Sea surface height retrieval method and system based on gated memory network and GNSS-IR

Abstract

The application discloses a sea surface height inversion method and system based on a gated memory network and GNSS-IR, and relates to the technical fields of ocean remote sensing and satellite navigation. The method comprises the following steps: acquiring GNSS-IR observation data, and extracting physical inversion process parameters therefrom; combining corresponding spatio-temporal matching meteorological parameters and measured sea surface height data of a tide station to construct a first multi-source feature data sequence; inputting the first multi-source feature data sequence into a gated memory network for training; the training process comprises grouping and independently pruning network weight gradients corresponding to different types of features in the first multi-source feature data sequence; and updating network weights related to meteorological parameters by using a differentiated learning rate strategy. GNSS-IR observation data to be inverted is constructed into a second multi-source feature data sequence, and the trained gated memory network is inputted, so that sea surface height inversion results are outputted. The second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

Description

Technical Field

[0001] This invention relates to the fields of marine remote sensing and satellite navigation technology, and in particular to a method and system for sea altitude inversion based on gated memory networks and GNSS-IR. Background Technology

[0002] Using GNSS-IR (Global Navigation Satellite System Interferometric Reflectometry) for sea height detection is a low-cost, wide-coverage, and important technical means with irreplaceable application value in fields such as marine environmental monitoring, tide forecasting, and sea level change research.

[0003] Currently, traditional GNSS-IR sea height inversion methods mainly rely on idealized physical models to calculate sea surface height by analyzing the interferometric phase relationship between direct and reflected signals. However, this type of method has significant limitations. Specifically, the real ocean and atmospheric environment is complex and variable. Error sources such as sea surface dynamics, complex atmospheric delays, and multipath effects are difficult to quantify accurately through physical models, directly leading to systematic biases and random noise at the centimeter to decimeter level in the inversion results. Furthermore, existing improvement methods mostly focus on fragmented corrections of single error sources, failing to systematically integrate multi-source information closely related to inversion quality, such as satellite geometry, signal quality, and meteorological conditions. They also lack effective modeling of the continuous temporal variation characteristics of sea surface height, resulting in limited accuracy improvement and insufficient generalization ability in complex scenarios such as weak reflection signals, extreme weather, and sea areas with different climates, making it difficult to meet the actual needs of high-precision sea height monitoring.

[0004] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] To address the limitations of traditional GNSS-IR sea height inversion methods, such as limited inversion accuracy, poor result stability, and insufficient environmental generalization ability, this invention proposes a sea height inversion method and system based on gated memory networks and GNSS-IR. This method integrates GNSS-IR physical inversion parameters with multi-source environmental data and leverages the temporal modeling capabilities and adaptive gradient management mechanism of gated memory networks to synergistically improve the accuracy, temporal stability, and environmental adaptability of sea height inversion.

[0006] The present invention achieves the above objectives through the following technical solutions:

[0007] Sea altitude inversion methods based on gated memory networks and GNSS-IR include:

[0008] Acquire GNSS-IR observation data and extract physical inversion process parameters from it;

[0009] Based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station, a first multi-source feature data sequence was constructed.

[0010] The first multi-source feature data sequence is input into the gated memory network for training; the training process includes: grouping and independently pruning the network weight gradients corresponding to different types of features in the first multi-source feature data sequence; and updating the network weights related to meteorological parameters using a differentiated learning rate strategy.

[0011] The GNSS-IR observation data to be inverted is constructed into a second multi-source feature data sequence and input into a trained gated memory network to output the sea height inversion result; the second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

[0012] As a preferred embodiment, the physical inversion process parameters include initial inversion sea height, peak frequency, spectral amplitude, peak noise, number of satellites, and satellite azimuth; the meteorological parameters include atmospheric temperature and atmospheric pressure; and the measured sea height data from the tide gauge station are observations calibrated to the tidal datum.

[0013] As a preferred embodiment, the acquisition of GNSS-IR observation data and the extraction of physical inversion process parameters from it includes:

[0014] Extract the signal-to-noise ratio (SNR) observations from the original GNSS observation data files;

[0015] Based on the satellite elevation angle, the signal-to-noise ratio observation sequence related to sea surface reflection was selected, and the satellite azimuth angle corresponding to each observation point and the number of visible satellites in the current time period were recorded.

[0016] Spectral analysis was performed on the selected signal-to-noise ratio observation sequence to extract the peak frequency and corresponding spectral amplitude, and the peak noise was calculated.

[0017] Based on the peak frequency, satellite elevation angle, and carrier wavelength, the initial inverted sea height is calculated using the following formula:

[0018]

[0019] In the formula, Peak frequency; The satellite's elevation angle; For GNSS carrier wavelength; This is the initial inversion of sea level.

