Construction method of sea depth inversion model based on two-stage clustering and space constraint and application thereof
By employing a partitioning strategy based on two-stage clustering and spatial constraints, along with a fully connected deep neural network, the problems of geological differences and data heterogeneity in ocean depth inversion were solved, achieving high-precision and efficient seabed topography modeling.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI UNIV OF SCI & TECH
- Filing Date
- 2026-03-19
- Publication Date
- 2026-06-16
AI Technical Summary
Existing ocean depth inversion methods are unable to effectively handle the geological differences and uneven data distribution in different sea areas, resulting in insufficient accuracy and low efficiency in large-area seabed topography inversion.
A partitioning strategy based on two-stage clustering and spatial constraints is adopted to divide the target sea area into multiple geographically continuous sub-regions. In each sub-region, a nonlinear mapping relationship between gravity anomalies and sea depth is trained through a fully connected deep neural network. Finally, the sub-regions are spliced together to form a partitioned sea depth inversion model.
It achieves high precision and efficiency in large-area ocean depth inversion, solves the problems of geological differences and uneven data distribution in different sea areas, and improves the accuracy and computational efficiency of seabed topography modeling.
Smart Images
Figure CN122220818A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of marine surveying and mapping technology, specifically relating to the construction method and application of a sea depth inversion model based on two-stage clustering and spatial constraints. Background Technology
[0002] Seafloor topography is fundamental data for studying marine geological structures, geophysical processes, seafloor resource exploration, maritime navigation safety, and marine environmental protection. Traditional shipborne depth sounding methods, such as single-beam and multi-beam sonar, while highly accurate, suffer from high costs, low efficiency, and limited coverage due to factors such as ship speed, navigation routes, and natural conditions. Data gaps still exist in most of the world's oceans, requiring depth prediction to fill these gaps.
[0003] The maturity of satellite altimetry technology has provided large-scale, high-precision ocean gravity data, becoming a primary data source for inverting global seabed topography. Among these, gravity anomaly data is widely used in ocean depth prediction due to its significant correlation with seabed topography within specific wavebands. Traditional ocean depth inversion methods include gravity geological methods and their improvements, the Smith & Sandwell method, the admittance function method, and the least squares collocation method. While these methods can provide reliable models to some extent, they still have limitations. For example, the gravity geological method is insufficient for handling complex topography and nonlinear effects, resulting in poor inversion accuracy in shallow sea areas; the admittance function method is suitable for large-scale ocean areas, but its accuracy is limited in regions with abrupt topographic changes.
[0004] In recent years, with the development of artificial intelligence technology, deep learning has been introduced into the field of seabed topography inversion. Commonly used methods include convolutional neural networks, fully connected neural networks, and backpropagation neural networks. Deep learning methods can better capture the nonlinear relationship between gravity anomalies and sea depth, achieving more accurate predictions for areas lacking ship-based survey data. However, for seabed topography inversion on a large regional or even global scale, the relationship between gravity anomalies and seabed topography varies due to differences in crustal type, density, and sedimentary layers in different sea areas; uneven distribution of ship-based survey data necessitates additional constraints in sparse data areas; and the complexity of gravity signals and topography within different sea depth ranges also affects inversion accuracy. Furthermore, the massive amounts of data place high demands on processing efficiency and model structure. Summary of the Invention
[0005] The purpose of this invention is to provide a method for constructing a deep-sea depth inversion model based on two-stage clustering and spatial constraints, and its application, so as to solve the problem that existing deep-sea depth inversion methods are difficult to effectively handle the geological differences and uneven data distribution in different sea areas, thereby achieving high-precision deep-sea depth inversion in large areas.
[0006] The present invention achieves the above objectives through the following technical solutions: Firstly, this invention proposes a method for constructing an ocean depth inversion model based on two-stage clustering and spatial constraints, the method comprising: Acquire raw observation data of the target sea area and preprocess it; the raw observation data includes shipborne depth sounding data and gravity anomaly data; Based on the processed raw observation data, a two-dimensional grid with equal intervals is constructed, and the partition characteristics on each grid cell are determined; wherein, the partition characteristics include single beam point density characteristics, average water depth characteristics, topographic relief characteristics, gravity anomaly characteristics, and offshore distance characteristics. Based on the partitioning characteristics, the target sea area is divided into multiple geographically contiguous sub-regions using a partitioning strategy that combines two-stage clustering with spatial constraints. Within each sub-region, the gravity anomaly components and their residuals are determined based on the processed raw observation data. The location information and corresponding gravity anomaly components and their residuals within each sub-region are used as inputs to train a fully connected deep neural network to obtain the ocean depth inversion prediction value for each sub-region. By stitching together the sub-regions, a regional ocean depth inversion model is obtained.
[0007] Furthermore, the preprocessing of the raw observation data includes: Outlier removal processing is performed on the shipborne depth sounding data; wherein the outlier removal processing includes: interpolating the reference depth model to obtain a reference depth value corresponding to the location of the shipborne depth sounding data, calculating the standard error between the shipborne depth sounding data and the reference depth value, and removing shipborne depth sounding data points whose difference exceeds a preset multiple of the standard error. The gravity anomaly data is subjected to component decomposition processing to obtain long-wave gravity anomaly components and short-wave gravity anomaly components.
[0008] Furthermore, the construction of the equally spaced two-dimensional grid includes: Starting from the boundary of the target sea area, the number of longitude and latitude nodes is determined by sequentially increasing the preset grid resolution. Perform a Cartesian product operation on the longitude nodes and the latitude nodes to generate a two-dimensional grid system covering the target sea area.
