. From the above analysis, in the previous research, a re-tracking method such as HALF, DER and model fitting is generally adopted to perform GNSS-R sea surface height inversion, and a corresponding error model is established by analyzing various errors involved in the inversion model to improve the height measurement inversion accuracy. The traditional re-tracking method is mostly an empirical model, the height measurement precision is low, and meanwhile, various error models are established, so that an inversion model is complex and difficult to realize.

Machine learning, especially Artificial Neural Network (ANN), can make full use of the self-learning and self-adaptive capacity of nerve cells to process complex nonlinear problems, compared with the prior inversion model, the Neural Network algorithm is simple, the relation between a plurality of observed quantities and the sea surface height can be established, physical quantities related to the sea surface height inversion can be fully utilized, and the defects of the traditional inversion method are made up to a certain extent. Machine learning algorithms have now gradually been incorporated into the GNSS-R domain and achieved excellent results. Liu and the like adopt a multi-hidden-layer neural network, chu and the like adopt a convolutional neural network, luo and the like adopt an inversion algorithm of a tree model, a mapping model from satellite-borne DDM observation data to European middle Weather forecast center (ECMWF) analysis field data is respectively established, and the obtained result is obviously superior to that of the traditional GNSS-R wind speed inversion method. Jia and the like invert soil moisture characteristics by adopting a machine learning XGboost algorithm based on shore-based GNSS-R data, and Yan and the like perform sea ice detection and sea ice density prediction by utilizing a convolutional neural network, so that good inversion effects are obtained.

At present, the research of machine learning in the GNSS-R sea surface height measurement field is still in a starting stage. Wang et al have constructed the new machine learning weighted average and fused the characteristic extraction method based on the airborne data of the sea of Polaroid, compare to DTU15 and verify that the mean absolute error of the model is about 0.25m, the correlation coefficient is about 0.75.Zhang and the like construct two different satellite-borne sea surface height inversion models which take height angle, time delay difference, antenna gain, signal to noise ratio, GPS satellite transmitter speed and TDS-1 satellite receiver speed as input and DTU15 verification model sea surface height as output, and combine Principal Component Analysis with Support Vector Regression (PCA-SVR) and Convolutional Neural Network (CNN), and successfully apply a machine learning algorithm to the field of satellite-borne GNSS-R sea surface height measurement. However, only the traditional HALF single-point re-tracking method is adopted to extract time delay from DDM to serve as machine learning input characteristics, data information of the DDM cannot be fully utilized, certain information loss is caused, the structures of the used PCA-SVR and one-dimensional CNN models are simple, and the relation between GNSS-R data and sea surface height is difficult to fully mine.

Disclosure of Invention

The technical problem of the invention is solved: the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network aims to improve the GNSS-R height finding inversion accuracy.

In order to solve the technical problem, the invention discloses a method for improving the GNSS-R height finding inversion accuracy based on a multi-hidden-layer neural network, which comprises the following steps:

acquiring multi-source data; wherein, multisource data includes: TDS-1 data, ionized layer data and EAR5 sea surface wind speed data; TDS-1 data is L1B-level data provided by a TDS-1 project; the ionosphere data is a global ionosphere map provided by the IGS; EAR5 sea surface wind speed data is an EAR5 analysis field data set of the ECMWF;

performing feature extraction on the multi-source data to obtain a feature parameter set;

constructing a sea surface height prediction model constructed based on a multi-hidden-layer neural network;

and taking the characteristic parameter set as the input of the sea surface height prediction model, and outputting the corresponding sea surface height through the sea surface height prediction model.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network, the characteristic extraction is performed on the multi-source data to obtain a characteristic parameter set, and the method comprises the following steps:

extracting DDM images and corresponding metadata from the L1B-level data;

denoising the DDM image to obtain a DDM image with the substrate noise removed; analyzing and processing the DDM image without the substrate noise to obtain an integral delay waveform IDW, a delay correlation curve forward-leaning edge slope LES, a delay-Doppler average value DDMA and a delay waveform peak value PCP; extracting the SNR and the ELE from the metadata to obtain the SNR and the ELE;

analyzing and processing ionization layer data to obtain a total atmospheric delay ATM;

analyzing and processing EAR5 sea surface wind speed data to obtain sea surface wind speed;

carrying out dimension reduction processing on the integral time delay waveform IDW;

and constructing to obtain a characteristic parameter set based on the integral time delay waveform IDW after dimensionality reduction, the forward-inclined edge slope LES of the delay correlation curve, the time delay-Doppler average value DDMA, the time delay waveform peak value PCP, the total atmospheric delay ATM, the sea surface wind speed, the signal-to-noise ratio SNR and the mirror surface point height angle ELE.

In the method for improving the GNSS-R altimetry inversion accuracy based on the multi-hidden-layer neural network, the DDM image from which the substrate noise is removed is analyzed and processed to obtain an integral delay waveform IDW, which includes:

obtaining an integral time delay waveform IDW by non-coherent integration calculation of a time delay waveform in a set Doppler frequency shift range in the DDM image after substrate noise removal:

wherein the content of the first and second substances,

indicating a set Doppler shift range, τ _j Which is representative of the time-delayed waveform,

showing the DDM image after removing the substrate noise, M showing the set Doppler shift range

