Global sea surface salinity inversion method fusing heterogeneous multi-mode satellite-borne BDS-3 / Galileo reflection signal data in complex marine environment

Abstract

The invention relates to a global sea surface salinity inversion method fusing heterogeneous multi-mode satellite-borne BDS-3 / Galileo reflection signal data in a complex marine environment. The method comprises the following steps: acquiring various marine environment related data; preprocessing the acquired marine environment related data; using the preprocessed data to train a network model fusing a convolutional neural network and a Transform, and obtaining a sea surface salinity inversion model; and performing global sea surface salinity inversion by using the sea surface salinity inversion model. According to the method, the sea surface salinity can be efficiently and accurately monitored in a global range by combining a satellite-borne GNSS-R technology, multi-modal data fusion and an advanced deep learning model.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of artificial intelligence and satellite navigation reflection ocean remote sensing, and in particular to a global sea surface salinity inversion method integrating heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment. Background Art

[0002] Sea surface salinity (SSS) is an important parameter to measure the salt concentration of ocean surface water, and has a profound impact on the global water cycle, climate change, marine ecological environment and other fields. With the intensification of climate change, the monitoring of the spatial and temporal distribution of sea surface salinity has become particularly important. Traditional sea surface salinity monitoring methods mostly rely on ocean buoys or ship observations, but these methods have limitations in global monitoring density and timeliness.

[0003] L-band microwave radiometers have great potential in ocean salinity inversion, especially in global ocean salinity monitoring, and their advantages of all-weather and global coverage cannot be ignored. However, due to the complexity of the marine environment, the inversion accuracy is affected by many factors, such as sea surface fluctuations, temperature changes, etc. Therefore, a single L-band microwave radiometer inversion is difficult to provide very accurate sea surface salinity data. In addition, the sea surface salinity inversion method of remote sensing satellites usually cannot directly obtain salinity, but instead measures the sea surface brightness temperature, performs a series of atmospheric and sea condition corrections, and finally estimates the salinity. The correction process needs to consider the influence of multiple ocean and meteorological factors (such as sea water temperature, sea surface wind speed, waves, etc.), which affect the intensity and characteristics of microwave radiation. The method of sea surface salinity inversion is usually indirect and requires a combination of multiple data sources (such as sea surface temperature, wind speed, etc.) and complex models and algorithms to complete.

[0004] In recent years, spaceborne remote sensing technology, especially spaceborne GNSS reflectometry (GNSS-R) technology, has provided a new solution for sea surface salinity inversion. Spaceborne GNSS-R is a technology that uses the interaction between signals transmitted by global navigation satellite systems (such as GPS, BDS, Galileo, etc.) and surface reflection signals. By receiving the reflection signal, information about the surface can be obtained. For the marine environment, the characteristics of the reflection signal are closely related to sea surface salinity, ocean surface fluctuations, wind speed, etc. GNSS-R technology has the following advantages: All-weather observation: GNSS signals can penetrate clouds and precipitation, are not affected by weather and lighting conditions, and can be observed in any weather conditions. High temporal resolution: The orbit and signal update frequency of the GNSS system is high, which can provide data with high temporal resolution, which is helpful for real-time monitoring of changes in sea surface salinity. Low cost: The reception cost of GNSS signals is low, and global coverage can be achieved through existing GNSS satellites, without the need for additional dedicated satellite launches.

[0005] At present, the inversion of sea surface salinity by satellite-based GNSS-R is mainly carried out on the GPS-R data of the CYGNSS mission. However, CYGNSS can only measure sea surface salinity in the area of ​​40° north and south latitude around the world, and cannot invert sea surface salinity in high latitudes. In addition, rainfall, as a common natural phenomenon on the sea surface, can not only reduce seawater salinity, but also cause changes in sea surface roughness. Existing studies have not taken this into account, however, taking rainfall correction into account is crucial under low wind speed conditions (less than 15m / s). The reasons are as follows: 1) The impact of rainfall on roughness: rainfall can cause changes in sea surface roughness, which mainly affects the scattering characteristics of the sea surface. At low wind speeds (usually less than 15m / s), the changes in sea surface roughness caused by rainfall are more significant, because when the wind speed is low, the fluctuations of the sea surface itself are small, and the impact of rainfall on the tiny roughness of the sea surface is easier to observe. 2) The impact under high wind speeds: Under high wind speed conditions, the sea surface roughness itself is large, and the effect of wind waves dominates the sea surface scattering. At this time, the effect of rainfall on roughness is relatively small, so the response of the GNSS-R signal is also weak, and the effect of rainfall on the inverted salinity is relatively low. 3) GNSS-R response capability: GNSS-R is more sensitive to changes in sea surface roughness under low wind speed conditions, while its response capability is limited under high wind speed conditions. Therefore, introducing rainfall correction under low wind speed can more effectively improve the inversion accuracy, especially when the scattering changes caused by rainfall are more obvious. Therefore, in order to improve the accuracy of sea surface salinity inversion, especially in low wind speed environments, rainfall correction is particularly important. In addition, sea surface currents can also cause changes in sea surface roughness, but so far, existing literature has not considered the impact of sea surface currents in sea surface salinity inversion. Therefore, the present invention takes into account its impact on sea surface salinity inversion.

