A Method for Improving the Accuracy of Ocean Gravity Anomaly Retrieval Based on Shortwave Topographic Multilayer Sensor
By extracting shortwave topographic components using Gaussian high-pass filtering and constructing the SWBMLP model, the problem of mismatch in the mid-frequency domain characteristics of ocean gravity anomaly inversion was solved, achieving higher prediction accuracy, especially in complex topographic regions where it was improved by 18.1%.
Patent Information
- Application Number
- CN202610820400.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-08
- Publication Date
- 2026-11-10
AI Technical Summary
In the inversion of ocean gravity anomalies, the frequency domain characteristics of the original seabed topography and the residual gravity anomaly do not match, which limits the prediction accuracy of deep learning models and makes it difficult to effectively integrate multi-source physical data.
The shortwave topographic component was extracted using Gaussian high-pass filtering technology, and a shortwave topographic multilayer perceptron (SWBMLP) model was constructed. The residual vertical deviation, shortwave topography and spherical angular distance were used as inputs, and the ocean gravity anomaly was recovered by combining it with a reference gravity anomaly model.
It significantly improves the accuracy of ocean gravity anomaly inversion, especially in complex terrain areas where it improves by 18.1%, and shows higher accuracy in high-frequency signal prediction.
Smart Images

Figure CN123066873A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary fields of marine gravimetry and artificial intelligence, specifically a method for improving the accuracy of marine gravity anomaly inversion based on a shortwave topographic multilayer sensor. Background Technology
[0002] High-precision ocean gravity anomaly information has significant applications in seafloor topography research, oil and gas resource exploration, polar ice cap monitoring, inertial navigation, and sea-level change monitoring. Satellite altimetry, due to its ability to precisely determine the ocean geoid, has become the primary means of acquiring ocean gravity anomalies. Commonly used classical algorithms for retrieving ocean gravity anomalies using satellite altimetry include the least squares collocation method, the inverse Stokes' theorem method, and the vertical deviation method. However, these traditional methods still have certain limitations in practical applications: while the least squares collocation method can establish a rigorous statistical model, the covariance function is difficult to determine accurately, and matrix operations are time-consuming. In contrast, the inverse Stokes' theorem method and the vertical deviation method calculate gravity anomalies based on the geoid and vertical deviation, respectively. Although they do not require prior determination of the covariance function and effectively improve computational efficiency, they suffer from inversion accuracy due to issues such as the interior zone effect in integral operations. Furthermore, traditional inversion algorithms struggle to effectively integrate seabed topographic data that is highly correlated with ocean gravity anomalies, which is precisely the core advantage of the rapidly developing deep learning methods in recent years.
[0003] With the development of deep learning technology, it has shown significant advantages in modeling nonlinear relationships among multi-source physical data, and has been successfully applied in recent years to various inversion fields such as ocean gravity anomaly, seabed topography, sea level height, and magnetotellurics. Because this technology can effectively fuse multi-source physical data such as vertical deviation and seabed topography, while traditional algorithms often only perform inversions from single data such as vertical deviation or sea level height, it shows great potential in the field of ocean gravity anomaly inversion. Existing research is based on the remove-restore method, which uses a reference field to remove long-wave background signals, and then uses deep learning to fuse multi-source data to predict residual gravity anomalies, ultimately recovering the ocean gravity anomaly. Within this framework, vertical deviation and seabed topography have been repeatedly confirmed as the most crucial input features.
[0004] However, current studies generally use raw seafloor topography directly as input to neural networks, ignoring the mismatch between its frequency domain characteristics and those of residual gravity anomalies. Physically, residual gravity anomalies processed by the removal-recovery method are primarily composed of high-frequency signals, while raw topographic data contains information across the entire frequency band. This inconsistency in frequency domain characteristics weakens the physical correlation between features and the target variable. Given that deep learning models heavily rely on the physical relationship between input features and the target variable, low-correlation input variables limit the model's prediction accuracy. Therefore, enhancing the correlation between topography and gravity anomalies from a data preprocessing perspective is a crucial issue that urgently needs to be addressed to improve the accuracy of model inversion. Summary of the Invention
[0005] This invention addresses the shortcomings and deficiencies of existing technologies by proposing a method for improving the accuracy of ocean gravity anomaly inversion based on a shortwave topographic multilayer sensor.
[0006] This invention achieves its purpose through the following measures: A method for improving the accuracy of ocean gravity anomaly inversion based on a shortwave topographic multilayer sensor (SWBMLP) is characterized by the following steps: First, a Gaussian high-pass filter is used to remove longwave background signals from the original seabed topography, extracting shortwave topographic components that have a higher correlation with residual gravity anomalies. Then, a SWBMLP model is constructed, using the residual vertical deviation, shortwave topographic components, and spherical angular distance as model inputs, and the residual gravity anomaly as model output for training. The trained SWBMLP model is used to predict the residual gravity anomalies at grid points in the study area, and combined with a reference gravity anomaly model to recover the final ocean gravity anomaly.