The number of Doppler points in the Doppler signal.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden layer neural network, the DDM image without the substrate noise is analyzed and processed to obtain a time delay correlation curve forward-leaning edge slope LES, which includes:

adopting the slope of a linear function fitted by an optimal first-order polynomial as the slope LES of the anteversion edge of the time delay correlation curve:

wherein, I (τ) _k ) Representing the leading edge function, tau, of the integrated time-delay waveform _k Representing the integrated delay waveform, and alpha and c represent the slope and intercept, respectively, of the best-fit line.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden layer neural network, the DDM image with the substrate noise removed is analyzed and processed to obtain the delay-Doppler average value DDMA, and the method comprises the following steps:

based on the DDM image after removing the substrate noise, the delay-doppler average DDMA is determined by the following formula:

wherein the content of the first and second substances,

The number of doppler points in the range, N, represents the number of delay points in the range of delays used.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network, the DDM image without the substrate noise is analyzed and processed to obtain a time delay waveform peak value PCP, and the method comprises the following steps:

calculating to obtain the peak value tau of the time delay waveform of the reflected signal by the following formula _spec ：

Wherein W (tau) represents the related power time delay waveform of the reflected signal, and tau represents time delay;

determining the peak value tau of time delay waveform of reflected signal _spec And (4) determining the corresponding delay position, namely the delay waveform peak value PCP.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden layer neural network, the ionization layer data is analyzed and processed to obtain the total atmospheric delay ATM, and the method comprises the following steps:

respectively calculating to obtain the ionospheric delay delta from GPS satellite to mirror reflection point based on the global ionospheric diagram provided by IGS ₁ Ionospheric delay delta from specular reflection point to TDS-1 ₂ And ionospheric delay delta from GPS satellite to TDS-1 receiver ₃ ；

Determining the total ionospheric delay delta of reflected signals relative to direct signals caused by the ionosphere _iono ：

δ _iono ＝(δ ₁ +δ ₂ )-δ ₃

Determining a total tropospheric delay delta of a reflected signal caused by the tropospheric layer relative to a direct signal _tro ：

δ _tro ＝(δ _tro1 +δ _tro2 )

Then, the total atmospheric delay ATM is:

ATM＝δ _iono +δ _tro

wherein, delta _tro1 Representing tropospheric correction, delta, of satellite transmitter to specular reflection point _tro2 Showing tropospheric correction of specular reflection points to TDS-1.

In the method for improving the GNSS-R altimetry inversion accuracy based on the multi-hidden-layer neural network, the method for analyzing and processing the EAR5 sea surface wind speed data to obtain the sea surface wind speed includes:

acquiring a TDS-1 satellite-borne data set;

and performing space-time matching on the TDS-1 satellite-borne data set and the EAR5 analysis field data set to obtain the corresponding sea surface wind speed.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network, the sea surface height prediction model constructed based on the multi-hidden-layer neural network is constructed, and the method comprises the following steps:

constructing a multi-hidden-layer neural network model;

determining a training set; wherein, include in the training set: sample data corresponding to an integral time delay waveform IDW, a forward slope LES of a delay correlation curve, a time delay-Doppler average value DDMA, a time delay waveform peak value PCP, total atmospheric delay ATM, sea surface wind speed, a signal-to-noise ratio SNR and a mirror plane point height angle ELE after dimensionality reduction processing;

training and verifying a multi-hidden-layer neural network model by using a 5-fold cross verification method based on the determined training set, and controlling the mode of dividing the training set and the test set each time by using random seed numbers to obtain optimal parameters;

and determining the sea surface height prediction model constructed based on the multi-hidden-layer neural network based on the obtained optimal parameters.

In the method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network, the output of the sea surface height prediction model constructed based on the multi-hidden-layer neural network is represented as follows:

h ₁ ＝σ(W ₁ Feature_set+b ₁ )

h _i ＝σ(W _i h _i-1 +b _i )

wherein h is _i Represents the output of the ith hidden layer, W _i Weight representing the ith hidden layer, b _i A bias term representing the ith hidden layer,

showing the output of a sea surface height prediction model built based on a multi-hidden-layer neural network, feature _ set showing the input of the sea surface height prediction model built based on the multi-hidden-layer neural network, H (W, b) showing a target function obtained by training, sigma (DEG) showing an activation function of a hidden layer, loss (DEG) showing a Loss function of the model, and DTU18 _SSH Representing the sea level height obtained by the verification model DTU 18.

The invention has the following advantages:

the invention discloses a method for improving GNSS-R height finding inversion accuracy based on a multi-hidden-layer neural network, which takes TDS-1 satellite-borne GNSS-R data as an example, provides a deeper and denser multi-hidden-layer neural network model for sea surface height inversion, and verifies the reliability of a novel MHL-NN model by using a global average sea surface DTU18 subjected to tide correction. In addition, in order to obtain a feature set more suitable for a sea surface height inversion model, 14 sets of feature sets with different information details are applied to respectively train a novel MHL-NN model, and the sensitivity of the sea surface height inversion performance to different input parameters is analyzed. The invention also realizes HALF single-point retracing algorithm, and compares the inversion result with the inversion result of the proposed novel MHL-NN model.

