Satellite altimetry coastal wtc correction method based on neural network

Abstract

The application belongs to the technical field of satellite altimetry, and specifically discloses a satellite altimetry coastal area WTC correction method based on a neural network. The application constructs a multi-source feature neural network fusion system for satellite altimetry coastal area WTC correction in view of the problems of insufficient model fitting capability of traditional linear / physical models and single model input feature dimension, fuses multi-dimensional meteorological parameters of ERA5 reanalysis data, satellite original WTC and off-shore distance of an observation point and the like features, takes high-precision GNSS_ZWD as a learning target, and realizes high-precision correction of the coastal area WTC by using an optimized neural network model. The application takes the satellite original WTC, the reanalysis multi-dimensional meteorological parameters and the off-shore distance of the observation point as joint inputs, constructs a correction model with stronger physical constraints by relying on the neural network, and significantly improves the precision and robustness of the WTC correction under the complex near-shore environment.

Description

Technical Field

[0001] This invention belongs to the field of satellite altimetry technology, and specifically relates to a satellite altimetry WTC correction method for coastal areas based on neural networks. Background Technology

[0002] Satellite radar altimetry is a core remote sensing tool for monitoring global and regional sea-level changes, playing a crucial role in sea-level change research and coastal marine environmental monitoring. During radar signal propagation, it is affected by various factors such as the dry / wet troposphere, ionosphere, sea surface roughness, and tides, leading to ranging errors. Therefore, distance correction and geophysical correction are necessary to extract the true water surface height. Wet tropospheric correction (WTC) is one of the most significant sources of error, with an absolute value reaching up to 0.5 meters and exhibiting extremely strong spatiotemporal variability, directly impacting observation accuracy.

[0003] To correct for this error, most mainstream altimeter satellites are equipped with microwave radiometers (MWRs). By measuring atmospheric brightness temperature (WTC), they retrieve WTC with centimeter-level accuracy in open sea areas. However, in coastal areas, when the satellite passes near the shore, the MWR antenna beam partially covers the land. The significant difference between the radiation characteristics of land and ocean severely contaminates the brightness temperature data, leading to a systematic bias in the retrieved WTC, making it difficult to meet the requirements for high-precision correction.

[0004] This WTC bias caused by land pollution restricts applications closely related to disaster prevention and mitigation, and ecological protection, such as nearshore sea level anomalies and storm surge warnings, becoming a major bottleneck for the application of satellite altimetry technology in nearshore areas. To address the issue of WTC accuracy degradation due to land pollution in coastal MWR observations, existing technologies mainly include: (1) Alternative methods based on reanalysis data, namely, directly estimating ZWD using global atmospheric reanalysis data such as ERA5. This type of method can provide a spatiotemporally continuous meteorological field, but its resolution (spatial: 0.25°×0.25°, temporal: 1 hour) is difficult to fully match the scale requirements of altimetry observations, and is affected by the errors of the reanalysis model itself.

[0005] (2) Based on the GNSS path delay moist tropospheric correction model (GPD / GPD+), this method integrates IGS global GNSS station data, MWR observations, and atmospheric models, and uses optimal interpolation to replace WTC. This method works well in nearshore waters, but the optimal interpolation is a linear method, which is difficult to characterize the nonlinear changes in water vapor. Traditional linear / physical models have insufficient model fitting ability, and nonlinear biases cannot be corrected, failing to characterize the nonlinear characteristics of water vapor distribution at the coastal-land interface, resulting in limited accuracy of land pollution bias correction.

[0006] (3) Based on machine learning modeling methods, existing studies have used neural networks to correct the WTC of specific satellites. However, the input feature dimension of the model is singular, and most of them only use the original satellite observation data. They do not make full use of meteorological parameters that are strongly correlated with wet delay, such as air pressure, temperature, humidity, and water vapor content, which limits the model's ability to fit complex water vapor changes.

