A method for characterizing complex geochemical spatial patterns based on deformable convolution

Abstract

The application provides a complex geochemical spatial pattern characterization method based on deformable convolution, relates to the field of geological mineral exploration and data processing, and comprises the following steps: collecting geochemical sampling point data to construct a data set; inputting the data set into a deformable convolution layer to obtain a sampling position of a convolution kernel with offset and a characteristic with offset; using small-step deformable convolution to extract small-scale characteristics and using large-step deformable convolution to extract large-scale characteristics and performing feature fusion on the characteristic with offset to obtain multi-scale spatial characteristics; and inputting the multi-scale spatial characteristics into an offset optimization circulation module and a full connection layer to output a complex geochemical spatial pattern that is accurately recognized. The application can adaptively adjust the sampling position of the convolution kernel, better capture complex spatial characteristics and patterns in geochemical data, and thus more accurately and effectively characterize the complex geochemical spatial pattern.

Description

Technical Field

[0001] This application relates to the fields of geological and mineral exploration and data processing, and in particular to a method for characterizing complex geochemical spatial models based on deformable convolution. Background Technology

[0002] Because mineralization processes originate from the interaction of multiple geological factors at different temporal and spatial scales, related geomorphic anomalies exhibit high complexity. Geochemical anomalies associated with mineralization are among the most important geological anomalies in mineral exploration, containing rich geological information. Accurate characterization of these anomalies is crucial for geological exploration, resource assessment, and other fields. These anomalies typically exhibit anisotropic spatial distribution characteristics, and their formation is controlled by ore-forming geological structures such as strata, faults, and magmatic intrusions, providing the necessary spatial, thermal, fluid, and material conditions for mineralization. Therefore, identifying complex spatial patterns in geochemistry is essential for accurately identifying important anomalies, thereby significantly improving the success rate of mineral exploration.

[0003] In the field of characterization methods for complex geochemical spatial models, common related technical solutions and their shortcomings are as follows: Traditional statistical methods, such as Kriging interpolation, rely on the assumption that the data conforms to a specific spatial distribution model. For complex geochemical data, the actual distribution often fails to meet this assumption, leading to significant deviations in the interpolation results and an inability to accurately reflect the true spatial model. Simple convolutional neural network methods use fixed convolution kernel sampling positions and lack adaptability to complex spatial models, making it difficult to effectively capture local variations and regional anomalies in the distribution of geochemical elements. Principal component analysis methods mainly focus on data dimensionality reduction and extraction of principal components, but are insufficient in exploring complex nonlinear relationships and spatial correlations between elements, and cannot adequately characterize complex geochemical spatial models. Currently, deep learning is a hot and cutting-edge method for mining and identifying geochemical spatial models, with convolutional neural networks and graph neural networks being the most popular. However, traditional convolutional methods have certain limitations when processing this type of data and are difficult to effectively capture complex spatial models. In graph neural networks, the construction of the topology graph used for model training is greatly influenced by subjective factors, significantly impacting the final results.

[0004] Deformable convolution is an emerging deep learning technique that introduces learnable offsets into traditional convolution, allowing the convolution kernel to adaptively adjust its sampling position to better adapt to the geometric deformation of the target. Therefore, applying deformable convolution to the characterization of complex geochemical spatial models, thereby fully mining the potential information in geochemical data and improving the ability to identify and analyze complex spatial models, is a pressing issue that needs to be addressed. Summary of the Invention

[0005] The purpose of this invention is to address the problem that existing methods have poor modeling capabilities for anisotropic and irregularly distributed spatial features in geochemical data caused by geological structures, resulting in low accuracy in anomaly pattern recognition. This invention provides a method for characterizing complex geochemical spatial patterns based on deformable convolution.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: S1. Collect geochemical sampling point data of the target area and preprocess it to obtain a dataset; S2. Construct a deformable convolutional layer by predicting the offset branch; input the dataset into the deformable convolutional layer to predict the spatial offset of each convolutional kernel position, and obtain the offset convolutional kernel sampling position and its offset features. S3. For the offset features output by the deformable convolutional layer, small-scale features for characterizing local element associations are extracted by small-stride deformable convolution, and large-scale features for characterizing regional anomalies are extracted by large-stride deformable convolution. The small-scale features and large-scale features are then fused to obtain multi-scale spatial features. S4. Input the multi-scale spatial features into the offset optimization loop module to obtain the first fused feature; S5. Input the first fusion feature into the fully connected layer to output a precisely identified complex geochemical spatial model.

