A geomagnetic map construction method and device based on an improved LAPSRN algorithm

Abstract

The application discloses a geomagnetic map construction method and device based on an improved LAPSRN algorithm, and belongs to the technical field of geomagnetic map construction. The method first carries out wavelet threshold denoising pretreatment on original geomagnetic data; then, the pretreated data is input into the improved LAPSRN network, multi-scale features are extracted by using a multi-scale convolution kernel in a feature extraction stage, and a local residual learning module is added to relieve gradient explosion; then, a bottom-up and top-down alternating bidirectional feature fusion module is used to fully fuse different scale features; finally, an L2 regularization term is added in a loss function to prevent overfitting. Through multi-scale feature extraction and fusion, the application enhances the perception ability of the network to the complex spatial features of the geomagnetic field, effectively improves the reconstruction accuracy and detail retention capability of the geomagnetic map, and can meet the actual needs of high-precision geomagnetic maps in scenes such as geomagnetic navigation, magnetic exploration and mineral exploration.

Description

Technical Field

[0001] This invention belongs to the field of geomagnetic map construction technology, specifically relating to a geomagnetic map construction method and apparatus based on an improved LAPSRN (Laplacian Pyramid Super-Resolution Network) algorithm. Background Technology

[0002] As a natural resource of the Earth, the geomagnetic field has both long-term stability and global coverage. In recent years, it has received increasing attention in fields such as navigation, magnetic exploration, and mineral exploration, becoming a stable and reliable positioning and detection solution in various scenarios.

[0003] However, obtaining high-precision geomagnetic maps usually requires manual field measurements, which is time-consuming and labor-intensive, resulting in high costs for the preparation and application of geomagnetic maps. Therefore, constructing high-precision geomagnetic maps using a limited amount of measured geomagnetic data has significant theoretical and practical value.

[0004] Existing methods for constructing geomagnetic maps mainly rely on interpolation, with commonly used methods including bicubic interpolation, Kriging interpolation, and particle swarm optimization (PSO) algorithms. While these algorithms offer improvements in metrics such as peak signal-to-noise ratio (PSNR) and root mean square error (RMSE) compared to traditional sparse geomagnetic maps, they struggle to accurately recover the detailed features between magnetic measurement points. Furthermore, conventional interpolation methods like Kriging and inverse distance weighting suffer from fitting smoothing issues, limiting the accuracy of the constructed geomagnetic maps.

[0005] In recent years, with the development of neural network technology, its application in various fields has gradually matured. Neural network-based algorithms have been introduced into the construction of geomagnetic maps, and experiments have shown that various objective evaluation indicators are higher than traditional geomagnetic map reconstruction methods, demonstrating the enormous potential of neural networks in geomagnetic map construction. However, some traditional neural network methods, such as basic CNNs and simple generative adversarial networks, suffer from problems such as shallow network layers, single feature extraction dimensions, and a lack of multi-scale geomagnetic information fusion design. These methods are insufficient in representing the spatial characteristics of complex geomagnetic fields, and the reconstructed geomagnetic maps easily lose local fine features, failing to meet the practical needs of high-precision geomagnetic maps.

[0006] Therefore, how to design a high-precision geomagnetic map construction method that can fully extract and integrate multi-scale geomagnetic features and maintain high-frequency detail information of the geomagnetic map has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a geomagnetic map construction method and apparatus based on an improved LAPSRN algorithm. Based on the improved LAPSRN algorithm, the invention enhances the spatial feature perception capability of the geomagnetic field through multi-scale feature extraction and bidirectional fusion modules, thereby achieving high-precision geomagnetic map reconstruction.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A method for constructing geomagnetic maps based on an improved LAPSRN algorithm, the method comprising:

[0010] Step 1: Perform wavelet thresholding denoising preprocessing on the raw geomagnetic data;

[0011] Step 2: Input the preprocessed data into the improved Laplacian Pyramid Super-Resolution Network (LAPSRN). The improved LAPSRN network includes a cascaded multi-scale feature extraction module and a multi-scale feature fusion module. The multi-scale feature extraction module uses multi-scale dilated convolutional kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers. The multi-scale feature fusion module uses an alternating bottom-up and top-down approach to fuse the multi-scale features of the geomagnetic map and outputs the fused multi-scale enhanced features.

