A coastal GNSS-IR sea surface level correction method based on cross-site machine learning
Patent Information
- Application Number
- CN202611488514.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-23
- Publication Date
- 2026-10-27
AI Technical Summary
[0005]本发明的目的是针对现有单站GNSS-IR机器学习改正方法需要长期积累目标站历史标签、邻近验潮站模型难以直接迁移、直接预测总水位时容易混入站点潮汐差异以及不规则采样条件下跨站特征难以统一构造的问题，提供一种基于跨站点机器学习的近岸GNSS-IR海面水位改正方法，利用多个已有站点的历史GNSS-IR和验潮资料建立可迁移关系，在目标站无需预先积累长期本地验潮标签的情况下，将目标站特征向量输入训练后的跨站点回归模型进行水位改正
[0039]1、本发明分别根据验潮站总水位和GNSS-IR总水位扣除同历元天文潮序列，得到含义明确的验潮站余水位标签和GNSS-IR余水位输入，降低不同站点潮差、潮相和垂直基准差异对模型迁移的干扰，同时仍以总水位作为最终输出。
Smart Images
Figure CN122982152A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) and machine learning for water level monitoring. Specifically, it relates to a nearshore GNSS-IR sea surface water level correction method based on cross-site machine learning, and more particularly to a GNSS-IR sea surface water level correction method that uses data from multiple existing source stations to train a cross-site regression model, predicts the residual water level at target stations that are not involved in the training, and reconstructs the corrected total water level at the target stations. Background Technology
[0002] GNSS-IR utilizes the interference effect between the direct signal and the reflected signal from the water surface simultaneously received by a shore-based GNSS receiver to extract the reflector height and invert the water level from the signal-to-noise ratio sequence. This method can reuse existing continuously operating GNSS stations and can also quickly deploy dedicated receiving equipment on the shoreline where monitoring is needed. It features non-contact operation, all-weather capability, and low construction cost. Therefore, GNSS-IR can not only supplement long-term tide gauge stations but is also suitable for temporary tide gauges during flood tides, storm surges, or engineering observations, as well as for adding water level monitoring points and densifying regional water level monitoring networks in areas where traditional tide gauge stations are insufficiently covered.
[0003] Current GNSS-IR water level inversion methods typically perform detrending and spectral analysis on low-elevation-angle satellite arcs and calculate the vertical distance from the antenna phase center to the water surface from the dominant frequency. Due to dynamic changes in the water surface, irregular observation intervals, signal attenuation, obstruction, satellite geometry, sea surface roughness, and differences in the reflection environment, the original inversion results often contain systematic biases, outliers, and time-varying errors. While sliding window robust estimation, dynamic water level correction, and spline fitting can reduce some errors, their fixed models struggle to simultaneously describe the nonlinear coupling of multiple error sources.
[0004] Machine learning can utilize GNSS-IR water level, signal quality, and environmental variables to learn nonlinear correction relationships. However, existing methods mostly model on a per-station basis, requiring the continuous accumulation of historical GNSS-IR and tide gauge data at the target station to form sufficient supervised learning samples before training and application. When the target station lacks local historical labels, data from nearby tide gauge stations or other hydrological stations are needed to participate in model training. The former approach prolongs the construction and application cycle of temporary and regional densification stations, while the latter approach, if directly applying neighboring station models, easily introduces tidal range, tidal phase, vertical reference, sampling characteristics, and local hydrodynamic conditions from neighboring stations into the target station, resulting in distribution shifts or even negative migration. Summary of the Invention
[0005] The purpose of this invention is to address the problems of existing single-station GNSS-IR machine learning correction methods, such as the need for long-term accumulation of historical labels for the target station, the difficulty in directly transferring models from neighboring tide gauge stations, the easy incorporation of tidal differences between stations when directly predicting total water level, and the difficulty in uniformly constructing cross-station features under irregular sampling conditions. This invention provides a nearshore GNSS-IR sea surface water level correction method based on cross-site machine learning. It establishes transferable relationships using historical GNSS-IR and tide gauge data from multiple existing stations. Without requiring the target station to accumulate long-term local tide gauge labels beforehand, the target station's feature vector is input into the trained cross-site regression model for water level correction. Furthermore, this invention clearly distinguishes between total water level, astronomical tide, and residual water level by utilizing the non-equidistant characteristics of raw GNSS-IR observations, enabling the model to predict the relatively more transferable residual water level between different stations, ultimately outputting a complete corrected total water level.
[0006] According to one aspect of the present invention, a nearshore GNSS-IR sea surface level correction method based on cross-site machine learning is provided, comprising:
[0007] The original GNSS-IR total water level of multiple source stations and at least one target station is obtained. At the same time, the reference total water level of the tide gauge corresponding to each source station, as well as the signal quality and interference status information, meteorological information and site environmental information corresponding to the source station and the target station are obtained.
[0008] Based on the original GNSS-IR total water level of each source station and target station, the reference total water level of each source station's corresponding tide gauge station, and their respective irregular observation epochs, a harmonic analysis is performed on the preset astronomical tide set to estimate the preset tide parameters and construct the source station tide gauge station astronomical tide sequence and the GNSS-IR astronomical tide sequence of each station respectively.