Drawings

FIG. 1 is a flowchart illustrating steps of a method for improving GNSS-R altimetry inversion accuracy based on a multi-hidden-layer neural network according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of an artificial neural network according to an embodiment of the present invention;

FIG. 3 is a correlation heatmap in accordance with an embodiment of the present invention; wherein 3 (a) is a correlation heat map of the time delay waveform and the sea surface height of the DTU 18; 3 (b) is a correlation heat map of the integrated delay waveform and the sea surface height of the DTU 18;

FIG. 4 is a diagram illustrating the calculation regions of the substrate noise and DDMA in a DDM according to an embodiment of the present invention; wherein, 4 (a) is a region for calculating the base noise in the DDM; 4 (b) is the region of the DDM used to calculate DDMA;

FIG. 5 is a diagram illustrating a preferred result of a hyper-parameter in an embodiment of the present invention; wherein, 5 (a) is the RMSE of the inversion sea surface height of the novel MHL-NN model with different layer numbers and neuron numbers; 5 (b) inverting correlation coefficients of the sea surface height by the novel MHL-NN model with different layer numbers and neuron numbers;

FIG. 6 is a schematic view of a DDM of a different reflective surface in an embodiment of the present invention; wherein 6 (a) is the sea surface reflected DDM;6 (b) DDM for sea ice reflection; 6 (c) DDM for land edge reflection; 6 (d) is a DDM containing noise;

FIG. 7 is a diagram illustrating the statistics of the delay and Doppler pixel position of the peak power of DDM data according to an embodiment of the present invention;

FIG. 8 is a schematic diagram of the global sea level distribution of TDS-1 experimental data in an embodiment of the present invention;

FIG. 9 is a diagram illustrating inversion results of a novel MHL-NN model according to an embodiment of the present invention; wherein, 9 (a) is a training and verification error curve diagram of the novel MHL-NN model; 9 (b) is a scattered point density graph of the inversion result of the novel MHL-NN model;

FIG. 10 is a diagram illustrating error statistics for a novel MHL-NN model according to an embodiment of the present invention; 10 (a) is a probability density function graph of an inversion result of the novel MHL-NN model relative to DTU18 verification data; 10 (b) an error distribution histogram of the inversion result of the novel MHL-NN model;

FIG. 11 is a schematic representation of the global sea level altitude for a different model in an embodiment of the present invention; wherein 11 (a) is a global sea level height inversion result of a HALF traditional method; 11 (b) obtaining a global sea level height inversion result of the novel MHL-NN model; 11 (c) Global sea level for DTU18 verification model

FIG. 12 is a schematic diagram of the global sea level height error statistics obtained by the HALF re-tracking method and the inversion of the novel MHL-NN model in the embodiment of the present invention.

Detailed Description

To make the objects, technical solutions and advantages of the present invention more apparent, the embodiments of the present invention will be described in detail with reference to the accompanying drawings.

As shown in fig. 1, in this embodiment, the method for improving GNSS-R altimetry inversion accuracy based on a multi-hidden-layer neural network includes:

step 101, obtaining multi-source data.

In this embodiment, the multi-source data may specifically include:

TDS-1 data: L1B level data provided by TDS-1 project; wherein the L1B level data is mainly composed of DDM images and corresponding metadata.

Ionospheric data: the Global Ionospheric Maps (GIM) provided by IGS.

EAR5 sea surface wind speed data: EAR5 of ECMWF analyzes the field data set.

And step 102, extracting the characteristics of the multi-source data to obtain a characteristic parameter set.

In this embodiment, the obtained multi-source data may be processed by using a feature construction and feature extraction method to obtain a feature parameter set that can be directly used by a subsequent model.

Preferably, a DDM image and corresponding metadata are extracted from the L1B level data; denoising the DDM image to obtain a DDM image with the substrate noise removed; analyzing and processing the DDM image without the substrate noise to obtain an integral delay waveform IDW, a delay correlation curve forward-leaning edge slope LES, a delay-Doppler average value DDMA and a delay waveform peak value PCP; extracting the SNR and the ELE from the metadata to obtain the SNR and the ELE; analyzing and processing ionization layer data to obtain a total atmospheric delay ATM; analyzing and processing EAR5 sea surface wind speed data to obtain sea surface wind speed; carrying out dimensionality reduction processing on the integral time delay waveform IDW; and constructing to obtain a characteristic parameter set based on an integral time delay waveform IDW after the dimensionality reduction, a forward-inclined edge slope LES of a delay correlation curve, a time delay-Doppler average value DDMA, a time delay waveform peak value PCP, a total atmospheric delay ATM, a sea surface wind speed, a signal-to-noise ratio SNR and a mirror surface point height angle ELE.

Specifically, the method comprises the following steps:

(1) The integrated delay waveform IDW after the dimensionality reduction processing, the forward slope of the delay correlation curve LES, the delay-Doppler average value DDMA and the delay waveform peak value PCP.

The base noise is the result of the internal noise effect of the receiver, does not contain characteristic information, and needs to be filtered from the DDM image. In this embodiment, a non-scattering signal region before a "flicker region" in a DDM image is selected as a substrate noise calculation region, for example, a rectangular frame region shown in fig. 4 (a), and a scattering power value of the region is averaged to obtain a substrate noise and a denoised DDM image.

1.1 Integral Delay waveform IDW (IDW) after dimension reduction processing

The integrated delay waveform IDW is calculated by non-coherent integration of the delay waveform in a set doppler shift range (e.g., -1000,1000 hz) in the DDM image after removing the substrate noise, and the calculation formula is as follows:

wherein the content of the first and second substances,

The number of Doppler points in the Doppler signal. Fig. 3 shows a heat map of the correlation of the Delay Waveform (DW) and the Integrated Delay Waveform (IDW) with the sea level height data of the DTU18 after tidal correction. From FIG. 3, IDW data and DT can be seenThe correlation of the U18 sea surface height data is stronger than the DW data.