[0007] The shortwave terrain multilayer sensor (SWBMLP) model described in this invention includes an input layer, a hidden layer, and an output layer, wherein the input layer consists of the residual vertical deviation meridional component. Residual vertical deviation (Mao-You component) Shortwave topography and spherical angular distance The composition, in which the residual vertical deviation and spherical angular distance are selected as inputs, is based on the analytical relationship between vertical deviation and gravity anomaly established by the inverse IVM formula, the expression of which is as follows: (3), in, It's a gravity anomaly. It is a calculation point, It is the flow point on the vertical deviation grid. It is normal gravity. and They are latitude and longitude, respectively. and These are the latitudinal and longitude intervals of the vertical deviation grid, respectively. and They are The perpendicular deviation of the point has the meridian component and the east-west component. yes and The azimuth between two points yes and The spherical angular distance between two points is calculated using the following formula: (4), It is the derivative of the kernel function, and its specific expression is: (5), All four types of input data are extracted from a 32′×32′ two-dimensional grid window centered on the calculation point. The input layer is represented by the following formula: (6), in, , , It is a calculation point Surrounding The residual perpendicular deviation meridional component, residual perpendicular deviation ramusoidal component, and shortwave topography of each grid point. It is a calculation point With the surrounding The spherical angular distance of each grid point.
[0008] The SWBMLP described in this invention comprises four hidden layers, each consisting of a fully connected layer, LayerNormalization (LayerNorm), and an activation function stacked sequentially. The number of neurons in the four fully connected layers are 2048, 1024, 512, and 256, respectively, designed as a progressive dimensionality reduction structure to aggregate a large number of low-level features into higher-level abstract features layer by layer. The number of neurons in each layer is a power of 2, also based on the aforementioned GPU hardware architecture efficiency considerations. The LayerNorm layer normalizes the output of each fully connected layer along the sample dimension, transforming the data into a distribution with a mean of 0 and a variance of 1, to alleviate the internal covariate shift problem. This normalization process helps accelerate model convergence and enhances the robustness of the training process. The activation function used is the ReLU function, and its calculation formula is as follows: (7), The output layer corresponds to the residual gravity anomaly at the current calculation point, and its expression is: (8).
[0009] In this invention, during the training of SWBMLP, the learning rate is set to 0.0001, the number of training epochs is set to 100, the batch size is set to 32, and stochastic gradient descent (SGD) is used for parameter optimization. After the model is built, the input data needs to be standardized, and the calculation formula is as follows: (9), in, It is standardized data. It is the original input data. and These represent the mean and variance of the original input data, respectively.
[0010] The specific method for extracting shortwave topographic components in this invention includes: designing a wavelength optimization algorithm based on the Pearson correlation coefficient, traversing candidate cutoff wavelengths within a preset wavelength range with a fixed step size; for each candidate cutoff wavelength, performing high-pass filtering on the original seabed topographic data using a two-dimensional Gaussian filter to obtain the corresponding scale of shortwave topographic components; calculating the Pearson correlation coefficient between the shortwave topographic components at each scale and the residual gravity anomaly; selecting the cutoff wavelength that makes the Pearson correlation coefficient reach its global maximum value as the optimal cutoff wavelength, and using this optimal cutoff wavelength to extract the final shortwave topographic components. The preset wavelength range of the wavelength optimization algorithm is 10km to 100km, and the search step size is 1km. The formula is shown below: (2), in, Represents the total number of samples. and Represents the first of two variables The values of each sample, and Representing variables and The sample mean.
[0011] Before model training, this invention also includes dataset construction and partitioning steps: gross errors are removed and systematic biases are corrected using the difference between shipborne gravity data and a reference gravity anomaly model; a terrain-based stratified sampling method is used to divide the corrected shipborne survey line data into training, validation, and test sets; the stratified sampling method uses the shipborne survey line as the smallest unit, classifying the survey lines into different intervals based on the average water depth and standard deviation of each survey line, and then proportionally distributing them to each dataset to ensure that each dataset has a consistent terrain distribution. Specifically, the slope and water depth distributions of the training, validation, and test sets are consistent, and the slope calculation formula is as follows: (1), in, , These represent the elevation changes of adjacent grid nodes in the GEBCO2024 along the east-west and south-north directions, respectively. , The slope value represents the actual physical horizontal distance in the corresponding direction. The larger the slope value, the more drastic the change in seabed topographic gradient, and the more significant the corresponding residual gravity anomaly signal.
[0012] In this invention, the residual vertical deviation includes the meridional component and the ramographic component of the residual vertical deviation; the reference gravity anomaly model adopts the XGM2019e-2159 model or a global ocean gravity anomaly model of equivalent level.
[0013] The specific steps for recovering the final ocean gravity anomaly in this invention are as follows: traverse each regular grid point within the study area, input the shortwave topography, residual vertical deviation, and spherical angular distance within the surrounding window into the trained SWBMLP model to obtain the residual gravity anomaly grid of that grid point; superimpose the residual gravity anomaly grid with the reference gravity anomaly grid calculated by the reference gravity anomaly model to obtain the final ocean gravity anomaly model.
[0014] In regions with drastic topographic changes, this invention can improve the inversion accuracy by up to 18.1% compared to the BMLP_GRA model, proving that this invention can effectively improve the accuracy of ocean gravity anomaly inversion. Attached Figure Description
[0015] Figure 1 This is a depth statistical result of the study area overview in this embodiment of the invention.
