Marine magnetic data fusion method based on multi-scale analysis and deep learning feature enhancement

Abstract

The invention discloses an ocean magnetic data fusion method based on multi-scale analysis and deep learning feature enhancement, and the method comprises the steps: obtaining satellite magnetic data and ship measurement magnetic data of the same target region, and carrying out the preprocessing; respectively performing two-dimensional wavelet decomposition on the satellite magnetic data and the ship measurement magnetic data to obtain a low-frequency component, an intermediate-frequency component and a high-frequency component; performing feature enhancement on the low-frequency component of the satellite magnetic data by using UNet and Attention networks; fusing low-frequency, intermediate-frequency and high-frequency components based on an adaptive weight strategy; the low frequency is an enhanced satellite component; and reconstructing the fused components of each frequency band through wavelet inverse transformation to generate a fused magnetic diagram. According to the method, multi-scale features are reserved by wavelet decomposition, and signal distortion is avoided; eliminating frequency domain aliasing through an adaptive weight strategy; uNet + Attention is selected to effectively enhance the low-frequency reliability under the condition of few samples, and the method is suitable for a deep sea sparse data scene.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of marine geophysics and artificial intelligence, and in particular to a method for fusing marine magnetic data based on multi-scale analysis and deep learning feature enhancement, which mainly realizes the fusion of satellite magnetic data and ship-based magnetic data. Background Technology

[0002] Marine magnetic data is a type of geophysical data that measures the intensity and variations of the Earth's magnetic field in marine areas using magnetometers. Its physical significance lies in reflecting the magnetic characteristics of seabed rocks and the structure of the Earth's crust, and it is widely used in seabed resource exploration, plate tectonics research, military target detection, and seabed engineering surveys. Single-source magnetic data (such as ship-based, airborne, satellite, and near-bottom magnetic data) suffers from incomplete spatial coverage, resolution differences, or noise interference, necessitating the use of multi-source data fusion techniques to improve data quality and interpretation accuracy.

[0003] Currently, ocean magnetic data mainly includes satellite magnetic data and ship-based magnetic data. Satellite magnetic data has reliable low-frequency signals and a wide distribution range, but lacks high-frequency details and has low spatial resolution. Ship-based magnetic data has high spatial resolution and rich high-frequency details, but it is subject to noise and has limited coverage. In applications, data fusion is required to obtain large-scale, high-resolution ocean magnetic data.

[0004] Current methods for fusing marine magnetic data primarily employ multi-source data collaborative analysis, including weighted averaging based on gridding and Kriging interpolation, aiming to integrate heterogeneous magnetic data from multiple sources to improve resolution and accuracy. The challenges of this fusing approach lie in:

[0005] 1) Magnetic data exhibits significant differences in spatiotemporal scales, necessitating the resolution of the benchmark unification issue;

[0006] 2) Magnetic data has multi-scale characteristics, and ignoring the fusion of multi-scale characteristics will lead to signal distortion and resolution loss;

[0007] 3) The sparsity of deep-sea data leads to high interpolation uncertainty.

[0008] Therefore, how to solve the problem of fusion of marine magnetic data is an urgent issue that needs to be addressed by those skilled in the art. Summary of the Invention

[0009] In view of this, the present invention provides a marine magnetic data fusion method based on multi-scale analysis and deep learning feature enhancement. By decomposing and separating frequency domain features through multi-scale analysis, combining UNet+Attention to enhance the reliability of satellite low-frequency signals, and adopting an adaptive weight fusion strategy, the above-mentioned technical problems can be solved at least partially.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] This invention provides a method for fusing marine magnetic data based on multi-scale analysis and deep learning feature enhancement, comprising the following steps:

[0012] S1. Acquire satellite magnetic data and ship-based magnetic data for the same target area and perform preprocessing;

[0013] S2. Perform two-dimensional wavelet decomposition on the satellite magnetic data and ship-based magnetic data respectively to obtain low-frequency, mid-frequency and high-frequency components;

[0014] S3. Feature enhancement of low-frequency components in satellite magnetic data is performed using UNet and Attention networks;

[0015] S4. Fuse low-frequency, mid-frequency, and high-frequency components based on an adaptive weighting strategy; the low-frequency components are the enhanced satellite components.