[0020] As a preferred embodiment, the step of independently pruning the network weight gradients corresponding to different types of features in the first multi-source feature data sequence includes:

[0021] The first multi-source feature data sequence was divided into a core observation group, a satellite geometry group, and an atmospheric correction group.

[0022] During the backpropagation process of the network, the Frobenius norm is calculated for the input weight gradient submatrix corresponding to each group of features, and independent pruning is performed based on the preset group pruning threshold.

[0023] As a preferred embodiment, the core observation group includes initial inverted sea height, peak frequency, spectral amplitude, and peak noise; the satellite geometry group includes satellite azimuth and number of satellites; and the atmospheric correction group includes atmospheric temperature and atmospheric pressure.

[0024] As a preferred embodiment, the differentiated learning rate strategy includes:

[0025] In the later stages of training, the learning rate of the network weight parameters connected to the input dimensions of atmospheric temperature and atmospheric pressure is adjusted to 0.1 times the learning rate of other network weight parameters.

[0026] As a preferred embodiment, the gated memory network consists of two stacked LSTM layers, with each layer having a hidden state dimension of 64. Dropout layers are set between the layers, and the output layer is a fully connected layer. The loss function for training the gated memory network is the sum of the mean squared error and the smoothing constraint term. The smoothing constraint term is the L2 norm of the difference between adjacent time points of the predicted sea level sequence.

[0027] A sea altitude inversion system based on gated memory networks and GNSS-IR includes:

[0028] The data acquisition module is used to acquire GNSS-IR observation data and extract physical inversion process parameters from it;

[0029] The feature construction module is used to construct the first multi-source feature data sequence based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station.

[0030] A network training module is used to input the first multi-source feature data sequence into a gated memory network for training; wherein, the network training module includes:

[0031] The gradient pruning submodule is used to independently prune the network weight gradients corresponding to different types of features in the first multi-source feature data sequence.

[0032] The learning rate update submodule is used to update the network weights related to meteorological parameters using a differentiated learning rate strategy;

[0033] The network inversion module is used to construct a second multi-source feature data sequence from the GNSS-IR observation data to be inverted, and input it into the trained gated memory network to output the sea height inversion result; the second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

[0034] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement steps of a sea altitude inversion method applicable to gated memory networks and GNSS-IR.

[0035] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements steps applicable to a sea altitude inversion method based on gated memory networks and GNSS-IR.

[0036] Compared with existing technologies, the beneficial effects of this invention are as follows: By extracting parameters from the GNSS-IR physical inversion process, a solid physical prior and interpretability foundation is provided for the model of this invention. Based on this, these physical parameters are further fused with meteorological parameters and measured sea height data from tide gauge stations to construct a unified and standardized multi-source feature sequence. This design fundamentally changes the limitations of traditional methods that rely on a single physical model for deduction or fragmented error correction, enabling the model to simultaneously perceive multi-dimensional key information such as signal quality, satellite geometry, and atmospheric environment, laying a comprehensive and reliable data foundation for high-precision sea height inversion. A gated memory network is used as the temporal modeling engine, achieving a deep understanding of the complex dynamics of sea surface height. This network possesses powerful temporal dependency capture capabilities, effectively learning the long-term and complex nonlinear mapping relationship between multi-source features and sea height, and is particularly capable of accurately modeling the coupling effect of time-varying interference factors such as atmospheric delay with the sea height signal. This design not only inherits the interpretability framework of the physical model but also possesses fitting capabilities and generalization performance that surpass traditional models. This significantly suppresses random noise and systematic bias in the inversion results, solving core problems such as local jitter and extreme value tracking lag inherent in traditional methods. An innovative adaptive optimization strategy for multi-source heterogeneous features is introduced, providing crucial assurance for stable model training. Addressing the differences in stability and importance among different feature groups, a group gradient independent pruning strategy is adopted to avoid excessive suppression of key feature gradients or interference from secondary feature gradients in the optimization direction, ensuring balanced and stable updates of network parameters. Simultaneously, considering the large fluctuations in meteorological parameters and their complex correlation with sea level, a late-stage learning rate decay strategy is implemented for their relevant network weights, achieving precise calibration of complex atmospheric influences. This combined optimization strategy effectively solves the gradient conflict and model oscillation problems in multi-source heterogeneous data training, ensuring robust convergence even under fluctuating data quality or extreme environmental interference. This is the core reason why the robustness of this invention is significantly superior to traditional methods. The second multi-source feature data sequence used for actual sea height prediction in this invention adopts the exact same feature structure and construction logic as the first multi-source feature data sequence in the model training stage. This design fundamentally avoids additional systematic errors or model performance degradation introduced by inconsistent feature engineering processes, ensuring that the advantages of the trained model can be fully realized in real application scenarios, and guaranteeing the reliability and accuracy stability of the inversion results.