[0007] Optionally, step S1 includes: Geochemical sampling point data of the target area are collected, the element content of each sampling point is preprocessed, and a fixed-size neighborhood window is constructed for each sampling point to form a training dataset that reflects the geochemical spatial model. The elemental content preprocessing includes: using the min-max normalization method to linearly map the original elemental content values ​​to... Interval.

[0008] Optionally, step S2 includes: An offset prediction branch is added to the conventional convolutional kernel structure of a deformable convolutional layer; the offset prediction branch is an additional convolutional layer, namely the offset prediction layer. The offset prediction layer outputs the offset field through convolution operations. To obtain the spatial offset, where Given the spatial feature matrix as input, For trainable parameters, output The dimension is , and The height and width of the feature map, The number of sampling points for the convolution kernel. This represents the two-dimensional offset corresponding to each sampling point.

[0009] Optionally, step S2 may further include: After obtaining the offset convolution kernel sampling positions, for sampling positions with non-integer coordinates, the offset feature at that position is calculated using bilinear interpolation.

[0010] Optionally, step S3 includes: A deformable convolutional layer with a stride of 1 is used to extract the small-scale features, and a deformable convolutional layer with a stride of 2 is used to extract the large-scale features. The large-scale feature map is upsampled and then concatenated with the small-scale feature map along the channel dimension, and then processed through a... The convolutional layer integrates features to obtain the multi-scale spatial features.

[0011] Optionally, step S4 includes: The offset optimization loop module uses the fused multi-scale spatial features to re-predict the convolution kernel offset and update the sampling position, repeating the process iteratively until the change in offset meets the preset stability condition. The offset optimization loop module is designed based on a gated loop unit; the preset stability condition is to calculate the norm of the offset change, and the iteration is terminated when it is less than a preset threshold.

[0012] Optionally, step S5 includes: The fully connected layer outputs accurately identified complex geochemical space models, including: A linear transformation is performed on the optimized first fusion feature to obtain the fractional vector. The softmax activation function is applied to convert the scores into a probability distribution. ,in The total number of categories of geochemical models. For the sampling point to belong to the first The probability of each pattern is used to determine the final identification result, i.e., the complex geochemical spatial model, based on the highest probability. The identified geochemical model labels are associated with the spatial coordinates of the sampling points, and geographic information system tools are used to generate a visualization map showing the spatial distribution patterns of the geochemical models.

[0013] An electronic device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform a complex geochemical spatial model characterization method based on deformable convolution.

[0014] A computer-readable storage medium storing instructions that, when executed, perform a method for characterizing complex geochemical spatial models based on deformable convolution.

[0015] The beneficial effects of the technical solution provided in this application are: This invention introduces learnable spatial offsets, enabling the convolution kernel to dynamically adjust its sampling position and adaptively focus on mineralization-related irregular anomaly regions, significantly improving the ability to characterize complex geological boundaries. Addressing the multi-scale characteristics of geochemical anomalies, a small-stride deformable convolution is used to extract local elemental co-occurrence relationships, preserving detailed information; a large-stride deformable convolution is used to capture regional anomaly backgrounds, expanding the receptive field. The organic fusion of multi-scale features achieves a comprehensive characterization from microscopic elemental combinations to macroscopic anomaly halos, avoiding information loss in single-scale analysis. A GRU-based offset iterative optimization module is introduced, iteratively correcting the convolution kernel sampling position until stable convergence. This mechanism overcomes the random errors of single-offset predictions, enabling the model to adaptively focus on the true core anomaly region, significantly improving the identification accuracy of weak and superimposed anomalies, and providing more reliable technical support for deep concealed mineral exploration prediction. Attached Figure Description

[0016] The present application will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a step diagram of an embodiment of this application; Figure 2 This is a visualization of the input data and offset of the window at the mining site in this embodiment of the application; Figure 3 This is a visualization of the input data, ordinary convolution, and variable convolution activation maps at the window in the embodiment of this application; Figure 4 This is a schematic diagram of the electronic device structure in the embodiments of this application. Detailed Implementation

[0017] To provide a clearer understanding of the technical features, objectives, and effects of this application, the specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0018] The embodiments of this application provide a method for characterizing complex geochemical spatial models based on deformable convolution.