[0016] S5. Reconstruct the fused frequency band components using inverse wavelet transform to generate a fused magnetograph.

[0017] In one embodiment, the preprocessing in step S1 includes: multi-source benchmark unification, discrete data gridding, and spatial information matching.

[0018] In one embodiment, the wavelet decomposition in step S2 uses the Haar basis function, with a decomposition layer of 3 layers. Each layer outputs one low-frequency component and three high-frequency components; the three high-frequency components are components in the horizontal direction, vertical direction, and diagonal direction, respectively.

[0019] In one embodiment, in step S3, the UNet and Attention network structure includes:

[0020] The encoder consists of two downsampling layers. The first layer is composed of a 3×3 convolution and a ReLU activation function. The input is the low-frequency component of the satellite magnetization data, and the output is a 64-channel feature map. The second layer is composed of a 2×2 max pooling layer, a 3×3 convolutional layer, and a ReLU activation function, and the output is a 128-channel feature map.

[0021] Decoder: Includes two upsampling layers; the first layer consists of 2×2 deconvolution, 3×3 convolution and ReLU activation function, which fuses context information and outputs a 64-channel feature map; the second layer is the output layer: it generates a single-channel enhanced low-frequency component by 1×1 convolution.

[0022] An attention gate module is embedded at the skip connection between the encoder and decoder to generate a spatial weight matrix through Sigmoid activation, which is then multiplied with the encoder features using the Hadamard product.

[0023] In one embodiment, step S4 includes: an adaptive weighting strategy, comprising:

[0024] The low-frequency frequencies use enhanced satellite components with a weight of 70% to 100%.

[0025] High-frequency measurements use ship-based components with a weighting of 90%–100%.

[0026] The intermediate frequency (IF) dynamically allocates satellite and shipborne measurement weights based on the local signal-to-noise ratio.

[0027] The formula for calculating the local signal-to-noise ratio is as follows:

[0028]

[0029] Among them, S i,j For signal components, N i,j X represents the noise component; Y represents the number of grid rows in the latitude direction; i and j represent the grid cell coordinates, indicating the row and column indices of the grid.

[0030] Satellite weight = [SNR] 卫星 / (SNR 卫星 +SNR 船测 )]×100%

[0031] The fusion calculation formula is as follows:

[0032] F i,j =w low ×L i,jsat-enhanced +w high ×H i,jship +w mid ×M i,j

[0033] Among them, F i,j L represents the fused component value. i,jsat-enhanced For the enhanced low-frequency component values ​​of satellite magnetic data, H i,jship For the high-frequency component values ​​of the magnetic field data measured on the ship, M i,j For the intermediate frequency component value, w low w high w mid The weights are assigned to low frequency, high frequency, and mid frequency, respectively.

[0034] In one embodiment, the reconstruction result in step S5 is quality-assessed using root mean square error (RMSE), correlation coefficient (CC), and signal-to-noise ratio (SNR).

[0035] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages:

[0036] (1) This invention can decompose data into components of different frequencies through multi-scale analysis of wavelet decomposition. During the fusion process, the features of different scales are processed, effectively preserving the multi-scale features of the data and overcoming the problem that traditional interpolation methods cannot preserve multi-scale features.

[0037] (2) The present invention adopts an adaptive weight setting and a multi-scale decomposition fusion strategy to perform fusion on different frequency components, thereby avoiding the frequency domain aliasing problem caused by direct weighted fusion and improving the frequency resolution and accuracy of the fused data.