Further, TDS-1 satellite-borne IDW data is a set of 128-dimensional data sets that contain many features that are redundant and highly uncorrelated with the sea surface, which increases the training time of the model and is prone to overfitting. Therefore, in order to improve the training efficiency and accuracy of the model, the IDW data set is screened by using a Pearson correlation coefficient method, the characteristic that the correlation coefficient of the IDW data set and the DTU18 verification model is less than 0.1 is eliminated, and the dimensionality of the screened IDW data set is reduced from 128 to 18; then, a 15-dimensional feature set with a cumulative contribution rate of 95% is extracted as a final IDW data set by using Principal Component Analysis (PCA).

1.2 Extend the Edge Slope LES (LES) of the forward curve

The anteversion edge slope LES of the delay correlation curve is used as the anteversion edge slope of IDW, and the first-order function slope of the best first-order polynomial fitting can be used as the anteversion edge slope LES of the delay correlation curve:

1.3 Delay-Doppler mean DDMA (DDMA)

In the present invention, DDMA can be calculated for a DDM image region having a chip range of-0.25, 0.25] chips, a Doppler shift range of-1000, 1000 Hz, centered on the point of specular reflection. The calculation region of a typical DDMA in a DDM image is shown in fig. 4 (b), where the maximum value of the scattered signal power in fig. 4 (b) represents a specular reflection point, and a rectangle with the specular reflection point as the center is 5 (doppler) × 3 (time delay) is the calculation region of the DDMA. Preferably, the specific calculation formula of the delay-doppler average DDMA is as follows:

where N represents the number of delay points in the range of delays used.

1.4 A peak PCP (PCP) of the delay of the peak correlation power)

The peak value PCP of the delay waveform is: normalizing the corresponding time delay position at the waveform peak value, thereby obtaining the time delay waveform peak value tau of the reflected signal _spec And determining the peak value PCP of the time delay waveform. Wherein, the peak value tau of the time delay waveform of the reflected signal can be calculated by the following formula _spec ：

Wherein, W (tau) represents the related power time delay waveform of the reflected signal, and tau represents the time delay.

(2) Total atmospheric delay ATM

Atmospheric delays are mainly composed of ionospheric and tropospheric delays.

2.1 Ionospheric delay)

Ionospheric delays along the direct and reflected signal paths can result in ranging errors of several meters. The invention can respectively calculate the ionospheric delay delta from the GPS satellite to the mirror reflection point based on the global ionospheric diagram provided by IGS ₁ Ionospheric delay delta from specular reflection point to TDS-1 ₂ And ionospheric delay delta from GPS satellite to TDS-1 receiver ₃ (ii) a The total ionospheric delay δ of the reflected signal caused by the ionosphere relative to the direct signal _iono Comprises the following steps: delta _iono ＝(δ ₁ +δ ₂ )-δ ₃ 。

2.2 Tropospheric delay)

Tropospheric delay can be calculated using the UNB3m model of New University of New Brunswick. The tropospheric delay effects are concentrated at heights below 10km from the ground, so tropospheric corrections are only applied to the downward delta below the receiver height _tro1 And upward delta _tro2 A reflected signal path. GPS satelliteThe direct signal path to the TDS-1 satellite receiver does not require tropospheric correction. Total tropospheric delay delta of reflected signals caused by tropospheric layers relative to direct signals _tro Comprises the following steps: delta _tro ＝(δ _tro1 +δ _tro2 ). Wherein, delta _tro1 Tropospheric correction, delta, representing the point of reflection of a satellite transmitter to a mirror _tro2 Showing tropospheric correction of specular reflection points to TDS-1.

From above, the total Atmospheric Delay ATM (ATM) can be expressed as: ATM = δ _iono +δ _tro 。

(3) Sea surface wind speed

In the invention, the TDS-1 satellite-borne data set is obtained, and the TDS-1 satellite-borne data set and an EAR5 analysis field data set are subjected to space-time (time: 1h; longitude and latitude: 0.25 degree) matching, so that the required sea surface wind speed is obtained.

(4) Signal-to-Noise Ratio (SNR) and specular point height angle ELE (ELE).

In the invention, two characteristics of signal-to-noise ratio SNR and specular point height angle ELE can be directly extracted from the metadata.

And 103, constructing a sea surface height prediction model constructed based on the multi-hidden-layer neural network.

In this embodiment, as shown in fig. 2, it can be seen that the neural network generally consists of an input layer, a Hidden layer and an output layer, wherein an artificial neural network including a plurality of Hidden layers is a Multi-Hidden layer neural network (MHL-NN). A novel multi-hidden-layer neural network algorithm is adopted to establish a sea surface height prediction model, the essence is a supervised learning regression problem, namely DDM data, metadata, atmospheric delay, sea surface wind speed and other information of a TDS-1 satellite are used as input data, corresponding sea surface height is used as an output data training model, the model is optimized by observing the expression effect of the trained model on a verification set, and finally, the prediction of unknown data is realized.

Firstly, constructing a multi-hidden-layer neural network model; then, a training set is determined, wherein the training set comprises: the acquisition mode of the sample data corresponding to the integral delay waveform IDW after the dimensionality reduction, the forward-leaning edge slope LES of the delay correlation curve, the delay-doppler average value DDMA, the delay waveform peak value PCP, the total atmospheric delay ATM, the sea surface wind speed, the signal-to-noise ratio SNR and the mirror point height angle ELE can refer to the feature extraction process of the step 102, and is not described herein again; secondly, training and verifying a multi-hidden-layer neural network model by using a 5-fold cross verification method based on the determined training set, and controlling the mode of dividing the training set and the test set each time by using random seed numbers to obtain optimal parameters; and finally, determining the sea surface height prediction model constructed based on the multi-hidden-layer neural network based on the obtained optimal parameters.