[0007] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0008] The purpose of this invention is to propose a satellite altimeter WTC correction method for coastal areas based on neural networks. By integrating multidimensional meteorological parameters from ERA5 reanalysis data, the original satellite WTC, and the distance of the observation point from the shore, and using high-precision GNSS_ZWD as the learning target, an optimized Elman neural network model is used to achieve high-precision correction of WTC in coastal areas.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A satellite altimeter WTC correction method for coastal areas based on neural networks includes the following steps: First, GNSS stations located in coastal areas that match the spatiotemporal range of the satellite's transit path are selected, and the ZWD at these stations is estimated. Then, the estimated ZWD is spatiotemporally matched with the original WTC value of the satellite. Finally, the station elevations are unified through elevation unification. place Elevation converted to satellite altimetry observation point place ; Step 2. Obtain meteorological parameters at satellite altimetry observation points based on ERA5 single-layer surface reanalysis data; First, based on ERA5 single-layer surface reanalysis data, meteorological parameters were extracted from four grid points adjacent to the satellite altimetry observation point; Then, each grid point is set at the grid elevation. Meteorological parameters at the location were corrected to the elevation of the satellite altimetry observation point. Then, using bilinear interpolation, the final air pressure at the satellite altimeter observation point is calculated. ,temperature and total column water vapor ; Reusing grid point temperature and dew point temperature Estimate relative humidity After elevation correction, bilinear interpolation is performed on the calculation results to estimate the water vapor pressure at the satellite altimetry observation point. and relative humidity ; Step 3. Construct a training dataset for training the neural network; The samples in the training dataset consist of model input features and labels; the model input features include ERA5 meteorological parameters. , , , , The original WTC value of the satellite and the distance D from the shore of the satellite altimetry observation point; The labels are obtained after processing in step 1, and they correspond one-to-one with the input features of the model mentioned above. ; Step 4. Build a WTC correction model for coastal areas based on Elman neural network. Train the model using the training dataset from Step 3, and then use the trained model to correct the original WTC values ​​of the satellite.

[0010] Furthermore, based on the aforementioned satellite altimeter WTC correction method for coastal areas based on neural networks, this invention also proposes a computer device that includes a memory and one or more processors.

[0011] The executable code is stored in memory. When the processor executes the executable code, it implements the steps of the aforementioned neural network-based satellite altimetry WTC correction method for coastal areas.

[0012] Furthermore, based on the aforementioned satellite altimeter WTC correction method for coastal areas based on neural networks, this invention also proposes a computer-readable storage medium storing a program that, when executed by a processor, implements the steps of the aforementioned satellite altimeter WTC correction method for coastal areas based on neural networks.

[0013] The present invention has the following advantages: As described above, this invention discloses a neural network-based method for correcting water vapor temperature (WTC) over coastal areas using satellite altimetry. This method integrates multidimensional meteorological parameters from ERA5 reanalysis data, the original satellite WTC, and the distance of observation points from the shore. Using high-precision GNSS_ZWD as the learning objective, it employs an optimized Elman neural network model to achieve high-precision correction of WTC over coastal areas. Compared to traditional optimal interpolation based on linear models, which struggles to characterize the drastic nonlinear changes in nearshore water vapor, this invention utilizes neural network nonlinear mapping to adaptively fit complex deviations under different coastal types and atmospheric conditions, overcoming linear limitations and adapting to complex coastal water vapor variations. Regarding model input, this invention enhances the physical consistency of the correction by incorporating atmospheric physical parameters. By introducing physical quantities such as air pressure, temperature, water vapor pressure, relative humidity, and total column water vapor from ERA5, a clear physical correlation is established between the input features and the wet delay, resulting in strong model interpretability. Furthermore, due to the model's input features being only 7-dimensional and the network parameters being small, single-point inference is less than 1ms, enabling efficient processing of hundreds of thousands of altimetry data points across the entire orbit, meeting operational post-processing requirements. This invention enables the training of multiple satellite data types, including Sentinel-3A (Ku / C-band) and Saral (Ka-band), within the same feature space and network structure. The model automatically learns the system biases of different payloads, eliminating the need to design a separate correction model for each satellite and exhibiting good multi-satellite and multi-frequency band compatibility. Furthermore, this invention requires no hardware modification, is easy to deploy, and all input data comes from publicly available operational products (satellite data products, ERA5, GNSS tropospheric products). It eliminates the need to obtain underlying MWR brightness temperature or modify onboard firmware, and can be directly integrated into existing satellite altimetry processing workflows. Attached Figure Description

[0014] Figure 1 This is a flowchart of the satellite altimeter WTC correction method for coastal areas based on neural networks in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the optimization of Elman neural network parameters using the sparrow algorithm in an embodiment of the present invention; Figure 3 This is a structural diagram of the coastal WTC correction model based on the Elman neural network in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Definitions: WTC (Wave Trace) is a key error source in satellite altimetry. WTC and ZWD (Zero-Wave Difference) are essentially the same, both reflecting the delay effect of atmospheric water vapor on the propagation path of electromagnetic waves. ZWD is a term from the GNSS field, while WTC is from satellite altimetry. The purpose of this invention is to correct the original WTC accuracy degradation caused by land pollution in coastal areas. ZWD is used only as a training label; the goal is to correct the systematic bias in WTC caused by land pollution, and the output is a corrected, high-precision WTC.

