Magnetic gradient tensor three-dimensional inversion method and system based on Unet3 +

Abstract

The invention belongs to the technical field of magnetic anomaly three-dimensional inversion, and particularly relates to a Unet3 +-based magnetic gradient tensor three-dimensional inversion method, which comprises the following steps of: establishing a dual-task Unet3 + three-dimensional inversion framework; and designing a loss function. According to the invention, the magnetic gradient tensor data contains abnormal information generated by geologic bodies with different depths and different scales. The Unet3 + full-scale connection mechanism can fully fuse the multi-scale features, so that the magnetic susceptibility model reflecting the complex underground structure can be constructed more accurately. According to the method, the boundary of a small-scale anomalous body can be accurately drawn by utilizing shallow layer features, and the overall form of a large-scale structure can be grasped by utilizing deep layer features.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of three-dimensional magnetic anomaly inversion technology, and particularly relates to a three-dimensional magnetic gradient tensor inversion method and system based on Unet3+. Background Technology

[0002] Three-dimensional magnetic anomaly inversion infers the spatial distribution and magnetic susceptibility structure of subsurface magnetic bodies by analyzing surface magnetic anomaly data. This technique has wide applications in mineral exploration, geological structure research, and underground facility detection. Particularly in complex subsurface environments, three-dimensional magnetic anomaly inversion can provide high-resolution subsurface structure models, aiding in the identification of the geometry and location of subsurface magnetic bodies. Magnetic gradient tensor data, compared to magnetic anomaly scalar or vector data, offers higher sensitivity and spatial resolution. This data can measure the gradient in three directions of the magnetic field, which is extremely helpful in capturing shallow and deep magnetic anomalies, while providing richer geological information for more accurate descriptions of complex geological structures.

[0003] To address the inherent challenges of 3D magnetic inversion, traditional methods primarily revolve around regularization theory and efficient optimization algorithms, aiming to find stable and geologically plausible solutions from ill-posed problems. Regularization methods mitigate ill-conditioning and multiple solutions by introducing a stable functional (penalty term) into the inversion objective function to constrain the solution space. These methods include smoothing constraints, focusing constraints, Gramian constraints, and hybrid regularization methods. Choosing an appropriate optimization algorithm is crucial for solving large-scale inversion problems. Commonly used algorithms include the conjugate gradient method and the LBFGS method, which can be used to solve large linear inversion equation systems. Additionally, methods such as the alternating direction multiplier method (ADMM) and the reweighted regularized conjugate gradient method (RRCG) can be used to solve complex inversion models with sparse constraints. Although traditional methods have made significant progress both theoretically and practically, they generally rely on iterative optimization, resulting in high computational costs. Furthermore, the quality of the inversion results is heavily dependent on the selection of regularization parameters and prior models, limiting their practical value.

[0004] Deep learning, as a data-driven approach, can automatically learn the complex nonlinear mapping relationship from observational data to subsurface models by constructing deep neural networks, providing a novel approach to overcoming the bottlenecks of traditional inversion methods. In recent years, deep learning technology has been widely applied in 3D magnetic anomaly inversion. Compared with traditional methods, deep learning significantly reduces computation time and improves the accuracy of inversion models. Among deep learning models, UNet and its variants are exemplary models of the great success of deep learning in image segmentation. Their "encoder-decoder" structure and "skip connection" mechanism can effectively fuse multi-scale features, accurately locating and segmenting targets. This characteristic also demonstrates great potential in geophysical inversion. Jiao et al. (2024) applied the UNet++ network to 3D magnetic data inversion and achieved good results. Compared with other variants, UNet3+'s full-scale connection mechanism can most fully fuse multi-scale features, and the connection method is more efficient. These advantages make it an ideal tool for processing magnetic gradient tensor data, especially in fast, large-scale 3D inversion tasks.