[0038] (3) This invention combines wavelet decomposition and deep learning feature enhancement, and adopts UNet+

[0039] Attention network models enhance the low-frequency components of satellite magnetic data without requiring extensive training with paired data. This makes them particularly effective for data fusion, especially in deep-sea areas where ship-based data samples are scarce, thus broadening the application scope of data fusion technology. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0041] Figure 1 The flowchart of the marine magnetic data fusion method based on multi-scale analysis and deep learning feature enhancement provided by this invention is shown.

[0042] Figure 2 The diagram shows the low-frequency, mid-frequency, and high-frequency components obtained by two-dimensional wavelet decomposition provided by this invention.

[0043] Figure 3 This is a diagram of the UNet+Attention network structure provided by the present invention.

[0044] Figure 4 Satellite magnetic anomaly diagram provided as an embodiment of the present invention.

[0045] Figure 5 The image shows a magnetic anomaly measured on a ship, as provided in an embodiment of the present invention.

[0046] Figure 6a This is a schematic diagram of the wavelet decomposition results of satellite data provided in an embodiment of the present invention.

[0047] Figure 6b This is a schematic diagram of wavelet decomposition results of ship survey data provided in an embodiment of the present invention.

[0048] Figure 7 The embodiment provided by the present invention is a fused and reconstructed marine magnetic anomaly map. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Reference Figure 1 As shown in the figure, this invention discloses a method for fusing marine magnetic data based on multi-scale analysis and deep learning feature enhancement, including the following steps:

[0051] S1. Acquire satellite magnetic data and ship-based magnetic data for the same target area and perform preprocessing. The preprocessing mainly includes multi-source benchmark unification, discrete data gridding, and spatial information matching.

[0052] 1. Unification of multi-source references:

[0053] Unify the calibration benchmarks and reference frames for magnetic data from different sources to eliminate systematic biases. The core is to ensure that all data (such as satellite and shipborne measurements) use the same normal field model to avoid errors caused by temporal variations in the geomagnetic field or differences in instruments.

[0054] For example, applying normal field correction (removing the Earth's main magnetic field) and standardizing timestamps and coordinate systems (such as WGS84).

[0055] 2. Discrete data gridding:

[0056] The discrete ship survey data points are converted into a regular grid format to ensure spatial consistency. The core issue is solving the problems of data sparsity and irregular sampling, and generating uniform grid data through interpolation algorithms.

[0057] Choose a grid resolution (e.g., 2 points) and use an interpolation method (e.g., Kriging or inverse distance weighting).

[0058] 3. Spatial Information Matching

[0059] To ensure that multi-source data are aligned within the same geographic coordinate system, projection bias and coordinate offset are addressed. The core principle is to match pixel or grid point locations through georeferencing. This can be achieved using control points (such as known landmarks) or automatic registration algorithms (such as mutual information).

[0060] S2. Perform two-dimensional wavelet decomposition on the satellite magnetic data and ship-based magnetic data respectively to obtain low-frequency, mid-frequency and high-frequency components;

[0061] In this step, two-dimensional wavelet decomposition is performed on the two types of preprocessed data. Wavelet analysis is a multi-resolution signal analysis method that decomposes data into components of different frequencies, including low-frequency and high-frequency components, by selecting appropriate wavelet basis functions and decomposition levels. The low-frequency components mainly contain the overall trend and general information of the data, while the high-frequency components contain the detailed features and noise information. Through multi-scale decomposition, the features of satellite magnetic data and shipborne magnetic data at different frequency scales can be separated, providing conditions for subsequent feature enhancement and fusion.

[0062] The principle of wavelet analysis is to represent a signal as a linear combination of wavelet functions, and its formula can be expressed as:

[0063] f(t)=∑ a,b C a,b ψ a,b (t)

[0064] Among them, C a,b For wavelet coefficients, ψ a,b (t) represents the wavelet basis function, a represents the scaling parameter, and b represents the translation parameter.