[0009] Furthermore, the method for determining the partitioning features includes: The single-beam point density feature is obtained by statistically analyzing the number of effective single-beam points falling into each grid cell. The average water depth characteristic is obtained by calculating the average of all water depths within each grid cell; The topographic relief feature is obtained by calculating the rate of change of water depth in the longitude and latitude directions within each grid cell, and by calculating the average gradient magnitude. The gravity anomaly characteristics are obtained by calculating the average value of gravity anomaly data within each grid cell; The offshore distance feature is obtained by calculating the Euclidean distance from the center point of each grid cell to the nearest coastline based on the coastline data.
[0010] Furthermore, the partitioning strategy, which combines two-stage clustering with spatial constraints, divides the target sea area into multiple geographically contiguous sub-regions, including: Using the grid cells as the processing object and the partition features as input, a clustering algorithm is used to perform preliminary clustering of the feature space, and the preliminary clustering result with the first label is output. Based on the preliminary clustering results, neighborhood relationships between grid cells are established. Using the first label as input, spatial continuity constraints are applied through an iterative algorithm to eliminate isolated points in space and output continuous spatial partitioning results with the second label. For each second label category in the continuous spatial partitioning results, a connectivity analysis is performed to split the label category containing multiple discontinuous regions into multiple independent regions, and the connectivity analysis results with a third label are output. Based on the data density, the connectivity analysis results are merged into regions. The region with the lowest data density is merged with the spatially adjacent region with the highest data density. Isolated regions with an area smaller than a preset threshold are merged into the nearest spatially adjacent region. The above merging operation is performed iteratively until the total number of partitions does not exceed the preset maximum number of partitions. The result is a geographically continuous sub-region division with a final label.
[0011] Furthermore, establishing neighborhood relationships between grid cells includes: The KD-tree data structure is used to find the nearest neighbor cells of each grid cell, and the neighborhood range includes directly adjacent cells and cells in the diagonal direction. An iterative algorithm is used to smooth the clustering of labels. In each iteration, the label of each grid cell is replaced with the label that appears most frequently in its neighborhood. This process is repeated until the label distribution is stable.
[0012] Furthermore, within each sub-region, determining the gravity anomaly components and their residuals based on the processed original observation data includes: Within each sub-region, the shipborne depth sounding data is divided into control points, check points, and verification points according to a preset ratio; Calculate the shortwave gravity anomaly at each control point using the Bouguer plate formula; The long-wave gravity anomaly is calculated based on the gravity anomaly data and the short-wave gravity anomaly. The long-wave gravity anomaly is meshed to construct a long-wave gravity anomaly field; A linear regression model was used to establish the relationship between shipborne depth sounding data and gravity anomalies, and the baseline values of short-wave gravity anomalies and long-wave gravity anomalies were calculated. Subtracting the shortwave gravity anomaly baseline value from the shortwave gravity anomaly yields the residual shortwave gravity anomaly; subtracting the longwave gravity anomaly baseline value from the longwave gravity anomaly yields the residual longwave gravity anomaly.
[0013] Furthermore, the step of training a fully connected deep neural network by using the location information and corresponding gravity anomaly components and their residuals within each sub-region as input includes: A fully connected deep neural network is constructed, comprising an input layer, multiple hidden layers, and an output layer. The input layer receives the longitude, latitude, shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly of control points within each sub-region as input features. The multiple hidden layers comprise a first to a fourth hidden layer connected sequentially, each hidden layer having a predetermined number of neurons. The output layer outputs the predicted ocean depth value. Initialize the weight parameters of the neural network; Enter the training loop and perform the following operations in each iteration: The input features of the control points are input into the neural network, propagated forward layer by layer through each hidden layer, and the predicted sea depth value is calculated by the output layer. The loss function value is calculated based on the predicted sea depth value and the actual water depth value of the control point; Based on the loss function value, the gradient of the network parameters of each layer is calculated using the backpropagation algorithm; The network weight parameters are updated based on the gradient using an optimization algorithm. The input features of the verification point are input into the updated neural network, the loss function value at the verification point is calculated, and it is determined whether the early stopping condition is met. When the early stopping condition is met or the preset maximum number of iterations is reached, the training loop is terminated, and the trained neural network model with the current network parameters is saved.
[0014] Furthermore, the predicted ocean depth values for each sub-region are stitched together to form a partitioned ocean depth inversion model covering the entire target sea area, including: Using the trained neural network models of each sub-region, ocean depth prediction is performed on the grid center point of a preset resolution within each sub-region to obtain the predicted ocean depth value of each sub-region. Then, the sub-region boundaries are extended outward by a preset distance and then stitched together. For the overlapping parts of adjacent sub-regions, the predicted ocean depth values of each sub-region are weighted and averaged to obtain a partitioned ocean depth inversion model covering the target sea area.
[0015] Secondly, this invention proposes a sea depth prediction method based on a partitioned sea depth inversion model, the method comprising: Obtain the location information of the points to be predicted in the target sea area; Based on the location information, the target sub-region to which the point to be predicted belongs is determined; wherein, the target sub-region is one of multiple geographically contiguous sub-regions obtained by pre-dividing the target sea area into zones using the above method; Obtain the neural network model corresponding to the target sub-region; wherein the neural network model is trained using the method described above; The gravity anomaly data at the point to be predicted is obtained, and the gravity anomaly data is decomposed into components to obtain shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly. The location information of the point to be predicted, along with the shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly, are input into the neural network model corresponding to the target sub-region to obtain the predicted sea depth value of the point to be predicted.