[0005] 3D inversion of magnetic gradient tensor is an important technique in geophysical exploration for characterizing the spatial distribution and physical properties of subsurface magnetic bodies, and it has been widely applied in mineral exploration, structural identification, and anomaly location. Existing research has attempted to introduce convolutional neural networks to perform 3D inversion of magnetic gradient tensor data, predicting magnetization distribution by learning the mapping relationship between magnetic gradient response and subsurface physical parameters. However, these methods typically employ a single-task regression framework, focusing on overall numerical fitting and lacking explicit constraints on the spatial location of anomalies, making them susceptible to background interference under complex geological conditions. Furthermore, existing network structures have limited ability to fuse multi-scale features, making it difficult to simultaneously consider shallow, small-scale anomalies and deep, large-scale structural features, resulting in blurred boundaries and insufficient stability in the inversion results, failing to meet the engineering requirements of high-precision 3D inversion. Summary of the Invention

[0006] To address the problems of existing technologies, current magnetic gradient tensor 3D inversion methods struggle to simultaneously achieve both spatial positioning accuracy of anomalies and numerical inversion accuracy of magnetization under complex underground structural conditions. Furthermore, they suffer from insufficient utilization of multi-scale features and are easily affected by background interference, resulting in blurred boundaries and poor stability in the inversion results, failing to meet the requirements for refined 3D inversion. This invention provides a magnetic gradient tensor 3D inversion method based on Unet3+.

[0007] This invention is implemented as follows: a three-dimensional inversion method based on the magnetic gradient tensor of Unet3+, the method comprising:

[0008] S1: Establish a dual-task Unet3+ 3D inversion framework;

[0009] S2: Loss function design.

[0010] Furthermore, the dual-task Unet3+ three-dimensional inversion framework in S1 specifically includes:

[0011] This framework uses an improved UNet3+ as the backbone network, constructs an encoder-decoder structure to achieve multi-scale feature extraction and fusion, and innovatively designs a dual-output head to predict the spatial position and magnetization intensity value of the anomaly body respectively. The segmentation task outputs a binary mask to identify the area where the magnetic body exists. The regression task predicts the magnetization intensity value based on the segmentation task, and finally outputs the three-dimensional distribution result of the magnetization intensity. The input of this network is five independent components of the magnetic gradient tensor, each component is 64×64 in size, and the output three-dimensional distribution size is 16×64×64, corresponding to the depth and horizontal direction lengths respectively. The inversion framework, its principle can be expressed as:

[0012] (1)

[0013] In the formula, C represents the convolutional fusion layer, represents the l-th feature fusion layer in the decoder, represents the k-th layer feature in the encoder, represents the feature transformation operation. When k < l, stride convolution is used for downsampling; when k = l, direct connection is used; when k > l, bilinear upsampling is used.

[0014] Furthermore, S2 specifically includes:

[0015] Flatten the predicted value of the segmentation task into a one-dimensional vector, which can be expressed as P={p i}, and the value of the true label can be expressed as G={g i}. Since the background voxels are much more than the anomaly voxels, the Dice coefficient is used to calculate the segmentation loss, and its expression is:

[0016] (2)

[0017] In the formula is the smoothing term, taking 1.0. To enhance the robustness to outliers, the regression task uses the SmoothL1 loss. Use Y={y i} to represent the one-dimensional vector after flattening the predicted value of the regression task, and T={t i} to represent the true label, then the calculation formula can be expressed as:

[0018] (3)

[0019] The total loss function of the model consists of two parts: segmentation loss and regression loss, and its expression is:

[0020] (4)

[0021] In the formula and represents the weights of the segmentation loss term and the regression loss term, respectively, with values ​​of 1.0 and 0.1; the Adam optimizer is used during training, with an initial learning rate of . The weight decays to The learning rate decays in increments of 0.5 every 10 rounds, with a total of 100 training rounds.

[0022] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, the computer program being executed by the processor causing the processor to perform the steps of the three-dimensional inversion method of the magnetic gradient tensor based on Unet3+.

[0023] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the three-dimensional inversion method based on the Unet3+ magnetic gradient tensor.

[0024] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0025] First, this invention fully utilizes the multi-scale anomaly information of the magnetic gradient tensor by introducing the full-scale feature fusion mechanism of Unet3+. Magnetic gradient tensor data simultaneously contains anomalous responses generated by shallow, small-scale magnetic bodies and deep, large-scale geological structures. Traditional single-scale or limited-scale fusion methods struggle to simultaneously capture anomaly features at different scales. This invention unifies and fuses features from various scales of the encoder during the decoding stage, enabling the network to simultaneously perceive shallow and deep information within the same feature space. This allows for accurate characterization of the spatial boundaries of small-scale anomalies and precise grasp of the overall morphology of large magnetic structures during 3D inversion, thus improving the comprehensive representation capability of complex underground structures.

