Shallow-temperature hydrothermal deposit modeling method based on short-wave infrared

Abstract

The invention discloses a shallow low-temperature hydrothermal deposit modeling method based on short-wave infrared, and relates to the technical field of artificial intelligence mineral product prediction, and the method comprises the steps: carrying out the weighted normalization processing of multi-source heterogeneous geological data, mapping the multi-source heterogeneous geological data to a unified feature space through a nonlinear fusion coding algorithm with a weight coefficient, and carrying out the modeling of the multi-source heterogeneous geological data. Generating a fusion feature vector; constructing a self-attention mechanism perceived by a geological structure, introducing a geological prior matrix reflecting terrain gradient and geochemical abnormal distribution on the basis of a key, query and value matrix, and dynamically adjusting local geological coupling weight parameters in an attention calculation process; and establishing a geological consistency loss mechanism, performing spatial constraint on the attention response and the alteration label, forming a composite loss structure through a nonlinear amplification term, a logarithmic penalty term and a smooth boundary constraint term, and enabling the geological consistency loss mechanism, classification loss and a regular term to participate in optimization together.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence mineral prediction technology, specifically a method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared radiation. Background Technology

[0002] Shallow low-temperature hydrothermal deposits are an important type of mineralization geology research. Their formation process involves complex multi-field coupling and multi-scale geological response characteristics. With the advancement of remote sensing, geochemistry, and topographic surveying technologies, geological information acquisition methods have been significantly improved. A large amount of multi-source heterogeneous geological data can support the analysis of deposit characteristics. Traditional methods based on statistical analysis, spectral matching, and geochemical anomaly identification can extract some mineralization clues, but they are difficult to effectively model when processing high-dimensional, multimodal data. In recent years, artificial intelligence and deep learning technologies have been gradually applied to the field of geological information identification. Through structures such as convolutional neural networks and Transformers, geological pattern mining and feature extraction are realized, providing a new research direction for intelligent deposit prediction.

[0003] Existing AI-based mineral deposit analysis methods generally suffer from limited multi-source data fusion capabilities, insufficient geological structure perception, and weak model interpretability. In terms of data fusion, traditional methods often employ linear weighting or simple splicing strategies, failing to consider the nonlinear correlation between remote sensing spectra, geochemical elements, and topographic features. This leads to biased feature representations and makes it difficult to reflect the multi-scale coupling effects in epithermal deposits. During feature identification, existing models mostly rely solely on global correlation calculations using self-attention mechanisms, without incorporating geological prior constraints. They lack effective perception of spatial information such as topographic gradient changes and geochemical anomaly distributions, making it difficult to accurately identify alteration zoning and mineralization boundaries. In the model optimization stage, existing training methods typically use a single loss function, failing to consider geological spatial continuity and structural consistency. This easily leads to overfitting or boundary blurring in complex geological contexts. Therefore, existing technologies cannot achieve accurate identification and structured analysis of alteration features in epithermal deposits through unified feature space representation, geological perception attention mechanisms, and spatial consistency optimization strategies. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing mineral deposit feature analysis technologies suffer from poor accuracy in multi-source geological data fusion, low spatial perception capability of geological structures, lack of geological consistency constraints in model training, and the problem of how to achieve high-precision identification and structured analysis of alteration zoning characteristics of shallow low-temperature hydrothermal deposits through artificial intelligence models.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a modeling method for shallow low-temperature hydrothermal deposits based on shortwave infrared radiation, comprising weighted normalization processing of multi-source heterogeneous geological data, mapping the multi-source heterogeneous geological data to a unified feature space through a nonlinear fusion coding algorithm with weighted coefficients, generating a fused feature vector; constructing a self-attention mechanism for geological structure perception, introducing a geological prior matrix reflecting the distribution of topographic gradients and geochemical anomalies based on the key, query, and value matrices, dynamically adjusting local geological coupling weight parameters during attention calculation; establishing a geological consistency loss mechanism, spatially constraining the attention response and alteration labels, constructing a composite loss structure through nonlinear amplification terms, logarithmic penalty terms, and smooth boundary constraint terms, and optimizing the geological consistency loss mechanism together with classification loss and regularization terms.