[0012] Step 3: Input the fused multi-scale enhanced features into the loss function of the improved LAPSRN network. The loss function includes an L2 regularization term. Optimize the parameters of the improved LAPSRN network through backpropagation, and finally output a high-resolution geomagnetic map.

[0013] Furthermore, step 1 includes: selecting wavelet basis functions based on support length, symmetry, and vanishing moment; determining the optimal number of wavelet decomposition layers; performing wavelet decomposition on the original geomagnetic data using an improved custom threshold function; performing threshold processing on the high-frequency coefficients after decomposition; retaining the effective low-frequency signals; and finally reconstructing the denoised geomagnetic data through inverse wavelet transform.

[0014] Furthermore, in step 2, the multi-scale feature extraction module adopts a parallel multi-branch structure, using 3×3 convolution kernels corresponding to dilated convolutions with dilation rates of 1, 2, and 3 to extract feature maps at different scales. After concatenating the outputs of the multi-branch structure, dimensionality reduction is performed using 1×1 convolution kernels. The local residual learning adds residual connections at the end of each convolutional layer to construct a direct information transmission channel from the front layer to the back layer.

[0015] Furthermore, in step 2, the multi-scale feature fusion module adopts a bidirectional feature fusion and weighted fusion mechanism. By constructing a bidirectional information interaction channel from top to bottom and from bottom to top, it uses a 1×1 convolutional layer to perform adaptive channel filtering and dimensionality reduction on features at different levels, and uses a 3×3 convolutional layer to perform deeper feature extraction, thus completing global residual learning.

[0016] Furthermore, in the multi-scale feature fusion module, the top-down path fuses low-level detail features with high-level detail features to increase the detail retention capability of high-level features, while the bottom-up path transmits high-level semantic features to the low level to enhance the semantic information of the low level.

[0017] Furthermore, in step 2, the improved Laplacian pyramid super-resolution network adopts a multi-level progressive upsampling architecture, constructing a network containing [a certain number of levels] for a task with an amplification factor of s. The system consists of several progressive levels, each performing a 2x geomagnetic map magnification. Each level includes a feature extraction branch and an image reconstruction branch. The image reconstruction branch employs a residual learning strategy, upsampling the input through bilinear initialization transpose convolution and then adding it pixel-by-pixel to the residual geomagnetic map generated by the feature extraction branch to obtain the current high-resolution output.

[0018] Furthermore, in step 3, the loss function is based on the Charill loss function, with an L2 regularization term added as a weight penalty term. Overfitting of the model is suppressed by controlling the regularization strength. The Charill loss function is used to calculate the difference between the predicted high-resolution geomagnetic map and the true high-resolution geomagnetic map.

[0019] On the other hand, the present invention provides a geomagnetic map construction device based on an improved LAPSRN algorithm, comprising:

[0020] The preprocessing module is used to perform wavelet threshold denoising preprocessing on the raw geomagnetic data;

[0021] The reconstruction module is used to input the preprocessed data into the improved Laplacian pyramid super-resolution network LAPSRN. The improved LAPSRN network includes a multi-scale feature extraction module and a multi-scale feature fusion module cascaded in sequence. The multi-scale feature extraction module uses multi-scale dilated convolution kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers. The multi-scale feature fusion module fuses the multi-scale features of the geomagnetic map in an alternating bottom-up and top-down manner, and outputs the fused multi-scale enhanced features.

[0022] The output module is used to input the fused multi-scale enhanced features into the loss function of the improved LAPSRN network. The loss function includes an L2 regularization term. The parameters of the improved LAPSRN network are optimized through backpropagation, and finally a high-resolution geomagnetic map is output.