[0006] The Tianmu-1 occultation meteorological detection constellation is China's first low-orbit meteorological satellite system built in a commercial mode. It is operated by Tianmu (Chongqing) Satellite Technology Co., Ltd., a subsidiary of China Aerospace Science and Industry Corporation. There are currently 23 satellites in orbit, mainly equipped with GNSS occultation and sea anti-detection payloads, supporting the four major navigation systems of BDS, GPS, Galileo, GLONASS and Japan's Quasi-Zero-Earth Satellite System (QZSS), and can obtain global distribution and all-weather atmospheric, surface and sea surface characteristic parameters. China launched Tianmu-1 BDS-3 (Beidou-3), as China's global satellite navigation system, BDS-3 provides high-precision positioning, navigation and timing services. The signal characteristics of BDS-3 are suitable for receiving reflection data, with high temporal resolution and strong signal strength. Galileo has a good coverage range. Similar to BDS-3, Galileo signals are also suitable for receiving reflection data in GNSS-R technology. The signals of these two systems complement each other and can provide more intensive and accurate data in different geographical areas and times, improving the accuracy of sea surface salinity inversion. However, to date, there is no literature reporting the results of inverting sea surface salinity using BDS-3 or Galileo reflection signals.

[0007] Commonly used methods for inverting sea surface salinity from space-borne GNSS-R include empirical models and machine learning models. The empirical model inverts sea surface salinity through empirical formulas based on physical models. This type of method is simple, but is greatly affected by the local environment, and is challenging to solve the complex nonlinear modeling problems of sea surface salinity inversion in complex marine environments. Machine learning models, especially artificial neural network (ANN) models, have been applied in CYGNSS GPS-R sea surface salinity inversion, which can solve the nonlinear modeling problems in sea surface salinity inversion and have better inversion accuracy than empirical models. However, there are still great challenges in dealing with complex tasks and large-scale data, mainly manifested as:

[0008] (1) Limited feature extraction capability: ANNs usually rely on manually extracted features or pre-designed features for learning. This makes it difficult for ANNs to automatically extract effective features from raw data when processing complex patterns and high-dimensional data, resulting in relatively weak feature extraction capabilities of ANNs.

[0009] (2) Poor learning ability: In complex ocean environments, when performing sea surface salinity inversion, the representation ability of ANN is limited by the depth and width of the network. In particular, when it is necessary to capture complex nonlinear relationships, the performance of ANN will be restricted. In addition, it is easy to fall into problems such as gradient vanishing or gradient exploding, making the optimization process more difficult.

[0010] (3) Poor generalization ability: When faced with sea surface salinity inversion in complex marine environments, traditional ANNs are prone to overfitting, especially when there are few training samples. Its generalization ability usually depends on feature selection and regularization methods, but in many cases, ANNs still find it difficult to achieve sufficiently good generalization effects.

[0011] (4) Poor processing capability for spatial structure: For image data (such as bistatic radar cross section (BRCS), effective scattering area (Ddm_effective_area), power DDM), ANN does not have the advantage of processing spatial features like CNN. Summary of the invention

[0012] In order to overcome the problems faced by the existing ANN model in inverting GNSS-R sea surface salinity, such as limited feature extraction capability, poor learning capability, poor generalization capability, and poor processing capability of spatial structure; the current situation that it only relies on GPS reflection signals and can only measure sea surface salinity in the area of ​​40° north and south latitude around the world, and cannot invert sea surface salinity in high latitude areas, the present invention provides a global sea surface salinity inversion method that fuses heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment. The method combines the features of multiple data sources and performs effective feature extraction and fusion through the deep learning modules of CNN and Transformer. The multi-modal data fusion method can capture more relevant information, thereby improving the prediction capability of the sea surface salinity inversion model.

[0013] To achieve the above object, the present invention provides the following solutions:

[0014] The global sea surface salinity inversion method under complex ocean environment is based on the integration of heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data, including:

[0015] Obtain various marine environment related data;

[0016] Pre-process the acquired marine environment related data;

[0017] Using the preprocessed data, the network model integrating convolutional neural network and Transformer is trained to obtain the sea surface salinity inversion model;

[0018] The global sea surface salinity inversion model is used to invert the sea surface salinity.

[0019] Optionally, various marine environment related data are obtained including:

[0020] Obtain L1 level data of BDS-3 / Galileo reflected signals onboard Tianmu-1;

[0021] Obtain ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data;

[0022] Obtain OSCAR ocean surface current data;

[0023] Get SMAP sea surface brightness temperature and salinity data.