Preferably, the sea surface height prediction model constructed based on the multi-hidden-layer neural network has K hidden layers, and given a set of input Feature _ sets, the outputs of the model layers can be represented as:

h ₁ ＝σ(W ₁ Feature_set+b ₁ )

h _i ＝σ(W _i h _i-1 +b _i )

showing the output of a sea surface height prediction model built based on a multi-hidden-layer neural network, feature _ set showing the input of the sea surface height prediction model built based on the multi-hidden-layer neural network, H (W, b) showing a target function obtained by training, sigma (DEG) showing an activation function of a hidden layer, loss (DEG) showing a Loss function of the model, and DTU18 _SSH Representing the sea level height obtained by the verification model DTU 18. The embodiment constructs based on multiple hidden spiritThe sea surface height prediction model constructed by the network adopts a ReLU function as an activation function, a mean square error function as a loss function and an Adam self-adaptive optimization algorithm as an optimization algorithm.

The number of Hidden layers and the number of neurons in each layer are two important hyper-parameters in a sea surface height prediction model (a novel MHL-NN model for short) constructed by a Multi-Hidden layer neural network, and the two important hyper-parameters need to be specified in advance in a model training stage. Fig. 5 shows the RMSE and PCC variation for different numbers of layers and different numbers of neurons. The result shows that as the number of layers and the number of neurons increase, the RMSE of the inversion result of the novel MHL-NN model gradually becomes smaller, and the PCC gradually increases. When the number of layers is increased to 3 and the number of neurons is increased to more than 200, the inversion performance of the novel MHL-NN model tends to be stable. Overall, an MHL-NN architecture with 4 hidden layers and 200 neurons per layer can be optimized.

And 104, taking the characteristic parameter set as the input of the sea surface height prediction model, and outputting the corresponding sea surface height through the sea surface height prediction model.

On the basis of the above embodiment, the verification process of the novel MHL-NN model is described below.

1.1 data set

1.1.1 time delay waveform data

A TDS-1 Satellite carrying a Space GNSS Receiver-Remote Sensing Instrument (SGR-ReSI) is launched and lifted off in 2014 by Surrey Satellite Technology Ltd (SSTL), and stops running at 2018, the track height is 635km, the inclination angle is 98 degrees, and the sampling rate is 16.367MHz. A large amount of measured data is accumulated during the operation of the TDS-1 satellite, and the data quality and the data quantity are both very suitable for sea surface height inversion research.

The experimental data of the invention uses all available DDM data and corresponding metadata acquired in 2018, 2-12 months of TDS-1 project, and the specular reflection point is located between 80 degrees of north latitude and 60 degrees of south latitude. Data was sourced from MERRByS co uk website, TDS-1 provides daily L1B data stored in four groups H00, H06, H12, H18 at 6 hour intervals. The DDM data consists of 128 delay pixels and 20 doppler pixels with a doppler resolution of 500Hz and a delay resolution of 0.25chip. And extracting a Delay waveform corresponding to a zero Doppler position in each DDM pixel to form a Delay waveform data set (DW).

1.1.2 DTU18 verification model

When TDS-1 satellite-borne GNSS-R sea surface height inversion is carried out, comparison verification is needed to be carried out with measured sea surface data, and satellite-borne sea surface height inversion accuracy is determined. Due to the lack of measured data, the present invention verifies the Sea Surface height inversion accuracy using a verification model consisting of the global Mean Sea Surface model developed by danish University of technology (DTU Sea Surface 18, DTU 18) and the TPXO8 global Sea tide model provided by Oregon State University (OSU). Sea Surface Height (SSH) DTU18 from the verification model _SSH Can be expressed as:

DTU18 _SSH ＝DTU18+TPXO _tide

wherein, TPXO _tide Representing the tidal correction calculated from the TPXO8 global tidal model.

1.1.3 data quality control

In the process of establishing a neural network height inversion model, data quality control is an essential part for reasonably utilizing data. Therefore, when the TDS-1 data is used for verifying the inversion algorithm, the quality of the data needs to be controlled, and the data quality and the screening method are analyzed.

Power signal accuracy: the Signal-to-Noise Ratio (SNR) and the Antenna Gain (AG) of the DDM can reflect the strength of the reflected Signal power, and the method screens data with SNR and AG both larger than 5dB for sea surface height inversion.

Sea ice removal: sea ice is a very serious disturbance for GNSS-R technology for sea level altimetry applications. Fig. 6 (a) and 6 (b) show the sea surface reflected DDM and the ice surface reflected DDM, respectively, and it can be seen from fig. 6 (b) that the reflection of the ice surface signal is specular reflection, and the signal in the DDM is concentrated in a very small frequency and code delay range. Since the sea surface generates scattering, the DDM can detect the reflected signal in a wide range of frequencies and code delays. To eliminate the sea ice effect, only data within ± 55 ° latitude were retained.

Land data elimination: since the GNSS-R technology can receive a wide area of reflected signals, when a specular reflection point approaches the land or the sea island, a part of the reflected signals from the land will be received. The scattering coefficient of land and the scattering coefficient of sea are different, so an asymmetric waveform as shown in fig. 6 (c) is observed in the DDM data. In order to eliminate land data, the invention uses High-resolution world coastline data provided by a Global High-resolution geographic Database (GSHHS), wherein the complete Database comprises 10222509 data points and the average point distance is about 178m.

When a GNSS satellite falls or rises behind the edge of the earth, the GNSS radio occultation (GNSS-RO) event can be effectively eliminated by selecting the altitude angle to be larger than 60 degrees.