[0023] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for constructing a geomagnetic map based on an improved LAPSRN algorithm.

[0024] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for constructing a geomagnetic map based on an improved LAPSRN algorithm.

[0025] The beneficial effects of this invention are as follows:

[0026] This invention employs an improved LAPSRN algorithm for super-resolution reconstruction of geomagnetic maps. By using a deep neural network to learn the end-to-end mapping from a small amount of magnetic measurement data to a high-resolution geomagnetic map, it can effectively capture the spatial distribution characteristics and non-stationary variation patterns of the geomagnetic field, significantly improve the detail preservation capability of the reconstructed geomagnetic map, and overcome the shortcomings of traditional interpolation methods that rely on the geometric relationship of neighboring measurement points for smooth estimation and are difficult to recover high-frequency information.

[0027] This invention introduces a multi-scale convolutional kernel structure in the feature extraction stage and uses dilated convolution to replace large-size convolutional kernels, which expands the receptive field while keeping the number of parameters unchanged, effectively reducing the model complexity; together with the local residual learning module, it alleviates the gradient explosion problem, making the model training more stable and reliable.

[0028] This invention employs a bidirectional feature fusion module that alternates between bottom-up and top-down approaches, achieving an effective combination of features at different scales. It can not only recover large-scale geomagnetic variation trends but also finely reconstruct high-frequency variation information between measurement points, outperforming ordinary neural network methods in both overall consistency and local detail richness.

[0029] This invention incorporates an L2 regularization term into the loss function, which effectively suppresses overfitting, improves the model's generalization performance, and further ensures the accuracy of geomagnetic map reconstruction. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the principle of a geomagnetic map construction method based on an improved LAPSRN algorithm according to the present invention.

[0031] Figure 2 This is a schematic diagram of the LAPSRN network structure;

[0032] Figure 3 This is a schematic diagram of the multi-scale feature extraction module;

[0033] Figure 4 These are schematic diagrams of regular convolution and dilated convolution;

[0034] Figure 5 A schematic diagram of a multi-scale dilated convolution block;

[0035] Figure 6 This is a schematic diagram of the local residual learning module;

[0036] Figure 7 This is a schematic diagram of a multi-scale feature fusion module. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] like Figure 1 As shown, this invention provides a geomagnetic map construction method based on an improved LAPSRN algorithm. It utilizes the LAPSRN algorithm, which ranks among the top neural network algorithms in terms of training speed and super-resolution geomagnetic map reconstruction quality. By further improving it, a multi-scale feature extraction module, a multi-scale feature fusion module, and a residual learning module are added to enhance its super-resolution reconstruction capability and improve the accuracy of geomagnetic map reconstruction.

[0039] Specifically, the method includes:

[0040] Step 1: Perform wavelet threshold denoising preprocessing on the raw geomagnetic data. The preprocessing includes determining the optimal wavelet basis function, wavelet decomposition level, and threshold function set for wavelet threshold denoising.

[0041] Step 2: Input the preprocessed data into the improved Laplacian pyramid super-resolution network LAPSRN. The LAPSRN network includes a multi-scale feature extraction module and a multi-scale feature fusion module cascaded in sequence. The multi-scale feature extraction module uses multi-scale dilated convolution kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers to mitigate gradient explosion. The multi-scale feature fusion module uses an alternating bottom-up and top-down approach to fully fuse the multi-scale features of the geomagnetic map and outputs the fused multi-scale enhanced features.

[0042] Step 3: Input the fused multi-scale enhanced features into the loss function of the LAPSRN network. The loss function includes an L2 regularization term. The improved LAPSRN network parameters are optimized through backpropagation, and finally a high-resolution geomagnetic map is output.

[0043] Furthermore, in step 1, the original geomagnetic sampling data is susceptible to interference from sensor hardware and environmental electromagnetic noise, and wavelet threshold denoising is required to achieve effective separation of signal and noise.