[0024] Optionally, the L1 level data of the BDS-3 / Galileo reflected signal onboard the Tianmu-1 satellite includes: image-related variables, normalized bistatic radar scattering cross section, leading edge slope, peak signal-to-noise ratio, signal-to-noise ratio at the mirror point and normalized signal-to-noise ratio data at the mirror point, and metadata variables that have an impact on sea surface salinity inversion;

[0025] The image-related variables include: bistatic radar cross section, effective scattering area and power DDM;

[0026] The metadata variables that affect the inversion of sea surface salinity include: UTC time corresponding to the intermediate moment of DDM collection, trajectory number of DDM data, X-component, Y-component and Z-component of GNSS satellite velocity, GNSS satellite PRN code, GNSS satellite batch code, altitude, longitude and latitude, roll angle and X-component, Y-component and Z-component of velocity of GNSS-R low-orbit satellite, altitude of mirror reflection point, receiver antenna gain of mirror reflection point, azimuth of mirror reflection point in antenna coordinate system, azimuth of mirror reflection point in satellite body coordinate system, azimuth of mirror reflection point in orbit coordinate system, azimuth of mirror reflection point in pattern coordinate system, incident angle of GNSS signal at mirror reflection point, longitude and latitude of mirror reflection point, altitude angle of mirror reflection point in antenna coordinate system, altitude angle of mirror reflection point in satellite body coordinate system, altitude angle of mirror reflection point in orbit coordinate system, altitude angle of mirror reflection point in pattern coordinate system and X-component, Y-component and Z-component of velocity of mirror reflection point.

[0027] Optionally, preprocessing the acquired marine environment related data includes:

[0028] The GNSS satellite batch code Gnss_block_flag variable in the Tianmu-1 GNSS-R L1 data is used to filter the BDS-3 and Galileo reflection signal data respectively;

[0029] The longitude and latitude of the specular reflection point in the BDS-3 and Galileo reflection signal data and the corresponding UTC time are used to perform spatiotemporal matching with the ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and sea surface salinity data;

[0030] Use the DDM quality identifier variable in the Tianmu-1 GNSS-R L1 data observations to filter out low-quality BDS-3 and Galileo reflection signal data.

[0031] Optionally, the network model integrating the convolutional neural network and the Transformer includes: a CNN module, a Transformer module, a fully connected layer and a fusion module, and an output module;

[0032] The CNN module is used to extract image features from input data;

[0033] The Transformer module is used to capture long-term dependencies in image features using a self-attention mechanism;

[0034] The fusion module is used to fuse and multiply the features that capture the long-term dependencies;

[0035] The output module is used to map and output the predicted value of sea surface salinity based on the multiplied features.

[0036] Optionally, training a network model integrating a convolutional neural network and a Transformer includes:

[0037] The BDS-3 / Galileo reflected signal L1 data onboard the Tianmu-1 satellite, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data and SMAP sea surface brightness temperature ocean environment auxiliary variable parameters are used as inputs of the model;

[0038] The SMAP sea surface salinity value is used as the target value for training the model;

[0039] The model is trained using the input and SMAP sea surface salinity target values. The model uses mean square error as the loss function and is optimized using the Adam optimizer; the optimization goal is to minimize the error between the predicted value and the target value.

[0040] Optionally, the fusion module includes: a fusion unit, a fully connected layer and a multiplication unit;

[0041] Using the sea surface salinity inversion model, inverting the global sea surface salinity includes:

[0042] Extracting first feature information from a bistatic radar cross section using the CNN module and the Transformer module;

[0043] Using the CNN module and the Transformer module, extracting second feature information from the effective scattering area;

[0044] Extracting third feature information from the power DDM using the CNN module and the Transformer module;

[0045] Using the fusion unit to fuse the first feature information, the second feature information and the third feature information;

[0046] Using fully connected layers, the fused features are further processed to capture and transmit complex patterns;

[0047] Using fully connected layers, complex nonlinear relationships between auxiliary variable parameters and sea surface salinity are constructed;

[0048] The complex nonlinear relationship is multiplied with the features further processed using a fully connected layer.

[0049] The beneficial effects of the present invention are:

[0050] The present invention first obtains a variety of marine environment-related data; secondly, preprocesses the obtained marine environment-related data; then, uses the preprocessed data to train a network model integrating a convolutional neural network and a Transformer to obtain a sea surface salinity inversion model; and finally, uses the sea surface salinity inversion model to perform global sea surface salinity inversion. The present invention provides an advanced deep learning sea surface salinity inversion model integrating CNN and Transformer, which effectively overcomes the influence of more potential influencing factors on the performance of satellite-borne GNSS-R sea surface salinity inversion, and for the first time realizes the inversion of high-precision and high-temporal-spatial resolution global sea surface salinity by integrating heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in complex marine environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0052] Figure 1 A schematic flow chart of a method for inverting global sea surface salinity by fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to an embodiment of the present invention;

[0053] Figure 2 This is a structural diagram of a sea surface salinity inversion model that integrates heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to an embodiment of the present invention;

[0054] Figure 3It is a scatter density map comparing the sea surface salinity inverted by the ANN model and the CNN-Transformer fusion model based on the BDS-3 reflection signal data and the SMAP data in an embodiment of the present invention;

[0055] Figure 4 A scatter density map comparing the sea surface salinity inverted by the ANN model and the CNN-Transformer fusion model based on the Galileo reflection signal data and the SMAP data according to an embodiment of the present invention;

[0056] Figure 5 The global sea surface salinity distribution of SMAP paired with BDS-3 reflection signal according to an embodiment of the present invention and the global deviation distribution between the sea surface salinity inverted by ANN model and CNN-Transformer fusion model based on BDS-3 reflection signal data and SMAP data; wherein (a) is the SMAP sea surface salinity paired with BDS-3 reflection signal / (psu), (b) is the sea surface salinity deviation inverted by BDS-3 reflection signal of ANN / (psu), and (c) is the sea surface salinity deviation inverted by BDS-3 reflection signal of CNN-Transformer / (psu);