[0007] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method of the present invention, the weighted normalization processing of multi-source heterogeneous geological data includes: normalizing spectral reflectance data by band, so that the gray values ​​of different image sources are in a unified numerical range after normalization; standardizing geochemical element concentration data according to the anomaly sensitivity of each element, so that the standardization results maintain a consistent proportion on the numerical scale; smoothing topographic gradient data according to the slope change rate, so that the smoothing results converge in high-slope areas to reduce numerical offset; after normalization, each type of data corresponds to a spectral feature weight coefficient, a geochemical feature weight coefficient, and a topographic feature weight coefficient, which are determined by preset or self-learning methods to adjust the proportional relationship of different data sources in the fusion feature calculation; before weighted fusion, the remote sensing spectral data, geochemical data, and topographic data are spatially registered, and a unified geographic projection coordinate system is used to resample each type of data to the same grid resolution so that the corresponding pixel positions correspond one-to-one in space; after weighting, the matrix is ​​superimposed to obtain a preliminary fusion feature matrix, and the obtained matrix is ​​used as the input for the nonlinear fusion coding step.

[0008] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method of the present invention, the generation of the fused feature vector includes: inputting weighted normalized spectral data, geochemical data, and topographic data into a fusion function; obtaining multi-source input feature data through linear superposition; transforming the multi-source input feature data using a nonlinear activation unit, wherein the nonlinear activation unit adopts a hyperbolic tangent function; and limiting the fused signal to a finite interval under the action of the activation function; automatically adjusting the dynamic weight of the input features according to the standard deviation change of each type of input feature; when the standard deviation fluctuation amplitude of the input feature is greater than the standard deviation fluctuation threshold of the input feature, the balance adjustment weight factor of the corresponding input feature automatically decreases; and outputting the fused encoding as a multi-dimensional feature vector, where each component corresponds to the response value after the combination of input features; and updating each parameter in the fusion process through backpropagation during the training phase.

[0009] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method described in this invention, the self-attention mechanism for constructing geological structure perception includes: establishing a key matrix, a query matrix, and a value matrix within the geological prior attention model; calculating the feature correlation between each spatial unit; generating a geological prior matrix through joint calculation of topographic gradient and geochemical anomaly degree to characterize the spatial differences in geological structure; during attention calculation, superimposing the geological prior matrix and the attention similarity matrix according to geological coupling weights to form a comprehensive attention score; and dynamically adjusting the attention weights based on the relative deviation between local values ​​and global statistical features in the geological prior matrix.

[0010] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method of the present invention, the dynamic adjustment of local geological coupling weight parameters during attention calculation includes: monitoring the local difference values ​​between the input features and the geological prior matrix during attention weighting operations, and using the local difference values ​​as the basis for adaptive adjustment of the local geological coupling weight parameters; when the local difference value is greater than the set geological difference sensitivity threshold, increasing the value of the local geological coupling weight parameters within the corresponding spatial unit range; continuously updating the local geological coupling weight parameters during the training phase, and determining that the local geological coupling weight parameters converge to a stable interval according to the optimization criteria during multiple iterations; through the dynamic adjustment mechanism, the attention matrix in the geological prior attention model is adapted according to the spatial distribution characteristics of the geological prior in each round of calculation, forming a weighted matrix structure that is updated in real time with changes in geological conditions.