[0037] The above general description and the description below are exemplary and illustrative only and are not intended to limit this application. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 A flowchart of the sea altitude inversion method based on gated memory network and GNSS-IR provided by the present invention; Figure 2 This is a schematic diagram of the GNSS-IR sea surface altimeter signal propagation and interferometry principle in an embodiment of the present invention; Figure 3 This is a diagram showing the correspondence between the Fresnel reflection zone observed by multi-station GNSS and the actual scene in an embodiment of the present invention; Figure 4 This is a comparison chart of the sea height time series retrieved by the traditional GNSS-IR method and the measured tide level series at the tide gauge station in this embodiment of the invention; Figure 5 This is a scatter plot showing the correlation between the traditional GNSS-IR inversion results and the measured values ​​at the tide gauge station in this embodiment of the invention. Figure 6 This is a comparison chart of the sea height time series retrieved by the method (GMSIM) of the present invention and the measured tide level series at the tide gauge station in this embodiment of the invention; Figure 7 This is a comparison diagram of the inversion error probability density distribution between the method of the present invention (GMSIM) and the traditional GNSS-IR method in this embodiment of the invention; Figure 8 A schematic diagram of the modular structure of the sea altitude inversion system based on gated memory network and GNSS-IR provided by the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0040] like Figure 1 As shown, this is an embodiment of the present invention, which provides a sea altitude inversion method based on gated memory networks and GNSS-IR. The method includes:

[0041] S1: Acquire GNSS-IR observation data and extract physical inversion process parameters from it.

[0042] The physical inversion process parameters include initial inversion sea height, peak frequency, spectral amplitude, peak noise, number of satellites, and satellite azimuth.

[0043] The specific steps include:

[0044] Extract the signal-to-noise ratio (SNR) observations from the original GNSS observation data files.

[0045] Specifically, the raw GNSS observation data file is the RINEX format observation file generated by the GNSS station receiver, which contains the signal-to-noise ratio observations of all visible satellites at each epoch.

[0046] Based on the satellite elevation angle, the signal-to-noise ratio observation sequence related to sea surface reflection was selected, and the satellite azimuth angle corresponding to each observation point was recorded.

[0047] Specifically, the satellite elevation angle and azimuth angle need to be calculated using GNSS station coordinates and satellite orbit information. The specific process is as follows:

[0048] If the GNSS station is a reference point with known precise coordinates, its geodetic coordinates can be directly input; if the coordinates are unknown, the approximate coordinates of the station need to be calculated by combining the GNSS observation file and the satellite orbit information file and using the receiver single-point positioning method.

[0049] Based on satellite orbit information files and established station coordinates, a standard satellite position calculation algorithm is used to solve for the position vector of each satellite relative to the station at each epoch, thereby calculating the satellite's elevation and azimuth angles. The satellite orbit information files can be real-time available RINEX format satellite navigation files, or post-hoc precise ephemeris files provided by international GNSS services or other organizations; the latter provides higher-precision orbit data to improve calculation accuracy.

[0050] According to the principles of GNSS-IR, reflected signals dominate under low elevation angle conditions. Therefore, based on the calculated satellite elevation angles, signal-to-noise ratio (SNR) data segments with elevation angles between 5 and 30 degrees need to be selected. Subsequently, for each effective SNR data segment, a high-order polynomial fitting method is used to remove the direct signal trend term that varies with time, resulting in an SNR oscillation residual sequence mainly caused by sea surface reflection. This sequence is the SNR observation sequence.

[0051] Spectral analysis was performed on the selected signal-to-noise ratio observation sequence to extract the peak frequency and corresponding spectral amplitude, and the peak noise was calculated.

[0052] Specifically, the Lomb-Scargle spectral analysis method is used for spectral estimation, extracting the peak frequency and spectral amplitude of the main peak position, and obtaining the peak noise by calculating the statistical quantity of the spectral amplitude outside the preset bandwidth on both sides of the main peak.

[0053] Based on the peak frequency, satellite elevation angle, and carrier wavelength, the initial inverted sea height is calculated using the following formula:

[0054]

[0055] In the formula, Peak frequency, For the satellite elevation angle, For GNSS carrier wavelength, This is the initial inversion of sea level.

[0056] Simultaneously record the satellite azimuth angle corresponding to the data segment and the number of visible satellites within the time period represented by the data segment.

[0057] Finally, for each valid signal-to-noise ratio reflectance data segment, a set of structured physical inversion process parameters is generated, including: initial inversion sea height, peak frequency, spectral amplitude, peak noise, satellite azimuth angle, and number of satellites.

[0058] like Figure 2 As shown in the figure, this diagram visually illustrates the propagation path and interference process of the GNSS direct signal and the sea surface reflected signal, providing a core physical model support for the derivation of the initial sea height inversion calculation formula. The basic principle of sea surface altimetry is that the direct signal and the signal reflected from the sea surface interfere with each other, and the phase difference between the two signals is used to determine the propagation path and interference process. By combining the peak frequency f obtained from spectral analysis, the above initial sea height inversion formula can be finally derived.