[0026] Compared to networks employing densely connected structures, the full-scale connection approach used in this invention effectively reduces redundant feature propagation while maintaining feature representation capabilities. By performing targeted scale alignment and fusion of features at different scales, a large number of repetitive convolution operations between ordinary-scale features and upsampled features are avoided, thereby reducing the network parameter size and computational complexity. In applications such as magnetic gradient tensor 3D inversion, which demands high computational efficiency, this structural design helps improve overall computational efficiency while ensuring inversion accuracy, enhancing the feasibility of the method in engineering applications.

[0027] Due to the full fusion of features across all scales, this invention offers significant advantages in the accuracy of identifying spatial boundaries of anomalies. By simultaneously utilizing high-resolution features and deep semantic features, the network can generate anomaly segmentation results with clearer boundaries and better continuity. In 3D inversion applications, this effect is directly reflected in the detailed characterization of magnetic body morphology, helping to improve the accuracy of ore body delineation, structural boundary identification, and anomaly spatial positioning, while reducing inversion uncertainties caused by boundary ambiguity.

[0028] This invention decouples the 3D inversion problem into two sub-tasks: spatial localization and numerical inversion, by setting up a dual-branch structure where segmentation and regression tasks work together. The segmentation task first determines the region where magnetic anomalies exist, providing clear spatial constraints for the regression task and thus suppressing the interference of background regions on magnetization inversion. This design not only improves the stability of magnetization prediction results but also enhances the model's robustness to noise and outliers, making the 3D inversion results more reliable.

[0029] This invention has achieved significant technical results in terms of multi-scale feature utilization, computational efficiency, spatial boundary characterization, and inversion stability, and is applicable to high-precision three-dimensional inversion scenarios of complex underground magnetic structures.

[0030] Secondly, as supporting evidence of the inventiveness of this invention, it is also reflected in the following important aspects:

[0031] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0032] The technical solution of this invention represents a major breakthrough in the field of geophysical intelligent interpretation. Its commercial application will reshape the working paradigm of magnetic exploration and create multi-dimensional commercial value.

[0033] ① Fostering a new generation of exploration software and service platforms: This method reduces the traditional 3D inversion interpretation time to the second level, and can directly generate high-resolution, high-fidelity subsurface 3D physical property models. This core algorithm can be integrated to form a new generation of "intelligent 3D magnetic interpretation software" or embedded as a core engine into existing exploration software and cloud platforms, creating sustainable high-value-added revenue through software licensing or module sales.

[0034] ② Significantly reduces exploration decision-making costs and risks: Rapid and intuitive 3D inversion imaging results can provide crucial information for drilling target location and resource estimation, effectively reducing ineffective drilling and improving exploration success rates. In the mineral and oil and gas exploration fields, a single project can be expected to save substantial engineering costs and improve resource discovery efficiency, bringing direct and significant economic benefits.

[0035] ③ Empowering the upgrading of the data service industry chain: Combining the increasingly popular high-precision magnetic gradient tensor measurement (such as UAV aeromagnetic gradient measurement), a one-stop solution of "data acquisition - intelligent 3D imaging - geological interpretation" can be created, which will greatly enhance the delivery capabilities and product value of data service providers and open up the high-end geological information technology service market.

[0036] ④ Expanding the ability to detect deep and complex environments: This method has excellent resolution capabilities for weak anomalies and overlapping anomalies, giving it unique technical advantages and broad application prospects in fields with urgent needs for fine imaging of concealed targets, such as deep mineral resource exploration, urban underground space exploration, archaeology, and national defense security.

[0037] In summary, this technology not only improves tool efficiency but also represents a qualitative leap in explanatory power. It is expected to lead to technological upgrades in related industries and create a market with an annual output value of tens of millions.