[0016] Figure 2 This is a statistical analysis of the water depth and slope distribution of each dataset in this embodiment of the invention, wherein the training set, validation set, and test set are marked in black, red, and blue, respectively. Figure 2 (a) represents the slope; Figure 2 (b) represents the water depth.
[0017] Figure 3 The following are the scatter density distributions and linear fitting results of gravity anomalies, residual gravity anomalies, and terrain in embodiments of the present invention: whereby... Figure 3 In the middle (a), the correlation coefficient between gravity anomaly and topography is 0.702; Figure 3 In the figure (b), the correlation coefficient between residual gravity anomaly and topography is 0.107; the black line in the figure represents linear fitting.
[0018] Figure 4The figure shows the variation trend of the Pearson correlation coefficient between residual gravity anomaly and shortwave topography with the filter cutoff wavelength in this embodiment of the invention. The black diamond mark in the figure corresponds to the global maximum correlation coefficient of 0.215, and the corresponding optimal cutoff wavelength is 29km.
[0019] Figure 5 This is the SWBMLP model structure in the embodiments of the present invention, wherein Figure 5 (a) shows the overall network structure: the input layer is composed of the meridian component, the ramus component, the shortwave topographic component, and the spherical angular distance of the residual vertical deviation around the calculation point, flattened into a one-dimensional vector; the output layer is the residual gravity anomaly. Figure 5 (b) is a schematic diagram of the input data: the red triangle marks the current calculation point, and the blue dots represent the grid points around the calculation point. The calculation point corresponds to the shipborne measurement point during the model training phase, and the regular grid points within the study area during the prediction phase.
[0020] Figure 6 This is the process for establishing the marine gravity anomaly model in this embodiment of the invention, taking SWBMLP_GRA as an example. Detailed Implementation
[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0022] The purpose of this invention is to develop a method for improving the accuracy of ocean gravity anomaly inversion based on a shortwave topographic multilayer sensor.
[0023] This invention proposes a novel shortwave topographic multilayer sensor method. This method innovatively employs Gaussian high-pass filtering and correlation analysis to extract the shortwave topographic features most strongly correlated with residual gravity anomalies. Based on this, residual vertical deviation and spherical angular distance are fused to construct the input of the MLP, with residual gravity anomaly as the output, aiming to improve the accuracy of ocean gravity anomaly inversion.
[0024] The study area selected for this example is the western Pacific Ocean, specifically the region between 139°E and 147°E and between 24°N and 32°N. The main reason for this selection is the complex seabed topography and wide range of water depths in this area. An overview of the study area is as follows: Figure 1 As shown, the seafloor topography of the study area, from west to east, includes the Izu-Ogasawara Island Arc, the Izu-Ogasawara Trench, and the deep-sea plain of the Pacific Plate. Within these macroscopic structures, numerous smaller structures such as seamounts, sea hills, mid-ocean ridges, and shoals are also distributed. Figure 1The histogram of water depth statistics for this region shows that the water depth ranges from 0 to -10000 m, exhibiting a distinct bimodal distribution at -6000 m and -3000 m. The complex topography, the vast range of water depths, and the irregular statistical distribution of water depth in this region provide ideal experimental conditions for evaluating the performance of the SWBMLP method. Furthermore, another important reason for choosing this region is its abundant shipborne gravity data.
[0025] The data for 130 shipborne gravity survey lines in this sea area were provided by the Japan Agency for Marine-Earth Science and Technology (JAMSTEC), which has completed lag correction, drift correction, and Eotvos correction for the free-air anomaly data. Due to long-wavelength errors in the shipborne gravity data, there will be systematic biases between the reference gravity anomaly model and the shipborne gravity measurements. To address this issue, this example first calculates the difference between the shipborne gravity and the reference gravity for each track segment. Then, it calculates the mean and standard deviation of these differences; data with differences exceeding three times the standard deviation are discarded. Finally, after removing the gross errors from each track segment, the average difference between the shipborne gravity data and the reference gravity is calculated again. This average is then added to the original shipborne data to obtain the corrected data.
[0026] Based on the corrected ship-based gravity data, this example further divides it for the training and evaluation of SWBMLP. The survey line data is divided into training, validation, and test sets, marked with gold, red, and green respectively. The division employs a terrain-based stratified sampling method to ensure consistent terrain distribution across datasets, thereby improving the model's generalization ability under various complex terrains. Simple random sampling was not used because adjacent measurement points within the same survey line have extremely strong spatial correlations, which can easily lead to data leakage issues in deep learning, ultimately reducing the credibility of the evaluation results.
[0027] This stratified sampling method uses shipborne survey lines as the smallest unit for data partitioning. First, based on the GEBCO2024 data, the average water depth and standard deviation of each survey line are extracted. These two indicators are then divided into three roughly equal sample sizes: 0% to 33.3%, 33.3% to 66.7%, and 66.7% to 100%. These two indicator intervals are then paired to form nine intervals. This stratification effectively represents the water depth distribution from 0m to -10000m while avoiding sample sparsity caused by excessive stratification. Subsequently, the 130 survey lines are assigned to these nine intervals according to their two indicators. The survey lines within each interval are then randomly shuffled and distributed sequentially to the training, validation, and test sets at ratios of 60%, 20%, and 20%, respectively, resulting in three datasets with consistent topographic distribution. This allocation ratio ensures that the neural network can fully learn the nonlinear mapping relationship while providing a sufficient sample size for independent evaluation of its generalization ability. After dividing the survey lines into 130 sections, to reduce the training time of the neural network, this invention downsampled the training and validation sets, taking one point every 10 points, while the test set was not downsampled. The number of ship survey data points in the training, validation, and test sets are 147,175, 29,175, and 328,170, respectively. The statistical distribution of slope and water depth for the above three datasets is as follows: Figure 3 As shown.