[0016] Example 1 This embodiment 1 describes a satellite altimeter WTC correction method for coastal areas based on neural networks. This method addresses the aforementioned technical problems in existing technologies by innovatively proposing a multi-source feature neural network fusion system for WTC correction in coastal areas based on satellite altimeters. By using the original satellite WTC, reanalysis multidimensional meteorological parameters, and the distance of the observation point from the shore as joint inputs, a correction model with stronger physical constraints is constructed based on the neural network. This fully captures the complex correlation between coastal land pollution and water vapor changes, significantly improving the accuracy and robustness of WTC correction in complex nearshore environments.

[0017] like Figure 1 As shown, the WTC correction method for coastal areas based on satellite altimetry using neural networks includes the following steps: Step 1. First, select GNSS stations located in coastal areas that match the spatiotemporal range of the satellite's transit path, and estimate the ZWD at each station. Then, perform a spatiotemporal double matching between the estimated ZWD and the original WTC value of the satellite. The station elevations after the double matching are then unified through elevation unification. place Elevation converted to satellite altimetry observation point place .

[0018] When Global Navigation Satellite System (GNSS) signals and altimeter signals propagate in the neutral atmosphere, they are affected by the same water vapor environment, resulting in physically consistent path delays.

[0019] The zenith wet delay (ZWD) retrieved by GNSS has millimeter-level observation accuracy, and WTC and ZWD have the same physical essence, both reflecting the delay effect of atmospheric water vapor on the propagation path of electromagnetic waves.

[0020] Therefore, GNSS ZWD is widely used as independent reference data for evaluating, verifying, and monitoring the WTC of coastal altimetry satellites.

[0021] Step 1.1. GNSS station selection and ZWD estimation.

[0022] I. GNSS station selection.

[0023] GNSS stations located within a first preset distance range on the landward side of the coastline are selected. This selection criterion ensures that the selected stations are located in the nearshore area where land pollution has the most significant impact, so as to truly reflect the characteristics of atmospheric water vapor changes along the coast.

[0024] Land pollution from satellite microwave radiometers occurs on nearshore sea surfaces.

[0025] However, to obtain reference ZWD data, there are no long-term operating GNSS reference stations on the sea surface. Therefore, it is necessary to select shore-based GNSS stations to indirectly reflect the atmospheric wet delay state of the nearshore sea area.

[0026] The nearshore area where the impact of land pollution is most significant refers to the spatial location represented by the GNSS station being near the sea area affected by land pollution, that is, the station is located in the nearshore land area within a certain distance from the coastline.

[0027] This processing method can accurately reflect the characteristics of atmospheric water vapor changes along the coast. The first preset distance can be set to 50km.

[0028] In this embodiment, a first preset distance (≤50 km) is set to screen nearshore GNSS stations to ensure that their water vapor environment is related to the water vapor environment of the nearshore sea area, serving as a basic constraint for the true data.

[0029] II. ZWD estimation.

[0030] For IGS stations providing ZTD, the Saastamoinen model is used to separate them. : ;in , The ground air pressure at the station. The latitude of the measuring station. This is the elevation of the measuring station.

[0031] Further based on The calculation of the station location The formula is as follows: .

[0032] Step 1.2. Perform a spatiotemporal double match between the estimated ZWD and the original WTC value of the satellite.

[0033] Because the location of ground GNSS stations is fixed, while satellite altimetry observation points are dynamic observation points that are discretely distributed along the orbit.

[0034] Ground-based GNSS stations and satellite altimetry observation points cannot completely overlap in spatial location and observation time. Strict spatiotemporal matching is necessary to ensure the spatiotemporal consistency and physical relevance of training samples.

[0035] Following the estimation in step 1.1 This is used to perform spatial and temporal matching with satellite altimetry WTC values ​​to ensure... It corresponds one-to-one with satellite observations and is consistent with their physical state.

[0036] I. Spatial Filtering: The distance between the satellite altimeter observation point and the GNSS station selected in step 1.1 must not exceed the second preset distance range to ensure that the atmospheric conditions of both are highly representative in space. The second preset distance is set, for example, to 100 km.

[0037] In this embodiment, the second preset distance (≤100 km) is used to limit the spatial distance between the satellite observation point and the GNSS station, ensuring that the two are in the same atmospheric water vapor field and that the water vapor state has strong spatiality.