[0065] The result of decomposition is the breakdown of data into low-frequency and high-frequency components at different levels. For example, in wavelet decomposition of two-dimensional data, each level yields one low-frequency component (LL) and three high-frequency components (LH, HL, HH), corresponding to detailed information in the horizontal, vertical, and diagonal directions, respectively. As the number of decomposition levels increases, more refined multi-scale features can be obtained. This avoids the smoothing distortion of Kriging interpolation (preserving local details of the HH component) and avoids frequency domain aliasing caused by direct fusion (processing frequency bands independently). Figure 2 As shown, performing three or more levels of two-dimensional wavelet decomposition yields high, medium, and low frequencies. This embodiment of the invention uses a three-level decomposition. The high-frequency components from the first level are designated as high frequencies, the high-frequency components from the second level are designated as medium-frequency components, and the low-frequency components from the third level are designated as low-frequency components. Because the components obtained from each level of decomposition are referred to as high-frequency and low-frequency data, these high, medium, and low frequencies are artificially defined.

[0066] S3. Feature enhancement of low-frequency components in satellite magnetic data is performed using UNet and Attention networks;

[0067] UNet is an image segmentation model based on convolutional neural networks, featuring an encoder and decoder structure. In this invention, a combined UNet+Attention structure is selected, such as... Figure 3As shown, it is mainly used for low-frequency feature enhancement. Its core architecture consists of two parts: an encoder (feature extraction) and a decoder (feature recovery and enhancement). The encoder extracts and downsamples the low-frequency components of the input through convolution and pooling operations, while the decoder upsamples and recovers the features through deconvolution and convolution operations, finally outputting the enhanced low-frequency components. An attention gate module is embedded at the connection between the encoder and decoder to dynamically focus on the low-frequency signal region and suppress the weights of the noise region according to the spatial distribution of the magnetic data.

[0068] (1) Encoder-decoder architecture:

[0069] ① Encoder (2-layer downsampling):

[0070] Layer 1 (enc1): Gradually extract abstract features and compress spatial dimensions.

[0071] The input is the low-frequency component of satellite magnetism data (single channel, dimension H×W), for example... Figure 3 The 1×256×256 feature map is processed by a 3×3 convolutional layer (Conv, padded with 1s to maintain spatial dimensions) and a ReLU activation function to extract basic features and output a 64-channel feature map (64×256×256). Its purpose is to capture basic edge and texture features.

[0072] Layer 2 (enc2): First, a 2×2 max-pooling layer (MaxPool, stride 2) halves the feature map size to 128×128. Then, a 3×3 convolutional layer (padding 1) and ReLU activation function are used to further extract abstract features, outputting a 128-channel feature map (128×128×128). Another 2×2 max-pooling layer (MaxPool, stride 2) is used for downsampling, outputting a 64×64 feature map (128×64×64). Its purpose is to extract high-level semantic features (such as regional magnetic field trends) while reducing the feature map size to lower computational complexity. The bottleneck layer expands the channels to 256, outputting a 256×64×64 feature map. Its purpose is to integrate multi-scale features, capture global magnetic field background trends, and provide high-dimensional feature selection for subsequent attention mechanisms.

[0073] ② Decoder (2-layer upsampling): Its function is to restore spatial details and output enhanced features.

[0074] Layer 1 (dec1): The feature map size is restored to the input size (128×128×128) through a 2×2 deconvolutional layer UpConv (stride 2), followed by a 3×3 convolutional layer (padding 1) and a ReLU activation function to fuse contextual information and output a 64-channel feature map (64×256×256). Its purpose is to fuse contextual information to reconstruct details.

[0075] Output layer (final): Compresses the feature map into a single channel using a 1×1 convolutional layer, outputting the enhanced low-frequency component (dimension consistent with the input, H×W) (64×256×256). Its function is to compress channels and match the original data format.

[0076] (2) Attention gate module:

[0077] An attention gate is embedded at the skip connection between encoding and decoding to fuse shallow details and deep semantic features through the skip connection, avoiding information loss; its structure includes:

[0078] Gated signal generation: 1×1 convolution Conv+Sigmoid;

[0079] Feature recalibration: Hadamard product;

[0080] Output: Attention weights × encoder features, i.e., low-frequency data augmentation results (1×256×256).