[0011] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method of the present invention, the geological consistency loss mechanism includes: in the multimodal geological feature recognition model, calculating the spatial difference between the attention response distribution output by the model and the alteration label, and constructing a loss function based on the spatial difference; the loss function consists of three types of sub-terms, including a nonlinear amplification term, a logarithmic penalty term, and a smooth boundary constraint term; the three types of sub-terms are combined into a composite loss structure by weighting, and the calculation can be performed using a fixed weight or an adaptive weight mode during training. The fixed weight is a preset constant, and the adaptive weight is dynamically adjusted according to the convergence rate of each sub-term loss; the multimodal geological feature recognition model performs parameter updates according to the convergence of the composite loss function, and the composite loss structure performs a complete backpropagation process in each iteration.

[0012] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method of the present invention, the optimization of the geological consistency loss mechanism in conjunction with the classification loss and regularization term includes: during the training phase of the multimodal geological feature recognition model, the geological consistency loss function, the sample classification loss function, and the regularization smoothing function are proportionally superimposed to form a comprehensive optimization objective; in each training iteration, the weighted sum of the three sub-items is minimized simultaneously; the training control unit dynamically adjusts the weights of each item according to the rate of change of loss, and when the rate of decrease of geological consistency loss is lower than that of classification loss, the weight of geological consistency loss is automatically increased to balance the convergence speed; when the gradient oscillation is greater than the preset gradient oscillation threshold, the weight of the regularization term is automatically increased; the joint optimization mechanism performs gradient normalization processing in each iteration, and all parameters of the multimodal geological feature recognition model are updated synchronously within a unified optimization framework.

[0013] Another objective of this invention is to provide a modeling system for shallow low-temperature hydrothermal deposits based on shortwave infrared radiation. This system can construct a self-attention mechanism for geological structure perception, introduce a geological prior matrix reflecting the distribution of topographic gradients and geochemical anomalies on the basis of key, query, and value matrices, and dynamically adjust local geological coupling weight parameters during the attention calculation process. This solves the problem that current geological feature recognition technologies ignore the spatial distribution patterns of geological features, leading to discontinuous identification of alteration zones.

[0014] As a preferred embodiment of the shortwave infrared-based shallow low-temperature hydrothermal deposit modeling system of the present invention, the system comprises: a multi-source feature mapping module, a geological prior guidance module, and a geological coupling optimization module. The multi-source feature mapping module performs weighted normalization processing on multi-source heterogeneous geological data, mapping the data to a unified feature space using a nonlinear fusion coding algorithm with weighted coefficients to generate a fused feature vector. The geological prior guidance module constructs a self-attention mechanism for geological structure perception, introducing a geological prior matrix reflecting the distribution of topographic gradients and geochemical anomalies based on the key, query, and value matrices, and dynamically adjusting local geological coupling weight parameters during attention calculation. The geological coupling optimization module establishes a geological consistency loss mechanism, spatially constraining the attention response and alteration labels, and constructing a composite loss structure using nonlinear amplification terms, logarithmic penalty terms, and smooth boundary constraint terms, integrating the geological consistency loss mechanism with classification loss and regularization terms for optimization.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared radiation.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared radiation.

[0017] The beneficial effects of this invention are as follows: The shortwave infrared-based shallow low-temperature hydrothermal deposit modeling method provided by this invention performs weighted normalization processing on multi-source heterogeneous geological data. Through a nonlinear fusion coding algorithm with weighted coefficients, the multi-source heterogeneous geological data is mapped to a unified feature space, generating a fused feature vector. This achieves scale unification and information fusion of multi-source heterogeneous geological data, reduces data bias, and enhances the expressive power of geological features. Furthermore, a self-attention mechanism for geological structure perception is constructed. Based on the key, query, and value matrices, a geological prior matrix reflecting the distribution of topographic gradients and geochemical anomalies is introduced. During the attention calculation process, local geological coupling weight parameters are dynamically adjusted. This invention achieves adaptive focusing and weight adjustment of the model for complex geological structures, improving the accuracy of identifying the main controlling factors of shallow low-temperature hydrothermal deposits. A geological consistency loss mechanism is established, spatially constraining the attention response and alteration labels. A composite loss structure is constructed using nonlinear amplification terms, logarithmic penalty terms, and smooth boundary constraint terms. This geological consistency loss mechanism, along with classification loss and regularization terms, participates in optimization, achieving simultaneous optimization of spatial continuity and semantic accuracy. This enhances the geological credibility and interpretability of the model output. The invention achieves better results in terms of multi-source information fusion accuracy, geological feature recognition capability, and accuracy of spatial modeling of mineralization zoning. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is an overall flowchart of the shallow low-temperature hydrothermal deposit modeling method based on shortwave infrared provided in Embodiment 1 of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figure 1 As an embodiment of the present invention, a method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared radiation is provided, comprising:

[0022] S1: Perform weighted normalization processing on multi-source heterogeneous geological data, and use a nonlinear fusion coding algorithm with weighted coefficients to map the multi-source heterogeneous geological data to a unified feature space to generate a fusion feature vector.

[0023] Furthermore, the weighted normalization process for multi-source heterogeneous geological data includes: normalizing spectral reflectance data by band, ensuring that the gray values ​​of different image sources are within a unified numerical range after normalization; standardizing geochemical element concentration data based on the anomaly sensitivity of each element, maintaining a consistent numerical scale in the normalization results; smoothing topographic gradient data based on the slope change rate, with the smoothing results showing a convergent distribution in high-slope areas to reduce numerical offset; after normalization, each type of data corresponds to a spectral feature weight coefficient, a geochemical feature weight coefficient, and a topographic feature weight coefficient, determined through preset or self-learning methods to adjust the proportional relationship of different data sources in the fusion feature calculation; before weighted fusion, spatial registration is performed on remote sensing spectral data, geochemical data, and topographic data, using a unified geographic projection coordinate system, resampling each type of data to the same grid resolution, ensuring a one-to-one spatial correspondence between corresponding pixel locations, and obtaining a preliminary fusion feature matrix through matrix superposition after weighting, which serves as the input for the nonlinear fusion coding step.

[0024] It should also be noted that a preferred scheme for weighted normalization of multi-source heterogeneous geological data specifically includes standardizing spectral reflectance data according to spectral bands and using a minimum-to-maximum normalization method to unify the grayscale values ​​of different satellite image sources to a uniform level. Numerical range; geochemical element concentration data are standardized based on the anomalous sensitivity of each element in the mineralization alteration zone. The anomalous sensitivity is obtained through geostatistical analysis and used to adjust the scale bias of different element concentrations; topographic gradient data are smoothed based on the slope change rate, and a Gaussian convolution smoothing kernel function is used for areas with abrupt slope changes to eliminate local numerical spikes caused by anomalous slopes; during the normalization process, spectral feature weight coefficients, geochemical feature weight coefficients, and topographic feature weight coefficients are set separately to control the proportional relationship of various data in the fusion process.

[0025] The initial weights can be preset empirically or learned automatically during model training. In self-learning mode, each weight parameter is treated as a trainable variable and updated through backpropagation. The iterative relationship is expressed as:

[0026]

[0027] in, Indicates the first Geological features in the model in the first Weight values ​​during secondary parameter optimization iterations Indicates the first Geological features in the model in the first Weight values ​​during secondary parameter optimization iterations Indicates the first Next parameter optimization iteration step Optimize the learning rate for geological features. This is the loss function for the fusion stage.

[0028] To ensure spatial consistency of multi-source data, a unified spatial registration was performed on the three types of data before weighted fusion. A unified geographic coordinate projection system was adopted, and bilinear interpolation or Kriging methods were used to resample data at different resolutions, ensuring a one-to-one correspondence of pixels among the various data types at the same grid resolution. After spatial registration, the preliminary fused integrated geological feature matrix was calculated. , is represented as:

[0029]

[0030] in, Represents the spectral feature weighting coefficient. Indicates the weighting coefficient of geochemical characteristics. Represents the terrain feature weighting coefficient. Represents the normalized spectral data, This represents the normalized geochemical data. This represents the normalized terrain data.