[0057] Figure 6 The global sea surface salinity distribution of SMAP paired with Galileo reflection signal and the global deviation distribution between the sea surface salinity inverted by ANN model and CNN-Transformer fusion model based on Galileo reflection signal data and SMAP data according to an embodiment of the present invention; wherein, (a) is the SMAP sea surface salinity paired with Galileo reflection signal / (psu), (b) is the sea surface salinity deviation / (psu) inverted by Galileo reflection signal of ANN, and (c) is the sea surface salinity deviation / (psu) inverted by Galileo reflection signal of CNN-Transformer. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] This embodiment proposes a global sea surface salinity inversion method that integrates heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment, including:

[0061] Obtain various marine environment related data;

[0062] Pre-process the acquired marine environment related data;

[0063] Using the preprocessed data, the network model integrating convolutional neural network and Transformer is trained to obtain the sea surface salinity inversion model;

[0064] The global sea surface salinity inversion model is used to invert the sea surface salinity.

[0065] Furthermore, a variety of marine environment related data are obtained, including:

[0066] Obtain L1 level data of BDS-3 / Galileo reflected signals onboard Tianmu-1;

[0067] Obtain ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data;

[0068] Obtain OSCAR ocean surface current data;

[0069] Get SMAP sea surface brightness temperature and salinity data.

[0070] Further, the L1-level data of the BDS-3 / Galileo reflected signal onboard the Tianmu-1 satellite includes: image-related variables, normalized bistatic radar cross section, leading edge slope, peak signal-to-noise ratio, signal-to-noise ratio at the mirror point and normalized signal-to-noise ratio data at the mirror point, and metadata variables that have an impact on the inversion of sea surface salinity;

[0071] The image-related variables include: bistatic radar cross section, effective scattering area and power DDM;

[0072] The metadata variables that affect the inversion of sea surface salinity include: UTC time corresponding to the intermediate moment of DDM collection, trajectory number of DDM data, X-component, Y-component and Z-component of GNSS satellite velocity, GNSS satellite PRN code, GNSS satellite batch code, altitude, longitude and latitude, roll angle and X-component, Y-component and Z-component of velocity of GNSS-R low-orbit satellite, altitude of mirror reflection point, receiver antenna gain of mirror reflection point, azimuth of mirror reflection point in antenna coordinate system, azimuth of mirror reflection point in satellite body coordinate system, azimuth of mirror reflection point in orbit coordinate system, azimuth of mirror reflection point in pattern coordinate system, incident angle of GNSS signal at mirror reflection point, longitude and latitude of mirror reflection point, altitude angle of mirror reflection point in antenna coordinate system, altitude angle of mirror reflection point in satellite body coordinate system, altitude angle of mirror reflection point in orbit coordinate system, altitude angle of mirror reflection point in pattern coordinate system and X-component, Y-component and Z-component of velocity of mirror reflection point.

[0073] Furthermore, the acquired marine environment related data is preprocessed including:

[0074] The GNSS satellite batch code Gnss_block_flag variable in the Tianmu-1 GNSS-R L1 data is used to filter the BDS-3 and Galileo reflection signal data respectively;

[0075] The longitude and latitude of the specular reflection point in the BDS-3 and Galileo reflection signal data and the corresponding UTC time are used to perform spatiotemporal matching with the ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and sea surface salinity data;

[0076] Use the DDM quality identifier variable in the Tianmu-1 GNSS-R L1 data observations to filter out low-quality BDS-3 / Galileo reflection signal data.

[0077] Furthermore, the network model integrating the convolutional neural network and the Transformer includes: a CNN module, a Transformer module, a fully connected layer and a fusion module, and an output module;

[0078] The CNN module is used to extract image features from input data;

[0079] The Transformer module is used to capture long-term dependencies in image features using a self-attention mechanism;

[0080] The fusion module is used to fuse and multiply the features that capture the long-term dependencies;

[0081] The output module is used to map and output the predicted value of sea surface salinity based on the multiplied features.

[0082] Furthermore, training the network model integrating the convolutional neural network and the Transformer includes:

[0083] The bistatic radar cross section, effective scattering area, power DDM, characteristic variables calculated from BRCS and raw count DDM (Ddm_raw_data), GNSS-R receiver-related and geometry-related variables, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data and SMAP sea surface brightness temperature ocean environment auxiliary variables are used as inputs to the model;

[0084] The SMAP sea surface salinity value is used as the target value for training the model;

[0085] The model is trained using the input and SMAP sea surface salinity target values. The model uses mean square error as the loss function and is optimized using the Adam optimizer; the optimization goal is to minimize the error between the predicted value and the target value.

