Multi-source submarine topography reconstruction method and system based on frequency domain attention neural network
By employing a multi-source seabed topography reconstruction method based on a frequency domain attention neural network, and utilizing fast Fourier transform and attention mechanisms, combined with fully connected layers and ReLU activation functions, the problem of insufficient small-scale feature capture in seabed topography mapping is solved, and high-precision seabed topography reconstruction is achieved.
Patent Information
- Application Number
- CN202511733092.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-02-27
AI Technical Summary
Existing seabed topographic mapping methods struggle to accurately capture small-scale features in the spatial domain, resulting in insufficient model accuracy and generalization ability. Furthermore, traditional methods face difficulties in integrating with physical theories in the frequency domain.
A multi-source seabed topography reconstruction method based on frequency domain attention neural network is adopted. By introducing fast Fourier transform and attention mechanism, combined with fully connected layer and ReLU activation function, a frequency domain attention model is constructed. Feature fusion is performed using GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient and vertical deviation data, and Parker formula is introduced as physical constraint.
It significantly improves the accuracy and precision of seabed topography reconstruction, effectively captures topographic features at multiple spatial scales, and enhances the physical interpretability and generalization ability of the model.
Smart Images
Figure CN121582489A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seabed topography reconstruction technology, and in particular to a multi-source seabed topography reconstruction method and system based on a frequency domain attention neural network. Background Technology
[0002] Accurate mapping of seabed topography is crucial for deep-sea exploration. Currently, seabed mapping methods can be broadly categorized into direct exploration using shipborne multibeam echo sounders and indirect inversion and reconstruction methods based on physical theories or data-driven models. Shipborne multibeam measurements provide high-precision depth measurements, but to date, only about 25% of the global seabed has been mapped with a resolution better than 400 meters, making the measurements costly and time-consuming. To address the limitations of direct measurements, indirect depth inversion methods have been developed. While satellite gravity-based inversion methods offer global coverage, their spatial resolution and accuracy are limited. To improve the accuracy of indirect seabed mapping, purely data-driven methods based on artificial intelligence have emerged, improving the inversion results to some extent. All of the above models operate in the spatial domain. In contrast, traditional seabed gravity inversion research is fundamentally rooted in the frequency domain. Frequency domain analysis provides deeper insights into topographic features across multiple spatial scales. Seabed topographic data is characterized by a wide dynamic range and high variance, causing small-scale features to be masked by adjacent large-scale features in spatial domain models. This limitation restricts the ability of spatial domain methods to capture fine details, ultimately reducing the accuracy and generalization ability of the models. Furthermore, since most theoretical inversion techniques are essentially developed in the frequency domain, spatial domain models face the challenge of achieving tight integration with physical theories. Summary of the Invention
[0003] The purpose of this invention is to provide a multi-source seabed topography reconstruction method and system based on a frequency domain attention neural network to solve the problems mentioned in the background art. By introducing the Fourier transform mechanism into the neural network architecture, attention in the frequency domain is realized, thereby improving the accuracy of seabed topography reconstruction.
[0004] To achieve the above objectives, the present invention provides the following solution: On one hand, it provides a multi-source seabed topography reconstruction method based on a frequency domain attention neural network, the specific steps of which include the following: A five-channel input tensor is constructed based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. The five-channel input tensor is standardized to construct a dataset, and the dataset is randomly divided into a training set and a test set by uniform sampling. The training set is processed using a frequency domain attention model to obtain fused features, which are then input into a predictor to generate a water depth estimate. The frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.
[0005] Preferably, the specific steps for processing the training set using a frequency domain attention model are as follows: For each data source in the training set, a two-dimensional fast Fourier transform is performed to extract frequency domain features; Each frequency domain feature is processed by an independent encoder and then spliced and fused to form a comprehensive feature representation; The comprehensive feature representation is input into the attention module, and a five-dimensional weight vector is calculated through a fully connected layer and a Softmax activation function; The fused features are obtained by weighting and summing the five encoded features using the weight vector and then connecting the residuals.
[0006] Preferably, the frequency domain attention model uses mean squared error as the loss function and is trained using the Adam optimizer.
[0007] Preferably, the method further includes embedding the Parker formula into the loss function of the frequency domain attention model, wherein the total loss function is a weighted sum of the mean squared error loss and the Parker constraint loss; the Parker constraint loss is calculated based on gravity anomalies in the frequency domain.
[0008] Preferably, a five-channel input tensor is constructed within an 11×11 grid.
[0009] Preferably, each of the independent encoders is implemented by two fully connected layers, containing 256 and 128 neurons respectively; the attention module is composed of fully connected layers, generating a five-dimensional weight vector followed by a Softmax activation function; the predictor maps the weighted features to the final output through an additional fully connected layer.