[0009] Using the original GNSS-IR observation epochs of each station as a unified time reference, time registration was performed on the reference total water level of the tide gauge corresponding to each source station and the external dynamic variables of the source and target stations.
[0010] A window is established with the central epoch of each station. Based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state characteristics, a feature vector is constructed. The astronomical tide sequence of the source station's tide gauge is subtracted from the reference total water level of the tide gauge corresponding to each source station to obtain the residual water level of the tide gauge and it is used as a supervised learning label.
[0011] A cross-site training sample set is constructed based on the feature vectors and supervised learning labels of multiple source sites. The cross-site regression model is trained using the cross-site training sample set, and the feature preprocessing parameters and model hyperparameters are determined based on the cross-site training sample set.
[0012] The target station feature vector is input into the trained cross-site regression model to obtain the predicted residual water level of the target station. The predicted residual water level of the target station is then added to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
[0013] As a further technical solution, the raw GNSS-IR total water level of multiple source stations and at least one target station is obtained, including:
[0014] For multiple source stations and at least one target station, acquire the GNSS observation files recorded by the GNSS receivers of each station, as well as the broadcast ephemeris files corresponding to the observation period;
[0015] The satellite elevation and azimuth angles are calculated based on the GNSS observation files and broadcast ephemeris files, and the signal-to-noise ratio observation sequence is extracted based on the GNSS observation files;
[0016] Based on the signal-to-noise ratio observation sequence, a detrended signal-to-noise ratio sequence dominated by water surface reflection interference oscillations is obtained through trend separation;
[0017] Based on the detrended signal-to-noise ratio sequence dominated by water surface reflection interference oscillation, the Lom-Skagler periodogram analysis was used to extract the main frequency of the spectral peak, and the height of the reflecting surface was calculated in combination with the carrier wavelength.
[0018] The original GNSS-IR total water level is obtained based on the height of the reflector surface and the elevation of the antenna phase center under a unified reference datum.
[0019] As a further technical solution, time registration is performed, including:
[0020] Search for the neighboring valid epochs of the variable to be registered, and the corresponding observations, on both sides of each GNSS-IR observation epoch.
[0021] Linear interpolation is performed under the condition of satisfying the preset maximum allowable time interval constraint to complete the time registration between the variable to be registered and the original epoch of GNSS-IR observation;
[0022] If there are no adjacent valid epochs or the interval between adjacent epochs exceeds the preset maximum allowable interval, the sample corresponding to the current GNSS-IR raw observation epoch will be marked as invalid and deleted.
[0023] As a further technical solution, the interference state information includes one or more of the following: peak-to-noise ratio, full width at half maximum (FWHM), amplitude, damping coefficient, phase parameter, and satellite elevation angle change rate.
[0024] As a further technical solution, the irregular sampling state characteristics include at least the number of window observations, the number of observations before and after the central epoch, the most recent time interval, the maximum time interval within the window, the window time coverage, and the time balance index.
[0025] As a further technical solution, the GNSS-IR residual water level is obtained by subtracting the GNSS-IR astronomical tide sequence of each station from the original total GNSS-IR water level of each station.
[0026] As a further technical solution, the expression for time registration is:
[0027]
[0028] Where j is the index of the variable to be registered. Let j be the j-th variable to be registered, and t be the target epoch. and This indicates the most recent valid epoch before or after t.
[0029] According to one aspect of the present invention, a nearshore GNSS-IR sea surface level correction system based on cross-site machine learning is provided, for implementing the aforementioned nearshore GNSS-IR sea surface level correction method based on cross-site machine learning, comprising:
[0030] The data acquisition module is used to acquire the raw GNSS-IR total water level of multiple source stations and at least one target station, and at the same time acquire the reference total water level of the tide gauge station corresponding to each source station, as well as the signal quality and interference status information, meteorological information and station environmental information corresponding to the source station and the target station.
[0031] The astronomical tide construction module is used to perform harmonic analysis on the preset astronomical tide set based on the original GNSS-IR total water level of each source station and target station, the reference total water level of the tide gauge station corresponding to each source station, and their respective irregular observation epochs, to estimate the preset tide parameters and construct the astronomical tide sequence of the source station tide gauge station and the GNSS-IR astronomical tide sequence of each station respectively.
[0032] The epoch registration module is used to time-register the reference total water level of the tide gauge station corresponding to each source station and the external dynamic variables of the source and target stations, using the original epochs of GNSS-IR observations at each station as a unified time reference.
[0033] The window sample construction module is used to establish windows based on the central epoch of each station. It constructs feature vectors based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state characteristics. The astronomical tide sequence of the source station's tide gauge is subtracted from the reference total water level of the tide gauge corresponding to each source station to obtain the residual water level of the tide gauge and use it as a supervised learning label.