[0086] Furthermore, the fusion module includes: a fusion unit, a fully connected layer and a multiplication unit;

[0087] Using the sea surface salinity inversion model, inverting the global sea surface salinity includes:

[0088] Extracting first feature information from a bistatic radar cross section using the CNN module and the Transformer module;

[0089] Using the CNN module and the Transformer module, extracting second feature information from the effective scattering area;

[0090] Extracting third feature information from the power DDM using the CNN module and the Transformer module;

[0091] Using the fusion unit to fuse the first feature information, the second feature information and the third feature information;

[0092] Using fully connected layers, the fused features are further processed to capture and transmit complex patterns;

[0093] Using fully connected layers, complex nonlinear relationships between auxiliary variable parameters and sea surface salinity are constructed;

[0094] The complex nonlinear relationship is multiplied with the features further processed using a fully connected layer.

[0095] In order to verify the accuracy and reliability of the method proposed in this embodiment, the L1-level BDS-3 and Galileo reflection signal observation data of Tianmu-1 covering March to July 2024, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and salinity data were collected. Among them, the L1-level data of BDS-3 and Galileo reflection signals of Tianmu-1 are provided by Aerospace Tianmu (Chongqing) Satellite Technology Co., Ltd. and Aerospace Science and Industry (Beijing) Space Information Application Co., Ltd.

[0096] like Figure 1 As shown in FIG. 1 , the global sea surface salinity inversion method using heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in complex ocean environments includes the following steps:

[0097] Step S1, obtaining the L1 level data of BDS-3 / Galileo reflected signals onboard Tianmu-1, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and salinity data;

[0098] Step S2, preprocessing the acquired data and performing quality control on the acquired data;

[0099] Step S3, dividing the preprocessed data set into a training set and a validation set, further performing spatiotemporal matching on the data set obtained after quality control, and then dividing it into a training set and a validation set according to a certain ratio;

[0100] Step S4, constructing a sea surface salinity inversion model integrating a convolutional neural network (CNN) and a Transformer, and training the model using a training set;

[0101] Step S5, using the SMAP sea surface salinity data as a reference, using the test data set to verify the accuracy of the sea surface salinity inversion model, and comparing this model with the sea surface salinity inversion results of the existing ANN model.

[0102] As an implementation of this embodiment, the L1 level data of the BDS-3 / Galileo reflected signal onboard the Tianmu-1 satellite in step S1 includes three categories. The first category is image-related variables, including bistatic radar cross section (BRCS), effective scattering area (Ddm_effective_area) and power DDM; the second category is the normalized bistatic radar cross section (Ddm_sp_nbrcs) and leading edge slope (Ddm_sp_les) calculated from BRCS, the peak signal-to-noise ratio (Ddm_peak_snr) and the signal-to-noise ratio at the mirror point (Ddm_sp_snr) calculated from the raw count DDM (Ddm_raw_data), and the mirror The third category is metadata variables that affect the inversion of sea surface salinity, including: the UTC time corresponding to the middle moment of DDM collection (Ddm_time_utc), the track number of DDM data (Ddm_track_id), the X component, Y component, and Z component of GNSS satellite velocity (Tx_vel_x, Tx_vel_y, and Tx_vel_z), the PRN code of GNSS satellite (Gnss_prn_code), the batch code of GNSS satellite (Gnss_block_flag), the altitude (Rx_alt), longitude, and latitude of GNSS-R low-orbit satellite (Rx_lon, Rx_lat), roll angle (Rx_roll) and velocity X, Y, and Z components (Rx_vel_x, Rx_vel_y, Rx_vel_z), the height of the mirror reflection point (Sp_alt), the receiver antenna gain of the mirror reflection point (Sp_antenna_gain), the azimuth of the mirror reflection point in the antenna coordinate system (Sp_az_antenna), the azimuth of the mirror reflection point in the satellite body coordinate system (Sp_az_body), the azimuth of the mirror reflection point in the orbital coordinate system (Sp_az_orbit), and the azimuth of the mirror reflection point in the pattern coordinate system (Sp_az_pattern) , the incident angle of the GNSS signal at the mirror reflection point (Sp_inc_angle), the longitude and latitude (Sp_lon, Sp_lat) of the mirror reflection point, the altitude angle of the mirror reflection point in the antenna coordinate system (Sp_theta_antenna), the altitude angle of the mirror reflection point in the satellite body coordinate system (Sp_theta_body), the altitude angle of the mirror reflection point in the orbital coordinate system (Sp_theta_orbit), the altitude angle of the mirror reflection point in the pattern coordinate system (Sp_theta_pattern) and the X component, Y component, and Z component of the velocity of the mirror reflection point (Sp_vel_x, Sp_vel_y, Sp_vel_z).

[0103] The ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, and SMAP sea surface brightness temperature are used as model auxiliary input parameters for sea surface salinity inversion modeling, aiming to consider the impact of multiple factors on sea surface roughness in a complex marine environment, thereby achieving high-precision inversion of sea surface salinity.

[0104] As an implementation of this embodiment, the step S2 of preprocessing the acquired data includes the following sub-steps:

[0105] Step S2.1, using the GNSS satellite batch code (Gnss_block_flag) variable in the Tianmu-1 GNSS-R L1 data, filter the BDS-3 and Galileo reflection signal data respectively. In Gnss_block_flag, for BDS satellites, the first number indicates the system type: 1 = BD-3S, 2 = BD-2, 3 = BD-3, and the second number indicates the orbit type: 1 = GEO, 2 = IGSO, 3 = MEO. For Galileo satellites: 10 = IOV, 20 = FOC.