[0038] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0039] This invention overcomes a core bottleneck in the intelligent transformation of magnetic exploration, achieving a revolutionary advancement in the efficient and accurate mapping from data to three-dimensional geological models. Its innovation is mainly reflected in filling the following two technological gaps:

[0040] ① A complete end-to-end 3D inversion framework based on Unet3+ deep neural network magnetic gradient tensor was constructed. This completely changes the traditional inversion mode that relies on iterative optimization, has high computational cost, and is heavily dependent on the initial model and regularization selection. It enables the direct and rapid output of 3D physical properties from the input observation tensor data, realizing a paradigm shift from "iterative solution" to "intelligent mapping" in inversion methodology.

[0041] ② This invention resolves the core contradiction in 3D inversion where resolution and efficiency are mutually exclusive. Traditional linear or iterative inversion methods often sacrifice resolution for stable solutions. This invention utilizes Unet3+'s unique full-scale jump connections and deep supervision mechanism to simultaneously capture anomalous macroscopic morphology and microscopic details, achieving high-resolution, high-fidelity 3D inversion results in an extremely short time—something previously impossible. Attached Figure Description

[0042] Figure 1 This is a flowchart of the three-dimensional inversion method of magnetic gradient tensor based on Unet3+ provided in the embodiments of the present invention;

[0043] Figure 2 This is the dual-task Unet3+ three-dimensional inversion framework provided in the embodiments of the present invention;

[0044] Figure 3 This is a training loss curve provided in an embodiment of the present invention;

[0045] Figure 4 The test set three-dimensional inversion results provided in this embodiment of the invention (the left figure is the network prediction result, and the right figure is the actual three-dimensional distribution).

[0046] Figure 5 Validation set 3D inversion results (left figure shows network prediction results, right figure shows the actual 3D distribution);

[0047] Figure 6 Forward modeling results for the test set (the top image shows the actual values, the bottom image shows the network predictions, and the black borders in the images represent the horizontal outlines of underground objects).

[0048] Figure 7 Validation set forward modeling results (the black border in the figure represents the horizontal outline of the underground object). Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0050] To ensure that the Unet3+-based three-dimensional inversion method for magnetic gradient tensors described in this invention can be fully understood and implemented by those skilled in the art, the technical solution of this invention is further described below in conjunction with specific embodiments. It should be understood that the following embodiments are merely illustrative and are not intended to limit the scope of protection of this invention.

[0051] In one exemplary embodiment, the magnetic gradient tensor data, after being acquired by a magnetic measurement device, forms five independent gradient components. Each component is represented in the form of a two-dimensional matrix, and its spatial resolution can be set according to the detection requirements. For example, in shallow detection scenarios, a higher-resolution two-dimensional grid can be used; in deep detection scenarios, the resolution can be appropriately reduced to balance computational complexity. The above settings are used to illustrate the composition of the input data, and the present invention does not limit the specific resolution format.

[0052] In terms of network structure, the 3D inversion network adopts a structure combining an encoder and a decoder. The encoder part can be composed of multiple convolutional units, used to extract feature information of the magnetic gradient tensor at different scales step by step; the decoder part achieves cross-scale feature fusion by introducing features from different levels of the encoder. For example, the feature fusion can be accomplished by convolution operations, or it can be combined with normalization operations or activation functions to further enhance the feature representation capability. All of the above methods can be used as optional implementation schemes of the present invention.

[0053] In a preferred embodiment, the network is configured with a segmentation output branch and a regression output branch. The segmentation branch is used to generate mask information for the spatial distribution of anomalies, and the regression branch is used to predict magnetization values ​​within the spatial range defined by the mask. This configuration allows spatial positioning information and numerical inversion information to be collaboratively optimized within the same framework, effectively reducing the interference of background regions on the magnetization inversion results. It should be understood that the segmentation and regression branches may share some feature layers in specific implementations, or they may be separated at higher feature stages; this invention does not limit this.

[0054] In constructing the loss function, the segmentation task preferably adopts a loss calculation method based on the degree of overlap between the predicted and ground truth regions to alleviate the class imbalance problem caused by the number of outlier voxels being much smaller than that of background voxels. The regression task preferably adopts a loss function form that is robust to outlier errors, employing different error calculation strategies when the prediction error falls within different intervals, thereby improving model stability while ensuring inversion accuracy. The above loss functions can be combined according to preset weights, and the weight ratios can be adjusted according to specific application scenarios.