Influence of sea surface wind speed: as wind speed increases, the sea surface becomes rougher, resulting in a weaker reflected signal power. Reynolds and the like utilize DDM data to perform sea surface wind speed inversion, and the result shows that a better inversion result can be obtained in a medium-low wind speed interval, and the inversion result of the wind speed above 12m/s still has larger deviation. Based on the method, in order to improve the inversion capability of the model, the TDS-1 data of the medium and low speed wind speed (< 12 m/s) with more concentrated data is selected to carry out sea surface height inversion.

And (3) removing other noises: hu et al reported 14 DDM data containing anomalies, which contain bright spots of anomalies and other faint signals in addition to the normal DDM waveform. Such data can affect the inversion and need to be culled. Figure 6 (d) shows DDM data containing anomalous signals, generally, a sea surface reflected DDM is horseshoe shaped and power peaks occur near 0chip delay and 0Hz doppler, but the DDM pattern of figure 6 (d) is cluttered with no apparent shape. FIG. 7 shows statistics of the time delay and Doppler pixel position of the peak power of DDM data, and it can be seen that the time delay of the peak power is mainly concentrated at [63,73] pt, and the Doppler is mainly concentrated at [9,14] pt. Based on this, in order to further improve the data quality, the invention eliminates the obvious abnormal data with the peak power outside the threshold value.

1.1.4 data matching

After quality control and data filtration, about 60 ten thousand groups of TDS-1 time delay waveform data sets are obtained. As TDS-1 satellite-borne data is continuous time-varying data, the DTU18 average sea surface model is grid data with 1' longitude and latitude. Therefore, the method comprises the steps of firstly carrying out space matching on a TDS-1 satellite-borne data set and a DTU18 average sea surface model (the difference between the longitude and the latitude of sample points of the two data sets is within 0.5'), extracting the DTU18 average sea surface height corresponding to the TDS-1 data, then utilizing a TPXO8 global tide model to calculate tide correction with the same time and the same longitude and latitude as a satellite-borne time delay waveform, and overlapping the tide correction with the satellite-borne time delay waveform on a DTU18 average sea surface to obtain the sea surface height value DTU18 of a DTU18 verification model _SSH . TDS-1 satellite-borne time delay waveform data set and corresponding sea surface height value DTU18 _SSH An original sample set is constructed. FIG. 8 shows the global distribution of TDS-1 time delay waveform data sets after screening, with sea level height values of [ -100, +80 [)]m is between.

And (3) matching the obtained original sample set according to the ratio of 8: the ratio of 2 is divided into a training set and a test set. Namely, 80% of data samples are randomly selected from an original sample set to serve as training data for optimization of model hyper-parameters and preliminary evaluation of model performance. And the rest 20% of data is taken as test data, and the test data does not participate in the establishment of the model and is only used for evaluating the final precision and generalization capability of the model.

1.2 feature sensitivity analysis

In order to evaluate the contribution of different combinations of the constructed time delay waveform DW, the integral time delay waveform IDW, the time delay-Doppler average value DDMA, the forward inclination slope LES of the time delay waveform, the height angle ELE of a mirror point, the signal-to-noise ratio SNR of the DDM, the atmospheric delay ATM, the sea surface wind speed EAR5 and the characteristics of a retracing algorithm (PCP, PCP70 and PFD) in the sea surface height inversion of the novel MHL-NN model, the invention uses 14 sets of characteristic sets with different information details to respectively train the novel MHL-NN model and verifies the precision on a test set, namely:

Set1：DW

Set2：IDW

Set3：IDW、ELE

Set4：IDW、SNR

Set5：IDW、ELE、SNR

Set6：IDW、ELE、SNR、ATM

Set7：IDW、ELE、SNR、EAR5

Set8：IDW、ELE、SNR、ATM、EAR5

Set9：IDW、ELE、SNR、ATM、EAR5、LES

Set10：IDW、ELE、SNR、ATM、EAR5、PCP

Set11：IDW、ELE、SNR、ATM、EAR5、PFD

Set12：IDW、ELE、SNR、ATM、EAR5、PCP70

Set13：DDMA、ELE、SNR、ATM、EAR5、PCP

Set14：IDW、ELE、SNR、ATM、EAR5、LES、PCP、DDMA

in order to verify the robustness of the fusion model, the random seed number of 5-fold cross validation is replaced twice, and the experiment is carried out again, wherein the replacement of the random seed number is equivalent to the re-division of the training set and the validation set. The novel MHL-NN model has good robustness, the MAD, the RMSE and the PCC of each model in the three experiments are basically not obviously changed, and the results of the three experiments are basically consistent.

As can be seen from Set1 and Set2, compared with the direct DW waveform data, the IDW waveform data can utilize more DDM information, thereby obtaining a better inversion result. After addition of SNR and ELE features alone, the RMSE of the inversion results was only slightly reduced and not significantly improved (Set 2, set3, set 4). However, after the SNR and ELE features are added simultaneously, the sea surface height inversion accuracy can be improved remarkably, PCC is increased by 7.1%, and RMSE is reduced by 26.7% (Set 2 and Set 5). Set6 and Set7 show that the sea surface wind speed characteristic and the atmospheric delay characteristic can provide positive contribution to the sea surface height inversion of the novel MHL-NN model, wherein the improvement of the atmospheric delay characteristic is particularly remarkable (relative to Set 5). Set8 gives the inversion result of the sea surface height after simultaneously increasing the sea surface wind speed characteristic and the atmospheric delay characteristic. PCC increase by 6.6% and RMSE decrease by 44.3% relative to Set 5. This improvement is mainly due to the fact that atmospheric delay characteristics can provide additional signal propagation path information, and sea surface wind speed provides additional sea surface roughness information. Set9-12 analyzed the effect of LES, PCP, PFD, PCP70 features constructed on DDM on sea level height inversion, and it can be seen that LES, PCP provide positive contributions (Set 8-10), while PFD and PFD70 provide negative contributions (Set 8, 11, 12). Set13 replaces the whole IDW delay waveform data with DDMA, and it can be seen that the inversion accuracy is slightly lower than that of the whole IDW delay waveform (Set 13 and Set 10) directly adopted. By adding all positively contributing features as inputs to the neural network (Set 14), the RMSE of the inversion height can be reduced to 4.23m and the pcc increased to 0.98.