[0038] II. Time Filtering: The time difference between the satellite transit time and the GNSS observation time is set to not exceed a preset time window (in this embodiment, the preset time window is set to, for example, 30 minutes) to ensure that the water vapor field is approximately stable in the time dimension.

[0039] Data outside this time window is discarded to avoid introducing mismatch errors caused by rapid atmospheric changes.

[0040] For those that meet the condition that the time difference is less than the preset time window Then, linear interpolation is used (the process is fairly standard and will not be elaborated further) to extrapolate the time to the satellite altimetry observation point, thus obtaining the synchronization between the station and satellite observations. .

[0041] Step 1.3. Elevation unification.

[0042] Step 1.1 The estimated GNSS_ZWD is the elevation of the station. The WTC satellite corresponds to the sea surface (satellite altimeter observation point) elevation, and the inconsistency between the two elevations will introduce a significant deviation.

[0043] Because the elevation of the GNSS station differs from the elevation of the satellite altimetry observation point (sea level), this invention employs an exponential model to adjust the elevation of the station. Elevation converted to satellite altimetry observation point place The formula is as follows: .

[0044] in and For the elevation of the station and the elevation of the satellite altimetry observation point, The empirical attenuation coefficient is taken as the global average of 2000.

[0045] Step 2. Obtain meteorological parameters at satellite altimetry observation points based on ERA5 single-layer surface reanalysis data.

[0046] First, based on ERA5 single-layer surface reanalysis data, meteorological parameters were extracted from four grid points adjacent to the satellite altimetry observation point.

[0047] Then, each grid point is set at the grid elevation. Meteorological parameters at the location were corrected to the elevation of the satellite altimetry observation point. Then, using bilinear interpolation, the final air pressure at the satellite altimeter observation point is calculated. ,temperature and total column water vapor .

[0048] Reusing grid point temperature and dew point temperature Estimate relative humidity After elevation correction, bilinear interpolation is performed on the calculation results to estimate the water vapor pressure at the satellite altimetry observation point. and relative humidity .

[0049] Step 2.1. In order to obtain meteorological parameters that match the satellite altimetry observation location and elevation, this invention extracts meteorological parameters from four grid points adjacent to the satellite altimetry observation point based on ERA5 single-layer surface reanalysis data.

[0050] Based on the relationship between meteorological parameters and elevation, the four grid points were respectively located at the grid elevation. Meteorological parameters at the location were corrected to the elevation of the satellite altimetry observation point. The formula is as follows: ; .

[0051] Among them, grid elevation Meteorological parameters at the location include air pressure. ,temperature and total column water vapor Then grid elevation The total column of water vapor Corrected to satellite altimetry observation elevation The formula is as follows: .

[0052] The above formulas are used to obtain the air pressure, temperature, and elevation correction values ​​of the total column water vapor at four adjacent grid points.

[0053] Then, bilinear interpolation is performed on all meteorological parameters of four adjacent grid points after elevation correction to obtain the final air pressure at the satellite altimetry observation point. ,temperature and total column water vapor .

[0054] Step 2.2. Utilize the temperature of the grid points and dew point temperature Estimate relative humidity After elevation correction, the water vapor pressure at the satellite altimeter observation point is estimated. and relative humidity .

[0055] This invention relates to the elevation of satellite altimetry observation points. Water vapor pressure at the location and relative humidity Similarly, the estimation process is performed using four grid points for calculation, followed by bilinear interpolation of the results. The estimation process is as follows: First, utilize the grid point temperature and dew point temperature Estimate relative humidity After elevation correction, it is used to estimate the elevation of the altimeter observation point. The specific estimation is shown in the following formula: ; ; ; ; .

[0056] In the formula For saturated water vapor pressure, relevant parameters , , The value is determined by the grid point temperature. Scope is determined.

[0057] When temperature For temperatures above 0℃, take R2 = 6.112 hPa, R3 = 17.502 K, and R4 = 32.19 K.

[0058] When temperature When the temperature is below -23℃, take R3 = 22.587K and R4 = -0.7K.

[0059] When -23℃ < When < 0℃, use the following formula for estimation; .

[0060] This formula represents the pressure of saturated water vapor above 0°C. With saturated water vapor pressure below -23℃ Estimate the saturated vapor pressure for this temperature range.

[0061] After the above processing, the elevations of four adjacent grid points are obtained after elevation correction to the satellite altimetry observation points. Water vapor pressure at the location and relative humidity .