[0059] S2, based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station, constructs the first multi-source feature data sequence.

[0060] Among them, meteorological parameters include atmospheric temperature and atmospheric pressure, and the measured sea height data of the tide gauge station are the observation values ​​after being calibrated with the tidal datum.

[0061] Specifically, meteorological parameters are taken from automatic observation data of meteorological stations near the GNSS station deployment location, or from spatial reanalysis data. To ensure data consistency, the temporal resolution of the raw meteorological data needs to be processed to match the GNSS observation epoch. The measured sea height data from the tide gauge station are derived from continuous tide level observation data from the tide gauge station, and the raw observation values ​​must be uniformly calibrated to the specified tidal datum.

[0062] Given the differences in data sources and sampling rates, spatiotemporal alignment is necessary. This spatiotemporal alignment includes temporal alignment and spatial alignment, as detailed below:

[0063] For time alignment, the core observation time corresponding to the physical inversion process parameters is used as the benchmark, and the meteorological parameters and measured sea height data from tide gauge stations that are closest to this benchmark in time are selected. If the data timestamps are not completely consistent, small time deviations can be handled by linear interpolation, and a maximum time tolerance is set. Samples exceeding this tolerance must be discarded. The setting of the maximum time tolerance needs to comprehensively consider the tidal change rate and the regular data sampling interval, allowing for reasonable processing delays while ensuring data timeliness. In a specific embodiment, the maximum time tolerance is set to 30 minutes.

[0064] For spatial alignment, to ensure the spatial representativeness of the sea surface height signal, if the GNSS station and the tide gauge station are not located at the same site, their distance must be within an acceptable range. This range is set based on the spatial correlation attenuation scale of short-scale sea surface height signals. Once this distance is exceeded, differences in local meteorological and hydrological conditions may lead to significant deviations in the sea surface height signal, thereby affecting the effectiveness of model learning and validation. Therefore, it is recommended to keep the distance within 10 kilometers. If meteorological station data is used, the station closest to the GNSS station should be selected first; if reanalysis data is used, data from the grid points where the GNSS station is located can be extracted directly.

[0065] The specific process of constructing the feature sequence is as follows:

[0066] The spatiotemporally aligned multi-source feature data are concatenated into a multi-source fusion feature vector in a preset fixed order. ,in, The total feature dimension. In a preferred embodiment, the total feature dimension is... Set to 8 dimensions, the multi-source fusion feature vector is composed of the following parameters in sequence: initial inversion sea height, peak frequency, spectral amplitude, peak noise, satellite azimuth, number of satellites, atmospheric temperature, and atmospheric pressure.

[0067] The measured sea height values ​​of the tide gauge stations after spatiotemporal alignment are used as labels.

[0068] Arrange all samples in chronological order, construct a feature matrix and label vector, and form the first multi-source feature data sequence.

[0069] To improve model training efficiency and stability, the constructed feature sequences can be standardized before being input into the network. Specifically, for each column of the feature matrix, the mean and standard deviation are calculated, and Z-score standardization is performed.

[0070] S3, input the first multi-source feature data sequence into the gated memory network for training; wherein, the training process includes: grouping and independently pruning the network weight gradients corresponding to different types of features in the first multi-source feature data sequence; and updating the network weights related to meteorological parameters using a differentiated learning rate strategy.

[0071] Among them, the Gated Memory Network (GMSIM) consists of two stacked LSTM layers, with each layer having a hidden state dimension of 64. Dropout layers are set between the layers, and the output layer is a fully connected layer.

[0072] Specifically, the gated memory network employs a two-layer stacked LSTM structure, aiming to learn temporal dependencies in multi-source features from shallow to deep. The detailed forward propagation process of this network is as follows:

[0073] The network input is a standardized sequence of first-source feature data. , ,..., },in, The time step is the total number of samples arranged chronologically within a continuous observation period. The network uses a two-layer stacked LSTM structure, with each layer having a hidden state dimension of 64.

[0074] The first layer of LSTM at each time step Receive input and the hidden state of the previous time step and cell state The current state is calculated using the following gating mechanism, and the output of this layer is a sequence of hidden states. .

[0075] Forgotten Gate:

[0076] Input Gate:

[0077] Candidate values:

[0078] Cell status update:

[0079] Output gate:

[0080] Hidden output:

[0081] in, It is the Sigmoid activation function. This represents element-wise multiplication. , , , , , , , These are the trainable parameters for this layer.

[0082] The output sequence of the first LSTM layer The input is fed into the Dropout layer, where some neurons' outputs are randomly set to zero with a probability of 0.2 during training to alleviate overfitting and enhance the model's generalization ability. The probability value of 0.2 is a commonly used empirical value in recurrent neural networks to balance regularization strength and model capacity, and it is particularly suitable for modeling high-volume time-series data with both continuity and periodicity in this task.