[0034] The cross-site regression model training module is used to construct a cross-site training sample set based on the feature vectors and supervised learning labels of multiple source sites, train the cross-site regression model using the cross-site training sample set, and determine the feature preprocessing parameters and model hyperparameters based on the cross-site training sample set.
[0035] The target station total water level acquisition module is used to input the target station feature vector into the trained cross-site regression model to obtain the predicted residual water level of the target station, and add the predicted residual water level of the target station to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
[0036] According to one aspect of the present invention, an electronic device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the nearshore GNSS-IR sea surface level correction method based on cross-site machine learning.
[0037] According to one aspect of the present invention, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps of the nearshore GNSS-IR sea surface level correction method based on cross-site machine learning.
[0038] Compared with the prior art, the beneficial effects of the present invention are:
[0039] 1. This invention subtracts the astronomical tide sequence of the same epoch from the total water level of the tide gauge station and the total water level of GNSS-IR, respectively, to obtain the residual water level label of the tide gauge station and the residual water level input of GNSS-IR with clear meaning. This reduces the interference of differences in tidal range, tidal phase and vertical reference of different stations on model migration, while still using the total water level as the final output.
[0040] 2. This invention adopts time registration and sliding window feature construction with the original GNSS-IR observation epoch as the unified time reference, without the need to force the original sequence to be resampled into a regular sequence, and explicitly describes the sparsity, gap and window asymmetry states.
[0041] 3. This invention constructs a cross-site training sample set by aggregating multiple existing source stations, and trains the cross-site regression model. The target station does not need to accumulate long-term local tide data in advance, and can input the trained cross-site regression model to obtain the corrected total water level. It is applicable to temporary tide gauge stations, regional densified GNSS-IR stations, and other stations that lack long-term co-located tide data.
[0042] 4. This invention describes the error occurrence conditions by using signal quality, interference state, meteorological and site environmental characteristics, enabling the regression model to learn the nonlinear relationships between multiple error sources, and retaining the ability to expand the training set according to the number and similarity of source stations. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art 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.
[0044] Figure 1 The flowchart of a nearshore GNSS-IR sea surface level correction method based on cross-site machine learning provided by the present invention is shown.
[0045] Figure 2 The results of water level correction verification at eight stations provided by this invention.
[0046] Figure 3 This invention provides a case diagram of water level correction at the TSAQ site.
[0047] Figure 4 This invention provides a case diagram of water level correction at the AMER site.
[0048] Figure 5 This invention provides a case diagram of water level correction at the SBCH site. Detailed Implementation
[0049] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0050] like Figure 1As shown, this embodiment of the invention proposes a nearshore GNSS-IR sea surface level correction method based on cross-site machine learning, including: Step S1, acquiring multi-site data. Acquiring the original GNSS-IR total water level of multiple source stations and at least one target station; simultaneously acquiring the reference total water level of the tide gauge corresponding to each source station, as well as the signal quality and interferometric state information, meteorological information, and site environmental information corresponding to the source and target stations. Step S2, constructing astronomical tide sequences station by station. Based on the original GNSS-IR total water level of each source and target station, the reference total water level of the tide gauge corresponding to each source station, and their respective irregular observation epochs, a harmonic analysis is performed on the preset astronomical tide set to estimate the preset tide parameters and construct the source station tide gauge astronomical tide sequence and the GNSS-IR astronomical tide sequence for each station. Step S3, completing time registration. Using the original GNSS-IR observation epoch of each station as a unified time reference, the reference total water level of the tide gauge corresponding to each source station and the external dynamic variables of the source and target stations are time-registered to the original GNSS-IR observation epoch. Step S4: Construct window features and residual water level labels. Establish windows based on the central epoch of each station. Construct feature vectors based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state features. Subtract the astronomical tide sequence of the source station's tide gauge from the reference total water level of each source station to obtain the residual water level of the tide gauge station, which is then used as a supervised learning label. Step S5: Perform cross-site training. Select multiple existing source stations based on spatial distance, overlap of effective observation periods, similarity of station environments, or data availability. Collect the features of each available source station and the residual water level labels of the tide gauge stations to form a cross-site training sample set, and train the XGBoost model. Feature preprocessing parameters and model hyperparameters are determined based on the cross-site training sample set. Step S6: Perform target station prediction and water level reconstruction. Input the target station feature vector into the trained cross-site regression model to obtain the predicted residual water level of the target station. Add the predicted residual water level of the target station to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
[0051] Specifically, step S1 includes the following:
[0052] For multiple source stations and at least one target station, acquire the GNSS observation files recorded by the GNSS receivers at each station, as well as the broadcast ephemeris files corresponding to the observation periods. For GNSS data acquired in the field, if the receiver or acquisition software synchronously records navigation messages, the broadcast ephemeris files for the corresponding observation periods can be directly exported. For cases where broadcast ephemeris files are not synchronously saved or where supplementary data is required, broadcast ephemeris files for the corresponding dates and periods can be downloaded from the IGS data center or other public GNSS data service platforms, such as the CDDIS data center or a university GNSS data center.