[0062] Then, the final water vapor pressure at the satellite altimetry observation point was calculated using bilinear interpolation. and relative humidity .

[0063] Step 3. Construct a training dataset for training the neural network.

[0064] The samples in the training dataset consist of model input features and labels. Model input features include ERA5 meteorological parameters. , , , , The original WTC value of the satellite and the distance D from the shore of the satellite altimetry observation point.

[0065] The labels are obtained after processing in step 1, and they correspond one-to-one with the input features of the model mentioned above. .

[0066] The original WTC value of the satellite is the brightness temperature directly observed by the onboard MWR of the altimeter satellite and the wet tropospheric correction obtained by inversion through the satellite's official algorithm. It is the original correction item in the satellite's Level-2 standard product used to eliminate the delay of radar ranging caused by water vapor, and has not undergone any post-processing optimization.

[0067] The purpose of this invention is to correct the original WTC (Wide-Time Chart) accuracy loss caused by land pollution in coastal areas. ZWD (Zero-Wide Dimensioning) is used only as a training label. The learning objective is to correct the WTC system bias caused by land pollution. The output is the corrected, high-precision WTC.

[0068] Specifically, the sample used for neural network training is represented as follows: Sample={ , , , , ,MWR_WTC,D, }

[0069] Step 4. Build a coastal WTC correction model based on Elman neural network. Train the model using the training dataset from Step 3, and then use the trained model to correct the original WTC values ​​of the satellite.

[0070] To ensure the effectiveness and generalization ability of the model, a time series segmentation strategy is used for sample allocation.

[0071] Specifically, the earlier consecutive time periods in the acquired training dataset are used as the training set for Elman model training and validation, while subsequent consecutive time periods are used as the test set for independent testing. Within the training set, the first 80% in chronological order is used as the training subset, and the last 20% is used as the validation subset to monitor the training process and adjust hyperparameters.

[0072] In the data preprocessing stage, the Elman model uses Z-score standardization to standardize the input features and labels of all samples in the training dataset in order to eliminate the influence of units and improve model stability.

[0073] In addition, existing neural network-based correction methods rely heavily on manual experience or grid search to determine key hyperparameters, which is not only time-consuming but also makes it difficult to guarantee global optimality and easily causes the model to get trapped in local minima.

[0074] To address the problem that neural network hyperparameters rely on human experience and trial and error and lack an adaptive optimization mechanism, this invention utilizes the Sparrow Search Algorithm (SSA) to adaptively optimize the hyperparameters of Elman neural networks.

[0075] This approach significantly improves the generalization ability and convergence efficiency of Elman neural networks in complex coastal environments.

[0076] Hyperparameter vector to be optimized This includes: the number of hidden layer nodes H, the initial learning rate lr, the weight decay coefficient ω, the Dropout ratio d, and the number of hidden layers L, i.e., the hyperparameter vector: =[H, lr, ω, d, L]. The optimal number of hidden layer nodes is obtained through optimization. Optimal learning rate lr best Optimal weight decay coefficient ω best Optimal Dropout ratio d best And the optimal number of hidden layers L best .

[0077] like Figure 2 As shown, the core mechanism of SSA optimization is to simulate the role division and position update strategy of "discoverer-follower-early warning" in a sparrow population. The specific implementation steps are as follows: S1: Initialize the sparrow population and algorithm parameters.

[0078] Set the sparrow population size to N (8), and the maximum number of iterations to Tmax (10). Randomly initialize the positions of N sparrows in the solution space, and create a position vector for each sparrow. This corresponds to a set of hyperparameters to be optimized.

[0079] The boundary constraints of the solution space are: .

[0080] S2: Calculate the initial fitness value.

[0081] For each sparrow in the initial population Perform the following operations: I. Construct an Elman neural network based on the hyperparameters it represents.

[0082] II. Use the training dataset constructed in step 3 to perform initial training on the network.

[0083] III. Evaluate the trained model using the validation set, and use the mean squared error (MSE) on the validation set as the fitness value of the sparrow optimization algorithm. The smaller the fitness value, the better the performance of the hyperparameter group.

[0084] S3: Iteratively update the sparrow's position.

[0085] In the t-th iteration, based on the sparrow's role, its position is updated according to the following rules: Discoverer location update: The discoverer is responsible for providing foraging directions for the entire population. Its location update formula is: ; in This represents the position coordinates of the i-th sparrow (individual) in the j-th dimension (search space dimension) during the t-th iteration. Let represent the position coordinates of the i-th sparrow in the j-th dimension at the (t+1)-th iteration (i.e., the updated position).