[0010] On the other hand, a multi-source seabed topography reconstruction system based on a frequency domain attention neural network is provided, including a data acquisition module, a preprocessing module, and a prediction module; wherein, The data acquisition module is used to construct a five-channel input tensor based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. The preprocessing module is used to standardize the five-channel input tensor, construct a dataset, and randomly divide the dataset into a training set and a test set by uniform sampling. The prediction module is used to process the training set using a frequency domain attention model to obtain fused features, and input the fused features into the predictor to generate a water depth estimate; wherein, the frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.
[0011] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: Fourier transform mechanism is introduced into neural network architecture to realize attention in the frequency domain, and a multi-source fusion model based on frequency domain attention mechanism is proposed for seabed topography reconstruction. By introducing a frequency domain transform layer, topographic features at multiple spatial scales can be effectively captured. Frequency domain processing is used to improve multi-source data fusion, which significantly improves reconstruction accuracy. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This invention provides a complete flowchart of data organization and multi-source data fusion for constructing seabed topography. Figure 3 This is a structural diagram of the frequency domain attention model of the present invention. Detailed Implementation
[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example
[0015] The purpose of this invention is to provide a multi-source seabed topography reconstruction method based on a frequency domain attention neural network, such as... Figure 1 As shown, the specific steps include the following: S1. Construct a five-channel input tensor based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. S2. Standardize the five-channel input tensor to construct a dataset, and randomly divide the dataset into a training set and a test set by uniform sampling. S3. The training set is processed using a frequency domain attention model to obtain fused features, which are then input into the predictor to generate a water depth estimate. The frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.
[0016] Furthermore, the data obtained in S1 includes the following: Seafloor topographic grid data: GEBCO2024 seafloor topographic grid data was used as the background data for the model. GEBCO (General Bathymetric Chart of the Oceans) is a project jointly supported by the International Hydrographic Organization (IHO) and the Intergovernmental Oceanographic Commission (IOC) of UNESCO, aiming to provide global seafloor topographic data and continuously update it to improve the accuracy of global seafloor topographic mapping.
[0017] Ocean gravity parameter data: SWOT satellite-optimized version of gravity anomaly data.
[0018] The Vertical Gravity Gradient (VGG) dataset also plays a crucial role in constructing detailed seabed topography through multi-source data fusion in this invention. VGG is highly sensitive to high-frequency topographic changes, making it particularly valuable for simulating complex or steep terrain. Incorporating VGG into model training helps improve prediction accuracy and enhances the physical interpretability of the results.
[0019] The data is derived from the SWOT dataset, specifically the vertical deviation (DOV) data, which includes north and east components with a resolution of 1 arcminute. The vertical deviation measures the angle between the actual gravity vector and the normal to the reference ellipsoid, reflecting variations in subsurface density. The east component captures lateral density variations, while the north component reveals crustal tectonic features. This data enables the model to detect subtle topographic and density variations.
[0020] Compared to conventional image processing tasks, seabed topography reconstruction faces unique challenges. First, seabed data is characterized by high variance and a wide dynamic range, which can lead to the masking or smoothing of small-scale features during normalization. Second, seabed topography boundaries often exhibit gradual features, with fluctuating gradients rather than steep transitions. To address these challenges, this invention employs a field-to-point strategy, prioritizing high granularity for each prediction point while maintaining spatial relevance. The dataset construction process is as follows: For each target location, a five-channel input tensor is constructed within an 11×11 grid, containing GEBCO depth data, gravity anomalies, vertical gravity gradients, and the north and east components of the vertical deviation. The 11×11 window size is chosen to strike a balance between local feature capture and computational efficiency, consistent with the typical spatial scale of seabed morphology. To mitigate boundary effects, all out-of-bounds values are filled with the global mean of the corresponding channel. All input data are standardized to zero mean and unit variance to alleviate the smoothing effect caused by variance. Figure 2 It demonstrates the complete data organization process.
[0021] To better capture periodic patterns in multi-source data, a frequency-domain attention model (FA) was developed, which combines Fast Fourier Transform (FFT) preprocessing with an attention mechanism. This model processes GEBCO depth data, gravity anomalies, vertical gravity gradients, and vertical deviation data within an 11×11 window, converting the input to the frequency domain before applying the attention mechanism. As an extension of the FA model, CNN-based encoding and multi-head self-attention can be introduced to form a frequency-domain spatially enhanced attention model (FEA). This adds convolutional network layers to enhance spatial feature extraction. However, experimental results show that the FEA model does not perform as expected because the enhanced spatial feature extraction of the convolutional neural network does not benefit the signal processing within the frequency-domain neural network; instead, it causes redundant model parameters and some overfitting. Therefore, overall, the frequency-domain attention model is the core of this invention and offers the best technical performance.