[0028] Figure 2 This indicates that the slope and water depth distributions of the training, validation, and test sets are basically consistent, which can well meet the requirements of data distribution consistency for deep learning experiments. The slope is calculated as follows: (1), in, , These represent the elevation changes of adjacent grid nodes in the GEBCO2024 along the east-west and south-north directions, respectively. , This represents the actual physical horizontal distance in the corresponding direction. The larger the slope value, the more drastic the change in seabed topographic gradient, and the more significant the corresponding residual gravity anomaly signal.
[0029] GEBCO2024 is the 2024 version of the global bathymetric dataset released by the General Bathymetric Chart of the Oceans (GEBCO). This dataset integrates gravity-inverted bathymetry, shipborne multibeam and single-beam field data, providing high-precision bathymetry information covering global ocean areas. This example primarily utilizes this data for regional topographic feature analysis and mapping.
[0030] This example uses XGM2019e-2159 as the reference gravity anomaly model. Reference vertical deviation and reference gravity anomaly are calculated based on this model and used in the removal-recovery method. SIO_V32.1 is a global ocean gravity anomaly model obtained by the Scripps Institution of Oceanography based on multi-source altimeter data from Geosat, ERS-1, Jason-1, Jason-2, CryoSat-2, and SARAL / Altika. This invention uses the vertical deviation and gravity anomaly of this model, subtracting them from the reference field to obtain the residual vertical deviation and residual gravity anomaly, and using the gravity anomaly as a reference benchmark for accuracy assessment. The calculation method for residual vertical deviation and residual gravity anomaly is as follows: First, the reference vertical deviation and reference gravity anomaly are calculated based on the XGM2019e-2159 model; then, the reference values are subtracted from the vertical deviation and gravity anomaly of the SIO_V32.1 model to obtain the residual vertical deviation and residual gravity anomaly used in this invention.
[0031] This example demonstrates the use of the Pearson correlation coefficient. The formula for calculating linear correlation is shown below: (2), in, Represents the total number of samples. and Represents the first of two variables The values of each sample, and Representing variables and The sample mean. Pearson correlation coefficient. The value range is always between -1 and 1. Indicates a positive correlation. Indicates a negative correlation. This indicates that there is no linear correlation between the two. The closer the absolute value is to 1, the stronger the linear correlation.
[0032] This example calculates the Pearson correlation coefficients between gravity anomalies and topography, and between residual gravity anomalies and topography, based on gravity anomalies, residual gravity anomalies, and topography data at all two-dimensional grid points within the study area. The results are as follows: Figure 4 As shown.
[0033] Figure 3 In the middle (a) and (b), the scatter density distributions of gravity anomaly-topography and residual gravity anomaly-topography, respectively, and their Pearson correlation coefficients are shown. Figure 3In (a), the correlation coefficient between gravity anomaly and topography is 0.702. This is because both the original gravity anomaly and the original topography contain long-wavelength signals reflecting macroscopic landforms, and their overall trends are highly consistent, thus exhibiting a strong positive correlation. In contrast, Figure 3 (b) The correlation coefficient between the residual gravity anomaly and the terrain is only 0.107, indicating that there is almost no significant linear relationship between them. This weak correlation reveals a frequency band mismatch problem between the data: after removing the reference field XGM2019e-2159, the residual gravity anomaly mainly reflects the shortwave characteristics of the local terrain; if the original terrain containing all frequency band information is used directly for analysis, the longwave components in the terrain will mask the true connection between the two in local undulations, resulting in a significant reduction in correlation. Directly using such frequency band mismatched data as neural network input will inevitably interfere with the model's learning of key features. Therefore, this invention requires high-pass filtering of the original terrain to remove the longwave signals and extract the shortwave terrain components consistent with the frequency band of the residual gravity anomaly, which can then be used as network input to effectively improve the accuracy of the inversion.
[0034] This example proposes a wavelength optimization algorithm based on Pearson correlation coefficient. The algorithm traverses and searches within a wavelength range of 10-100 km in 1 km increments. For each candidate cutoff wavelength, a two-dimensional Gaussian filter is used to perform high-pass filtering on the original water depth topography, extracting the shortwave topography components at the corresponding scale. Subsequently, the Pearson correlation coefficient between residual gravity anomaly and shortwave topography at each scale is calculated, and the wavelength that maximizes the global correlation coefficient is determined as the optimal cutoff wavelength for the current study area. The trend of the Pearson correlation coefficient between residual gravity anomaly and shortwave topography with cutoff wavelength is shown below. Figure 4 As shown.