[0031] It should be noted that generating the fused feature vector includes inputting weighted and normalized spectral data, geochemical data, and topographic data into a fusion function, obtaining multi-source input feature data through linear superposition, transforming the multi-source input feature data using a nonlinear activation unit (HMU) employing a hyperbolic tangent function, and confining the fused signal within a finite interval under the action of the activation function; automatically adjusting the dynamic weights of the input features based on the standard deviation changes of each type of input feature; when the standard deviation fluctuation of the input feature exceeds the standard deviation fluctuation threshold, the corresponding balance adjustment weight factor of the input feature automatically decreases; and outputting the fused encoded feature vector as a multi-dimensional feature vector, with each component corresponding to the response value after combining the input features. The parameters in the fusion process are updated through backpropagation during the training phase.

[0032] It should also be noted that a preferred scheme for generating the fused feature vector specifically includes introducing a dynamic weight adjustment mechanism, which adaptively corrects the weights of each input feature based on the degree of fluctuation in the standard deviation of each feature. Let the first feature be... Class input features in the th The standard deviation at the nth training iteration is Then the standard deviation fluctuation range is defined. , is represented as:

[0033]

[0034] when When the input features fluctuate slightly, it means that the fluctuation is within an acceptable range of random perturbations. Appropriate slight weight adjustments can improve the fusion balance.

[0035] when When this occurs, it indicates that the distribution of the input features has changed significantly, and the magnitude of this change exceeds the range of natural noise. The model needs to trigger a weight decay mechanism to weaken the dominant influence of this feature in the fusion process.

[0036] when If the current input feature fluctuates drastically, it is very likely caused by anomalies or local anomalies. In order to prevent the training results from becoming unbalanced, the weight of this feature needs to be significantly reduced.

[0037] It should also be noted that by performing weighted normalization on multi-source heterogeneous geological data such as spectral, geochemical, and topographic gradient data, alignment and feature balance of data from different sources are achieved at a unified numerical scale. A nonlinear fusion coding algorithm with weighted coefficients is used to map geological information with different physical properties to a unified feature space, thereby generating a fused feature vector. During this process, spectral band normalization, geochemical anomaly standardization, and topographic slope smoothing enable the model to simultaneously preserve the differences and comparability of geological information. This achieves a unified expression of multi-source geological information, providing high-quality input for deep feature learning and effectively avoiding feature shift problems caused by differences in resolution, scale, or numerical values ​​between multi-source data. Through nonlinear coding processing that integrates various geological weights, the integrity and stability of geological feature representation are improved, laying the foundation for the model to extract key features such as fracture structures, alteration zoning, and geochemical anomalies in shallow low-temperature hydrothermal deposits, achieving efficient fusion of multi-source geological data and enhanced geological information expression.

[0038] S2: Construct a self-attention mechanism for geological structure perception. Based on the key, query and value matrices, introduce a geological prior matrix that reflects the distribution of topographic gradient and geochemical anomalies. Dynamically adjust the local geological coupling weight parameters during the attention calculation process.

[0039] Furthermore, the self-attention mechanism for geological structure perception includes: establishing a key matrix, query matrix, and value matrix within the geological prior attention model; calculating the feature correlation between spatial units; generating a geological prior matrix by jointly calculating the topographic gradient and geochemical anomaly degree to characterize the spatial differences in geological structures; superimposing the geological prior matrix and the attention similarity matrix according to geological coupling weights during attention calculation to form a comprehensive attention score; and dynamically adjusting the attention weights based on the relative deviation between local values ​​and global statistical features in the geological prior matrix.