[0086] t represents the current iteration number, and j=1,2,…,5 represents the dimension of the hyperparameters. ∈(0,1] is a random number. ∈[0,1] is the warning value. ∈[0.5,1] is the safety threshold (set to 0.6 in this invention). For random numbers that follow a normal distribution, It is a matrix of all 1s.

[0087] when When the population is safe, the discoverer conducts a wide search; otherwise, all sparrows fly to a safe area.

[0088] Follower location update: Followers update their positions based on the discoverer's guidance, and their update strategy is as follows: .

[0089] in, The current position of the best discoverer. The position of the sparrow with the worst overall score. A matrix whose elements are randomly assigned the value 1 or -1, and .

[0090] Location update of the alerter: A random selection of sparrows was chosen as early warning systems; when they sensed danger, they engaged in predatory behavior. .

[0091] in, These are step size control parameters. ∈[-1,1] is a random number. Based on the current adaptability of sparrows, and These are the current global best fitness and worst fitness, respectively. It is a very small constant.

[0092] S4: Update the global optimal solution.

[0093] Calculate the fitness of each sparrow in the new generation population and update the global optimum position in the current iteration. and .

[0094] S5: Convergence test.

[0095] Repeat steps S3 and S4 until the preset maximum number of iterations Tmax is reached or the optimal fitness no longer decreases after multiple consecutive rounds. Output the final global optimum position. This corresponds to a set of optimal Elman neural network hyperparameters.

[0096] This invention constructs an SSA-optimized neural network model for WTC correction in coastal areas of satellite altimetry.

[0097] By leveraging the global parallel search capability of SSA, key hyperparameters such as the number of hidden layer nodes, learning rate, and regularization coefficient of Elman networks are adaptively optimized. This solves the technical problem of the difficulty in determining the optimal hyperparameters of deep learning models and their tendency to get trapped in local optima, and significantly improves the generalization ability and correction accuracy of the model.

[0098] like Figure 3 The network structure of the Elman neural network in this embodiment is shown.

[0099] The optimal hyperparameters obtained based on SSA optimization An Elman neural network was constructed as the correction model.

[0100] The Elman neural network introduces a support layer on the basis of the traditional feedforward network, which enables it to remember historical state information, thereby effectively capturing the continuous evolution of atmospheric water vapor content and the impact of land pollution in the time dimension.

[0101] An Elman neural network consists of four parts: an input layer, hidden layers, a continuation layer, and an output layer. Let the number of nodes in the input layer be I, and the number of hidden layers be L. best The number of hidden layer nodes is H best The number of output layer nodes is O=1, corresponding to the corrected WTC output.

[0102] The number of nodes in the receiving layer is the same as that in the hidden layer, and it is used to store the state of the hidden layer at the previous moment.

[0103] The number of nodes in the input layer is I, corresponding to the model input features, namely the air pressure at the satellite altimetry observation point. Temperature Water vapor pressure relative humidity Column water vapor The original WTC of the satellite and the distance D from the shore of the observation point.

[0104] Let the input vector at time t be... The hidden layer state vector is The state vector of the receiving layer is The output layer result is The present invention directly takes the hidden layer output from the previous time step as the receiving layer state, that is: .

[0105] The hidden layer state update uses the Tanh nonlinear activation function, the mathematical expression of which is: .

[0106] in, This is the weight matrix from the input layer to the hidden layer. This is the weight matrix from the receiving layer to the hidden layer. This is the hidden layer bias vector.

[0107] To prevent overfitting and accelerate convergence, this invention adds a batch normalization layer and a Dropout layer sequentially after the hidden layer, wherein the Dropout ratio is set to d. best The output layer uses a linear mapping to directly output the corrected WTC, and the estimation formula is as follows: ; in This is the weight vector from the hidden layer to the output layer. This is the output layer bias.

[0108] This invention establishes a universal correction framework for satellite altimetry WTC that does not rely on underlying satellite data. It uses only publicly available standardized data and is compatible with multiple types of altimetry satellites, thus solving the technical defects of traditional correction methods, such as poor universality and high deployment threshold.

[0109] During the model training phase, the Adam optimizer is used to optimize the network parameters θ={ , , , , Iterative updates are performed, with the initial learning rate set to lr. best The weight decay coefficient is set to ω. best .

[0110] The loss function uses mean squared error (MSE) for a batch of samples. loss function Defined as: ; in Input features to the model, This represents the model's predicted value. This indicates the number of samples in each batch.