[0035] Figure 4 The results show that as the cutoff wavelength increases, the correlation coefficient between the two exhibits a trend of first rapidly increasing and then slowly decreasing. When the cutoff wavelength increases from 10km to 29km, the correlation coefficient increases significantly, reaching a global peak of 0.215 at 29km; thereafter, as the cutoff wavelength continues to increase, the correlation coefficient begins to slowly decrease. This pattern indicates that when the cutoff wavelength is around 29km, the extracted shortwave topographic components match the frequency band of the residual gravity anomaly most closely, effectively suppressing the interference of longwave signals on the correlation. Therefore, this invention selects 29km as the optimal cutoff wavelength for Gaussian high-pass filtering in the current sea area to extract shortwave topographic features and uses it as one of the core inputs of SWBMLP.
[0036] This example presents a multilayer perceptron method based on shortwave topography—SWBMLP—for the inversion of ocean gravity anomalies. This invention will describe the design of the input layer, hidden layer, and output layer of this model in sequence, with the overall structure and specific composition of the input layer as follows: Figure 5 (a) and Figure 5 As shown in (b).
[0037] Depend on Figure 5 As shown in (a), the input layer consists of the residual vertical deviation meridional component. Residual vertical deviation (Mao-You component) Shortwave topography and spherical angular distance The composition is as follows: The selection of residual vertical deviation and spherical angular distance as inputs is based on the analytical relationship between vertical deviation and gravity anomaly established by the inverse IVM formula, which is expressed as follows: (3), in, It's a gravity anomaly. It is a calculation point, It is the flow point on the vertical deviation grid. It is normal gravity. and They are latitude and longitude, respectively. and These are the latitudinal and longitude intervals of the vertical deviation grid, respectively. and They are The perpendicular deviation of the point has the meridian component and the east-west component. yes and The azimuth between two points yes and The spherical angular distance between two points is calculated using the following formula: (4), It is the derivative of the kernel function, and its specific expression is: (5), like Figure 5As shown in (b), all four types of input data are extracted from a 32′×32′ two-dimensional grid window centered on the computation point. The selection of this window size is based on both physical and computational considerations. Physically, since the grid spacing is 1′×1′, and MLP establishes the relationship between residual vertical deviation, spherical distance, and residual gravity anomaly, this mechanism is formally similar to the inverse IVM formula. Existing research indicates that the optimal integration radius of IVM is between 30′ and 60′, and the window size selected in this invention falls within this range. Computationally, setting the grid side length to 32 is to adapt to the GPU hardware architecture and the underlying alignment mechanism, which provides higher processing efficiency for data dimensions of powers of 2. After flattening the 32×32 two-dimensional grid into a 1024-dimensional one-dimensional vector, the four types of features are sequentially concatenated to form a 4096-dimensional one-dimensional input vector. The input layer can be represented by the following formula: (6), in, , , It is a calculation point Surrounding The residual perpendicular deviation meridional component, residual perpendicular deviation ramusoidal component, and shortwave topography of each grid point. It is a calculation point With the surrounding The spherical angular distance of each grid point.
[0038] Depend on Figure 5 As shown in (a), SWBMLP contains four hidden layers, each consisting of a fully connected layer, LayerNormalization (LayerNorm), and an activation function stacked sequentially. The number of neurons in the four fully connected layers are 2048, 1024, 512, and 256 respectively, designed as a progressive dimensionality reduction structure to aggregate a large number of low-level features into higher-level abstract features layer by layer. The number of neurons in each layer is a power of 2, also based on the aforementioned consideration of GPU hardware architecture efficiency. The LayerNorm layer normalizes the output of each fully connected layer along the sample dimension, transforming the data into a distribution with a mean of 0 and a variance of 1, to alleviate the internal covariate shift problem. This normalization process is independent of batch size and remains stable even under small batch training conditions, which helps to accelerate model convergence and enhance the robustness of the training process. The ReLU function is selected as the activation function, which is one of the most widely used activation functions in deep learning. It introduces nonlinear transformation capabilities into the network, enabling the model to learn and express complex nonlinear mapping relationships. Its calculation formula is as follows: (7), like Figure 5As shown in (a), the output layer corresponds to the residual gravity anomaly at the current calculation point, and its expression is: (8), During the training of SWBMLP, the learning rate was set to 0.0001, the number of training epochs was set to 100, the batch size was set to 32, and stochastic gradient descent (SGD) was used for parameter optimization. The loss function was mean squared error. These hyperparameters were determined through pre-experiments to ensure the stability of training convergence. Specifically, in 10... -1 Up to 10 -5 Among the candidate values, 0.0001 was chosen as the learning rate because this value, when combined with SGD, allows the loss function to decrease smoothly. 100 training epochs provide the model with sufficient iterations to fully learn the nonlinear mapping relationship and achieve stable convergence without overfitting. A batch size of 32 balances low GPU memory usage with high computational efficiency while ensuring normal model convergence.