[0053] Based on the GNSS observation files and broadcast ephemeris files, the satellite elevation and azimuth angles are calculated, and the signal-to-noise ratio (SNR) observation sequences at corresponding frequencies are extracted from the GNSS observation files. For each source and target station, low-elevation-angle ascending and descending satellite arc segments in the water surface reflection direction are selected from the SNR observation sequences according to the on-site reflection environment. Trend separation is performed on the selected ascending and descending satellite arc segments. This can be achieved using low-order polynomials or variational mode decomposition to remove low-frequency direct radiation components, resulting in a detrended SNR sequence dominated by water surface reflection interferometric oscillations.
[0054] Subsequently, based on the detrended signal-to-noise ratio sequence dominated by water surface reflection interferometry oscillations, Lomb-Scargle periodogram (LSP) analysis was performed with sin(e) as the independent variable to extract the dominant frequency of the spectral peak. For each effective satellite arc segment, the vertical distance from the antenna phase center corresponding to the observation epoch t to the water surface reflector was taken as the reflector height, calculated as follows:
[0055] (1)
[0056] in, λ is the dominant frequency of the spectral peak, and λ is the carrier wavelength. Indicates the height of the reflective surface.
[0057] Let the elevation of the antenna phase center under a unified reference datum be... Calculate the GNSS-IR water level for the effective observation epochs, and use it as the original GNSS-IR total water level. The expression is:
[0058] (2)
[0059] During the peak extraction or interferometric signal fitting process, the peak-to-noise ratio (PNR), full width at half maximum (FWHM), fitting amplitude, damping coefficient, initial phase, fitting error, and satellite elevation angle change rate are simultaneously saved as candidate variables for subsequent model descriptions of signal quality and interferometric state.
[0060] Specifically, step S2 includes the following:
[0061] For any source station s, the reference total water level sequence of the tide gauge station corresponding to the source station is used as input. The preset astronomical tide set is harmonic analyzed by the UTide harmonic analysis tool to estimate the parameters of each preset tide and construct the astronomical tide sequence of the tide gauge station corresponding to the source station.
[0062] For any station q, which includes a source station and a target station, the original GNSS-IR total water level sequence of the station and its irregular observation epochs are used as input. The same preset astronomical tide set is harmonic analyzed using the UTide harmonic analysis tool to estimate the parameters of each preset tide and construct the corresponding GNSS-IR astronomical tide sequence for the station.
[0063] Therefore, the astronomical tide sequence of the source station's tide gauge and the GNSS-IR astronomical tide sequence of each station are obtained independently through harmonic analysis of their respective water level sequences. Specifically, the astronomical tide sequence of the source station's tide gauge is used to construct supervised learning labels from the reference total water level of the tide gauge; the GNSS-IR astronomical tide sequence is used to construct the GNSS-IR residual water level from the original GNSS-IR total water level of each station, and is then used for total water level reconstruction after predicting the residual water level of the target station.
[0064] Specifically, step S3 includes the following:
[0065] Reference total water level at tide gauge stations corresponding to each source station Furthermore, the meteorological and other external dynamic variables corresponding to the source and target stations have their own observation epochs, which are not completely consistent with the original observation epochs of GNSS-IR. Therefore, time registration is required for the reference total water level of the tide gauge corresponding to each source station and the external dynamic variables of the source and target stations. The process is as follows:
[0066] For any variable to be registered Taking each GNSS-IR raw observation epoch t as the target epoch, in the variables Searching for the nearest valid epoch before and after t in a time series and and the corresponding observations and Under the condition of satisfying the preset maximum allowable time interval constraint, linear interpolation is performed according to the following formula to achieve time registration between the variable to be registered and the original epoch of GNSS-IR observation:
[0067] (3)
[0068] Where j is the index of the variable to be registered. This represents the j-th variable to be registered; for the source station, the variable to be registered includes the corresponding tide gauge reference total water level and meteorological and other external dynamic variables; for the target station, the variable to be registered includes meteorological and other external dynamic variables. Each time formula (4) is executed, Each represents a specific variable to be registered. For any variable to be registered, when or Does not exist, or If the interval exceeds the preset maximum allowable interval, the variable to be registered cannot be effectively registered at the current GNSS-IR original observation epoch t, and the sample corresponding to the GNSS-IR original observation epoch will be marked as invalid and deleted.
[0069] Specifically, step S4 includes the following:
[0070] A window W(s,i) is established using the i-th central epoch t(s,i) of any station q. The window can be a symmetrical retrospective window, or, in real-time applications, a causal window containing only current and historical observations. Statistical characteristics of the window, including mean, standard deviation, minimum, maximum, quartiles, most recent effective value, and local rate of change, are calculated for the original GNSS-IR water level, GNSS-IR residual water level, PNR, FWHM, amplitude, damping, phase, and meteorological variables.