[0040] It should be noted that a preferred scheme for constructing a self-attention mechanism for geological structure perception specifically includes superimposing the attention matrix and the geological prior matrix proportionally according to geological coupling weight parameters to form a geologically guided feature matrix. , is represented as:

[0041]

[0042] in, Represents the original fused feature matrix. For geological prior matrix, For geological coupling weights, it represents the weights in the first... The dynamic balance coefficient between data-driven features and geological prior features during each training iteration controls the intensity of geological constraints in the model, and its value ranges from [value range missing]. The model in the first During the next training iteration According to geological learning rate Gradual adjustment, represented as:

[0043]

[0044] in, The model represents the first time. The geological coupling weights during the next training iteration, and the geological-guided feature matrix after the iteration stabilizes. Spatially, it exhibits an adaptive distribution to geological structures. Finally, the geological guidance feature matrix is ​​normalized and compared with the value matrix. Multiplying them together yields the geological attention output matrix. , is represented as:

[0045]

[0046] in, To represent the feature dimensions of the attention space, This represents the original content of the features of interest, which includes multidimensional information about geological spatial units, such as alteration intensity, geochemical concentration, slope direction, etc.

[0047] It should be noted that the dynamic adjustment of local geological coupling weight parameters during attention calculation includes: monitoring the local difference values ​​between the input features and the geological prior matrix during attention weighting operations, and using these local difference values ​​as the basis for adaptive adjustment of the local geological coupling weight parameters; increasing the value of the local geological coupling weight parameters within the corresponding spatial unit range when the local difference value exceeds the set geological difference sensitivity threshold; continuously updating the local geological coupling weight parameters during the training phase to ensure that the local geological coupling weight parameters converge to a stable range according to the optimization criteria during multiple iterations; and through the dynamic adjustment mechanism, the attention matrix in the geological prior attention model is adapted to the spatial distribution characteristics of the geological prior in each round of calculation, forming a weighted matrix structure that is updated in real time with changes in geological conditions.

[0048] It should also be noted that a preferred scheme for dynamically adjusting the local geological coupling weight parameters during attention calculation specifically includes, during the dynamic adjustment of the local geological coupling weight parameters, the geological guidance feature matrix... With geological prior matrix The degree of matching is measured by local difference values. To achieve dynamic adjustment, the local difference matrix is ​​calculated in each iteration. And extract the average geological difference index. As an overall difference indicator; when At that time, automatically increase the local geological coupling weight parameter of the current iteration. To enhance the strength of geological constraints; when At that time, reduce Expressed using the balance characteristic, it is represented as:

[0049]

[0050] in, Indicates the first During the next training iteration, the location is in geological space coordinates. Local geological coupling weight parameters at the location, For geological learning rate, The threshold for sensitivity to geological differences The local average geological difference index is represented by the index on the first day. In each training iteration, the model prediction and the geological prior matrix are in position... The average degree of difference is used to determine the model's predictions. A large difference indicates that the model has not fully learned geological patterns; a small difference indicates that the model's predictions are close to the actual geological structure. Indicates the first During the next training iteration, the location is in geological space coordinates. Local geological coupling weight parameters at the location, updated local geological coupling weight parameters In the next iteration, this is passed back to the fusion calculation formula, expressed as:

[0051]

[0052] in, The updated geological guidance feature matrix shows that the closer the value is to the actual alteration distribution, the more accurately the model can capture the ore-controlling features of shallow low-temperature hydrothermal deposits.

[0053] It should also be noted that by constructing a self-attention mechanism for geological structure perception, the model achieves proactive perception and dynamic response to geological prior information during feature learning. Based on the key, query, and value matrices, a geological prior matrix reflecting topographic gradients and geochemical anomaly distributions is introduced, and geological coupling weights are superimposed during attention calculation, enabling the model to focus spatially on geological anomaly areas. By monitoring local differences between input features and the geological prior matrix, and using a geological difference sensitivity threshold as a criterion, the local geological coupling weight parameters are dynamically adjusted, achieving adaptive control over different geological units. This mechanism enables the model to possess recognition logic similar to that of a geological expert, automatically strengthening the feature response of key tectonic zones in complex geological backgrounds. The model no longer passively relies on data distribution but actively adapts to changes in geological structure, thereby improving the spatial resolution and geological interpretability of fault identification and alteration zone extraction, and enhancing the model's learning stability and robustness in complex mineralized environments.