[0044] Based on characteristics such as support length, symmetry, and vanishing moment, db and sym series wavelet basis functions were selected; the number of decomposition layers was initialized to ensure noise separation effect and avoid excessive decomposition leading to loss of effective signals; an improved custom threshold function was adopted to solve the problems of discontinuous hard thresholds and blurred details in soft thresholds, and a fixed threshold estimation method was used to achieve accurate suppression of noise coefficients; wavelet decomposition was performed on the original geomagnetic data, and high-frequency coefficients of each layer were thresholded to retain low-frequency effective signals; finally, the denoised geomagnetic data was reconstructed through inverse wavelet transform, providing high-quality input for super-resolution reconstruction.

[0045] Furthermore, in step 2, as Figure 2 As shown, the original LAPSRN network uses a sparse magnetograph as initial input and achieves super-resolution reconstruction results with different magnifications in a single network through a multi-level progressive upsampling architecture. This network employs a staged learning strategy, targeting magnification factors of... The task is to build a system containing The structure is a progressive hierarchy, with each level performing a 2x magnetograph magnification. Its dual-branch architecture includes a feature extraction branch and a magnetograph reconstruction branch, based on a Laplacian pyramid design, achieving low-resolution to high-resolution conversion through multi-level cascading. In the feature extraction stage, each... The hierarchical unit adopts a multi-scale feature fusion strategy: by It consists of one convolutional layer and one transposed convolution. The output of the transposed convolution is used to generate... The residual magnetograph also provides input for high-level feature extraction. During the magnetograph reconstruction stage, each... The level unit employs a residual learning strategy: the input is upsampled after being initialized by a bilinear transpose convolution (parameters can be optimized), and then compared with the input generated by the feature extraction branch. The residual magnetograms are added pixel by pixel to obtain the current high-resolution output and then passed to the next level.

[0046] After introducing a multi-scale feature extraction module, the receptive field of the feature map in the convolutional neural network changes significantly with the scale of the convolutional kernel: small-scale convolutional kernels have small receptive fields, which can effectively capture local details of the magnetogram, but may weaken the ability to perceive global features; large-scale convolutional kernels, although limited in their ability to represent geometric details, have larger receptive fields that can acquire global contextual information, thus improving the model's generalization performance. Based on this, such as Figure 3As shown, this invention incorporates a parallel multi-scale convolutional kernel structure into the multi-scale feature extraction module. By integrating feature information of different receptive field sizes, it achieves comprehensive extraction and fusion of multi-scale features of the magnetograph. It uses a parallel multi-branch structure to extract features of the input magnetograph at multiple scales. Preferably, it uses... , , The convolutional kernels extract feature maps at different scales. Let... , , Indicating the third of the three feature extraction branches One convolutional kernel, , , Indicates the first The magnetograph feature output of each convolutional module, If we represent the input, then its simplified mathematical model can be expressed as:

[0047] ,

[0048] ,

[0049] ,

[0050] In the formula, This represents the ReLU activation function. After each convolution operation, the input is fed into an activation function, and then the outputs of the three branches are concatenated to obtain the concatenated feature map. :

[0051] ,

[0052] In the formula, Concat represents the concatenation function, which is used after concatenation. Dimensionality reduction of convolutional kernels on spliced ​​feature maps:

[0053] ,

[0054] In the formula, express The convolution kernel.

[0055] But the introduction and The convolutional kernel inevitably increases the number of model parameters and computational complexity. To reduce the number of model parameters and improve the model's computation speed, this invention uses convolutional kernels with different dilation rates. A dilated convolution kernel replaces the original one. , Convolution kernel. Taking convolution kernels as an example, dilated convolution in Filling the convolution kernel with zeros achieves essentially the same function as standard convolution; it can be viewed as another form of regular convolution, such as... Figure 4 The diagram shows a conventional convolution and a dilated convolution. The left diagram shows the conventional convolution. Convolution, the right figure shows dilation rate 2. Hollow convolution.