[0106] Step S2.2, use the longitude and latitude (Sp_lon and Sp_lat) of the specular reflection point in the BDS-3 and Galileo reflection signal data and the corresponding UTC time (Ddm_time_utc) to perform spatiotemporal matching with the ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and sea surface salinity data, respectively.

[0107] Step S2.3, use the DDM quality flag variable in the Tianmu-1 GNSS-R L1 data observation data to filter low-quality reflection signal data. In addition, observations with poor overall quality in other external datasets are also discarded, marked by the quality flag in the data product.

[0108] As an implementation of this embodiment, step S3 divides the preprocessed data set into a training set and a validation set, wherein the training set and the validation set account for 60% and 40% of the total number of samples, respectively.

[0109] As an implementation manner of this embodiment, the input data for constructing the sea surface salinity inversion model integrating convolutional neural network (CNN) and Transformer described in step S4 includes: bistatic radar cross section (BRCS), effective scattering area (Ddm_effective_area), power DDM, characteristic variables calculated from BRCS and raw count DDM (Ddm_raw_data), GNSS-R receiver-related and geometry-related variable parameters, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data and SMAP sea surface brightness temperature ocean environment auxiliary variable parameters.

[0110] Furthermore, the sea surface salinity inversion model framework consists of four input lines. The first input line previously used convolutional neural networks and Transformer to extract key feature information from the bistatic radar cross section (BRCS), the third input line previously used convolutional neural networks and Transformer to extract key feature information from the effective scattering area (Ddm_effective_area), the fourth input line previously used convolutional neural networks and Transformer to extract key feature information from the power DDM, and the second input line used a fully connected network (FCN) to construct a complex nonlinear relationship between the auxiliary variable parameters and the sea surface salinity. As shown in the attached figure, Figure 2 shown.

[0111] CNN module: Applied to the input 2D data, its purpose is to extract image features. Each CNN module contains two convolutional layers (Conv2D), using the ReLU activation function, and reducing the dimension of the features through global average pooling (GlobalAveragePooling2D). The output of each CNN module is further processed by a fully connected layer (Dense) to extract high-level features.

[0112] Transformer module: The self-attention mechanism (Self-Attention) in the Transformer model is used to capture the long-term dependency relationship in the input data. Specifically, the TransformerBlock function defines a basic Transformer layer, which includes the following operations: Multi-Head Self-Attention (MultiHeadAttention): used to perform multiple different attention calculations on the input, thereby enhancing the model's ability to pay attention to different features. Residual Connection (Add): Add the input and the output of the self-attention layer to alleviate the gradient disappearance problem. Feed Forward Network (Feed Forward Network): After the Transformer layer, a fully connected layer is further added to increase the expressive power of the model. The Transformer module helps to capture the contextual information and long-term dependencies of sequence data.

[0113] Fully connected layers and fusion: The features from different inputs (first input, third input, fourth input) are fused through the Add layer. The features are then further processed through the fully connected layer (Dense) to capture and transmit complex patterns. The Multiply operation multiplies the vector of the second input variable processed by the fully connected method with the fused features. This represents the mutual influence between different data sources.

[0114] Output layer: Finally, the model is progressively mapped to a single output (sea surface salinity prediction) through a fully connected layer, using a linear activation function.

[0115] Model training: The model constructed using the training set is trained. The model uses mean square error (MSE) as the loss function and is optimized using the Adam optimizer. The optimization goal is to minimize the error between the predicted value and the target value.

[0116] The model combines the features of multiple data sources and performs effective feature extraction and fusion through the deep learning modules of CNN and Transformer. This multimodal data fusion method can capture more relevant information, thereby improving the predictive ability of the sea surface salinity inversion model.

[0117] As an implementation of this embodiment, the SMAP sea surface salinity data is used as a reference in step S5, and the accuracy of the sea surface salinity inversion model is verified using a test data set, and the sea surface salinity inversion results of this model and the existing ANN model are compared. There are four accuracy indicators selected, namely: root mean square error (RMSE), bias, mean absolute percentage error (MAPE) and correlation coefficient (CC):

[0118]

[0119] In the formula, m is the number of data samples, y i,M , is the sea surface salinity estimated by the model, y i,T is the sea surface salinity value from the reference dataset, y i,M and i,T The average value of .

[0120] In order to evaluate the performance of the method provided in this embodiment in inverting sea surface salinity in a complex marine environment, the SMAP sea surface salinity is used as reference data, and the inversion results of the existing ANN model are compared and verified with the inversion results of the CNN-Transformer fusion model proposed in the present invention. Table 1 shows the accuracy comparison of the inverted sea surface salinity of the ANN model and the CNN-Transformer fusion model proposed in the present invention. Figure 3 The scatter density maps of the sea surface salinity inverted by the ANN model and the CNN-Transformer fusion model based on the BDS-3 reflection signal data and the SMAP data are given. Figure 4 The scatter density maps of sea surface salinity inverted by ANN model and CNN-Transformer fusion model based on Galileo reflection signal data are compared with SMAP data.