[0019] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating the steps of a complex geochemical spatial model characterization method based on deformable convolution, as described in an embodiment of this application, including: S1. Collect geochemical sampling point data of the target area and preprocess it to obtain a dataset; S2. Construct a deformable convolutional layer by predicting the offset branch; input the dataset into the deformable convolutional layer to predict the spatial offset of each convolutional kernel position, and obtain the offset convolutional kernel sampling position and its offset features. S3. For the offset features output by the deformable convolutional layer, small-scale features for characterizing local element associations are extracted by small-stride deformable convolution, and large-scale features for characterizing regional anomalies are extracted by large-stride deformable convolution. The small-scale features and large-scale features are then fused to obtain multi-scale spatial features. S4. Input the multi-scale spatial features into the offset optimization loop module to obtain the first fused feature; S5. Input the first fusion feature into the fully connected layer to output a precisely identified complex geochemical spatial model.

[0020] As an example, deformable convolution can dynamically adjust the sampling position, deforming the receptive field to match irregular and complex structures. This flexibility allows the convolutional kernel to adapt to irregular structures, extracting features from key regions of regularity more accurately. Therefore, deformable convolution is more robust to geometric changes such as orientation, scale, or deformation, and can maintain the consistency of feature representation in different modal states. This is particularly evident when dealing with tasks involving irregular geochemical spatial structures.

[0021] Step S1 includes: Geochemical sampling point data of the target area are collected, the element content of each sampling point is preprocessed, and a fixed-size neighborhood window is constructed for each sampling point to form a training dataset that reflects the geochemical spatial model. The elemental content preprocessing includes: using the min-max normalization method to linearly map the original elemental content values ​​to... Interval.

[0022] As one example, the data collection involves obtaining geochemical datasets of the target area from geological databases or field sampling activities. The normalization process uses a min-max normalization method to linearly map the elemental abundance values ​​to... The interval and the neighborhood window are circular regions centered on the target point. The initial spatial feature matrix integrates the normalized content and neighborhood statistical features.

[0023] As one embodiment, constructing a fixed-size neighborhood window specifically involves using the target sampling point as the center and a preset radius... A circular region; by calculating the Euclidean distance between two points. To determine whether other sampling points are within the neighborhood window, where and These are the spatial coordinates of the target point and its neighboring points, respectively.

[0024] Step S2 includes: An offset prediction branch is added to the conventional convolutional kernel structure of a deformable convolutional layer; the offset prediction branch is an additional convolutional layer, namely the offset prediction layer. The offset prediction layer outputs the offset field through convolution operations. To obtain the spatial offset, where Given the spatial feature matrix as input, For trainable parameters, output The dimension is , and The height and width of the feature map, The number of sampling points for the convolution kernel. This represents the two-dimensional offset corresponding to each sampling point.

[0025] As an example, when implementing deformable convolutional layers, we extend the conventional convolutional kernel structure. The conventional convolution operation is defined as: for any position on the output feature map... Its value The formula is calculated by applying the convolution kernel weights to local regions of the input feature map, and is expressed as follows: .here, It is the size of the convolution kernel (e.g.) hour ), The weight parameters of the convolution kernel (which need to be learned through training). It is a predefined fixed sampling position offset (e.g., for...) nuclear, cover arrive (grid) It is the feature matrix of the input space. ( For height, For width, (This refers to the number of channels). This operation assumes a fixed sampling location and cannot adapt to spatial deformations of the input features.

[0026] An offset prediction branch is added to enhance spatial adaptability. This branch consists of an additional convolutional layer called the offset prediction layer. The initial spatial feature matrix is ​​input. The offset prediction layer uses a convolution operation, with an output dimension of... ,in Indicate each position and each sampling point 2D offset The offset prediction formula is:

[0027] in It is a convolution operation (usually used) nuclear), These are the trainable parameters (weights and biases) of this layer, and the output is... This is the offset field. Offset amount. directly from Extract from, for example, for position and index , correspond A specific slice.