[0083] The second LSTM layer takes the output of the Dropout layer as its input sequence and outputs the final hidden state sequence. The dimension is also 64. This layer is used to learn higher-order temporal patterns from the abstract features extracted from the first layer.

[0084] Take the second LSTM layer at the last time step Hidden state The final sea height prediction value is mapped through a fully connected layer. ,in, The weight matrix of the output layer. This is the bias term for the output layer.

[0085] The training process uses the Adam optimizer to minimize the composite loss function:

[0086]

[0087] in, This refers to the batch size, which is the number of samples input into the network for each forward and backward propagation. For the first Measured sea height label for each sample The predicted value corresponds to the model. The first term of the loss function is the mean squared error, which measures the deviation between the predicted value and the true value; the second term is a first-order smoothing constraint term, which penalizes sharp fluctuations between adjacent predicted values, aiming to improve the temporal continuity of the output sequence. The weighting coefficient for the smoothing term is used to balance accuracy and smoothness, and is typically optimized within the range of 0.01 to 0.1 through cross-validation. This range ensures that the smoothing term suppresses random noise in GNSS-IR observations without excessively weakening the true amplitude of tidal variations.

[0088] During backpropagation, a gradient pruning strategy is applied. Specifically, based on the core observation group, satellite geometry group, and atmospheric correction group to which the input features belong, the Frobenius norm (the square root of the sum of the squares of all elements in the matrix) is calculated for the input weight gradient submatrix of the first-layer LSTM. The core observation group includes initial inverted sea height, peak frequency, spectral amplitude, and peak noise. The satellite geometry group includes satellite azimuth and number of satellites. The atmospheric correction group includes atmospheric temperature and atmospheric pressure. If the norm of a gradient exceeds a preset threshold for that group, all gradient values ​​in that submatrix are scaled proportionally to make their norm equal to the threshold. The preset group thresholds are differentiated based on the importance and stability of the feature groups: the core observation group gradient is the most critical, and its threshold can be set higher, such as 1.0 to 2.0; the satellite geometry group is next, such as 0.5 to 1.0; the atmospheric correction group gradient may require stricter constraints to prevent overfitting, such as 0.1 to 0.5. Specific values ​​need to be fine-tuned on the validation set.

[0089] In the initial training phase, all parameters use the same learning rate, such as 0.001. This initial learning rate is a typical recommended value for the Adam optimizer for medium-sized temporal regression tasks, which helps ensure stable convergence of training. When the training epochs reach a certain proportion of the total number of epochs (e.g., 70%), the learning rate of the weight parameters associated with the atmospheric correction features is reduced to 0.1 times their original value. The reason for this design is that using a uniform learning rate in the early stages of training helps the network quickly grasp the basic relationships between all features; in the later stages of training, reducing the learning rate for atmospheric correction features that are highly volatile and have complex relationships helps the network make more refined and robust adjustments to these features, avoiding oscillations around the optimal value due to an excessively high learning rate. The learning rate decay factor of 0.1 is a commonly used empirical value in the field of deep learning, which can achieve a relatively significant decay effect.

[0090] S4. The GNSS-IR observation data to be inverted is constructed into a second multi-source feature data sequence, which is then input into the trained gated memory network to output the sea height inversion result. The second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

[0091] Specifically, in the inversion stage, since there is no corresponding true sea height value, the second multi-source feature data sequence does not contain tide gauge station label data, but only physical inversion process parameters extracted from the observation data and matching meteorological parameters.

[0092] To objectively evaluate the model's generalization ability on unknown data, the sea height inversion results output by the network need to be compared with the measured data from tide gauge stations that have undergone strict spatiotemporal matching during the same period, in order to verify the sea height inversion performance.

[0093] like Figure 3As shown, this figure includes schematic diagrams and real-world photographs of the Fresnel reflector zones of four experimental stations, corresponding to: (a) TAR0 station, (b) PTLD station, (c) HKQT station, and (d) SC02 station. The core information and verification significance of each station are shown in Table 1. Furthermore, the schematic diagrams of the reflector zones in the figure clearly indicate the effective observation azimuth of each station, perfectly matching the coastal orientation of the stations in the actual photographs. This intuitively explains the specific implementation of setting the azimuth observation range based on geographical location and coastal environment. For example, the schematic diagram of the TAR0 station shows the open sea surface reflector zone to the southwest, and the schematic diagram of the PTLD station shows the western sea surface reflector zone. This design ensures that the reflected signal mainly comes from the open sea surface, reducing land clutter interference and providing environmental assurance for the accuracy of subsequent feature extraction.