[0081] In this embodiment, a lightweight design is adopted: only two layers of downsampling / upsampling are used to reduce computational overhead (suitable for large-scale ocean data); attention-guided processing is used to overcome the shortcomings of traditional UNet in equal fusion and improve the fidelity of low-frequency signals; it can achieve low-sample adaptation, for example, training can be performed with 50% satellite data, solving the problem of scarce deep-sea data.

[0082] S4. Fuse low-frequency, mid-frequency, and high-frequency components based on an adaptive weighting strategy; the low-frequency components are the enhanced satellite components.

[0083] During the fusion process, adaptive weight settings are used for fusion based on the characteristics and data properties of different frequency components:

[0084] (1) Low-frequency satellite magnetic data components enhanced by UNet are selected, with the initial weight set at 70% to 100%, which can be adjusted, and 70% is preferred. Since the low-frequency signal of satellite magnetic data is reliable, UNet enhancement can better preserve the overall trend of the data and supplement certain high-frequency details, so it is given a higher weight.

[0085] (2) The high-frequency component of the ship-based magnetic data is selected, with the weight initially set at 90% to 100%, which can be adjusted, and 90% is preferred. The high-frequency details of the ship-based magnetic data are rich and can provide fine features of the data. Therefore, the ship-based data is mainly used in the high-frequency part to make full use of its high-frequency detail advantage.

[0086] (3) Intermediate frequency (IF) dynamic weights are set based on local signal-to-noise ratio (SNR). The formula for calculating local SNR is:

[0087]

[0088] Among them, S i,j For signal components, N i,j X represents the noise component; Y represents the number of grid rows in the latitude direction; i and j represent the grid cell coordinates, indicating the row and column indices of the grid.

[0089] By calculating the local signal-to-noise ratio, the weights of satellite magnetic data and ship-based magnetic data in the intermediate frequency component can be dynamically adjusted according to the signal and noise conditions in different regions, thereby improving the quality of the fused data.

[0090] Satellite weight = [SNR] 卫星 / (SNR 卫星 +SNR 船测 )]×100%

[0091] The fusion calculation formula is as follows:

[0092] F i,j =w low ×L i,jsat-enhanced +w high ×H i,jship +w mid ×M i,j

[0093] Among them, F i,j L represents the fused component value. i,jsat-enhanced For the enhanced low-frequency component values ​​of satellite magnetic data, H i,jship For the high-frequency component values ​​of the magnetic field data measured on the ship, M i,j For the intermediate frequency component value, w low w high w mid The weights are assigned to low frequency, high frequency, and mid frequency, respectively.

[0094] S5. Reconstruct the fused frequency band components using inverse wavelet transform to generate a fused magnetograph.

[0095] Based on the principle of wavelet reconstruction, the fused frequency components are reconstructed to obtain the fused magnetic data. Wavelet reconstruction is the inverse process of wavelet decomposition; by combining different frequency components in the reverse order of decomposition, the original data form is restored. The reconstructed fused magnetic data can be visualized to intuitively observe the data distribution and characteristics, and quality assessment can be performed to verify the fusion effect.

[0096] The marine magnetic data fusion method provided by this invention, based on multi-scale analysis and deep learning feature enhancement, uses wavelet decomposition to preserve multi-scale features and avoid signal distortion; it eliminates frequency domain aliasing through an adaptive weighting strategy; and it selects UNet+Attention to effectively enhance low-frequency reliability under few sample conditions, making it suitable for deep-sea sparse data scenarios.