[0016] The beneficial effects of this invention are as follows: 1. This invention uses a partitioning strategy that combines two-stage clustering with spatial constraints to divide the target sea area into multiple geographically continuous sub-regions, ensuring internal consistency in the geological features and data distribution within each sub-region. This effectively solves the problem of inconsistent gravity anomalies and seabed topography caused by differences in crustal types, density, and sedimentary layers in different sea areas.
[0017] 2. This invention determines the gravity anomaly component and its residual within each sub-region, fully exploring the response characteristics between the local sea area's gravity field and seabed topography. A fully connected deep neural network is used to train each sub-region separately, enabling adaptive learning of the nonlinear mapping relationship between gravity anomalies and sea depth in different sea areas. The sub-regions are then stitched together to form a partitioned sea depth inversion model covering the entire target sea area, avoiding the computational resource consumption caused by directly processing massive amounts of data. Furthermore, outward training and weighted average fusion ensure the continuity and stability of the stitched model. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating a method for constructing a sea-depth inversion model according to an embodiment of the present invention. Figure 2 This is another flowchart illustrating the method for constructing a sea-depth inversion model according to an embodiment of the present invention; Figure 3 This is a processed shipborne depth sounding trajectory distribution map in one implementation example of the present invention; Figure 4 This is a partitioning result diagram in one implementation example of the present invention; Figure 5 This is a schematic diagram of the ocean depth model obtained by the method for constructing the ocean depth inversion model according to an embodiment of the present invention. Detailed Implementation
[0019] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content. Example 1
[0020] See Figure 1 and Figure 2 A specific embodiment of the present invention proposes a method for constructing an ocean depth inversion model based on two-stage clustering and spatial constraints, the method comprising: S101. Obtain raw observation data of the target sea area and perform preprocessing.
[0021] Specifically, raw observation data for the target sea area is acquired, including shipborne bathymetry data and gravity anomaly data. The shipborne bathymetry data may be single-beam shipborne bathymetry data published by the National Centers for Environmental Information (NCEI), and the gravity anomaly data may be gravity anomaly model data published by the Scripps Institution of Oceanography (SIO) at the University of California, San Diego.
[0022] The raw observation data is preprocessed, including outlier removal from shipborne depth sounding data and component decomposition of gravity anomaly data.
[0023] Outlier removal is performed on the shipborne sounding data using a reference depth model. For example, the reference depth model can be the GEBCO_2025 model. Specifically, the reference depth model is interpolated to obtain the reference depth value corresponding to the location of the shipborne sounding data. The standard error between the shipborne sounding data and the reference depth value is calculated, and shipborne sounding data points with a difference exceeding a preset multiple of the standard error are removed. For example, the preset multiple can be three times, i.e., the 3σ principle is used for outlier removal. This process effectively removes outlier data caused by measurement errors or environmental interference.
[0024] The gravity anomaly data is decomposed into component components: the original gravity anomaly data is decomposed into long-wave gravity anomaly components and short-wave gravity anomaly components. Short-wave gravity anomalies mainly reflect local seafloor variations, such as fault zones and sediments; long-wave gravity anomalies mainly reflect macroscopic geological features, such as mid-ocean ridges and continental slopes. Component decomposition allows for a better characterization of the response relationship between geological features at different scales and ocean depth.
[0025] S102. Based on the processed raw observation data, construct a two-dimensional grid with equal intervals and determine the partition characteristics on each grid cell.
[0026] Constructing an equally spaced two-dimensional grid: Based on the distribution range of the original observation data and a preset grid resolution, an equally spaced two-dimensional grid system is created. Specifically, starting from the boundary of the target sea area, the number of longitude and latitude nodes is determined by sequentially increasing the preset grid resolution; the longitude nodes and latitude nodes are then subjected to a Cartesian product operation to generate a two-dimensional grid system covering the target sea area. As an example, the preset grid resolution can be 0.5° × 0.5°, equivalent to approximately 56 kilometers of actual distance. This resolution balances computational efficiency with the need to represent the spatial heterogeneity of seabed topography, while also being compatible with the original resolutions of various global physical data products.
[0027] The partitioning features on each grid cell are determined, including single-beam point density features, mean water depth features, topographic relief features, gravity anomaly features, and offshore distance features. The specific methods for determining each feature are as follows: The single-beam point density feature is obtained by statistically analyzing the number of effective single-beam points falling within each grid cell. This feature reflects the richness of ship-based survey data in the region and is of great significance for subsequent zoning and model training.
[0028] The average water depth characteristic is obtained by calculating the average of all water depths within each grid cell. This characteristic can distinguish the different geophysical properties of deep water areas from those of shallow water areas.
[0029] The topographic relief feature is obtained by calculating the rate of change of water depth in the longitude and latitude directions within each grid cell and then calculating the average gradient magnitude. Specifically, the gradients of water depth in the east-west (longitude) and north-south (latitude) directions are first calculated, and then the average gradient magnitude is calculated. This feature helps distinguish between flat deep-sea plains and complex tectonic regions, such as submarine mountains and fault zones.
[0030] The gravity anomaly characteristic is obtained by calculating the average value of gravity anomaly data within each grid cell. This characteristic reflects the overall characteristics of the regional gravity field.
[0031] The offshore distance feature is obtained by calculating the Euclidean distance from the center point of each grid cell to the nearest coastline based on coastline data. This feature can effectively distinguish the different geological features and data distribution characteristics between nearshore and offshore areas.