[0056] Compared to traditional receptive field expansion methods, dilated convolutions specifically address their inherent limitations. Traditional methods for expanding the receptive field mainly employ three strategies: increasing the depth of convolutional layers, using large-size convolutional kernels, and implementing downsampling. The first two strategies significantly increase the number of model parameters and are prone to overfitting; while downsampling can reduce the dimensionality of feature maps, it results in the loss of texture details and requires additional upsampling modules. In contrast, dilated convolutions can effectively expand the receptive field while maintaining feature map resolution, without increasing the parameter burden, and also allows for efficient control of model complexity.

[0057] like Figure 5 As shown, if variables The dilation coefficient characterizes dilated convolution. The actual kernel size obtained after dilated convolution and the original kernel size exhibit the following mathematical relationship:

[0058] ,

[0059] in, The size of the original convolution kernel. The size of the expanded convolutional kernel. For the convolution dilation rate, the original , , The convolution kernels are all replaced with dilated convolutions with dilation rates of 1, 2, and 3. When using... When the convolution kernel has an inflation rate of 1, its receptive field is the same as that of a traditional convolution kernel. When the expansion rate increases to 2, the receptive field expands to When the expansion rate is 3, the receptive field increases to Despite the significant differences in receptive fields, the number of parameters remains constant in all three cases, which is 9.

[0060] Gradient explosion and vanishing gradients are common problems during deep neural network training. To address this issue, this invention employs a local residual learning mechanism. A direct information transmission channel is constructed from the previous layer to the next, and a local residual learning module is added at the end of each convolutional layer. The added local learning module is as follows: Figure 6 As shown, the features of each layer are input into the convolutional layer of that layer and used as input to the convolutional layer of the next layer.

[0061] The efficiency of feature transfer is enhanced by residual connections, while local features are extracted by stacked convolutional layers, making the scale feature extraction module more stable and reliable when learning features at different scales.

[0062] like Figure 7 As shown, the multi-scale feature fusion module employs bidirectional feature fusion and weighted fusion mechanisms. The main idea is to improve accuracy through multi-level feature pyramids and bidirectional information transmission. By constructing bidirectional information interaction channels from top to bottom and bottom to top, cross-level features are efficiently integrated. Compared to traditional unidirectional feature pyramid networks, this architecture improves multi-scale feature fusion efficiency while controlling computational complexity. Its core uses bidirectional paths as independent feature processing units, achieving multi-level feature interaction through modular stacking. This simplifies the network topology and enhances the extraction and fusion capabilities of features at different scales, effectively improving overall network performance.

[0063] This module uses multi-scale feature extraction followed by... Feature maps obtained by convolution dimensionality reduction For input, using Convolution operations and The combination of convolutional operations achieves feature optimization, and feature fusion and local residual learning are realized through alternating bidirectional cross-scale paths. Convolutional layers employ adaptive mechanisms to selectively extract features at different levels, performing adaptive channel filtering and dimensionality reduction on features at different levels. This effectively removes redundant information, reduces computational burden, and maintains consistent feature map spatial scale. The convolutional layers then perform deeper feature extraction on top of this foundation, thereby completing the global residual learning process. In the model, the multi-scale feature fusion module tightly integrates features from various scales using both top-down and bottom-up approaches. The mathematical models for these two fusion methods can be expressed as follows:

[0064] ,

[0065] ,

[0066] ,

[0067] Will The fusion of low-level and high-level detail features enhances the preservation of detail in high-level features. In the formula, , , This represents the features of the current layer after top-down fusion (when the index is n-1, it represents the features of the previous layer after bottom-up fusion; when the index is n, it represents the features of the current layer after bottom-up fusion). and express and convolution kernel, Convolutional kernels compress the number of channels to a suitable dimension. The convolution kernels are used to fuse and splice the features while preserving the number of channels.