[0097] Example:

[0098] (I) Data Preparation

[0099] Taking satellite and shipborne magnetic data from the Western Pacific Ocean as examples, data fusion processing was performed to verify the effectiveness and feasibility of the method proposed in this invention. The data range is 152°-157°E and 23.5°-26.5°N. The satellite magnetic data mainly comes from EMAG2_V3, released by the US NCEI in 2017, with a grid resolution of 2 mins. Figure 4 As shown, the shipborne magnetic data mainly comes from the shipborne magnetic data released by NCEI, with the horizontal axis representing longitude and the vertical axis representing latitude.

[0100] (II) Data Preprocessing

[0101] For multi-source shipborne magnetic field data, unified normal field correction and gridding are performed to ensure data quality and consistency, resulting in data with uniform grid intervals, such as... Figure 5 As shown, the horizontal axis represents longitude, and the vertical axis represents latitude.

[0102] For example, both types of data are uniformly calibrated to the IGRF-13 model (international standard) of the 2020.0 epoch, eliminating magnetic field drift caused by different measurement years (e.g., satellite data is from 2017, while ship-based data may be from different years). This makes the residual magnetic anomalies (after removing the main magnetic field) of satellite and ship-based data comparable under the same physical reference. It also avoids signal distortion caused by inconsistent references during fusion (e.g., smoothing of seamount magnetic anomalies).

[0103] For example, the characteristics of the raw data:

[0104] Ship survey data: discrete points (such as measurement route data, with uneven density and sparse distribution in deep-sea areas).

[0105] Satellite data: gridded (2-minute resolution, approximately 3.7km × 3.7km grid).

[0106] The meshing process is as follows:

[0107] Step 1: Set the target grid to a 2-minute resolution (matching satellite data).

[0108] Step 2: Apply the Kriging interpolation algorithm to generate a regular mesh based on the spatial correlation of ship survey points. Simplified formula:

[0109]

[0110] Where Z is the magnetic anomaly value, λ i The weighting coefficients are determined by the variogram model.

[0111] Ship survey data was converted into a regular grid within the range of 152°–157°E and 23.5°–26.5°N, such as... Figure 5 As shown, this can eliminate uneven coverage issues (such as gaps between shipping lines) and ensure spatial alignment during subsequent fusion (avoiding holes in deep-sea areas).

[0112] In addition, spatial information matching is processed as follows:

[0113] Satellite data: global grid, but local coordinate offsets may exist (such as grid projection errors in EMAG2_V3).

[0114] Shipborne survey data: After gridding, it needs to be precisely overlapped with the satellite grid.

[0115] The matching process is as follows:

[0116] Step 1: Select a common control point (such as the known top point of Haishan Mountain or the location of the magnetic anomaly peak).

[0117] Step 2: Apply affine transformation to adjust the coordinates of the ship survey grid, ensuring that each grid point (i,j) corresponds to the same latitude and longitude. Transformation formula:

[0118]

[0119] Where a, b, c, d are rotation / scaling parameters, and t x ,t y This represents the translation amount.

[0120] For example, the achievable result is complete alignment of satellite and shipborne data within the 152°-157°E and 23.5°-26.5°N region (error < 0.01 degrees). This prevents spatial misalignment during fusion (such as blurred magnetic anomaly boundaries) and improves the accuracy of subsequent wavelet decomposition.

[0121] In applications in the Western Pacific, the RMSE of preprocessed data was reduced by 30%, laying the foundation for multi-scale decomposition and fusion. This method is applicable to the fusion of global ocean magnetic data (such as in scenarios like the Arctic or South China Sea).

[0122] (III) Multiscale Decomposition

[0123] By selecting appropriate wavelet basis functions (Haar is used in this embodiment) and a decomposition level of 3, two-dimensional wavelet decomposition is performed on the preprocessed satellite magnetic data and ship-based magnetic data, respectively, to obtain the frequency components as follows: Figures 6a-6b As shown, the horizontal axis represents longitude, and the vertical axis represents latitude.