[0055] During model training, an adaptive gradient optimization algorithm can be used to update network parameters, combined with a weight decay and learning rate gradual decay strategy, to improve the stability and generalization ability of model convergence. The number of training epochs, the initial learning rate, and the decay step size can all be flexibly set according to the data scale and inversion accuracy requirements.

[0056] This invention provides a three-dimensional inversion method based on the magnetic gradient tensor of Unet3+, the method comprising:

[0057] S1: Establish a dual-task Unet3+ 3D inversion framework;

[0058] S2: Loss function design.

[0059] The dual-task Unet3+ 3D inversion framework in S1 specifically includes:

[0060] like Figure 2 As shown, this framework uses an improved UNet3+ as the backbone network, constructing an encoder-decoder structure to achieve multi-scale feature extraction and fusion. It innovatively designs dual output heads to predict the spatial location and magnetization value of the anomalous body, respectively. The segmentation task outputs a binary mask to identify the region where the magnetic body exists. The regression task predicts the magnetization value based on the segmentation task, and finally outputs the three-dimensional distribution result of the magnetization intensity. The network input consists of five independent components of the magnetic gradient tensor, each component being 64×64 in size. The output three-dimensional distribution is 16×64×64 in size, corresponding to the depth and horizontal length, respectively. The principle of the inversion framework can be expressed as follows:

[0061] (1)

[0062] In the formula, C represents the convolutional fusion layer, represents the l-th feature fusion layer in the decoder, represents the k-th layer feature in the encoder, represents the feature transformation operation. When k < l, stride convolution is used for downsampling; when k = l, direct connection is used; when k > l, bilinear upsampling is used.

[0063] The specific content of S2 includes:

[0064] Flattening the predicted value of the segmentation task into a one-dimensional vector can be expressed as P = {p i}, and the value of the true label can be expressed as G = {g i}. Since the background voxels are much more than the abnormal voxels, the Dice coefficient is used to calculate the segmentation loss, and its expression is:

[0065] (2)

[0066] In the formula is the smoothing term, taking 1.0. To enhance the robustness to outliers, the SmoothL1 loss is used for the regression task. Using Y = {y i} to represent the one-dimensional vector after flattening the predicted value of the regression task, and T = {t i} to represent the true label, then the calculation formula can be expressed as:

[0067] (3)

[0068] The total loss function of the model consists of two parts: segmentation loss and regression loss, and its expression is:

[0069] (4)

[0070] In the formula and represent the weights of the segmentation loss term and the regression loss term respectively, taking 1.0 and 0.1 respectively; the Adam optimizer is used during training, the initial learning rate is , the weight decay is , the learning rate decays step by step by multiplying 0.5 every 10 epochs, and the total number of training epochs is 100.

[0071] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the three-dimensional inversion method of magnetic gradient tensor based on Unet3+.

[0072] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the three-dimensional inversion method based on the Unet3+ magnetic gradient tensor.

[0073] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0074] A Cartesian coordinate system was established, with x pointing north, y pointing east, and z pointing vertically downward. The 64m×256m×256m underground space was divided into a 16×64×64 three-dimensional grid. A random walk model was used to generate 2000 sets of underground three-dimensional magnetization intensity distribution data. The forward modeling formula for a vertical cuboid was used to calculate the magnetic gradient tensor value corresponding to each cuboid element in the grid, as shown in equations (5) to (9), where M represents the magnetization intensity. The direction cosine of magnetization, Let (x, y, z) represent the coordinates of the center of the cuboid, (x, y, z) represent the ground observation point, and a, b, c represent the lengths of the cuboid along the x, y, and z directions, respectively. Representing a point in underground space, the magnetic gradient tensor value of each point on the observation surface (height 11m) is calculated using the superposition theorem, and this value is used as the label data to calculate the data fitting loss. After dividing the dataset into training and validation sets in a 9:1 ratio, 100 sets of regular spheres are generated as the validation set. The loss curves for the training, testing, and validation sets during training are shown below. Figure 3 As shown.

[0075] (5)

[0076] (6)

[0077] (7)

[0078] (8)

[0079] (9)

[0080] A portion of the test set and validation set samples were extracted and visualized. The 3D distribution plot of the test set inversion results is shown below. Figure 3 As shown, the three-dimensional distribution of the validation set inversion results is as follows: Figure 5 As shown.