[0094] Table 1. Core Information and Verification Significance of Each Station

[0095] like Figure 4 The figure shows a comparison and error statistical analysis of the sea height time series retrieved using the traditional GNSS-IR method and the measured tide level series from tide gauge stations. Subplots (a) to (d) sequentially show the comparison between the retrieved sea height time series (curves) and the measured tide level series (scatter plots) from the TAR0, PTLD, HKQT, and SC02 stations. The horizontal axis represents time in years, the left vertical axis represents the sea height value (in meters), and the right vertical axis represents the absolute difference between the retrieved and measured values ​​(in meters). Gray areas or curves represent the changes in the absolute error value. Subplots (e) to (h) correspond to the histograms of the absolute error distribution for the four stations mentioned above, with the horizontal axis representing the error value (in meters) and the vertical axis representing the sample size. The results shown in the figure indicate that although the traditional GNSS-IR method can roughly reflect the trend of tidal changes, it has significant drawbacks: First, there is significant local fluctuation, such as the dramatic fluctuation of the curve at the PTLD station from 2020.00 to 2020.05 in (b), which deviates significantly from the stability of the measured scatter points; second, there is a lag in extreme value tracking, such as the visible time difference between the peak or trough occurrence times at the TAR0 station in (a) and the SC02 station in (d); third, there is amplitude deviation, with the vertical deviation between the inverted curve and the measured scatter points exceeding 0.5 meters in some periods. This figure visually reveals the insufficient accuracy and stability of the traditional GNSS-IR method due to model errors and noise interference from both temporal variations and statistical distribution perspectives.

[0096] like Figure 5The figure shows the correlation analysis of scatter plot density between traditional GNSS-IR inversion results and measured values ​​from tide gauge stations. Each subplot corresponds to stations (a) TAR0, (b) PTLD, (c) HKQT, and (d) SC02, respectively. The horizontal axis represents the sea height value retrieved by traditional GNSS-IR, i.e., the traditional GNSS-IR inversion result, and the vertical axis represents the sea height value measured by the tide gauge station, i.e., the tide level observation value. The color intensity in the figure indicates the data point density. The black dashed line y = x is the ideal fitting line, and the gray dashed line represents the linear relationship of the scatter plot fitting. The root mean square error (RMSE) and Pearson correlation coefficient (PCC) for each station are also labeled. The results show that the data points at each station are basically distributed along the ideal fitting line. The closeness of the gray fitting line to the black ideal line is highly consistent with the PCC value, indicating that the traditional method has basic inversion capabilities. The RMSE value directly quantifies the inversion error level. Meanwhile, the error distribution of different stations showed significant differences: (a) The point clouds of TAR0 and (d) SC02 stations were concentrated along the diagonal, with the density peak closely following the ideal fitting line, and the gray dashed line almost overlapping the black dashed line, indicating small systematic deviation and good control of random errors; (b) The point cloud of PTLD station was relatively dispersed, and lateral diffusion occurred in the range of larger values ​​of the abscissa corresponding to the high tide level, with the gray dashed line slightly deviating from the black dashed line. The Pearson correlation coefficient (PCC) was 0.82, the lowest among the four stations, indicating that the inversion error of this station increased under high water conditions, possibly due to multipath effects or satellite geometry; (c) The point cloud of HKQT station was generally distributed along the diagonal, but there were a few discrete points. The gray dashed line and the black dashed line were still relatively close, and the Pearson correlation coefficient (PCC) was 0.94, which was relatively high, reflecting both the local interference of extreme weather on traditional methods and the good overall correlation. This figure systematically reveals the accuracy level and limitations of traditional GNSS-IR methods from two aspects: statistical correlation and error distribution. It also clarifies the key areas that subsequent deep learning models need to focus on improving, such as error suppression under high water levels and complex sea conditions.

[0097] like Figure 6 The figure shows a comparison and error statistical analysis of the sea height time series retrieved by the method of this invention (GMSIM) and the measured tide level series from tide gauge stations. In the figure, subplots (a) to (d) are comparison plots of the retrieved sea height time series (curves) and measured tide level series (scatter plots) from stations TAR0, PTLD, HKQT, and SC02, respectively. The coordinate axes have the same meaning as... Figure 4 Consistent; subplots (e) to (h) are histograms of the absolute error distribution for each station, with the root mean square error (RMSE) and Pearson correlation coefficient (PCC) labeled for each station. The results show that the method of this invention (GMSIM) exhibits superior quantization performance at all stations, with the RMSE generally better than the traditional GNSS-IR method (compared to...). Figure 5 This confirms that the systematic bias was effectively suppressed; at the same time, the PCC of all stations was higher than that of the traditional GNSS-IR method (compared to...). Figure 5 The high accuracy indicates that the retrieved sequence is highly synchronized with the measured tide level trend. From the time-series curve morphology, firstly, the sequence is smooth and continuous, with no abnormal fluctuations in the curves at any station; secondly, extreme value tracking is accurate, with the peak or trough inflection points at (a) TAR0 station and (d) SC02 station highly aligned with the measured scatter points; and thirdly, trend consistency is high, with the rising and falling trends of the curves completely synchronized with the convergence and dispersion trends of the measured scatter points, even during extreme weather periods at (c) HKQT station, no significant deviation was observed. This figure, combined with time-series comparison and quantitative indicators (RMSE and PCC), visually verifies that the method of this invention (GMSIM) can achieve high-precision, high-consistency, and high-stability sea-elevation retrieval under various environments.