[0039] After the model is built, the input data needs to be standardized. The calculation formula is as follows: (9), Among them, is the standardized data. It is the original input data. and These represent the mean and variance of the original input data, respectively. Standardization transforms all input features into a distribution with a mean of 0 and a standard deviation of 1, effectively eliminating dimensional differences between different features, making the training process more stable, and accelerating the convergence speed of gradient descent, thereby improving the training efficiency and performance of the network. Considering that the learning of model parameters is based on the statistical distribution of the training set, after standardizing the training set, the mean of the dataset is recorded. with standard deviation The validation and test sets are standardized using the same parameters to ensure that the feature scaling baseline remains consistent across datasets.
[0040] The performance of this example will be analyzed below: To conduct comparative experiments, this example constructs three ocean gravity anomaly models: SWBMLP_GRA, BMLP_GRA, and MLP_GRA. All three use residual vertical deviation and spherical angular distance as inputs, differing only in whether and what type of terrain features are incorporated: SWBMLP_GRA uses shortwave topography, BMLP_GRA uses raw topography, and MLP_GRA contains no topographic information. All other settings and modeling procedures remain consistent. This invention uses SWBMLP_GRA as an example, and its specific construction process is as follows: Figure 6 As shown. The first part is data preprocessing: First, shortwave topographic grid, residual vertical deviation grid, and ship-borne gravity data need to be obtained. Then, the shortwave topography, residual vertical deviation, and spherical angular distance within a 32′×32′ window around each ship's measuring point are used as input, and the residual gravity anomaly of that ship's measuring point is used as output, thus constructing a dataset. The dataset is then divided and standardized to obtain a training set, a validation set, and a test set. The second part is neural network training: The input data of the training set is fed into SWBMLP in batches, and the mean square error between the network's predicted value and the actual residual gravity anomaly is used as the loss function. If the preset maximum number of training rounds is not reached, the network weights are updated using the SGD algorithm to gradually approximate the measured residual gravity anomaly of the ship's measuring point. After each training round, the loss function is calculated on the validation set, and its decrease is observed to ensure that the model has a certain generalization ability. After 100 rounds of training, the model parameters with the smallest loss function value on the validation set are retained, thus obtaining the trained SWBMLP model. The third part involves generating the gravity anomaly model: The shortwave topography, residual vertical deviation, and spherical angular distance within a 32′×32′ window surrounding the center point of each grid are input into the trained SWBMLP to obtain the residual gravity anomaly at that point. This process is repeated across the entire study area to obtain a residual gravity anomaly grid. The residual gravity anomaly grid is then overlaid with the XGM2019e-2159 reference gravity anomaly grid to obtain the final ocean gravity anomaly model, SWBMLP_GRA. The model's predicted values are interpolated to the coordinates of the measurement points on the test vessel to obtain the predicted values of the ship's gravity measurements. Subsequently, accuracy evaluation metrics are calculated based on the true and predicted values. This example demonstrates the inversion results of the SWBMLP_GRA model constructed according to the above process.
[0041] It is known that high-value areas of ocean gravity anomalies are concentrated in island arcs and island regions, while low-value areas are mainly distributed in trenches. The extreme value difference of gravity anomalies in the study area reaches 674.7 mGal. This large-span anomaly variation is shaped by the complex and diverse seabed topography of the area, making it a suitable region for comprehensively evaluating the accuracy of model inversion.
[0042] This example uses four models—SWBMLP_GRA, BMLP_GRA, MLP_GRA, and SIO_V32.1—to perform bilinear interpolation to the measurement points on the test vessel, and then performs statistical analysis after subtracting them from the ship's gravity anomalies. The RMSE of the SWBMLP_GRA model is reduced by 0.03 mGal, 0.07 mGal, 0.39 mGal, and 0.31 mGal compared to the BMLP_GRA, MLP_GRA, SIO_V32.1, and DTU21 models, respectively. For the overall study area, SWBMLP_GRA demonstrates superior accuracy compared to other ocean gravity anomaly models, with a particularly significant improvement over traditional models, confirming the outstanding capabilities of deep learning methods in the nonlinear fusion of multi-source geophysical data. However, the accuracy difference between SWBMLP_GRA and BMLP_GRA and MLP_GRA is relatively limited. To further reveal the differences between the models, this example calculates the difference between SWBMLP_GRA and each comparative model.
[0043] The mean difference between SWBMLP_GRA and each comparative model approaches 0mGal, indicating that this model is highly consistent with existing mainstream models on the overall gravity field baseline, and there is no systematic bias. Furthermore, the difference between SWBMLP_GRA and each comparative model shows a clear trend of gradual decrease. First, the difference between SWBMLP_GRA and SIO_V32.1 is the largest, reflecting the fundamental difference in computational principles between deep learning algorithms and traditional physics methods. Second, in comparisons of deep learning architectures, the difference between SWBMLP_GRA and MLP_GRA is greater than the difference between SWBMLP_GRA and BMLP_GRA, mainly due to differences in feature selection. MLP_GRA lacks terrain data and cannot supplement terrain constraints on fitting the relationship between residual vertical deviation and residual gravity anomaly; while both BMLP_GRA and SWBMLP_GRA incorporate terrain features, the only difference being whether long-wavelength signals are filtered out. This is a difference at the data preprocessing level, hence the closer difference between them. To visually present the spatial distribution of model differences and focus on areas with large differences, this invention plots the spatial distribution of the differences between SWBMLP_GRA and each comparative model.