[0028] Obtained through offset prediction branch Next, the sampling positions of the offset convolution kernel are calculated. For the output position... and each sampling point The new sampling location is defined as .because Possibly non-integer coordinates (i.e., not grid aligned), requiring bilinear interpolation to calculate input features. The value at that point. The interpolation formula is:

[0029] in yes The four nearest integer neighbors are located around the same point. It is a bilinear kernel function:

[0030] Ensure the interpolation is smooth and differentiable. Finally, calculate the output features using the offset sampling positions: This formula integrates predicted offsets, enabling the convolutional kernel to adapt and deform, thus improving robustness to geometric transformations. The entire layer optimizes its weights through end-to-end training. and offset parameters This ensures that offset learning and feature extraction are carried out in tandem.

[0031] Step S2 also includes: After obtaining the offset convolution kernel sampling positions, for sampling positions with non-integer coordinates, the offset feature at that position is calculated using bilinear interpolation.

[0032] As one embodiment, an offset prediction branch is added to the conventional convolutional kernel structure of a deformable convolutional layer. Training data reflecting geochemical spatial patterns is input into this layer, and the spatial offset of the convolutional kernel is obtained through the offset field output by this layer. For sampling positions with non-integer coordinates, bilinear interpolation is used to calculate the input feature values. The entire layer optimizes the weights and offset parameters through end-to-end training. By predicting the spatial offset of each convolutional kernel position through the offset branch, the offset convolutional kernel sampling positions are obtained.

[0033] Step S3 includes: A deformable convolutional layer with a stride of 1 is used to extract the small-scale features, and a deformable convolutional layer with a stride of 2 is used to extract the large-scale features. The large-scale feature map is upsampled and then concatenated with the small-scale feature map along the channel dimension, and then processed through a... The convolutional layer integrates features to obtain the multi-scale spatial features.

[0034] As one implementation example, small stride deformable convolution ensures that the feature map resolution remains unchanged, focusing on pixel-level features, while large stride deformable convolution increases the receptive field, effectively detecting global abnormal patterns and reducing computational cost. The specific implementation method of this step is as follows: Input feature map From the preceding deformable convolutional layer, where Indicates the number of channels. and These represent the height and width, respectively, with the learned offsets attached to the feature map. This offset allows the convolutional kernel to dynamically adapt to the input structure. Apply stride. Deformable convolutional layers with kernel size of The weight is This is used to extract small-scale features related to local elements. Output features The calculation is as follows:

[0035] here This indicates the spatial coordinates of the output feature map. It is the standard offset position of the convolution kernel (e.g., for nuclear, cover arrive ), These are additional learnable offsets, optimized through backpropagation. These are the convolution weight parameters. A small stride ensures the feature map resolution remains constant. Experience the wildness of small details, focusing on pixel-level details such as edges, textures, and local correlations to avoid information loss.

[0036] Use step size Deformable convolutional layers process the same input features Convolution kernels of the same size are Weight Independent optimization is used to extract large-scale features of regional anomalies. Output features. The calculation is as follows:

[0037] in This indicates the downsampling of the input position (e.g., Output size is ), and The definition is the same as above. A larger stride increases the receptive field, covering a wider area (e.g., ...). (pixel), effectively detects global abnormal patterns such as object-level inconsistencies or contextual deviations, while reducing computational load.

[0038] small-scale features and large-scale features Integration, forming multi-scale spatial features First of all, Perform an upsampling operation and use bilinear interpolation to restore the original size. Then, stitch them together along the channel dimension and apply... Convolutional layer Perform feature integration:

[0039] here This represents a convolution operation, where weights are learned through training, and the output is:

[0040] By combining local details and global context, the model's robustness to multi-scale spatial information is enhanced, making it suitable for subsequent tasks such as anomaly detection.

[0041] Step S4 includes: The offset optimization loop module uses the fused multi-scale spatial features to re-predict the convolution kernel offset and update the sampling position, repeating the process iteratively until the change in offset meets the preset stability condition. The offset optimization loop module is designed based on a gated loop unit; the preset stability condition is to calculate the norm of the offset change, and the iteration is terminated when it is less than a preset threshold.