[0121] Table 1 Inversion performance of sea surface salinity inverted by different models compared with SMAP data

[0122]

[0123] As can be seen from Table 1, the accuracy of inverting sea surface salinity using the CNN-Transformer fusion model based on BDS-3 and Galileo reflection signal data is better than that of the existing ANN model. Compared with the ANN model, the accuracy in RMSE is improved by 33.45% and 34.64% respectively. The CNN-Transformer fusion model proposed in the present invention is used to significantly improve the performance of sea surface salinity inversion. Figure 3 and attached Figure 4 It can also be seen that the correlation between the sea surface salinity inverted by the CNN-Transformer fusion model based on the BDS-3 and Galileo reflection signal data and the SMAP data is significantly higher than that of the ANN model, especially the correlation coefficient between the sea surface salinity inverted by the CNN-Transformer fusion model based on the BDS-3 reflection signal and the SMAP data is as high as 0.87.

[0124] The embodiment of the present invention further verifies the global performance of sea surface salinity inverted by ANN model and CNN-Transformer fusion model. Figure 5The global sea surface salinity distribution of SMAP paired with BDS-3 reflection signal and the global deviation distribution between sea surface salinity inverted from BDS-3 reflection signal data using ANN model and CNN-Transformer fusion model and SMAP data are given. Figure 6 The global sea surface salinity distribution of SMAP paired with Galileo reflection signal and the global deviation distribution between sea surface salinity inverted by ANN model and CNN-Transformer fusion model based on Galileo reflection signal data and SMAP data are given. Figure 5 and attached Figure 6 It can be clearly seen that the CNN-Transformer fusion model proposed in the present invention performs significantly better than the existing ANN model for sea surface salinity inversion globally and has a smaller global sea surface salinity inversion deviation.

[0125] Compared with ANN, the CNN-Transformer fusion model proposed in this embodiment has significant advantages in the following aspects for sea surface salinity inversion:

[0126] (1) Automatic feature extraction capability: CNN can effectively capture local features, while Transformer can capture global dependencies, enabling the fusion model to handle more complex tasks.

[0127] (2) Time series data modeling capability: Transformer can handle long-distance dependencies, which enhances the performance of the model on time series data.

[0128] (3) Learning ability and expressiveness: The CNN-Transformer fusion model has stronger learning ability and can extract rich features from data and adapt to complex tasks.

[0129] (3) Computational efficiency and training effect: The fusion model extracts local features through convolution and Transformer parallel computing to accelerate the training process and optimize the model efficiency.

[0130] Therefore, although ANN may still have certain advantages in simple problems and small-scale data, the performance and flexibility of the CNN-Transformer fusion model are significantly better than traditional ANN when faced with complex tasks and large-scale data.

[0131] In summary, this embodiment provides a global sea surface salinity inversion method that integrates heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment. By combining satellite-borne GNSS-R technology, multi-modal data fusion and advanced deep learning models, sea surface salinity can be monitored efficiently and accurately on a global scale.

[0132] In existing research work, CYGNSS observations can only be used to invert sea surface salinity in low and mid-latitude areas. Currently, there is not enough GNSS-R observation data to perform sea surface salinity inversion in high latitudes. This invention uses Tianmu-1 data for the first time to carry out global sea surface salinity inversion, and the technical solution of the invention can effectively solve this problem.

[0133] Existing studies have limited the inversion performance of sea surface salinity under high wind speed conditions (greater than 15m / s) using CYGNSS observations. The technical solution of the present invention can outperform the inversion performance of existing methods under higher wind speeds (less than 20m / s); in addition, rainfall, as a common natural phenomenon on the sea surface, can both reduce seawater salinity and cause changes in sea surface roughness. Existing studies have not taken this into account, however, it is crucial to take rainfall correction into account under low wind speed conditions (less than 15m / s). The reasons are as follows: 1) The impact of rainfall on roughness: rainfall can cause changes in sea surface roughness, which mainly affects the scattering characteristics of the sea surface. At low wind speeds (usually less than 15m / s), the changes in sea surface roughness caused by rainfall are more significant, because when the wind speed is low, the fluctuations of the sea surface itself are small, and the impact of rainfall on the tiny roughness of the sea surface is easier to observe. 2) Impact at high wind speeds: Under high wind speed conditions, the sea surface roughness itself is large, and the effect of wind waves dominates the sea surface scattering. At this time, the effect of rainfall on roughness is relatively small, so the response of the GNSS-R signal is also weak, and the effect of rainfall on the inverted salinity is low. 3) GNSS-R response capability: GNSS-R is more sensitive to changes in sea surface roughness under low wind speed conditions, while its response capability is limited under high wind speed conditions. Therefore, introducing rainfall correction under low wind speed can more effectively improve the inversion accuracy, especially when the scattering changes caused by rainfall are more obvious. Therefore, in order to improve the accuracy of sea surface salinity inversion, especially in low wind speed environments, rainfall correction is particularly important, and the use of the technical solution of the present invention can overcome the impact of sea surface rainfall on sea surface salinity inversion.