[0071] To characterize the irregular sampling state of raw GNSS-IR observations, the total number of window observations (n) and the number of observations before and after the central epoch are further calculated. and Recent time interval Maximum internal spacing within the window The window time coverage (C) and time balance index (B) are used to characterize the interferometric state information. One or more of the following parameters characterize the peak-to-noise ratio (PNR), full width at half maximum (FWHM), amplitude, damping coefficient, phase parameter, and satellite elevation angle change rate. All of these parameters are directly obtained statistically from the effective GNSS-IR observation epochs within the window. Furthermore, window statistical features and irregular sampling state features are used to construct the feature vector corresponding to the central epoch, serving as inputs for subsequent cross-site regression model training and target station prediction. Specifically, the window statistical features constructed based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, and meteorological information, along with the irregular sampling state features, are concatenated with static environmental attributes of the station, such as station latitude and longitude, antenna elevation, local water depth, seabed slope, and nearshore water distance, to form the feature vector corresponding to the central epoch. .
[0072] The time balance index B can be defined as:
[0073] (4)
[0074] in, and These represent the observations before and after the central epoch, respectively. For source station s, the reference total water level of the tide gauge station will be used. Excluding the astronomical tides of the same epoch at this station The residual water level at the tide gauge station is obtained; for any station q, the original GNSS-IR total water level is obtained. Excluding the astronomical tides of the same epoch at this station The GNSS-IR residual water levels for all stations are obtained. The residual water level at the tide gauge station is constructed only for the source station and serves as the supervised learning label for the cross-site regression model; the GNSS-IR residual water level is constructed for all stations, with the GNSS-IR residual water level at the source station serving as the model training input and the GNSS-IR residual water level at the target station serving as the model prediction input. They are defined as follows:
[0075] (5)
[0076] (6)
[0077] in, The reference total water level of the tide gauge at the source station s in epoch t. For any station q, the original GNSS-IR total water level at the same epoch. and These are the astronomical tide sequences obtained by UTide harmonic analysis of the tide gauge water level sequence based on source station s and the original GNSS-IR total water level at any station q, respectively. and These are the residual water levels at the tide gauge station and the GNSS-IR residual water levels, respectively.
[0078] Therefore, the prediction target of the subsequent cross-site regression model is the residual water level, which has a clear physical meaning. This construction allows the cross-site regression model to focus on GNSS-IR observation errors, non-tidal water level changes, and their relationship with signal and environmental conditions, rather than repeatedly learning the main astronomical tidal cycle across different sites.
[0079] Specifically, step S5 includes the following:
[0080] Multiple existing source stations are selected as usable source stations based on spatial distance, overlap of effective observation periods, similarity of site environments, or data availability. Let the set of usable source stations be denoted as . , Represents the set of source sites Any source station in the dataset. A feature vector constructed from each source station according to a unified data structure. and the residual water level label at the tide gauge station Construct a cross-site training sample set, as shown in the following expression:
[0081] (7)
[0082] in, As the source site The number of effective training samples.
[0083] Let the target station be ,and The target station does not participate in the construction of the source station's training sample set; it only constructs feature vectors using the same data structure as the source station to form the target station's feature set to be predicted. The expression is as follows:
[0084] (8)
[0085] in, For the target station The number of valid samples to be predicted.
[0086] When the target station lacks tide gauge data as a reference or lacks long-term historical data accumulation, its tide gauge reference water level is not a necessary input for the cross-site regression model. In this embodiment of the invention, the tide gauge reference water level is only used for source station supervision label construction, so the target station's tide gauge reference water level is not obtained. The method proposed in this embodiment of the invention only utilizes existing data from multiple source stations to complete cross-site model training. The target station does not need to participate in model training, nor does it need to provide local tide gauge residual water level labels. It only needs to input its own GNSS-IR observations and related feature data to obtain the target station's predicted residual water level and reconstruct the corrected total water level, thus avoiding reliance on the target station's historical tide gauge data before the cross-site regression model is put into application.
[0087] A cross-site regression model is trained using a cross-site training sample set consisting of multiple source stations. The feature vectors of the source stations are used as model inputs, and the remaining water level of the corresponding tide gauge station is used as the supervised learning label. The cross-site regression model can employ supervised regression models such as XGBoost, Random Forest, or Gradient Boosting Tree; the specific model form does not constitute a limitation on the technical solution of this invention. After the cross-site regression model is trained, it is used to predict the remaining water level of the target station's feature data.
[0088] Specifically, step S6 includes the following:
[0089] target station In the observation epoch eigenvectors Input the trained cross-site regression model to obtain the predicted residual water level of the target station in that epoch. The predicted residual water level at the target station is compared with the astronomical tide level at the target station in the same epoch. Add them together to get the corrected total water level at the target station. The expression is as follows:
[0090] (9)
[0091] Therefore, astronomical tides are used both to construct a more transferable learning space between sites and to restore the physically interpretable total water level at the output. If a small number of labeled samples are subsequently obtained at the target station, samples from the initial time period can be selected in chronological order and added to the cross-site training sample set for adaptation. The adapted cross-site regression model can then be used for water level correction at the target station.