[0068] ,

[0069] ,

[0070] ,

[0071] In the formula, , , This represents the features after bottom-up fusion. After convolution, feature maps of different scales are stitched together to form a richer feature representation, allowing the network to utilize information from different scales simultaneously, thereby enhancing its ability to perceive details and structure in magnetographs.

[0072] Furthermore, in step 3, the enhanced features obtained through bidirectional multi-scale feature fusion are used to learn the residual magnetogram layer by layer. and upsampled low-resolution magnetogram The pixels are added one by one to obtain the current high-resolution magnetograph output. LAPSRN proposes using the Charbonnier loss function as a new optimization objective to improve the accuracy of magnetograph super-resolution reconstruction. A low-resolution input magnetograph is given. and the set of network parameters to be optimized The model learns a mapping function. To generate near-realistic high-resolution magnetograms Prediction results Each level will use bi / triple interpolation to obtain the reference high-resolution magnetograph. Adjust to match the current level output For matching sizes, the mathematical model of the loss function is:

[0073] ,

[0074] ,

[0075] In the formula, superscript Indicates the first There are training samples, where N represents the total number of training samples, and the subscript is... Represents the first in the pyramid structure Level-wise branching, where S represents the total number of branches. The Charbonnier loss function is used. It is a very small positive number, usually taken as .

[0076] To address the overfitting problem during training, L2 regularization is introduced, adding the L2 norm of the weights as a penalty term to the objective function. This reduces model complexity and improves generalization performance. The formula is as follows:

[0077] ,

[0078] In the formula, For the model's data loss, It is a regularization parameter used to control the strength of regularization. During model optimization, the weight vector... By introducing an L2 regularization term, the loss function simultaneously considers both data fitting error and parameter constraints. This dual optimization mechanism not only effectively controls the magnitude of model parameters but also significantly improves the model's generalization performance, thereby reducing model complexity. The final loss function is:

[0079] .

[0080] On the other hand, the present invention provides a geomagnetic map construction device based on an improved LAPSRN algorithm, the various modules of which can implement the various steps of the aforementioned method, specifically including:

[0081] The preprocessing module is used to perform wavelet threshold denoising preprocessing on the raw geomagnetic data;

[0082] The reconstruction module is used to input the preprocessed data into the improved Laplacian pyramid super-resolution network LAPSRN. The improved LAPSRN network includes a multi-scale feature extraction module and a multi-scale feature fusion module cascaded in sequence. The multi-scale feature extraction module uses multi-scale dilated convolution kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers. The multi-scale feature fusion module fuses the multi-scale features of the geomagnetic map in an alternating bottom-up and top-down manner, and outputs the fused multi-scale enhanced features.

[0083] The output module is used to input the fused multi-scale enhanced features into the loss function of the improved LAPSRN network. The loss function includes an L2 regularization term. The parameters of the improved LAPSRN network are optimized through backpropagation, and finally a high-resolution geomagnetic map is output.

[0084] Thirdly, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for constructing a geomagnetic map based on an improved LAPSRN algorithm.

[0085] Fourthly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for constructing a geomagnetic map based on an improved LAPSRN algorithm.

[0086] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing geomagnetic maps based on an improved LAPSRN algorithm, characterized in that, The method includes: Step 1: Perform wavelet thresholding denoising preprocessing on the raw geomagnetic data; Step 2: Input the preprocessed data into the improved Laplacian Pyramid Super-Resolution Network (LAPSRN). The improved LAPSRN network includes a cascaded multi-scale feature extraction module and a multi-scale feature fusion module. The multi-scale feature extraction module uses multi-scale dilated convolutional kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers. The multi-scale feature fusion module uses an alternating bottom-up and top-down approach to fuse the multi-scale features of the geomagnetic map and outputs the fused multi-scale enhanced features. Step 3: Input the fused multi-scale enhanced features into the loss function of the improved LAPSRN network. The loss function includes an L2 regularization term. Optimize the parameters of the improved LAPSRN network through backpropagation, and finally output a high-resolution geomagnetic map.