[0042] In one embodiment, the offset optimization loop module initializes the sampling position offset as a zero vector, and uses a fully connected layer to perform nonlinear mapping to predict the convolutional kernel offset increment, thereby updating the sampling position. The specific implementation method of this step is as follows: Multi-scale spatial features are extracted from the input data, and feature maps of different scales are generated using a pre-trained convolutional neural network. ,in Represents scale index (e.g.) (Corresponding to low, medium, and high resolution). These features are concatenated and fused along the channel dimension to form a unified feature representation. ,in The operation ensures that feature information at each scale is complementary, enhancing spatial context awareness. The fused feature dimension is [missing value]. , This is the total number of channels. and These are the feature map height and width. Initialize the sampling position offset. A zero vector indicates that the initial convolution kernel sampling points have no offset. and The input offset optimization loop module, based on a gated recurrent unit (GRU) design, handles sequence dependencies. It calculates the current time step. Hidden state ,in It is the hidden state of the previous time step (initial) (Set to zero), the GRU unit dynamically fuses feature and offset information by updating and resetting gates, and the output dimension matches the hidden layer size. Predict the kernel offset increment using fused features and hidden states. Nonlinear mapping is achieved through fully connected layers. , It is the weight matrix (dimensions) ), It is a bias vector (dimension) ), Indicates the offset dimension (usually 2D or 3D coordinates). Update sampling position is... The new positions are directly applied to adjust the sampling point coordinates of the deformable convolution kernel, ensuring that feature extraction adapts to local structural changes. This iterative process is repeated to calculate the offset change norm. That is, the Euclidean distance. If ,in If the preset stability threshold (e.g., 0.001) is met, the iteration terminates; otherwise, it increments. Continue the loop until the condition is met or the maximum number of iterations is reached. To avoid infinite loops, the entire process involves end-to-end training and optimization of module parameters to ensure that the offset converges to a stable state. During this process, the magnitude and direction of the offset at the mining point window are visualized and analyzed, and compared with the input data visualization to obtain... Figure 2 .

[0043] Step S5 includes: The fully connected layer outputs accurately identified complex geochemical space models, including: A linear transformation is performed on the optimized first fusion feature to obtain the fractional vector. The softmax activation function is applied to convert the scores into a probability distribution. ,in The total number of categories of geochemical models. For the sampling point to belong to the first The probability of each pattern is used to determine the final identification result, i.e., the complex geochemical spatial model, based on the highest probability. The identified geochemical model labels are associated with the spatial coordinates of the sampling points, and geographic information system tools are used to generate a visualization map showing the spatial distribution patterns of the geochemical models.

[0044] As an example, the optimized fusion features are input into a fully connected layer, outputting geochemical model labels for each sampling point. These labels are then mapped to geographic space using spatial coordinates to generate a model visualization. The specific implementation method for this step is as follows: Optimized fusion feature vector As input, where This is the feature dimension (e.g., the dimension of the output from a feature extraction network), and this vector is fed into the fully connected layer. The fully connected layer consists of a weight matrix. and bias vector definition, This represents the total number of geochemical model categories (e.g., different mineral assemblage types). The linear transformation output is calculated as follows: ,in It is an inactive score vector, each element Corresponding category The original score.

[0045] The scores are converted into a probability distribution by applying the softmax activation function:

[0046] in Is the sampling point the first The probability of a geochemical model, The range is from 1 to Output probability vector This represents the confidence level across multiple categories. Based on the probability distribution, it predicts the geochemical model label for each sampling point. : ,here These are discrete label values; select the category index corresponding to the highest probability.

[0047] Integrate spatial coordinate data of sampling points, such as longitude and latitude ,in Indicates the sampling point index, which will predict the label. With coordinates This process creates a geospatial mapping dataset. Using Geographic Information System (GIS) tools or Python libraries such as geopandas, a spatial point layer is generated based on coordinates and labels, where the label values... Used as a categorical attribute for coloring.

[0048] Generate pattern visualizations by overlaying point layers onto a geographic base map using plotting functions such as matplotlib's scatter or a GIS rendering engine, based on... Different color codes are assigned to create scatter plots or heatmaps, visually displaying the spatial distribution patterns of geochemical models. Output images are saved in standard formats such as PNG or GeoTIFF, ensuring a consistent spatial reference frame to support geological analysis. During this process, the input data at the mineral occurrence window, along with ordinary and variable convolution activation maps, are visualized and analyzed to obtain… Figure 3 .