[0032] Since the five features mentioned above have different magnitudes and numerical ranges, standardization is required to eliminate scale differences. As an example, Z-score standardization can be used, subtracting the global mean from each feature and dividing by the global standard deviation. The standardized feature matrix facilitates subsequent clustering algorithms in considering the influence of each dimension equally.
[0033] Step S103: Based on the partitioning characteristics, the target sea area is divided into multiple geographically continuous sub-regions using a partitioning strategy that combines two-stage clustering with spatial constraints.
[0034] The partitioning strategy, which combines two-stage clustering with spatial constraints, specifically includes the following sub-steps: Sub-step S1031: Preliminary clustering in the feature space: Using the grid cells as the processing object and the partition features as input, a clustering algorithm is used to perform preliminary clustering in the feature space, outputting preliminary clustering results with the first label. As an example, the K-means clustering algorithm can be used, and the maximum number of partitions can be preset according to factors such as computational efficiency and macroscopic classification of seabed landforms. K-means++ is used for multiple initializations to ensure the stability of the results.
[0035] Sub-step S1032, Spatial continuity constraint: Based on the preliminary clustering results, establish the neighborhood relationship between grid cells, take the first label as input, perform spatial continuity constraint through an iterative algorithm, eliminate isolated points in space, and output a continuous spatial partitioning result with a second label.
[0036] Establishing neighborhood relationships between grid cells specifically includes: using a spatial index structure to find multiple nearest neighbor cells for each grid cell. As an example, a KD-tree data structure can be used to find multiple nearest neighbor cells for each grid cell. The neighborhood range includes directly adjacent cells and cells in the diagonal direction, ensuring sufficient neighborhood coverage. An iterative algorithm is used to smooth the clustering labels. In each iteration, the label of each grid cell is replaced with the label that appears most frequently in its neighborhood. This iteration is repeated until the label distribution meets a preset stability condition, for example, 15-20 iterations. This process effectively eliminates spatially isolated points, forming continuous spatial partitions.
[0037] Sub-step S1033, Connectivity Splitting: Perform connectivity analysis on each second label category in the continuous spatial partitioning result, splitting label categories containing multiple discontinuous regions into multiple independent regions, and outputting the connectivity analysis results with third labels. Specifically, a connectivity analysis algorithm is used to check whether each label category forms a single continuous region. For label categories containing multiple discontinuous regions, they are split into multiple independent regions and assigned new label numbers. This step ensures that each partition is ultimately a geographically continuous spatial unit.
[0038] Sub-step S1034, Region Merging Optimization: Based on data density, the connectivity analysis results are merged into regions. The region with the lowest data density is merged with the spatially adjacent region with the highest data density. Isolated regions with an area smaller than a preset threshold are merged into the nearest spatially adjacent region. This merging operation is iteratively performed until the total number of partitions does not exceed the preset maximum number of partitions. The result is a geographically continuous sub-region division with final labels. This step prioritizes merging sparse data regions to avoid excessively small partitions.
[0039] As an example, the final partitioning results can be saved in NetCDF format, including the partition label matrix, latitude and longitude coordinate axes, and five original feature values. The Marching Squares algorithm is used to extract the boundary contours of each region from the partition label matrix, and the Douglas-Puk algorithm is used to simplify the boundaries, retaining key nodes to ensure the efficiency of subsequent inversion calculations.
[0040] Step S104: Within each sub-region, determine the long-wavelength and short-wavelength components of the gravity anomaly and their residuals based on the processed original observation data.
[0041] Specifically, it includes the following sub-steps: Sub-step S1041, Data Division: Within each sub-region, the shipborne depth sounding data is divided into control points, check points, and verification points according to a preset ratio. As an example, a random division in a 3:1:1 ratio can be used, where control points are used for training the fully connected neural network, check points are used for accuracy evaluation, and verification points are used for an early stopping mechanism to prevent overfitting.
[0042] Sub-step S1042: Calculate the shortwave gravity anomaly: Calculate the shortwave gravity anomaly at each control point using the Bouguer plate formula. The calculation formula is as follows: (1) In the formula, Let be the shortwave gravity anomaly at point i; G is the gravitational constant. is the density difference constant; E(i) is the sea depth at control point i; D is the reference depth, which is generally the maximum water depth covered by the study area.
[0043] Sub-step S1043: Calculate the long-wave gravity anomaly: Based on the gravity anomaly data and the short-wave gravity anomaly, calculate the long-wave gravity anomaly. The calculation formula is as follows: (2) In the formula, This is a gravity anomaly; This refers to the long-wavelength component of the gravity anomaly. This represents the shortwave component of gravity anomalies.
[0044] Sub-step S1044: Constructing a long-wavelength gravity anomaly field: The long-wavelength gravity anomaly is meshed to construct a long-wavelength gravity anomaly field. For unknown points, the long-wavelength gravity anomaly component at that point can be obtained by interpolating the long-wavelength gravity anomaly field.
[0045] Sub-step S1045: Calculate the baseline value and residuals: A linear regression model is used to establish the relationship between shipborne bathymetry data and gravity anomalies, calculating the baseline values for short-wave and long-wave gravity anomalies. The baseline values for short-wave gravity anomalies are subtracted from the short-wave gravity anomalies to obtain the residual short-wave gravity anomalies; similarly, the baseline values for long-wave gravity anomalies are subtracted from the long-wave gravity anomalies to obtain the residual long-wave gravity anomalies. The residual components contain more complex geological information and local nonlinear characteristics, including parts of the original gravity anomaly signal that cannot be explained by a simple linear relationship; therefore, they are also used as input features for subsequent neural network training.