[0124] (iv) Deep Learning Feature Enhancement

[0125] We collected 50% of the low-frequency components of satellite magnetic data as training samples, divided them into a training set using a sliding window, and interpolated the ship-measured low-frequency magnetic data onto the satellite data grid as ground truth labels to train the UNet+Attention model. We then input the low-frequency components of the satellite magnetic data into the trained UNet+Attention model to obtain the enhanced low-frequency components of the satellite magnetic data.

[0126] (V) Adaptive Fusion

[0127] According to the adaptive fusion strategy, the weight of low-frequency satellite data is set to 70%, the weight of ship survey data is set to 30%, the weight of high-frequency satellite data is set to 10%, the weight of ship survey data is set to 90%, and the weight of mid-frequency data is calculated based on the local signal-to-noise ratio to fuse the decomposed frequency components.

[0128] (vi) Magnetic anomaly reconstruction

[0129] Using the wavelet reconstruction algorithm, the fused frequency components are reconstructed through wavelet inverse operation to obtain the fused magnetic data, such as... Figure 7 As shown, the horizontal axis represents longitude, and the vertical axis represents latitude. The quality of the reconstructed magnetic data was evaluated using metrics including root mean square error (RMSE), correlation coefficient (CC), and signal-to-noise ratio (SNR). The results indicate that the fused data has high quality and accuracy.

[0130] 1. Root Mean Square Error (RMSE)

[0131] This measure assesses the degree of deviation between the fused data and the actual data (ship survey data), reflecting the overall accuracy. A lower value indicates a smaller error and that the fused data is closer to the actual value.

[0132] The formula is as follows:

[0133]

[0134] Among them, y i The magnetic anomaly value at grid point i is the true value of the ship survey data. The magnetic anomaly value at grid point i is used to fuse the data; N is the total number of grid points (e.g., in the Western Pacific Ocean at a resolution of 2 minutes, the number of grid points is approximately 90 rows × 150 columns = 13,500 points).

[0135] Root mean square error (RMSE) quantifies the global error between fused and ship-based survey data, especially for interpolation uncertainties in deep-sea areas. The significantly reduced RMSE after fusion indicates that this method reduces the smoothing distortion of traditional interpolation. This method preserves details (such as seamount boundaries) and reduces local errors through wavelet frequency division. It directly addresses the problem that traditional interpolation methods cannot preserve multi-scale features, ensuring spatial consistency of the fused data.

[0136] 2. Correlation coefficient (CC)

[0137] Assess the degree of linear correlation between the fused data and the real data, with a value range of [-1, 1], where the closer to 1, the higher the consistency.

[0138] General formula (Pearson correlation coefficient):

[0139]

[0140] in, This is the average of the ship survey data. To merge the data mean.

[0141] The correlation coefficient (CC) verifies whether the fused data retains the spatial pattern of the original magnetic anomaly (such as the trend of seamount tectonics) and avoids frequency domain aliasing.

[0142] Example Result: CC is close to 1, proving that the adaptive fusion strategy is effective.

[0143] The low-frequency portion (UNet augmented satellite data) preserves regional trends (such as the basement magnetic field).

[0144] The high-frequency component (ship survey data weighted 90%) enhances the correlation of details (such as the direction of fault zones).

[0145] Comparative analysis: Direct weighted fusion results in lower correlation coefficient (CC) (due to signal distortion caused by frequency domain aliasing). This method improves correlation through frequency band fusion, avoiding frequency domain aliasing and ensuring reliability for applications such as seabed resource exploration and seabed engineering surveys.

[0146] 3. Signal-to-noise ratio (SNR)

[0147] The ratio of signal power to noise power; a higher value indicates better data quality (less noise interference).

[0148] The above content explicitly defines the local SNR formula:

[0149]

[0150] The SNR was significantly improved after fusion, especially in the deep-sea area where ship survey data is sparse. The SNR of the HH high-frequency component of the ship survey data was improved by 37% after wavelet decomposition, and the fusion reconstruction was further optimized.