[0049] This application also discloses an electronic device. (See reference...) Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of this application. The electronic device 500 may include: at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0050] The communication bus 502 is used to enable communication between these components.

[0051] The user interface 503 may include a display screen, and optionally, the user interface 503 may also include a standard wired interface or a wireless interface.

[0052] The network interface 504 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0053] This application also discloses a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute the aforementioned method for characterizing complex geochemical spatial models based on deformable convolution.

[0054] The above are merely exemplary embodiments of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure.

[0055] This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described in this disclosure. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

Claims

1. A method for characterizing complex geochemical spatial models based on deformable convolution, characterized in that, The method includes the following steps: S1. Collect geochemical sampling point data of the target area and preprocess it to obtain a dataset; S2. Predict branches using offsets and construct deformable convolutional layers; Input the dataset into the deformable convolutional layer to predict the spatial offset of each convolutional kernel position, and obtain the offset convolutional kernel sampling position and its offset features; S3. For the offset features output by the deformable convolutional layer, small-scale features for characterizing local element associations are extracted by small-stride deformable convolution, and large-scale features for characterizing regional anomalies are extracted by large-stride deformable convolution. The small-scale features and large-scale features are then fused to obtain multi-scale spatial features. S4. Input the multi-scale spatial features into the offset optimization loop module to obtain the first fused feature; S5. Input the first fusion feature into the fully connected layer to output a precisely identified complex geochemical spatial model.

2. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 1, characterized in that, Step S1 includes: Geochemical sampling point data of the target area are collected, the element content of each sampling point is preprocessed, and a fixed-size neighborhood window is constructed for each sampling point to form a training dataset that reflects the geochemical spatial model. The elemental content preprocessing includes: using the min-max normalization method to linearly map the original elemental content values ​​to... Interval.

3. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 1, characterized in that, Step S2 includes: An offset prediction branch is added to the conventional convolutional kernel structure of a deformable convolutional layer; the offset prediction branch is an additional convolutional layer, namely the offset prediction layer. The offset prediction layer outputs the offset field through convolution operations. To obtain the spatial offset, where Given the spatial feature matrix as input, For trainable parameters, output The dimension is , and The height and width of the feature map, The number of sampling points for the convolution kernel. This represents the two-dimensional offset corresponding to each sampling point.

4. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 3, characterized in that, Step S2 also includes: After obtaining the offset convolution kernel sampling positions, for sampling positions with non-integer coordinates, the offset feature at that position is calculated using bilinear interpolation.

5. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 1, characterized in that, Step S3 includes: A deformable convolutional layer with a stride of 1 is used to extract the small-scale features, and a deformable convolutional layer with a stride of 2 is used to extract the large-scale features. The large-scale feature map is upsampled and then concatenated with the small-scale feature map along the channel dimension, and then processed through a... The convolutional layer integrates features to obtain the multi-scale spatial features.

6. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 1, characterized in that, Step S4 includes: The offset optimization loop module uses the fused multi-scale spatial features to re-predict the convolution kernel offset and update the sampling position, repeating the process iteratively until the change in offset meets the preset stability condition. The offset optimization loop module is designed based on a gated loop unit; the preset stability condition is to calculate the norm of the offset change, and the iteration is terminated when it is less than a preset threshold.

7. The method for characterizing complex geochemical spatial models based on deformable convolution as described in claim 1, characterized in that, Step S5 includes: The fully connected layer outputs accurately identified complex geochemical space models, including: A linear transformation is performed on the optimized first fusion feature to obtain the fractional vector. The softmax activation function is applied to convert the scores into a probability distribution. ,in The total number of categories of geochemical models. For the sampling point to belong to the first The probability of each pattern is used to determine the final identification result, i.e., the complex geochemical spatial model, based on the highest probability. The identified geochemical model labels are associated with the spatial coordinates of the sampling points, and geographic information system tools are used to generate a visualization map showing the spatial distribution patterns of the geochemical models.

8. An electronic device, characterized in that, The device includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to perform the complex geochemical spatial model characterization method based on deformable convolution as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by a computer, perform the complex geochemical spatial model characterization method based on deformable convolution as described in any one of claims 1-7.

Images