1.3 height measurement accuracy analysis of novel MHL-NN model

The invention evaluates the sea-level height inversion performance of the proposed novel MHL-NN model with 4 hidden layers and 200 neurons per layer. And establishing a sea surface height inversion model by taking the feature set14 with the highest inversion accuracy as the input of the novel MHL-NN model. And inputting the data in the test set into the trained model to obtain the corresponding sea surface height, and comparing the sea surface height with the DTU18 sea surface height value after the tide correction to calculate the main performance index.

Fig. 9 (a) shows a training and verification error graph of the novel MHL-NN inversion model, and it can be seen that, in the initial stage of training, with the increase of the number of iterations (Epoch), the training and verification errors both decrease significantly, the rate of decrease gradually slows down after about 100 iterations, and the training and verification errors both tend to stabilize gradually after 400 iterations. Meanwhile, after the error function is stable, the RMSE on the verification set is obviously higher than that of the training set. This is mainly due to the fact that the model can be trained using all samples on the training set during the training process, while the samples on the validation set are only used as feedback to adjust the parameters.

FIG. 9 (b) shows a scatter density plot of the new MHL-NN inversion results. It can be seen visually that the inversion result of the novel MHL-NN model and the DTU18 model has strong correlation, and the PCC is 0.98. However, because the main purpose of the TDS-1 satellite data is not sea surface height inversion, the satellite receiver does not optimize the height inversion, the inversion error is large, the MAD is about 4.23m, and the RMSE is about 5.94m. The main causes of errors are: (1) For the sea surface height inversion task, TDS-1L1B level data is influenced by limited DDM delay resolution (224.3942 ns), low receiver bandwidth (2 MHz) and low peak antenna gain (13.3 dB), and the accuracy is poor. (2) Because TDS-1 can only acquire L1 single-frequency measurement signals, atmospheric delay errors (ionospheric delay, tropospheric delay) cannot be directly eliminated by means of dual-frequency (L1 and L2) measurements, but can only be attenuated by establishing corresponding tropospheric and ionospheric models. (3) Some abnormal DDM data (residual sea ice data, DDM data containing abnormal noise and signals, etc.) existing in the data and not removed also affect the sea surface height inversion result. The above problem leads to an error of 3.5-7 m in the inversion results as a whole.

FIG. 10 (a) shows the Probability Density Function (PDF) of the sea height of the novel MHL-NN model and the DTU18 verification model, and it can be seen that the data distribution of the sea height inverted by the novel MHL-NN model is generally consistent with that of the DTU18 verification model. Meanwhile, the probability that the sea surface height inverted by the novel MHL-NN model is between-10 m to 0m, -60 m to-20 m and 45 m to 90m is obviously lower than that of the DTU18 verification model, which is mainly caused by uneven data distribution of different sea surface height values in a training data set. FIG. 10 (b) shows an error distribution histogram of the inversion result of the novel MHL-NN model relative to the sea height of the DTU18 verification model, and it can be seen that the statistical result is close to a normal distribution, wherein the inversion error of 69.78% is between-5 m and 5m, and the inversion error of 92.13% is between-10 m and 10 m.

Based on the above embodiments, the following description is provided for the specific application of the novel MHL-NN model

In order to verify the superiority of the novel MHL-NN model relative to the traditional satellite-borne GNSS-R sea surface inversion method, the inversion accuracy of the two methods is compared. In the traditional sea surface height inversion, the best-precision HALF characteristic point method (a point where 70% of energy value of the leading edge of the normalized waveform peak is calculated is a re-tracking point) in the single-point re-tracking method is selected by the re-tracking method, so that the time delay difference of a reflected signal relative to a direct signal is estimated, and various errors in time delay measurement are corrected.

11 (a), 11 (b) show the results of the global sea height inversion of the HALF conventional sea height inversion method and the novel MHL-NN model on the partitioned test sets, respectively; fig. 11 (c) shows the sea level results of the corresponding DTU18 verification model. It can be seen from fig. 11 that both models have good inversion results, and the obtained sea heights are consistent with the sea heights of the verification models in the global range. Meanwhile, as can be seen from the position of the rectangular frame in fig. 11, compared with the traditional inversion method of the sea surface height of the HALF, the novel MHL-NN model obviously has a better inversion result and is more consistent with the sea surface height of the DTU18 verification model.

FIG. 12 shows the error statistics of the global sea height inverted by the HALF re-tracking method and the new MHL-NN model relative to the DTU18 validation model. It can be seen that the inversion result error of the novel MHL-NN model is smaller, about 92.13% of the inversion error is between-10 and 10m, and about 67.63% of the inversion error of the HALF re-tracking method is between-10 and 10 m.