[0134] In summary, this embodiment provides an advanced deep learning sea surface salinity inversion model that integrates CNN and Transformer, which effectively overcomes the influence of more potential influencing factors on the performance of space-borne GNSS-R sea surface salinity inversion, and for the first time realizes the inversion of high-precision and high spatiotemporal resolution global sea surface salinity by integrating heterogeneous multi-modal space-borne BDS-3 / Galileo reflection signal data in complex ocean environments.

[0135] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A global sea surface salinity inversion method based on the integration of heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in complex ocean environments, characterized by: include: Obtain various marine environment related data; Pre-process the acquired marine environment related data; Using the preprocessed data, the network model integrating convolutional neural network and Transformer is trained to obtain the sea surface salinity inversion model; The global sea surface salinity inversion model is used to invert the sea surface salinity.

2. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 1 is characterized in that: Obtain a variety of marine environment related data including: Obtain L1 level data of BDS-3 / Galileo reflected signals onboard Tianmu-1; Obtain ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data; Obtain OSCAR ocean surface current data; Get SMAP sea surface brightness temperature and salinity data.

3. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 2 is characterized in that: The L1 level data of the BDS-3 / Galileo reflected signal carried by the Tianmu-1 satellite include: image-related variables, normalized bistatic radar cross section, front slope, peak signal-to-noise ratio, signal-to-noise ratio at the mirror point and normalized signal-to-noise ratio data at the mirror point, and metadata variables that affect the inversion of sea surface salinity; The image-related variables include: bistatic radar cross section, effective scattering area and power DDM; The metadata variables that affect the inversion of sea surface salinity include: UTC time corresponding to the intermediate moment of DDM collection, trajectory number of DDM data, X-component, Y-component and Z-component of GNSS satellite velocity, GNSS satellite PRN code, GNSS satellite batch code, altitude, longitude and latitude, roll angle and X-component, Y-component and Z-component of velocity of GNSS-R low-orbit satellite, altitude of mirror reflection point, receiver antenna gain of mirror reflection point, azimuth of mirror reflection point in antenna coordinate system, azimuth of mirror reflection point in satellite body coordinate system, azimuth of mirror reflection point in orbit coordinate system, azimuth of mirror reflection point in pattern coordinate system, incident angle of GNSS signal at mirror reflection point, longitude and latitude of mirror reflection point, altitude angle of mirror reflection point in antenna coordinate system, altitude angle of mirror reflection point in satellite body coordinate system, altitude angle of mirror reflection point in orbit coordinate system, altitude angle of mirror reflection point in pattern coordinate system and X-component, Y-component and Z-component of velocity of mirror reflection point.

4. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 1 is characterized in that: The preprocessing of the acquired marine environment related data includes: The GNSS satellite batch code Gnss_block_flag variable in the Tianmu-1 GNSS-R L1 data is used to filter the BDS-3 and Galileo reflection signal data respectively; The longitude and latitude of the specular reflection point in the BDS-3 and Galileo reflection signal data and the corresponding UTC time are used to perform spatiotemporal matching with the ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data, SMAP sea surface brightness temperature and sea surface salinity data; Use the DDM quality identifier variable in the Tianmu-1 GNSS-R L1 data observations to filter out low-quality BDS-3 and Galileo reflection signal data.

5. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 1 is characterized in that: The network model integrating convolutional neural network and Transformer includes: CNN module, Transformer module, fully connected layer and fusion module, and output module; The CNN module is used to extract image features from input data; The Transformer module is used to capture long-term dependencies in image features using a self-attention mechanism; The fusion module is used to fuse and multiply the features that capture the long-term dependencies; The output module is used to map and output the predicted value of sea surface salinity based on the multiplied features.

6. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 3 is characterized in that: Training the network model that integrates convolutional neural network and Transformer includes: The BDS-3 / Galileo reflected signal L1 data onboard the Tianmu-1 satellite, ERA5 sea surface wind speed and direction, significant wave height, rainfall and sea surface temperature data, OSCAR sea surface current data and SMAP sea surface brightness temperature ocean environment auxiliary variable parameters are used as inputs of the model; The SMAP sea surface salinity value is used as the target value for training the model; The model is trained using the input and SMAP sea surface salinity target values. The model uses mean square error as the loss function and is optimized using the Adam optimizer; the optimization goal is to minimize the error between the predicted value and the target value.

7. The global sea surface salinity inversion method of fusing heterogeneous multi-modal satellite-borne BDS-3 / Galileo reflection signal data in a complex ocean environment according to claim 5, characterized in that: The fusion module includes: a fusion unit, a fully connected layer and a multiplication unit; Using the sea surface salinity inversion model, inverting the global sea surface salinity includes: Extracting first feature information from a bistatic radar cross section using the CNN module and the Transformer module; Using the CNN module and the Transformer module, extracting second feature information from the effective scattering area; Extracting third feature information from the power DDM using the CNN module and the Transformer module; Using the fusion unit to fuse the first feature information, the second feature information and the third feature information; Using fully connected layers, the fused features are further processed to capture and transmit complex patterns; Using fully connected layers, complex nonlinear relationships between auxiliary variable parameters and sea surface salinity are constructed; The complex nonlinear relationship is multiplied with the features further processed by the fully connected layer, and finally a comprehensive feature reflecting the global feature fusion and nonlinear relationship is obtained, which is used for the inversion model to output the global sea surface salinity.

Images