[0098] like Figure 7 The figure shows a comprehensive comparison of the inversion error distribution between the method of this invention (GMSIM) and the traditional GNSS-IR method. The horizontal axis represents the difference between the inverted value and the measured value at the tide gauge station, i.e., the error value (unit: m); the vertical axis represents the number of samples. The figure displays the error distribution histograms and their fitting curves for the two methods, and labels the mean μ and standard deviation σ of their respective error distributions. The results show that the traditional GNSS-IR method has a wider error distribution range, a more dispersed histogram, a relatively flat fitting curve, and exhibits significant negative systematic bias (μ deviates significantly from zero) and large random dispersion (σ is large), resulting in poor prediction stability. In contrast, the error of the method of this invention (GMSIM) is highly concentrated within ±0.5 meters, the fitting curve is steep and symmetrical, its mean μ approaches zero, and the standard deviation σ is reduced to about 1 / 5 of that of the traditional GNSS-IR method. These data demonstrate that the method of this invention (GMSIM) almost eliminates systematic bias and reduces the dispersion of random errors by an order of magnitude, thereby achieving higher overall inversion accuracy, stability, and robustness.

[0099] like Figure 8 As shown, another embodiment of the present invention provides a sea altitude inversion system based on a gated memory network and GNSS-IR. The system includes a data acquisition module, a feature construction module, a network training module, and a network inversion module.

[0100] The data acquisition module is used to acquire GNSS-IR observation data and extract physical inversion process parameters from it.

[0101] The feature construction module is used to construct the first multi-source feature data sequence based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station.

[0102] The network training module is used to input the first multi-source feature data sequence into the gated memory network for training.

[0103] The network training module includes a gradient pruning submodule and a learning rate update submodule.

[0104] The gradient pruning submodule is used to independently prune the network weight gradients corresponding to different types of features in the first multi-source feature data sequence.

[0105] The learning rate update submodule is used to update the network weights related to meteorological parameters using a differentiated learning rate strategy.

[0106] The network inversion module is used to construct a second multi-source feature data sequence from the GNSS-IR observation data to be inverted, and input it into the trained gated memory network to output the sea height inversion result. The second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

[0107] This embodiment also provides a computer device applicable to a sea altitude inversion method based on a gated memory network and GNSS-IR, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement all or part of the steps of the method described in the above embodiments of the present invention.

[0108] This embodiment also provides a storage medium on which a computer program is stored. When the computer program is executed by a processor, it performs the method in any optional implementation of the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0109] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0110] In summary, this invention proposes a sea level inversion method and system based on gated memory networks and GNSS-IR. Specifically, at the data level, the initial inversion sea level and peak frequency are extracted based on the core principles of GNSS-IR, preserving the physical mechanisms of signal propagation and interference, effectively avoiding the blindness of purely data-driven models. Subsequently, through rigorous spatiotemporal alignment techniques, the aforementioned physical parameters are organically integrated with meteorological parameters and tide gauge data, enabling the model to simultaneously perceive multi-dimensional information such as signal quality, satellite geometry, and atmospheric environment. This addresses the shortcomings of single data being susceptible to atmospheric interference and spatial heterogeneity, ultimately constructing a comprehensive and reliable multi-source feature system. At the algorithm level, a two-layer stacked LSTM gated memory network is innovatively adopted as the core of temporal modeling. Its hidden state dimension and inter-layer dropout parameters are specially designed to achieve a balance between model capacity and generalization ability. This network can deeply mine the complex nonlinear temporal mapping relationship between multi-source features and sea level, and can especially accurately learn the dynamic correction law of time-varying interference such as atmospheric delay. Meanwhile, a first-order smoothing constraint term is innovatively introduced into the loss function, explicitly encoding the key physical prior knowledge of continuous sea level change as the optimization objective. This ensures that the output sequence is smooth and free of physical distortion while pursuing high accuracy, fundamentally solving the jitter, lag, and bias problems of traditional methods. At the training mechanism level, an innovative strategy combining independent gradient pruning and differentiated learning rates based on feature physical attributes is proposed. This strategy sets gradient pruning thresholds differently according to feature importance, ensuring balanced and stable updates of network parameters. In the later stages of training, learning rate decay is applied to the weights of meteorological parameters with large fluctuations, achieving fine calibration for the influence of complex atmospheres. This strategy effectively solves the gradient conflict and oscillation problems in training with multi-source heterogeneous data, which is the key to maintaining high robustness of the model in complex scenarios. At the systems engineering level, an innovative design is adopted that unifies the training and inversion feature construction logic. That is, the feature structure, spatiotemporal alignment standard, and data processing flow of the second multi-source feature data sequence are completely consistent with those of the first multi-source feature data sequence, with only the tide gauge station label data being removed. This fundamentally avoids the additional system errors or performance degradation introduced by the inconsistency of feature engineering in traditional methods, ensuring that the model advantages in the training phase are fully realized in real application scenarios, thereby guaranteeing the reliability and accuracy stability of the inversion results.