In conclusion, the invention discloses a method for improving the GNSS-R height finding inversion accuracy based on a multi-hidden-layer neural network,

firstly, in order to make up for the defects of the traditional method, a novel multi-hidden-layer neural network model MHL-NN which takes information such as Delay-Doppler Map (DDM) data, metadata, atmospheric Delay, sea surface wind speed and the like of a TechDemosat-1 (TDS-1) satellite as input and sea surface height as output is constructed based on an artificial neural network algorithm, and the sensitivity of the inversion performance of the novel MHL-NN model to different input characteristics is analyzed. The elevation angle of the specular reflection point, the signal-to-noise ratio of the DDM, the atmospheric delay and the sea surface wind speed can provide important contributions to sea surface height inversion. Second, the reliability of the new MHL-NN model was verified using the tide corrected global average sea surface height model DTU 18. Experimental results show that the novel MHL-NN model has excellent inversion performance, the average absolute error of the model is 4.23m compared with that of a DTU18 verification model, the correlation coefficient is 0.98, and the inversion accuracy is obviously superior to the TDS-1 sea surface height inversion accuracy (about 6.4 m) in the existing literature. Thirdly, the inversion accuracy of the novel MHL-NN model and the traditional single-point re-tracking method is analyzed in a comparison mode. Experimental results show that the novel MHL-NN model can effectively improve the inversion accuracy of the sea surface height, and the accuracy is improved by about 32.86%. The novel MHL-NN model provides new theory and method reference for future GNSS-R satellite sea surface highly accurate inversion.

Although the present invention has been described with reference to the preferred embodiments, it is not intended to limit the present invention, and those skilled in the art can make variations and modifications of the present invention without departing from the spirit and scope of the present invention by using the methods and technical contents disclosed above.

Those skilled in the art will appreciate that the invention may be practiced without these specific details.

Claims

1. A method for improving GNSS-R height finding inversion accuracy based on a multi-hidden layer neural network is characterized by comprising the following steps:

acquiring multi-source data; wherein the multi-source data includes: TDS-1 data, ionized layer data and EAR5 sea surface wind speed data; TDS-1 data is L1B-level data provided by a TDS-1 project; the ionosphere data is a global ionosphere map provided by the IGS; the EAR5 sea surface wind speed data is an EAR5 analysis field data set of the ECMWF;

extracting the characteristics of the multi-source data to obtain a characteristic parameter set;

2. The method for improving the GNSS-R altimetry inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 1, wherein the step of performing feature extraction on the multi-source data to obtain a feature parameter set comprises:

extracting a DDM image and corresponding metadata from the L1B-level data;

3. The method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 2, wherein the step of analyzing and processing the DDM image without the substrate noise to obtain an integrated time delay waveform IDW comprises:

obtaining an integral delay waveform IDW by incoherent integration calculation of a delay waveform in a set Doppler frequency shift range in the DDM image after substrate noise removal:

wherein, the first and the second end of the pipe are connected with each other,

indicates the set Doppler shift range, τ _j Which is representative of the time-delayed waveform,

representing the DDM image after removing substrate noise, M represents the imageFixed Doppler shift range

The number of Doppler points in the Doppler signal.

4. The method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network according to claim 2, wherein the step of analyzing and processing the DDM image without the substrate noise to obtain a pre-tilt edge slope LES of the time delay correlation curve comprises the following steps:

5. The method for improving the GNSS-R altimetry inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 2, wherein the step of analyzing and processing the DDM image without the substrate noise to obtain the time delay-Doppler mean value DDMA comprises:

wherein the content of the first and second substances,

6. The method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 2, wherein the step of analyzing and processing the DDM image without the substrate noise to obtain the time delay waveform peak value PCP comprises the following steps:

Wherein W (tau) represents the related power delay waveform of the reflected signal, tau represents the time delay;

7. The method for improving GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 2, wherein analyzing and processing the ionosphere data to obtain the total atmospheric delay ATM comprises:

respectively calculating to obtain the ionospheric delay delta from GPS satellite to mirror reflection point based on the global ionospheric map provided by IGS ₁ Ionospheric delay delta from specular reflection point to TDS-1 ₂ And ionospheric delay delta from GPS satellite to TDS-1 receiver ₃ ；

δ _iono ＝(δ ₁ +δ ₂ )-δ ₃

DeterminingTotal tropospheric delay delta of reflected signals caused by tropospheric layers relative to direct signals _tro ：

δ _tro ＝(δ _tro1 +δ _tro2 )

Then, the total atmospheric delay ATM is:

ATM＝δ _iono +δ _tro

8. The method for improving the GNSS-R height finding inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 2, wherein the step of analyzing and processing EAR5 sea surface wind speed data to obtain the sea surface wind speed comprises the following steps:

acquiring a TDS-1 satellite-borne data set;

9. The method for improving the GNSS-R altimetry inversion accuracy based on the multi-hidden-layer neural network as claimed in claim 1, wherein the constructing of the sea level height prediction model based on the multi-hidden-layer neural network comprises:

constructing a multi-hidden-layer neural network model;

training and verifying a multi-hidden-layer neural network model by using a 5-fold cross verification method based on the determined training set, and controlling a mode of dividing the training set and the test set each time by using random seed numbers to obtain optimal parameters;

10. The method for improving the accuracy of GNSS-R altimetry inversion based on the multi-hidden-layer neural network as claimed in claim 9, wherein the output of the sea-level altitude prediction model constructed based on the multi-hidden-layer neural network is represented as follows:

h ₁ ＝σ(W ₁ Feature_set+b ₁ )

h _i ＝σ(W _i h _i-1 +b _i )