[0046] Through the above processing, high-quality features highly correlated with sea depth changes were obtained in each sub-region, including shortwave gravity anomalies, longwave gravity anomalies, residual shortwave gravity anomalies, and residual longwave gravity anomalies.
[0047] Step S105: Use the location information and corresponding gravity anomaly components and their residuals within each sub-region as inputs to train a fully connected deep neural network, thereby obtaining a neural network model for each sub-region.
[0048] Specifically, it includes the following sub-steps: Sub-step S1051: Constructing the neural network architecture: Constructing a fully connected deep neural network, the network including an input layer, multiple hidden layers, and an output layer. The input layer receives the longitude, latitude, shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly of control points within each sub-region as input features; the multiple hidden layers include a first hidden layer to a fourth hidden layer connected in sequence, each hidden layer having a preset number of neurons; the output layer outputs the predicted sea depth value.
[0049] As an example, the first hidden layer can be configured with 16 neurons, the second with 32 neurons, the third with 256 neurons, and the fourth with 512 neurons. A modified linear unit (MRU) is chosen as the activation function to better capture the complex nonlinear relationships between the data.
[0050] Sub-step S1052: Initialize network parameters: Initialize the weight parameters of the neural network. The weights and biases of each layer can be set using a random initialization method.
[0051] Sub-step S1053, Enter the training loop: Perform the following operations in each iteration: The input features of the control points are input into the neural network, propagated forward layer by layer through each hidden layer, and the predicted sea depth value is calculated by the output layer. The loss function value is calculated based on the predicted sea depth value and the actual water depth value of the control point; Based on the loss function value, the gradient of the network parameters of each layer is calculated using the backpropagation algorithm; The network weight parameters are updated based on the gradient using an optimization algorithm. As an example, the Adam optimization algorithm can be used, with an initial learning rate of 0.001. The input features of the verification point are input into the updated neural network, the loss function value at the verification point is calculated, and it is determined whether the early stopping condition is met.
[0052] Sub-step S1054: Terminate training and save the best neural network model: When the early stopping condition is met or the preset maximum number of iterations is reached, terminate the training loop and save the trained neural network model with the current network parameters. As an example, the maximum number of iterations can be set to 40 epochs, and the early stopping mechanism is to stop training when the loss function does not decrease significantly within 10 epochs.
[0053] To reduce potential edge effects when assembling the overall model, the boundaries of each sub-region are extended outward by a preset distance during training. For example, this preset distance can be 0.1°.
[0054] Step S106: The sub-regions are stitched together, and the overlapping regions are weighted averaged to obtain the zoned ocean depth inversion model of the entire target sea area.
[0055] Specifically, it includes the following sub-steps: Sub-step S1061: Obtain the neural network model of each sub-region: Obtain the neural network model of each sub-region trained in step S105, wherein the neural network model of each sub-region is trained by extending the boundary of the sub-region outward by a preset distance during the training phase.
[0056] Sub-step S1062: Perform sub-region prediction: Using the neural network model of each sub-region, predict the ocean depth at the center point of the grid at a preset resolution within each sub-region to obtain the predicted ocean depth value for each sub-region. As an example, the preset resolution can be 1′×1′.
[0057] Sub-step S1063: Merging Overlapping Parts: Weighted average fusion is performed on the overlapping parts of adjacent sub-regions to obtain a zoned ocean depth inversion model covering the target sea area. The weighted average fusion method effectively eliminates edge effects, ensuring a smooth and continuous overall model after stitching.
[0058] Through the above steps, this embodiment of the invention constructs a regional ocean depth inversion model covering the entire target sea area. This model fully considers the geological differences, data distribution characteristics, and topographic complexity of different sea areas. Example 2
[0059] Another specific embodiment of the present invention proposes the application of a partitioned ocean depth inversion model, specifically involving an ocean depth prediction method based on the partitioned ocean depth inversion model, the method comprising the following steps: Step S201: Obtain the location information of the points to be predicted in the target sea area.
[0060] Specifically, obtain the latitude and longitude coordinates of the point to be predicted, either input by the user or specified by the system.
[0061] Step S202: Determine the target sub-region to which the point to be predicted belongs. Based on the location information, determine the target sub-region to which the point to be predicted belongs. The target sub-region is one of multiple geographically contiguous sub-regions pre-divided by partitioning the target sea area using the method described in Example 1. Specifically, the target sub-region to which the point belongs can be determined by finding the sub-region grid cell to which the point to be predicted falls based on its latitude and longitude coordinates.
[0062] Step S203: Obtain the neural network model corresponding to the target sub-region.
[0063] A neural network model corresponding to the target sub-region is obtained, wherein the neural network model is trained using the method described in Example 1. Each sub-region corresponds to an independently trained neural network model, which fully learns the response relationship between gravity anomalies and sea depth within that sub-region.
[0064] Step S204: Obtain gravity anomaly data at the point to be predicted and perform preprocessing.
[0065] Gravity anomaly data at the point to be predicted is acquired, and the gravity anomaly data is preprocessed. Specifically, following the method described in step S104 of Embodiment 1, the gravity anomaly data is decomposed into components to obtain shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly. This preprocessing process is consistent with the data preprocessing during model training to ensure consistency in the feature space of the input data.
[0066] Step S205: Input the model to predict ocean depth.