[0111] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A sea altitude inversion method based on gated memory networks and GNSS-IR, characterized in that, include: Acquire GNSS-IR observation data and extract physical inversion process parameters from it; Based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station, a first multi-source feature data sequence was constructed. The first multi-source feature data sequence is input into the gated memory network for training; the training process includes: grouping and independently pruning the network weight gradients corresponding to different types of features in the first multi-source feature data sequence; and updating the network weights related to meteorological parameters using a differentiated learning rate strategy. The GNSS-IR observation data to be inverted is constructed into a second multi-source feature data sequence and input into a trained gated memory network to output the sea height inversion result; the second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

2. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 1, characterized in that, The physical inversion process parameters include initial inversion sea height, peak frequency, spectral amplitude, peak noise, number of satellites, and satellite azimuth; the meteorological parameters include atmospheric temperature and atmospheric pressure; and the measured sea height data from the tide gauge station are observation values ​​calibrated to the tidal datum.

3. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 1, characterized in that, The process of acquiring GNSS-IR observation data and extracting physical inversion parameters from it includes: Extract the signal-to-noise ratio (SNR) observations from the original GNSS observation data files; Based on the satellite elevation angle, the signal-to-noise ratio observation sequence related to sea surface reflection was selected, and the satellite azimuth angle corresponding to each observation point and the number of visible satellites in the current time period were recorded. Spectral analysis was performed on the selected signal-to-noise ratio observation sequence to extract the peak frequency and corresponding spectral amplitude, and the peak noise was calculated. Based on the peak frequency, satellite elevation angle, and carrier wavelength, the initial inverted sea height is calculated using the following formula: ； In the formula, Peak frequency; The satellite's elevation angle; For GNSS carrier wavelength; This is the initial inversion of sea level.

4. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 1, characterized in that, The step of independently pruning the network weight gradients corresponding to different types of features in the first multi-source feature data sequence includes: The first multi-source feature data sequence was divided into a core observation group, a satellite geometry group, and an atmospheric correction group. During the backpropagation process of the network, the Frobenius norm is calculated for the input weight gradient submatrix corresponding to each group of features, and independent pruning is performed based on the preset group pruning threshold.

5. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 4, characterized in that, The core observation group includes initial inverted sea height, peak frequency, spectral amplitude, and peak noise; the satellite geometry group includes satellite azimuth and number of satellites; and the atmospheric correction group includes atmospheric temperature and atmospheric pressure.

6. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 1, characterized in that, The differentiated learning rate strategy includes: In the later stages of training, the learning rate of the network weight parameters connected to the input dimensions of atmospheric temperature and atmospheric pressure is adjusted to 0.1 times the learning rate of other network weight parameters.

7. The sea altitude inversion method based on gated memory network and GNSS-IR according to claim 1, characterized in that, The gated memory network consists of two stacked LSTM layers, with each layer having a hidden state dimension of 64. Dropout layers are placed between the layers, and the output layer is a fully connected layer. The loss function for training the gated memory network is the sum of the mean squared error and the smoothing constraint term. The smoothing constraint term is the L2 norm of the difference between adjacent time points of the predicted sea level sequence.

8. A sea height inversion system based on gated memory networks and GNSS-IR, used to execute the sea height inversion method based on gated memory networks and GNSS-IR as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to acquire GNSS-IR observation data and extract physical inversion process parameters from it; The feature construction module is used to construct the first multi-source feature data sequence based on the physical inversion process parameters, the corresponding spatiotemporally matched meteorological parameters, and the measured sea height data from the tide gauge station. A network training module is used to input the first multi-source feature data sequence into a gated memory network for training; wherein, the network training module includes: The gradient pruning submodule is used to independently prune the network weight gradients corresponding to different types of features in the first multi-source feature data sequence. The learning rate update submodule is used to update the network weights related to meteorological parameters using a differentiated learning rate strategy; The network inversion module is used to construct a second multi-source feature data sequence from the GNSS-IR observation data to be inverted, and input it into the trained gated memory network to output the sea height inversion result; the second multi-source feature data sequence adopts the same feature structure and construction logic as the first multi-source feature data sequence.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the sea altitude inversion method based on gated memory network and GNSS-IR as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the sea altitude inversion method based on gated memory network and GNSS-IR as described in any one of claims 1 to 7.

Images