2. The geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 1, characterized in that, Step 1 includes: selecting wavelet basis functions based on support length, symmetry, and vanishing moment; determining the optimal number of wavelet decomposition layers; performing wavelet decomposition on the original geomagnetic data using an improved custom threshold function; performing threshold processing on the high-frequency coefficients after decomposition; retaining the effective low-frequency signals; and finally reconstructing the denoised geomagnetic data through inverse wavelet transform.

3. The geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 1, characterized in that, In step 2, the multi-scale feature extraction module adopts a parallel multi-branch structure and uses 3×3 convolution kernels corresponding to dilated convolutions with dilation rates of 1, 2 and 3 to extract feature maps at different scales. After concatenating the outputs of the multi-branch structure, dimensionality reduction is performed using 1×1 convolution kernels. The local residual learning adds residual connections at the end of each convolutional layer to build a direct information transmission channel from the front layer to the back layer.

4. The geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 1, characterized in that, In step 2, the multi-scale feature fusion module adopts a bidirectional feature fusion and weighted fusion mechanism. By constructing a bidirectional information interaction channel from top to bottom and bottom to top, it uses a 1×1 convolutional layer to perform adaptive channel filtering and dimensionality reduction on features at different levels, and uses a 3×3 convolutional layer to perform deeper feature extraction to complete global residual learning.

5. A geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 4, characterized in that, In the multi-scale feature fusion module, the top-down path fuses low-level detail features with high-level detail features to increase the detail retention capability of high-level features, while the bottom-up path transmits high-level semantic features to the low level to enhance the semantic information of the low level.

6. The geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 1, characterized in that, In step 2, the improved Laplacian pyramid super-resolution network adopts a multi-level progressive upsampling architecture, constructing a network containing [a specific architecture] for a task with a magnification factor of s. The system consists of several progressive levels, each performing a 2x geomagnetic map magnification. Each level includes a feature extraction branch and a geomagnetic map reconstruction branch. The geomagnetic map reconstruction branch employs a residual learning strategy, upsampling the input through bilinear initialization transpose convolution and then adding it pixel-by-pixel to the residual geomagnetic map generated by the feature extraction branch to obtain the current high-resolution output.

7. The geomagnetic map construction method based on the improved LAPSRN algorithm according to claim 1, characterized in that, In step 3, the loss function is based on the Charill loss function, with an L2 regularization term added as a weight penalty term. Overfitting of the model is suppressed by controlling the regularization strength. The Charill loss function is used to calculate the difference between the predicted high-resolution geomagnetic map and the actual high-resolution geomagnetic map.

8. A geomagnetic map construction device based on an improved LAPSRN algorithm, characterized in that, include: The preprocessing module is used to perform wavelet threshold denoising preprocessing on the raw geomagnetic data; The reconstruction module is used to input the preprocessed data into the improved Laplacian pyramid super-resolution network LAPSRN. The improved LAPSRN network includes a multi-scale feature extraction module and a multi-scale feature fusion module cascaded in sequence. The multi-scale feature extraction module uses multi-scale dilated convolution kernels to extract multi-scale features of the geomagnetic map and introduces local residual learning between convolutional layers. The multi-scale feature fusion module fuses the multi-scale features of the geomagnetic map in an alternating bottom-up and top-down manner, and outputs the fused multi-scale enhanced features. The output module is used to input the fused multi-scale enhanced features into the loss function of the improved LAPSRN network. The loss function includes an L2 regularization term. The parameters of the improved LAPSRN network are optimized through backpropagation, and finally a high-resolution geomagnetic map is output.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When one or more programs are executed by the one or more processors, the one or more processors implement the geomagnetic map construction method based on the improved LAPSRN algorithm as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, enable the processor to implement the geomagnetic map construction method based on the improved LAPSRN algorithm as described in any one of claims 1-7.

Images