[0067] The location information of the point to be predicted, along with the shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly, are input into the ocean depth inversion model corresponding to the target sub-region. The predicted ocean depth value of the point to be predicted is obtained through forward propagation calculation of the model.
[0068] Through the above steps, this embodiment of the invention achieves rapid prediction of sea depth at any location within a target sea area. Due to the adoption of a partitioned training strategy, the model for each sub-region is optimized for the geological characteristics and data distribution of that region, resulting in high accuracy in the prediction results. Furthermore, since prediction only requires calling the model for the corresponding sub-region, the resource consumption of complex calculations across the entire region is avoided, thus improving prediction efficiency.
[0069] Application Cases To further illustrate the technical effects of the present invention, a sea area was used as an example to construct a sea depth inversion model and verify its accuracy.
[0070] (1) Data preparation: Obtain shipborne depth sounding data and gravity anomaly data for the sea area, and perform preprocessing and outlier removal according to the method described in Example 1. Construct a 0.5°×0.5° two-dimensional grid and calculate the partitioning characteristics of each grid cell. Figure 3 The distribution of the processed shipborne depth sounding trajectory is shown, and it can be seen that the data is unevenly distributed in space, requiring processing through a partitioning strategy.
[0071] (2) Zoning Results: Using a zoning strategy combining two-stage clustering and spatial constraints, the sea area was divided into four geographically contiguous sub-regions, denoted as Region B to Region E, with Region A being the nearshore area not included in the zoning. Figure 4 It can be seen that the zoning results fully consider the connectivity of the sea area and the data distribution characteristics, forming continuous spatial units.
[0072] (3) Model Training: Within each sub-region, control points, check points, and verification points are divided in a 3:1:1 ratio. The Bouguer plate formula is used to calculate short-wave gravity anomalies, thereby obtaining long-wave gravity anomalies and residual components. A fully connected deep neural network with four hidden layers is constructed, using longitude, latitude, short-wave gravity anomalies, long-wave gravity anomalies, residual short-wave gravity anomalies, and residual long-wave gravity anomalies as input features, and each sub-region is trained separately. During training, each sub-region is extended outward by 0.1° to avoid edge effects during subsequent stitching.
[0073] (4) Model splicing: The center point of the 1′×1′ grid in each sub-region is predicted, and the weighted average is used to fuse the overlapping areas to obtain the partitioned ocean depth inversion model FCD_Depth_sub that covers the entire target sea area.
[0074] (5) Accuracy verification: In order to evaluate the effect of the intelligent partitioning inversion method in this invention, standard deviation (STD), root mean square error (RMSE), and mean absolute percentage error (MAPE) were selected as accuracy evaluation indicators. The results of each partition were calculated and compared with the results of no partitioning and the results of existing global ocean depth models GEBCO_2025, SIOv27.1, DTU18 and ocean depth model GGM_Depth obtained by gravity geological method.
[0075] STD reflects the dispersion of the inverted depth values and is used to measure the consistency of the inversion results. RMSE measures the absolute deviation between the inverted depth and the true depth; a smaller value indicates higher accuracy. MAPE measures the average relative error percentage of the inverted depth relative to the true depth and is used to assess overall relative accuracy. The MAPE calculation formula is as follows: (3) In the formula, n is the total number of checkpoints in the corresponding area; It is the water depth inversion value at the i-th checkpoint; It is the actual water depth value at the i-th checkpoint.
[0076] (6) Results and Analysis: Table 1 shows the accuracy comparison results of the stitched partitioned model FCD_Depth_sub and the non-partitioned model FCD_Depth over the entire region. As can be seen from Table 1, the overall accuracy of the partitioned model FCD_Depth_sub is better than that of the non-partitioned model FCD_Depth and other global ocean depth models over the entire study area. Specifically, the RMSE of FCD_Depth_sub is 44.401 m and the MAPE is 1.151%, both of which are better than that of FCD_Depth (RMSE: 45.054 m, MAPE: 1.281%), indicating that the model has higher overall water depth inversion accuracy after partitioned training. Compared with other global models (such as GGM_Depth, GEBCO_2025, SIOv27.1, DTU18), FCD_Depth_sub performs best in the three key indicators of STD, RMSE and MAPE, and the partitioning strategy effectively improves the model's adaptability to regional water depth changes.
[0077] Table 1 shows the comparison results of the FCD_Depth_sub, FCD_Depth, GGM_Depth, GEBCO_2025, SIOv27.1 and DTU18 models across the entire region.
[0078] Table 1: Table 2 shows the performance of the zoning model FCD_Depth_sub in each sub-region (BE) and in the offshore region A, which was not included in the zoning. In region BD, the zoning model outperformed other comparative models in both RMSE and MAPE, exhibiting better performance and lower error. Although the RMSE in region E (19.526 m) was slightly higher than that of FCD_Depth (19.054 m), its MAPE (0.868%) was better than that of FCD_Depth (1.194%), and the overall error was still lower than that of most global ocean depth models. Meanwhile, in region A, FCD_Depth_sub's RMSE was 29.473 m and its MAPE was 4.114%, significantly better than FCD_Depth (RMSE: 37.644 m, MAPE: 5.572%) and other models. The obtained zoning model is also applicable to complex offshore areas, demonstrating good stability and applicability.
[0079] Table 2 shows the performance of the FCD_Depth_sub, FCD_Depth, GGM_Depth, GEBCO_2025, SIOv27.1, and DTU18 models on different partitions and in region A, which is not involved in the partitioning.