[0092] Specifically, the multiple source stations proposed in this embodiment refer to at least two source stations. Furthermore, the training samples for the cross-site model are not individual samples for each site, but rather consist of feature vectors corresponding to different GNSS-IR observation epochs at each source station, along with supervised learning labels. Therefore, even with only two source stations, as long as both source stations have sufficient effective observation data, a cross-site training sample set containing a large number of samples can still be formed, and model training can be completed. The use of the remaining seven sites as source stations in subsequent embodiments is to verify the effectiveness of each of the eight sites individually using the leave-one-site method, and is not a necessary limitation on the number of source stations in this invention. The number of source stations and their similarity to the target station will affect the model transfer effect, but will not affect the implementation of the method of this invention under conditions of two or more source stations.
[0093] Specifically, this embodiment of the invention uses eight shore-based GNSS stations—AMER, CALC, N301, ROPS, SBCH, TSAQ, TXNJ, and TXPO—located in a northern bay to verify the method proposed in this embodiment. The leave-one-site method is only used as a verification scheme in this embodiment and is not a necessary step in the method of this invention. Specifically, the leave-one-site method is as follows: each time, one station is used as the verification target station, and the remaining seven stations are used as source stations to train the cross-site model; this is repeated eight times, so that each station is independently used as a target station once. The tide gauge water level of the current target station is not involved in the model training, preprocessing parameter fitting, and hyperparameter selection for this iteration; it is only used to calculate the verification index.
[0094] SBCH is located at the easternmost end of the station network, approximately 168 km in a straight line from the nearest research station, AMER, while the other six stations are mainly concentrated along the western coast. SBCH is also the only station in Table 1 to show negative migration throughout the entire time period, indicating that weak spatial proximity and its accompanying environmental differences may increase the distribution differences of cross-station data; this spatial relationship is only used to illustrate the characteristics of the results and is not the sole cause of negative migration.
[0095] At the same effective epoch for each target station, the original GNSS-IR water level, robust sliding window estimation, B-spline fitting, and the cross-site machine learning results of this invention were compared. The root mean square error was calculated after removing the average error for each evaluation period; the improvement rate was taken as the traditional baseline, using the smaller error between the robust estimation and the B-spline fitting. The full-time results for the eight stations are shown in Table 1.
[0096] Table 1. Validation results of the eight target stations throughout the entire time period (root mean square error in cm)
[0097]
[0098] As shown in Table 1, the average root mean square error of the eight stations decreased from 9.4 cm to 4.5 cm from the original GNSS-IR water level. The average improvement rate relative to the better traditional baseline of each station was 9.1%, with seven stations achieving positive improvement. The ROPS station had the highest improvement rate at 21.4%.
[0099] As can be seen from the station distribution map, AMER, CALC, ROPS, N301, TXNJ, TXPO, and TSAQ are mainly distributed in relatively continuous coastal areas, while SBCH is located at the eastern end of the study area, showing a significant spatial separation from the other stations, and is considered an edge target station in the station network. Furthermore, the local shoreline and reflection environment of the section where SBCH is located differ significantly from most source stations. Therefore, when migrating the cross-site model based on the other stations to SBCH, the available similar source station information is relatively limited, resulting in a root mean square error of 4.3 cm over all time periods, slightly higher than the 4.2 cm of the traditional baseline, corresponding to an improvement rate of [missing information]. The results indicate that when the target station differs significantly from the source station in spatial location and local reflection environment, the cross-site model may exhibit slight negative transfer.
[0100] To further analyze the correction effects across different time periods, 2023 at the TSAQ station, 2021 at the AMER station, and 2022 at the SBCH station were selected as representative annual cases, and the same effective epoch selection and error calculation principles as the full-time evaluation were adopted. The results are shown in Table 2 and... Figures 2-5 As shown.
[0101] Table 2. Water level correction results for three representative sites in annual cases (root mean square error, unit: cm)
[0102]
[0103] In the three annual cases, the root mean square error of the method of this invention was 3.3 cm, 3.2 cm, and 3.0 cm, respectively, with correlation coefficients of 97.6%, 98.9%, and 99.1%. Among them, the correction results of TSAQ and AMER stations were better than the traditional baseline for the corresponding years; the root mean square error of SBCH station in 2022 further decreased from 3.2 cm of the traditional better baseline to 3.0 cm, indicating that the slight negative migration that occurred at SBCH throughout the entire period did not continue in all periods, but may be related to the characteristics of water level changes, observation quality, and changes in local reflection environment in different periods.
[0104] Based on comprehensive data from all time periods and representative annual cases, the method of this invention can utilize accumulated GNSS-IR observations and tide gauge reference data from multiple source stations to correct water levels at target stations lacking long-term historical labels, achieving results superior to traditional smoothing or fitting methods on most target stations. For marginal target stations that are spatially distant from the source stations or have significantly different shoreline morphology or reflective environments, the model transfer effect may be weakened. If necessary, this can be addressed by adding source stations with similar environments, introducing a small number of labeled samples from the target station for model adaptation, or screening the source stations to improve the stability of the target station correction results.