[0044] It can be seen that regions A, B, C, and D with significant differences correspond to areas with dramatic seabed topographic relief. This example further statistically analyzes these four local regions. The WBMLP_GRA model reduced the RMSE in these four regions by 0.05–0.99 mGal, 0.17–1.00 mGal, 0.07–0.18 mGal, and 0.17–0.54 mGal, respectively, compared to the other three models. The model achieved improved accuracy in all complex sea areas, with the most significant improvement in region B. This indicates that shortwave topography effectively enhances the model's predictive ability for local high-frequency gravity signals. Furthermore, the deep learning model shows significant improvement over SIO_V32.1 in regions A and B, which may be closely related to the topographic conditions: the water depth in these two regions is mostly between 0 m and -4000 m, and several islands and reefs are distributed within them. The surrounding waters of these islands and reefs are complex, and radar echoes are easily interfered with by land reflection signals, leading to a decrease in the accuracy of satellite altimetry data in these areas, thus limiting the accuracy of traditional inversion methods.
[0045] The impact of slope on prediction accuracy is examined below: the accuracy improvement of the SWBMLP_GRA model is particularly significant in areas with dramatic seabed topographic relief. Based on this, this example further explores the relationship between model accuracy and the degree of topographic change. The test set shipboard gravity data is divided into four intervals according to their corresponding slope values, and the accuracy index of each model is statistically analyzed separately. Across all slope ranges, SWBMLP_GRA exhibited the lowest RMSE and STD, indicating that its inversion accuracy under varying terrain gradients was superior to the comparative models. Specifically, in the slope ranges of 0.20–0.25 and 0.25–0.50, SWBMLP_GRA's RMSEs were 2.29 mGal and 2.40 mGal, respectively, representing reductions of 8.8% and 18.1% compared to BMLP_GRA. This result confirms that high-pass filtering effectively suppressed interference from long-wavelength signals, maintaining a match between the input and target variables in the frequency domain, thereby significantly enhancing the model's predictive ability for local high-frequency gravity anomalies in areas with steep slopes and dramatic terrain changes. In contrast, within the 0.20–0.50 slope range, BMLP_GRA's error increased significantly, even falling below the accuracy of the MLP_GRA model, which did not incorporate any terrain data. This indicates that directly inputting raw full-band terrain information will cause long-wave components to interfere with the neural network's learning of high-frequency residual gravity anomalies, thus causing more significant errors in locally complex terrain areas.
[0046] The following examines the impact of terrain on prediction accuracy: To further reveal the correlation between model accuracy and topography, this example divides the water depth range of 0 to -10000m into five intervals of 2000m, and statistically analyzes the accuracy indicators of each ocean gravity anomaly model at different depths. The prediction accuracy of the four models generally shows a trend of first increasing and then decreasing with increasing water depth: in the shallow water area with drastic topography (0 to -2000m) and the ultra-deep trench area (-10000 to -8000m), the errors of each model are relatively large; while in the deep-sea plain area with gentle topography (-8000 to -6000m), the accuracy reaches its optimal level. In a horizontal comparison, the SWBMLP_GRA model has the lowest RMSE and STD in most water depth ranges. The overall accuracy of MLP_GRA is better than SIO V32.1, confirming the advantage of deep learning methods in nonlinear mapping capabilities; BMLP_GRA and SWBMLP_GRA further improve accuracy on this basis, fully demonstrating that topographic information plays an irreplaceable physical constraint role in the process of extrapolating residual gravity anomalies from residual vertical deviation. It is particularly noteworthy that in the main ocean area from 0 to -8000m, SWBMLP_GRA outperforms BMLP_GRA in accuracy across the board, confirming the effectiveness of the frequency matching strategy from a physical mechanism perspective: shortwave topography and residual gravity anomalies, as high-frequency signals, are highly compatible in spatial frequency, thus more effectively assisting the model in capturing subtle gravity changes caused by local topographic disturbances. However, in the extremely deep ocean area from -10000 to -8000m, BMLP_GRA is slightly more accurate than SWBMLP_GRA. This local reversal may be because this depth range typically corresponds to large-scale tectonic units such as trenches and subduction zones. Some medium- and long-wave gravity response features contained in the original topography still have certain physical indicative significance for such extreme deep structures. High-pass filtering may have filtered out this effective low-frequency information when extracting shortwave topography, causing SWBMLP_GRA to fail to maintain its advantage in this local range.
Claims
1. A method for improving the accuracy of ocean gravity anomaly inversion based on a shortwave topographic multilayer sensor, characterized in that, First, the long-wave background signal in the original seabed topography is removed by using Gaussian high-pass filtering technology, and then the short-wave topographic component with higher correlation to the residual gravity anomaly is extracted. Then, a shortwave topographic multilayer sensor (SWBMLP) model is constructed, with the residual vertical deviation, shortwave topographic components, and spherical angular distance as model inputs and the residual gravity anomaly as model output for training. The trained SWBMLP model is used to predict the residual gravity anomaly at grid points in the research area, and the final ocean gravity anomaly is obtained by combining it with a reference gravity anomaly model.
2. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topography multilayer sensing machine according to claim 1, characterized in that, The shortwave terrain multilayer sensor (SWBMLP) model comprises an input layer, a hidden layer, and an output layer, wherein the input layer consists of the residual vertical deviation meridional component. Residual vertical deviation (Mao-You component) Shortwave topography and spherical angular distance The composition, in which the residual vertical deviation and spherical angular distance are selected as inputs, is based on the analytical relationship between vertical deviation and gravity anomaly established by the inverse IVM formula, the expression of which is as follows: (3), in, It's a gravity anomaly. It is a calculation point, It is the flow point on the vertical deviation grid. It is normal gravity. and They are latitude and longitude, respectively. and These are the latitudinal and longitude intervals of the vertical deviation grid, respectively. and They are The perpendicular deviation of the point has the meridian component and the east-west component. yes and The azimuth between two points yes and The spherical angular distance between two points is calculated using the following formula: (4), It is the derivative of the kernel function, and its specific expression is: (5), All four types of input data are extracted from a 32′×32′ two-dimensional grid window centered on the calculation point. The input layer is represented by the following formula: (6), in, , , It is a calculation point Surrounding The residual perpendicular deviation meridional component, residual perpendicular deviation ramusoidal component, and shortwave topography of each grid point. It is a calculation point With the surrounding The spherical angular distance of each grid point.
3. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topography multilayer sensing machine according to claim 2, characterized in that, The SWBMLP consists of four hidden layers, each composed of a fully connected layer, LayerNormalization, and an activation function stacked sequentially. The number of neurons in the four fully connected layers are 2048, 1024, 512, and 256 respectively, designed as a progressive dimensionality reduction structure to aggregate a large number of low-level features into higher-level abstract features layer by layer. The number of neurons in each layer is a power of 2, also based on the efficiency considerations of GPU hardware architecture. The LayerNorm layer standardizes the output of each fully connected layer along the sample dimension, transforming the data into a distribution with a mean of 0 and a variance of 1 to alleviate the internal covariate shift problem. The ReLU activation function is used, and its calculation formula is as follows: (7), The output layer corresponds to the residual gravity anomaly at the current calculation point, and its expression is: (8)。 4. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topographic multilayer sensing machine according to claim 3, characterized in that, During the training of SWBMLP, the learning rate was set to 0.0001, the number of training epochs was set to 100, and the batch size was set to 32. Parameter optimization used stochastic gradient descent (SGD). After model building, the input data needed to be standardized, and the calculation formula was as follows: (9), in, It is standardized data. It is the original input data. and represent the mean and variance of the original input data, respectively.
5. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topography multilayer sensing machine according to claim 1, characterized in that, The specific method for extracting shortwave topographic components includes: designing a wavelength optimization algorithm based on the Pearson correlation coefficient, traversing candidate cutoff wavelengths within a preset wavelength range with a fixed step size; for each candidate cutoff wavelength, performing high-pass filtering on the original seabed topographic data using a two-dimensional Gaussian filter to obtain the corresponding scale of shortwave topographic components; calculating the Pearson correlation coefficient between the shortwave topographic components at each scale and the residual gravity anomaly; selecting the cutoff wavelength that makes the Pearson correlation coefficient reach its global maximum as the optimal cutoff wavelength, and using this optimal cutoff wavelength to extract the final shortwave topographic components. The preset wavelength range of the wavelength optimization algorithm is 10km to 100km, and the search step size is 1km. The formula is shown below: (2), in, Represents the total number of samples. and Represents the first of two variables The values of each sample, and Representing variables and The sample mean.
6. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topography multilayer sensing machine according to claim 1, characterized in that, Before model training, the process includes dataset construction and partitioning: Gross errors are removed and systematic biases are corrected using the difference between the shipborne gravity data and the reference gravity anomaly model; a terrain-based stratified sampling method is used to divide the corrected shipborne survey line data into training, validation, and test sets; this stratified sampling method uses the shipborne survey line as the smallest unit, assigning the survey lines to different intervals based on the average water depth and standard deviation of each line, and then distributing them proportionally to each dataset to ensure consistent terrain distribution across datasets. Specifically, the slope and water depth distribution are consistent across the training, validation, and test sets, with the slope calculated using the following formula: (1), in, , These represent the elevation changes of adjacent grid nodes in the GEBCO2024 along the east-west and south-north directions, respectively. , The slope value represents the actual physical horizontal distance in the corresponding direction. The larger the slope value, the more drastic the change in seabed topographic gradient, and the more significant the corresponding residual gravity anomaly signal.
7. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topography multilayer sensing machine according to claim 1, characterized in that, The residual vertical deviation includes the meridional component and the ramographic component of the residual vertical deviation; the reference gravity anomaly model adopts the XGM2019e-2159 model or a global ocean gravity anomaly model of equivalent level.
8. The method for improving the accuracy of ocean gravity anomaly inversion based on shortwave topographic multilayer sensing machine according to claim 1, characterized in that, The specific steps to recover the final ocean gravity anomaly are as follows: traverse every regular grid point in the study area, input the shortwave topography, residual vertical deviation and spherical angular distance within the surrounding window into the trained SWBMLP model, and obtain the residual gravity anomaly grid of that grid point. The residual gravity anomaly grid is superimposed with the reference gravity anomaly grid calculated by the reference gravity anomaly model to obtain the final ocean gravity anomaly model.