[0080] Table 2: To further analyze the performance of the method of the present invention under different conditions, the accuracy results were statistically analyzed for different sea depth ranges and different offshore distances.
[0081] Table 3 shows the accuracy comparison across different ocean depths. As can be seen from Table 3, in shallow sea areas with a depth of 500 meters or less, the root mean square error (RMSE) of the proposed partitioning model FCD_Depth_sub is 26.274 m, an improvement of approximately 17.0% compared to the 31.662 m of the non-partitioned model FCD_Depth, and significantly better than other comparative models. This indicates that the partitioning strategy of this invention can better capture local features in shallow sea areas with complex topography. In deep sea areas with a depth greater than 500 meters, the accuracy of FCD_Depth_sub is close to that of FCD_Depth, proving that the partitioning strategy does not introduce significant accuracy loss.
[0082] Table 3 shows the statistical results of the FCD_Depth_sub, FCD_Depth, GGM_Depth, GEBCO_2025, SIOv27.1 and DTU18 models at different sea depths at the checkpoint.
[0083] Table 3: Table 4 shows the accuracy comparison across different offshore distance ranges. As can be seen from Table 4, the accuracy of each model generally improves with increasing offshore distance. Within the distance range of ≥0km to ≥500km, the root mean square error and mean absolute percentage error of the model FCD_Depth_sub are both better than or close to the optimal levels. This advantage is particularly pronounced in open ocean areas above ≥200km.
[0084] Table 4 shows the statistical results of the FCD_Depth_sub, FCD_Depth, GGM_Depth, GEBCO_2025, SIOv27.1 and DTU18 models at different distances from the coastline at the checkpoint.
[0085] Table 4: Based on the above statistical results, the following conclusions can be drawn: At the checkpoints in each sub-region, the mean absolute percentage error of the model of this invention is significantly better than that of the comparative model. The root mean square error and standard deviation in multiple sub-regions are also superior to other comparative models, and it is also superior to other comparative models in peripheral region A, demonstrating higher accuracy and better stability. Particularly in shallow sea areas with complex topography and water depths less than 500 meters, the root mean square error of FCD_Depth_sub is improved by approximately 17.0% compared to FCD_Depth, and is significantly better than other comparative models. This indicates that the intelligent partitioning strategy of this invention can adaptively enhance the extraction of local features in sparse data and complex terrain regions, effectively solving the difficulties of traditional ocean depth inversion.
[0086] In deep water and most offshore distances, the accuracy of FCD_Depth_sub and FCD_Depth is very close, proving that the partitioning strategy does not introduce significant accuracy loss or edge effects due to region division and stitching. Although there are slight fluctuations in the MAPE index at medium offshore distances (≥100km), the overall error level is still within the optimal range, ensuring the overall geographical continuity and stability of the model. Meanwhile, partitioned training divides the entire regional data into multiple sub-regions, each with a moderate amount of data, significantly reducing training time and improving computational efficiency, providing a feasible technical solution for achieving high-precision ocean depth inversion over large areas.
[0087] In summary, this invention effectively solves the accuracy problem caused by geological differences and uneven data distribution in large-area ocean depth inversion by combining intelligent partitioning strategy and deep neural network, while improving computational efficiency and providing a feasible technical solution for seabed topography modeling.
[0088] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0089] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0090] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A method for constructing an ocean depth inversion model based on two-stage clustering and spatial constraints, characterized in that, The method includes: Acquire raw observation data of the target sea area and preprocess it; the raw observation data includes shipborne depth sounding data and gravity anomaly data; A two-dimensional grid with equal spacing is constructed, and the partition characteristics on each grid cell are determined based on the original observation data; wherein, the partition characteristics include single beam point density characteristics, average water depth characteristics, topographic relief characteristics, gravity anomaly characteristics, and offshore distance characteristics. Based on the partitioning characteristics, the target sea area is divided into multiple geographically contiguous sub-regions using a partitioning strategy that combines two-stage clustering with spatial constraints. Within each sub-region, the gravity anomaly components and their residuals are determined based on the processed raw observation data. The location information and corresponding gravity anomaly components and their residuals within each sub-region are used as inputs to train a fully connected deep neural network to obtain the ocean depth inversion prediction value for each sub-region. By stitching together the sub-regions, a regional ocean depth inversion model is obtained.
2. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, The preprocessing of the raw observation data includes: Outlier removal processing is performed on the shipborne depth sounding data; wherein the outlier removal processing includes: interpolating the reference depth model to obtain a reference depth value corresponding to the location of the shipborne depth sounding data, calculating the standard error between the shipborne depth sounding data and the reference depth value, and removing shipborne depth sounding data points whose difference exceeds a preset multiple of the standard error. The gravity anomaly data is subjected to component decomposition to obtain long-wave gravity anomaly components and short-wave gravity anomaly components, and the residuals are calculated.
3. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, The construction of the equally spaced two-dimensional grid includes: Starting from the boundary of the target sea area, the number of longitude and latitude nodes is determined by sequentially increasing the preset grid resolution. Perform a Cartesian product operation on the longitude nodes and the latitude nodes to generate a two-dimensional grid system covering the target sea area.
4. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, The methods for determining the partitioning features include: The single-beam point density feature is obtained by statistically analyzing the number of effective single-beam points falling into each grid cell. The average water depth characteristic is obtained by calculating the average of all water depths within each grid cell; The topographic relief feature is obtained by calculating the rate of change of water depth in the longitude and latitude directions within each grid cell, and by calculating the average gradient magnitude. The gravity anomaly characteristics are obtained by calculating the average value of gravity anomaly data within each grid cell; The offshore distance feature is obtained by calculating the Euclidean distance from the center point of each grid cell to the nearest coastline based on the coastline data.
5. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, The partitioning strategy, which combines two-stage clustering with spatial constraints, divides the target sea area into multiple geographically contiguous sub-regions, including: Using the grid cells as the processing object and the partition features as input, a clustering algorithm is used to perform preliminary clustering of the feature space, and the preliminary clustering result with the first label is output. Based on the preliminary clustering results, neighborhood relationships between grid cells are established. Using the first label as input, spatial continuity constraints are applied through an iterative algorithm to eliminate isolated points in space and output continuous spatial partitioning results with the second label. For each second label category in the continuous spatial partitioning results, a connectivity analysis is performed to split the label category containing multiple discontinuous regions into multiple independent regions, and the connectivity analysis results with a third label are output. Based on the data density, the connectivity analysis results are merged into regions. The region with the lowest data density is merged with the spatially adjacent region with the highest data density. Isolated regions with an area smaller than a preset threshold are merged into the nearest spatially adjacent region. The above merging operation is performed iteratively until the total number of partitions does not exceed the preset maximum number of partitions. The result is a geographically continuous sub-region division with a final label.
6. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 5, characterized in that, The establishment of neighborhood relationships between grid cells includes: The KD-tree data structure is used to find the nearest neighbor cells of each grid cell, and the neighborhood range includes directly adjacent cells and cells in the diagonal direction. An iterative algorithm is used to smooth the clustering of labels. In each iteration, the label of each grid cell is replaced with the label that appears most frequently in its neighborhood. This process is repeated until the label distribution is stable.
7. The method for constructing a sea-depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, Within each sub-region, the gravity anomaly components and their residuals are determined based on the processed original observation data, including: Within each sub-region, the shipborne depth sounding data is divided into control points, check points, and verification points according to a preset ratio; Calculate the shortwave gravity anomaly at each control point using the Bouguer plate formula; The long-wave gravity anomaly is calculated based on the gravity anomaly data and the short-wave gravity anomaly. The long-wave gravity anomaly is meshed to construct a long-wave gravity anomaly field; A linear regression model was used to establish the relationship between shipborne depth sounding data and gravity anomalies, and the baseline values of short-wave gravity anomalies and long-wave gravity anomalies were calculated. Subtracting the shortwave gravity anomaly baseline value from the shortwave gravity anomaly yields the residual shortwave gravity anomaly; subtracting the longwave gravity anomaly baseline value from the longwave gravity anomaly yields the residual longwave gravity anomaly.
8. The method for constructing a sea depth inversion model based on two-stage clustering and spatial constraints according to claim 7, characterized in that, The step of training a fully connected deep neural network by taking the location information of each sub-region and its corresponding gravity anomaly component and its residual as input includes: A fully connected deep neural network is constructed, comprising an input layer, multiple hidden layers, and an output layer. The input layer receives the longitude, latitude, shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly of control points within each sub-region as input features. The multiple hidden layers comprise a first to a fourth hidden layer connected sequentially, each hidden layer having a predetermined number of neurons. The output layer outputs the predicted ocean depth value. Initialize the weight parameters of the neural network; Enter the training loop and perform the following operations in each iteration: The input features of the control points are input into the neural network, propagated forward layer by layer through each hidden layer, and the predicted sea depth value is calculated by the output layer. The loss function value is calculated based on the predicted sea depth value and the actual water depth value of the control point; Based on the loss function value, the gradient of the network parameters of each layer is calculated using the backpropagation algorithm; The network weight parameters are updated based on the gradient using an optimization algorithm. The input features of the verification point are input into the updated neural network, the loss function value at the verification point is calculated, and it is determined whether the early stopping condition is met. When the early stopping condition is met or the preset maximum number of iterations is reached, the training loop is terminated and the trained neural network model is saved.
9. The method for constructing an ocean depth inversion model based on two-stage clustering and spatial constraints according to claim 1, characterized in that, The predicted ocean depths for each sub-region are stitched together to form a partitioned ocean depth inversion model covering the entire target sea area, including: Based on the neural network model trained in each sub-region, the ocean depth is predicted for the grid center point of the preset resolution in each sub-region, and the predicted ocean depth value of each sub-region is obtained. Then, the sub-regions obtained after training are stitched together. For the overlapping parts of adjacent sub-regions, the predicted sea depth values of each sub-region are weighted and averaged to obtain a partitioned sea depth inversion model covering the target sea area.
10. A method for ocean depth prediction based on a regional ocean depth inversion model, characterized in that, The method includes: Obtain the location information of the points to be predicted in the target sea area; Based on the location information, the target sub-region to which the point to be predicted belongs is determined; wherein, the target sub-region is one of a plurality of geographically contiguous sub-regions obtained by pre-dividing the target sea area into zones using the method described in any one of claims 1-9; Obtain a fully connected neural network model corresponding to the target sub-region; wherein the fully connected neural network model is trained by the method described in any one of claims 1-9; The gravity anomaly data at the point to be predicted is obtained, and the gravity anomaly data is decomposed into components to obtain shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly. The location information of the point to be predicted, along with the shortwave gravity anomaly, longwave gravity anomaly, residual shortwave gravity anomaly, and residual longwave gravity anomaly, are input into the fully connected neural network model corresponding to the target sub-region to obtain the predicted sea depth value of the point to be predicted.