[0105] Specifically, the embodiments of this invention illustrate the technical process and implementation effects of the cross-site machine learning water level correction method of this invention. Implementation results show that this method has good cross-site applicability to most target stations, but its specific correction accuracy is related to the source station data coverage, the environmental similarity between the source and target stations, the observation quality of the target station, and the model parameter settings. Therefore, the number of stations, research area, time range, feature combination, and model parameters proposed in the embodiments of this invention are merely examples and do not constitute a limitation on the scope of protection of this invention.
[0106] The implementation of the various embodiments of the present invention is based on programmed processing by a device with processor functionality. Therefore, in practical engineering, the technical solutions and functions of the various embodiments of the present invention are encapsulated into various modules. Based on this reality, and building upon the above embodiments, the embodiments of the present invention provide a nearshore GNSS-IR sea surface level correction system based on cross-site machine learning. This system is used to execute a nearshore GNSS-IR sea surface level correction method based on cross-site machine learning from the above method embodiments.
[0107] The system includes: a data acquisition module for acquiring the raw GNSS-IR total water level of multiple source stations and at least one target station, while also acquiring the reference total water level of the tide gauge corresponding to each source station, as well as the signal quality and interferometric state information, meteorological information, and site environmental information corresponding to the source and target stations; an astronomical tide construction module for performing harmonic analysis on a preset astronomical tide set based on the raw GNSS-IR total water level of each source and target station, the reference total water level of the tide gauge corresponding to each source station, and their respective irregular observation epochs, estimating preset tidal parameters, and constructing the astronomical tide sequence of the source station tide gauge and the GNSS-IR astronomical tide sequence of each station respectively; an epoch registration module for time registration of the reference total water level of the tide gauge corresponding to each source station and the external dynamic variables of the source and target stations, using the raw GNSS-IR observation epochs of each station as a unified time reference; and a window sample construction module for using the original GNSS-IR observation epochs of each station as a unified time reference. The central epoch establishes a window, constructing feature vectors based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state characteristics. The residual water level of each source station is obtained by subtracting the astronomical tide sequence of the source station's tide gauge from the reference total water level of the corresponding tide gauge station, and used as a supervised learning label. A cross-site regression model training module is used to construct a cross-site training sample set based on the feature vectors and supervised learning labels of multiple source stations. The cross-site regression model is trained using this cross-site training sample set, and feature preprocessing parameters and model hyperparameters are determined based on the cross-site training sample set. A target station total water level acquisition module is used to input the target station feature vector into the trained cross-site regression model to obtain the predicted residual water level of the target station, and then add the predicted residual water level of the target station to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
[0108] The nearshore GNSS-IR sea surface level correction system based on cross-site machine learning provided in this invention addresses the problems of existing single-site GNSS-IR machine learning correction methods, such as the need for long-term accumulation of historical labels for the target station, the difficulty in directly transferring models from neighboring tide gauge stations, the easy incorporation of tidal differences between stations when directly predicting total water level, and the difficulty in uniformly constructing cross-site features under irregular sampling conditions. This system employs several modules to comprehensively utilize historical GNSS-IR and tide gauge data from multiple existing stations to establish transferable relationships, enabling direct water level correction even when the target station lacks tide gauge labels. Furthermore, this method should adapt to the non-equidistant characteristics of raw GNSS-IR observations and clearly distinguish between total water level, astronomical tides, and residual water level, allowing the model to predict the relatively more transferable residual water level between different stations, ultimately outputting a complete corrected total water level.
[0109] Based on the same inventive concept as the foregoing embodiments, this embodiment of the invention also provides an electronic device, including 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 realize a nearshore GNSS-IR sea surface level correction method based on cross-site machine learning as proposed in the above embodiments.
[0110] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the program is used to process irregular sampling of GNSS-IR observations and implement cross-site water level correction.
[0111] Finally, it should be noted that the above specific embodiments are merely representative examples of the present invention. Obviously, the present invention is not limited to the above specific embodiments and many variations are possible. Any simple modifications, equivalent changes, and alterations made to the above specific embodiments based on the technical essence of the present invention should be considered within the protection scope of the present invention.