[0022] Furthermore, the specific steps for processing the training set using the frequency domain attention model in S3 are as follows: S31. For each data source in the training set, perform two-dimensional fast Fourier transform processing to extract frequency domain features; S32. Each frequency domain feature is processed by an independent encoder and then spliced and fused to form a comprehensive feature representation; S33. The comprehensive feature representation is input into the attention module, and a five-dimensional weight vector is calculated through a fully connected layer and a Softmax activation function. S34. The fused features are obtained by weighting and summing the five encoded features using the weight vector and connecting them with the residuals.
[0023] Figure 3 The diagram shows the structure of the frequency domain attention model. Each data source undergoes a two-dimensional FFT to extract frequency components. The computation process decomposes the spatial grid signal into complex frequency domain features. Since the input window size is 11×11 (i.e., 121 points), and the real and imaginary parts of the transformed complex numbers are concatenated, each data source generates a feature vector of size 242. Subsequently, each frequency domain feature is processed by an independent encoder. Each encoder contains two fully connected layers, using the ReLU activation function to map the feature dimensions to 128 and 64, respectively. Next, the features encoded from all five data sources are concatenated to form a comprehensive feature representation. This comprehensive feature is fed into an attention module, which computes a five-dimensional weight vector through fully connected layers and a Softmax activation function to measure the importance of each data source. The weight vector is then used to weight and sum the five encoded features to obtain a weighted feature representation. To enhance training stability, a residual connection is introduced, adding the weighted feature to the result of a linear transformation of the concatenated feature. Finally, the fused features are mapped to the final water depth prediction through a predictor (composed of a fully connected layer and a ReLU activation function). The frequency domain attention model uses mean squared error (MSE) as the loss function and is trained with an Adam optimizer with a learning rate of 0.0001. By introducing FFT, which effectively captures spectral features, this model improves the performance of traditional attention mechanisms in handling oscillating data.
[0024] Furthermore, this invention also integrates the Parker formula (a fundamental relationship between gravity anomalies and seafloor depth in the frequency domain) into the neural network framework. Specifically, the loss function of the frequency domain attention model needs to be modified by adding a physical constraint term derived from the Parker formula. The core idea is to ensure that the predicted seafloor depth conforms to the physical relationship between gravity anomalies and water depth, thereby enhancing physical interpretability and predictive rationality. The Parker formula can be expressed as an infinite series in the frequency domain. This formula describes the relationship between the Fourier transform of gravity anomalies and seafloor depth, involving parameters such as the gravitational constant, density difference, and wavenumber. This equation, through infinite series expansion, captures the relationship between gravity signals and subsurface structures.
[0025] Specifically, the Parker constraint is added as an auxiliary term to encourage predictions to conform to geophysical principles. The total loss function is defined as the weighted sum of the mean squared error loss and the Parker constraint loss, where the weight hyperparameter is set to 0.1 in this embodiment. The Parker constraint loss term is calculated based on gravity anomalies in the frequency domain. Considering the complex series involved in the Parker formula, a simplification is performed in this invention. The specific simplification method is as follows: first, a first-order approximation of the Parker formula is used; then, the magnitude of the Fourier transform of the gravity anomaly is averaged and multiplied by an empirical scaling factor (set to 1000 in this invention). This empirical scaling factor is determined based on the data volume and practical experience, and it replaces various complex constant terms in the original formula. The basic principle of this simplification is that this component only functions as a constraint term. By doing so, it is ensured that physical constraints can be seamlessly and efficiently integrated into the computational process of the neural network, while retaining a certain degree of physical information. This formulation ensures the physical rationality of the predictions, while allowing the model to achieve a balance between data fitting and constraint compliance.
[0026] The following experiments compare the frequency domain attention model (FA) with three classic deep learning models (MLP, CNN, and Transformer) and two spatial domain attention models (basic attention model and enhanced spatial multi-head self-attention model ES-MHSA) to evaluate the accuracy and robustness of reconstructing seabed topography in different geological regions.
[0027] The quantitative accuracy metrics of each model in different regions are summarized in Table 1. The results clearly show that the basic attention model and the FA model consistently outperform the GEBCO baseline data as well as the MLP, CNN, and Transformer models in all six test regions. This highlights the advantages of channel-level attention mechanisms over traditional deep learning architectures in seabed topography reconstruction.
[0028]
[0029]
[0030]
[0031]
[0032]
[0033]
[0034] Attention-based models outperform traditional attention-based models because they can dynamically prioritize and fuse information from heterogeneous data sources. In contrast, MLP and CNN models lack attention mechanisms: MLPs perform homogeneous fusion across all channels, while CNNs focus on extracting regional features within a single channel. While Transformer models emphasize interactions between labels, they cannot adaptively adjust the weights of different data channels. Notably, the frequency-domain attention model consistently outperforms basic attention models across all regions, demonstrating the added value of frequency-domain attention mechanisms compared to spatial-domain attention.