[0111] To improve training stability, a gradient pruning mechanism is introduced, limiting the maximum L2 norm of the gradient to 1.0. An early stopping strategy is also employed: if the validation set loss does not decrease for 50 consecutive rounds, training is terminated and the model parameters with the optimal validation loss are restored.

[0112] After the model is trained, the WTC of the coastal area is corrected using the trained model. Specifically, the collected 7-dimensional features are input into the trained Elman neural network model, and the corrected WTC is output.

[0113] Of course, this embodiment is not limited to using Elman neural networks as the WTC correction model. For example, a WTC correction model based on back propagation neural network (BPNN) can also be used.

[0114] This BPNN model uses the exact same input features and learning objectives as the Elman model.

[0115] The BPNN model can also achieve the core objective of this invention with different network architectures, namely, by fusing multi-source data, the neural network model can effectively correct the WTC of satellite altimetry in coastal areas.

[0116] Of course, those skilled in the art can also choose other neural networks such as recurrent neural networks (RNN) or long short-term memory networks (LSTM) as specific implementation models according to actual requirements for estimation resources, model training speed and accuracy.

[0117] This invention systematically corrects land-polluted WTC in coastal areas by integrating multidimensional meteorological parameters from GNSS high-precision ZWD and ERA5 reanalysis data with raw satellite WTC data and using a neural network model.

[0118] Example 2 This embodiment 2 describes a computer device including a memory and one or more processors. Executable code is stored in the memory. When the processor executes the executable code, it implements the steps of the satellite altimeter WTC correction method for coastal areas based on neural networks described in embodiment 1 above.

[0119] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0120] Example 3 This embodiment 3 describes a computer-readable storage medium storing a program that, when executed by a processor, is used to implement the steps of the satellite altimeter WTC correction method for coastal areas based on neural networks in embodiment 1 above.

[0121] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0122] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A satellite altimeter WTC correction method for coastal areas based on neural networks, characterized in that, Includes the following steps: Step 1. First, select GNSS stations located in coastal areas that match the spatiotemporal range of the satellite's transit path, and estimate the ZWD at each station. Then, perform a spatiotemporal double matching between the estimated ZWD and the original WTC value of the satellite. Finally, unify the elevation of the stations after the double matching. place Elevation converted to satellite altimetry observation point place ; Step 2. Obtain meteorological parameters at satellite altimetry observation points based on ERA5 single-layer surface reanalysis data; First, based on ERA5 single-layer surface reanalysis data, meteorological parameters were extracted from four grid points adjacent to the satellite altimetry observation point; Then, each grid point is set at the grid elevation. Meteorological parameters at the location were corrected to the elevation of the satellite altimetry observation point. Then, using bilinear interpolation, the result is calculated. Final air pressure at the location ,temperature and total column water vapor ; Reusing grid point temperature and dew point temperature Estimate relative humidity After elevation correction, bilinear interpolation is performed on the calculation results to estimate the water vapor pressure at the satellite altimetry observation point. and relative humidity ; Step 3. Construct a training dataset for training the neural network; The samples in the training dataset consist of model input features and labels; the model input features include ERA5 meteorological parameters. , , , , The original WTC value of the satellite and the distance D from the shore of the satellite altimetry observation point; The labels are obtained after processing in step 1, and they correspond one-to-one with the input features of the model mentioned above. ; Step 4. Build a WTC correction model for coastal areas based on Elman neural network. Train the model using the training dataset from Step 3, and then use the trained model to correct the original WTC values ​​of the satellite.

2. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 1, characterized in that, In step 1, the process of GNSS station selection and ZWD estimation is as follows: GNSS stations located within a first preset distance on the landward side of the coastline are selected to ensure that the selected stations are in the nearshore area where land pollution is most significant, and to truly reflect the characteristics of atmospheric water vapor changes along the coast. For IGS stations providing ZTD, the Saastamoinen model is used to separate them. : ;in , The ground air pressure at the station. The latitude of the measuring station. Elevation of the measuring station; Further based on The calculation of the station location The formula is as follows: .

3. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 2, characterized in that, In step 1, the process of performing spatiotemporal dual matching between the estimated ZWD and the original satellite WTC value is as follows: I. Spatial Filtering: The distance between the satellite altimetry observation point and the GNSS station selected in step 1 must not exceed the second preset distance range; II. Time Filtering: The time difference between the satellite transit time and the GNSS observation time shall not exceed a preset time window; Data exceeding the specified time window will be discarded; data meeting the time difference threshold of less than the preset time window will be discarded. Linear interpolation is then used to extrapolate the time to the satellite altimetry observation point, thus obtaining the synchronization between the station and satellite observations. .

4. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 3, characterized in that, In step 1, the specific process of elevation unification is as follows: Using an exponential model, the stations at the time of measurement Elevation converted to satellite altimetry observation point place The formula is as follows: ； in and These are the station elevation and the satellite altimetry observation point elevation, respectively. This is the empirical attenuation coefficient.

5. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 4, characterized in that, Step 2 specifically involves: Based on ERA5 single-layer surface reanalysis data, meteorological parameters were extracted from four grid points adjacent to the satellite altimetry observation point. Based on the relationship between meteorological parameters and elevation, the four grid points were respectively located at the grid elevation. Meteorological parameters at the location were corrected to the elevation of the satellite altimetry observation point. The formula is as follows: ； ； Among them, grid elevation Meteorological parameters at the location include air pressure. ,temperature and total column water vapor Then grid elevation The total column of water vapor Corrected to satellite altimetry observation elevation The formula is as follows: ； The above formulas are used to obtain the air pressure, temperature and total column water vapor elevation correction values ​​for four adjacent grid points respectively; Then, bilinear interpolation is performed on all meteorological parameters of four adjacent grid points after elevation correction to obtain the final air pressure at the satellite altimetry observation point. ,temperature and total column water vapor ; Water vapor pressure at satellite altimetry observation points and relative humidity Similarly, calculations are first performed on the four grid points, and then bilinear interpolation is applied to the results. The estimation process is as follows: First, utilize the grid point temperature and dew point temperature Estimate relative humidity After elevation correction, the elevation of the altimeter observation point is obtained. place and The specific estimation is shown in the following formula: ； ； ； ； ； In the formula For saturated water vapor pressure, relevant parameters , , The value is determined by the grid point temperature. Scope determined; Then, using bilinear interpolation, the final water vapor pressure at the satellite altimetry observation point was calculated. and relative humidity .

6. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 1, characterized in that, In step 4, the hyperparameters of the Elman neural network are adaptively optimized using the sparrow search algorithm. Hyperparameter vector to be optimized This includes: the number of hidden layer nodes H, the initial learning rate lr, the weight decay coefficient ω, the Dropout ratio d, and the number of hidden layers L, i.e., the hyperparameter vector: =[H, lr, ω, d, L]; By optimization, the optimal number of hidden layer nodes is obtained. Optimal learning rate lr best Optimal weight decay coefficient ω best Optimal Dropout ratio d best And the optimal number of hidden layers L best .

7. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 6, characterized in that, In step 4, the Elman neural network consists of four parts: an input layer, a hidden layer, a supporting layer, and an output layer. Let the number of input layer nodes be I, and the number of hidden layers be L. best The number of hidden layer nodes is H best The number of nodes in the output layer is O=1, corresponding to the output of the corrected WTC; the number of nodes in the receiving layer is the same as that in the hidden layer, and it is used to store the state of the hidden layer at the previous time step. The number of nodes in the input layer is I, corresponding to the model input features, namely the air pressure at the satellite altimetry observation point. Temperature Water vapor pressure relative humidity Column water vapor The original WTC value of the satellite and the distance D from the shore of the satellite altimetry observation point; Let the input vector at time t be... The hidden layer state vector is The state vector of the receiving layer is The output layer result is The state of the receiving layer is directly taken from the output of the hidden layer at the previous time step, that is: ； The hidden layer state update uses the Tanh nonlinear activation function, the mathematical expression of which is: ； in, This is the weight matrix from the input layer to the hidden layer. This is the weight matrix from the receiving layer to the hidden layer. This is the hidden layer bias vector; After the hidden layer, a batch normalization layer and a Dropout layer are added sequentially, with the Dropout ratio set to d. best ; The output layer uses a linear mapping to directly output the corrected WTC. The estimation formula is as follows: ； in This is the weight vector from the hidden layer to the output layer. This is the output layer bias.

8. The satellite altimeter WTC correction method for coastal areas based on neural networks according to claim 7, characterized in that, In step 4, the Adam optimizer is used during the model training phase to optimize the network parameters θ={ , , , , Iterative updates are performed, with the initial learning rate set to lr. best The weight decay coefficient is set to ω. best ; The loss function uses mean squared error for a batch of samples. loss function Defined as: ； in Input features to the model, This represents the model's predicted value. This indicates the number of samples in each batch.

Images