Claims
1. A nearshore GNSS-IR sea surface level correction method based on cross-site machine learning, characterized in that, include: The original GNSS-IR total water level of multiple source stations and at least one target station is obtained. At the same time, the reference total water level of the tide gauge corresponding to each source station, as well as the signal quality and interference status information, meteorological information and site environmental information corresponding to the source station and the target station are obtained. Based on the original GNSS-IR total water level of each source station and target station, the reference total water level of each source station's corresponding tide gauge station, and their respective irregular observation epochs, a harmonic analysis is performed on the preset astronomical tide set to estimate the preset tide parameters and construct the source station tide gauge station astronomical tide sequence and the GNSS-IR astronomical tide sequence of each station respectively. Using the original GNSS-IR observation epochs of each station as a unified time reference, time registration was performed on the reference total water level of the tide gauge corresponding to each source station and the external dynamic variables of the source and target stations. A window is established with the central epoch of each station. Based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state characteristics, a feature vector is constructed. The astronomical tide sequence of the source station's tide gauge is subtracted from the reference total water level of the tide gauge corresponding to each source station to obtain the residual water level of the tide gauge and use it as a supervised learning label. A cross-site training sample set is constructed based on the feature vectors and supervised learning labels of multiple source stations. The cross-site regression model is trained using the cross-site training sample set, and the feature preprocessing parameters and model hyperparameters are determined based on the cross-site training sample set. The target station feature vector is input into the trained cross-site regression model to obtain the predicted residual water level of the target station. The predicted residual water level of the target station is then added to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
2. The nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 1, characterized in that, Obtain the raw GNSS-IR total water level from multiple source stations and at least one target station, including: For multiple source stations and at least one target station, acquire the GNSS observation files recorded by the GNSS receivers of each station, as well as the broadcast ephemeris files corresponding to the observation period; The satellite elevation and azimuth angles are calculated based on the GNSS observation files and broadcast ephemeris files, and the signal-to-noise ratio observation sequence is extracted based on the GNSS observation files; Based on the signal-to-noise ratio observation sequence, a detrended signal-to-noise ratio sequence dominated by water surface reflection interference oscillations is obtained through trend separation; Based on the detrended signal-to-noise ratio sequence dominated by water surface reflection interference oscillation, the Lom-Skagler periodogram analysis was used to extract the main frequency of the spectral peak, and the height of the reflecting surface was calculated in combination with the carrier wavelength. The original GNSS-IR total water level is obtained based on the height of the reflector surface and the elevation of the antenna phase center under a unified reference datum.
3. The nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 1, characterized in that, Time registration includes: Search for the neighboring valid epochs of the variable to be registered, and the corresponding observations, on both sides of each GNSS-IR observation epoch. Linear interpolation is performed under the condition of satisfying the preset maximum allowable time interval constraint to complete the time registration between the variable to be registered and the original epoch of GNSS-IR observation; If there are no adjacent valid epochs or the interval between adjacent epochs exceeds the preset maximum allowable interval, the sample corresponding to the current GNSS-IR raw observation epoch will be marked as invalid and deleted.
4. The nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 1, characterized in that, The interference state information includes one or more of the following: peak-to-noise ratio, full width at half maximum (FWHM), amplitude, damping coefficient, phase parameter, and satellite elevation angle change rate.
5. The nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 1, characterized in that, The irregular sampling state characteristics include at least the number of window observations, the number of observations before and after the central epoch, the most recent time interval, the maximum time interval within the window, the window time coverage, and the time balance index.
6. The nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 1, characterized in that, The GNSS-IR residual water level is obtained by subtracting the GNSS-IR astronomical tide sequence of each station from the original total GNSS-IR water level of each station.
7. A nearshore GNSS-IR sea surface level correction method based on cross-site machine learning according to claim 3, characterized in that, The expression for time registration is: , Where j is the index of the variable to be registered. Let j be the j-th variable to be registered, and t be the target epoch. and This indicates the most recent valid epoch before or after t.
8. A nearshore GNSS-IR sea surface level correction system based on cross-site machine learning, characterized in that, To implement the nearshore GNSS-IR sea surface level correction method based on cross-site machine learning as described in any one of claims 1-7, comprising: The data acquisition module is used to acquire the raw GNSS-IR total water level of multiple source stations and at least one target station, and at the same time acquire the reference total water level of the tide gauge station corresponding to each source station, as well as the signal quality and interference status information, meteorological information and station environmental information corresponding to the source station and the target station. The astronomical tide construction module is used to perform harmonic analysis on the preset astronomical tide set based on the original GNSS-IR total water level of each source station and target station, the reference total water level of the tide gauge station corresponding to each source station, and their respective irregular observation epochs, to estimate the preset tide parameters and construct the astronomical tide sequence of the source station tide gauge station and the GNSS-IR astronomical tide sequence of each station respectively. The epoch registration module is used to time-register the reference total water level of the tide gauge station corresponding to each source station and the external dynamic variables of the source and target stations, using the original epochs of GNSS-IR observations at each station as a unified time reference. The window sample construction module is used to establish windows based on the central epoch of each station. It constructs feature vectors based on the original GNSS-IR total water level, GNSS-IR residual water level, signal quality, interferometric state information, meteorological information, station environmental information, and irregular sampling state characteristics. The astronomical tide sequence of the source station's tide gauge is subtracted from the reference total water level of the tide gauge corresponding to each source station to obtain the residual water level of the tide gauge and use it as a supervised learning label. The cross-site regression model training module is used to construct a cross-site training sample set based on the feature vectors and supervised learning labels of multiple source sites, train the cross-site regression model using the cross-site training sample set, and determine the feature preprocessing parameters and model hyperparameters based on the cross-site training sample set. The target station total water level acquisition module is used to input the target station feature vector into the trained cross-site regression model to obtain the predicted residual water level of the target station, and add the predicted residual water level of the target station to the astronomical tide sequence of the target station to obtain the corrected total water level of the target station.
9. An electronic 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 nearshore GNSS-IR sea surface level correction method based on cross-site machine learning as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the nearshore GNSS-IR sea surface level correction method based on cross-site machine learning as described in any one of claims 1 to 7.