[0035] To further investigate the model's performance at different scales, frequency domain error analysis was performed using band RMSE plots and power spectral density plots. The band RMSE plots show that the FA model exhibits the lowest RMSE across almost all frequency ranges in all regions. The FA model demonstrates a significant advantage across the entire spectrum, indicating its ability to accurately reconstruct features at different spatial scales. This implies that, in most cases, the FA model's frequency domain attention mechanism can accurately detect the frequency band information of the terrain, thereby achieving efficient and targeted feature reconstruction in different frequency domains.
[0036] The error PSD plot further confirms these findings: across all regions, the FA model maintains the lowest error spectral energy at all frequencies, highlighting its superiority in reconstructing terrain features at all scales. In contrast, the FEA model, which incorporates a CNN encoder in the frequency domain, does not achieve optimal results. Its relatively high error energy and parameter redundancy suggest that CNN-based spatial feature extraction may not be effectively transferred to frequency domain representation.
[0037] In summary, these results demonstrate the effectiveness of frequency-domain attention models in seabed mapping, particularly in complex and highly heterogeneous environments. By leveraging spectral characteristics, these models minimize errors and outperform traditional and spatial domain deep learning models, advancing the technological level of marine geology, resource exploration, and climate modeling.
[0038] On the other hand, a multi-source seabed topography reconstruction system based on a frequency domain attention neural network is provided, including a data acquisition module, a preprocessing module, and a prediction module; wherein, The data acquisition module is used to construct a five-channel input tensor based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. The preprocessing module is used to standardize the five-channel input tensor, construct a dataset, and randomly divide the dataset into a training set and a test set by uniform sampling. The prediction module processes the training set using a frequency domain attention model to obtain fused features, which are then input into the predictor to generate a water depth estimate. The frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.
[0039] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for reconstructing multi-source seabed topography based on a frequency domain attention neural network, characterized in that, The specific steps include the following: A five-channel input tensor is constructed based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. The five-channel input tensor is standardized to construct a dataset, and the dataset is randomly divided into a training set and a test set by uniform sampling. The training set is processed using a frequency domain attention model to obtain fused features, which are then input into a predictor to generate a water depth estimate. The frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.
2. The multi-source seabed topography reconstruction method based on a frequency domain attention neural network according to claim 1, characterized in that, The specific steps for processing the training set using the frequency domain attention model are as follows: For each data source in the training set, a two-dimensional fast Fourier transform is performed to extract frequency domain features; Each frequency domain feature is processed by an independent encoder and then spliced and fused to form a comprehensive feature representation; The comprehensive feature representation is input into the attention module, and a five-dimensional weight vector is calculated through a fully connected layer and a Softmax activation function; The fused features are obtained by weighting and summing the five encoded features using the weight vector and then connecting the residuals.
3. The multi-source seabed topography reconstruction method based on a frequency domain attention neural network according to claim 1, characterized in that, The frequency domain attention model uses mean squared error as the loss function and is trained using the Adam optimizer.
4. The multi-source seabed topography reconstruction method based on a frequency domain attention neural network according to claim 3, characterized in that, It also includes embedding the Parker formula into the loss function of the frequency domain attention model, whereby the total loss function is a weighted sum of the mean squared error loss and the Parker constraint loss; the Parker constraint loss is calculated based on gravity anomalies in the frequency domain.
5. The multi-source seabed topography reconstruction method based on a frequency domain attention neural network according to claim 1, characterized in that, Construct a five-channel input tensor within an 11×11 grid.
6. The multi-source seabed topography reconstruction method based on a frequency domain attention neural network according to claim 2, characterized in that, Each of the independent encoders is implemented by two fully connected layers, containing 256 and 128 neurons respectively; the attention module is composed of fully connected layers, generating a five-dimensional weight vector followed by a Softmax activation function; the predictor maps the weighted features to the final output through an additional fully connected layer.
7. A multi-source seabed topography reconstruction system based on a frequency domain attention neural network, characterized in that, It includes a data acquisition module, a preprocessing module, and a prediction module; among which, The data acquisition module is used to construct a five-channel input tensor based on the acquired data; the data includes GEBCO water depth data, ocean gravity parameter anomaly data, vertical gravity gradient data, northward component data of vertical deviation, and eastward component data of vertical deviation. The preprocessing module is used to standardize the five-channel input tensor, construct a dataset, and randomly divide the dataset into a training set and a test set by uniform sampling. The prediction module is used to process the training set using a frequency domain attention model to obtain fused features, and input the fused features into the predictor to generate a water depth estimate; wherein, the frequency domain attention model combines fast Fourier transform preprocessing with an attention mechanism, and the predictor consists of a fully connected layer and a ReLU activation function.