[0081] By performing forward modeling using formulas (5) to (9) on the above inversion results, contour maps of the magnetic gradient tensor data of the observation surface can be obtained. Figure 6 and Figure 7 These are the forward modeling results for the test set and the validation set, respectively.

[0082] As shown in the above inversion results, the Unet3+ 3D inversion network designed in this invention can effectively invert the 3D distribution and horizontal contours of underground objects. The magnetic gradient tensor data calculated from the inversion results has little difference from the actual data and can distinguish between different magnetic objects, thus effectively improving the accuracy of 3D inversion. Using the model trained by this network, efficient and accurate 3D inversion can be achieved, providing strong support for engineering practices in fields such as mineral exploration, geological structure research, and underground facility detection.

[0083] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A three-dimensional inversion method based on the magnetic gradient tensor of Unet3+, characterized in that, The method includes the following steps: A three-dimensional inversion network structure with an encoder and a decoder as the backbone is constructed. The network uses five independent components of the magnetic gradient tensor as input features and forms a unified feature expression through multi-scale feature extraction and cross-scale fusion. Within the same network framework, a segmentation task and a regression task are set up to cooperate with each other. The segmentation task is used to determine the existence region of the magnetic anomaly in three-dimensional space, and the regression task predicts the magnetization intensity value within the existence region. By constraining the spatial location through the segmentation task, the interference of non-abnormal regions on the magnetization inversion results is suppressed, thereby obtaining the three-dimensional spatial distribution results of magnetization.

2. The method according to claim 1, characterized in that, The input magnetic gradient tensor consists of five independent components, each of which is input into the network in the form of a two-dimensional grid to characterize the changes in magnetic field gradient in different directions.

3. The method according to claim 1, characterized in that, The output of the network is the voxelized distribution of magnetization in three-dimensional space, with one dimension corresponding to the underground depth direction and the other two dimensions corresponding to the horizontal direction.

4. A feature fusion method for three-dimensional inversion of magnetic gradient tensor, characterized in that: In each feature fusion layer of the decoding stage, feature information from multiple scales from the encoding stage is introduced; For coding features at different scales, based on their spatial resolution and the scale relationship of the current decoding layer, downsampling transformation, direct connection at the same scale, or upsampling transformation are performed respectively. Multi-source features that have been scale-aligned are convolved and fused along the channel dimension to form a unified feature representation containing multi-scale spatial information, which is used for spatial reconstruction of magnetic anomalies.

5. The method according to claim 4, characterized in that, When the scale of the encoded features is lower than the scale of the current decoding layer, downsampling is performed using convolution with stride to achieve scale matching.

6. The method according to claim 4, characterized in that, When the scale of the encoded features is higher than the scale of the current decoding layer, upsampling is performed using bilinear interpolation to achieve scale matching.

7. A method for constructing a loss function for training a three-dimensional inversion model using magnetic gradient tensor, characterized in that: To address the severe imbalance between the number of magnetic anomalies and background voxels, a segmentation loss function based on the overlap between the prediction results and the actual labels is used to impose key constraints on the anomaly region. For the numerical inversion task of magnetization, a regression loss function that is robust to outliers is adopted to reduce the impact of extreme errors on the model training process; By weighting and combining the segmentation loss and regression loss, a unified optimization objective function is formed, enabling the spatial positioning accuracy and numerical inversion accuracy to be optimized synergistically during the same training process.

8. The method according to claim 7, characterized in that, The segmentation loss is obtained by calculating the overlap ratio between the predicted abnormal region and the actual abnormal region, and a smoothing term is introduced in the calculation process to avoid numerical instability.

9. The method according to claim 7, characterized in that, The regression loss adopts the squared error form when the prediction error is less than the preset threshold, and adopts the linear error form when the prediction error is greater than the threshold, thus balancing accuracy and robustness.

10. A training method for a three-dimensional inversion network using magnetic gradient tensor, characterized in that: An adaptive optimization algorithm based on first-order moment estimation and second-order moment estimation is used to update the network parameters; A weight decay mechanism is introduced during training to suppress model overfitting; By gradually reducing the parameter update amplitude through a phased decay of the learning rate, the network can achieve stable convergence in the later stages of training, thereby improving the reliability of the three-dimensional inversion results of magnetization intensity.

Images