[0151] The fused data from the Western Pacific region, through indicator evaluation, confirms that "the resolution has been significantly improved, and the details of seamount structures are clear." This meets the high-precision requirements for military target detection, seabed engineering surveys, and other applications.

[0152] In this embodiment, wavelet decomposition is used to avoid resolution loss caused by global interpolation; in addition, frequency band fusion is used to eliminate aliasing effects; it is suitable for small sample adaptation, and UNet+Attention is used to achieve effective enhancement with 50% of training samples, breaking through the bottleneck of deep-sea data.

[0153] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0154] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for fusing marine magnetic data based on multi-scale analysis and deep learning feature enhancement, characterized in that, Includes the following steps: S1. Acquire satellite magnetic data and ship-based magnetic data for the same target area and perform preprocessing; S2. Perform two-dimensional wavelet decomposition on the satellite magnetic data and ship-based magnetic data respectively to obtain low-frequency, mid-frequency and high-frequency components; S3. Feature enhancement of low-frequency components of satellite magnetic data is performed using UNet and Attention networks; S4. Fusing low-frequency, mid-frequency, and high-frequency components based on an adaptive weighting strategy; The low frequency refers to the enhanced satellite component; S5. Reconstruct the fused frequency band components using inverse wavelet transform to generate a fused magnetograph.

2. The method according to claim 1, characterized in that, The preprocessing in step S1 includes: multi-source benchmark unification, discrete data gridding, and spatial information matching.

3. The method according to claim 1, characterized in that, In step S2, the wavelet decomposition uses the Haar basis function, with a decomposition layer of 3. Each layer outputs one low-frequency component and three high-frequency components; the three high-frequency components are the components in the horizontal direction, vertical direction and diagonal direction, respectively.

4. The method according to claim 1, characterized in that, In step S3, the UNet and Attention network structures include: The encoder consists of two downsampling layers. The first layer is composed of a 3×3 convolution and a ReLU activation function. The input is the low-frequency component of the satellite magnetization data, and the output is a 64-channel feature map. The second layer is composed of a 2×2 max pooling layer, a 3×3 convolutional layer, and a ReLU activation function, and the output is a 128-channel feature map. Decoder: Includes two upsampling layers; the first layer consists of 2×2 deconvolution, 3×3 convolution and ReLU activation function, which fuses context information and outputs a 64-channel feature map; the second layer is the output layer: it generates a single-channel enhanced low-frequency component by 1×1 convolution. An attention gate module is embedded at the skip connection between the encoder and decoder to generate a spatial weight matrix through Sigmoid activation, which is then multiplied with the encoder features using the Hadamard product.

5. The method according to claim 1, characterized in that, Step S4 includes: an adaptive weighting strategy, including: The low-frequency frequencies use enhanced satellite components with a weight of 70% to 100%. High-frequency measurements use ship-based components with a weighting of 90%–100%. The intermediate frequency (IF) dynamically allocates satellite and shipborne measurement weights based on the local signal-to-noise ratio. The formula for calculating the local signal-to-noise ratio is as follows: Among them, S i,j For signal components, N i,j X represents the noise component; Y represents the number of grid rows in the latitude direction; i and j represent the grid cell coordinates, indicating the row and column indices of the grid. Satellite weight = [SNR] 卫星 / (SNR 卫星 +SNR 船测 )]×100% The fusion calculation formula is as follows: F i,j =w low ×L i,jsat-enhanced +w high ×H i,jship +w mid ×M i,j Among them, F i,j L represents the fused component value. i,jsat-enhanced For the enhanced low-frequency component values ​​of satellite magnetic data, H i,jship For the high-frequency component values ​​of the magnetic field data measured on the ship, M i,j For the intermediate frequency component value, w low w high w mid The weights are assigned to low frequency, high frequency, and mid frequency, respectively.

6. The method according to claim 1, characterized in that, The reconstruction results in step S5 are evaluated for quality using root mean square error (RMSE), correlation coefficient (CC), and signal-to-noise ratio (SNR).

Images