[0054] S3: Establish a geological consistency loss mechanism, spatially constrain the attention response and alteration label, and construct a composite loss structure through nonlinear amplification term, logarithmic penalty term and smooth boundary constraint term, so that the geological consistency loss mechanism, classification loss and regularization term can participate in the optimization.

[0055] Furthermore, establishing a geological consistency loss mechanism involves calculating the spatial difference between the attention response distribution output by the model and the alteration label in the multimodal geological feature recognition model, and constructing a loss function based on the spatial difference. The loss function consists of three types of sub-terms: a nonlinear amplification term, a logarithmic penalty term, and a smooth boundary constraint term. These three types of sub-terms are combined into a composite loss structure through a weighted approach. During training, either a fixed weight or an adaptive weight mode can be used for calculation. The fixed weight is a preset constant, while the adaptive weight is dynamically adjusted according to the convergence rate of each sub-term's loss. The multimodal geological feature recognition model updates its parameters based on the convergence of the composite loss function, and the composite loss structure performs a complete backpropagation process in each iteration.

[0056] It should be noted that a preferred scheme for establishing a geological consistency loss mechanism specifically includes combining the nonlinear amplification term, the logarithmic penalty term, and the smoothing boundary constraint term into a composite loss structure through a weighted method, and then calculating the geological consistency loss function. , is represented as:

[0057]

[0058] in, This represents the degree of difference between the model output features and the geological prior features. Representing the geological prior matrix Spatial gradient, This is the geological nonlinear amplification factor. For the weights of nonlinear amplification terms, The weight of the logarithmic penalty term, The weights for the smoothing boundary constraint terms.

[0059] It should be noted that the geological consistency loss mechanism, along with the classification loss and regularization term, participates in the optimization. This includes: during the training phase of the multimodal geological feature recognition model, the geological consistency loss function, the sample classification loss function, and the regularization smoothing function are proportionally superimposed to form a comprehensive optimization objective; in each training iteration, the weighted sum of the three sub-items is minimized simultaneously; the training control unit dynamically adjusts the weights of each item according to the rate of loss change; when the rate of decrease in geological consistency loss is lower than that of classification loss, the weight of geological consistency loss is automatically increased to balance the convergence speed; when gradient oscillation exceeds a preset gradient oscillation threshold, the weight of the regularization term is automatically increased; the joint optimization mechanism performs gradient normalization in each iteration, and all parameters of the multimodal geological feature recognition model are updated synchronously within a unified optimization framework.

[0060] It should also be noted that the optimal optimization scheme that integrates the geological consistency loss mechanism with classification loss and regularization term specifically includes, for the multimodal geological feature identification task of shallow low-temperature hydrothermal deposits, a dynamic joint optimization mechanism is designed to balance the optimization relationship among geological consistency constraints, sample classification accuracy, and model regularization smoothness, and to calculate the comprehensive optimization objective function. , is represented as:

[0061]

[0062] in, This is the geological consistency loss term, reflecting the degree of spatial agreement between the model output and the alteration labels; The classification loss function; This is the regularization loss term; The dynamic weighting coefficient for the geological consistency loss term. The dynamic weighting coefficients of the sample classification loss term. This represents the dynamic weighting coefficient of the regularization loss term. During training, the rate of change of the three types of losses is monitored in real time. When the rate of decrease of the geological consistency loss is lower than that of the classification loss, the weight of the former is automatically increased to strengthen the geological structure constraint.

[0063] Calculate the rate of change of geological consistency gradient , is represented as:

[0064]

[0065] in, The model represents the first time. The magnitude of gradient change during each training iteration The model represents the first time. Geological learning direction during each training iteration The difference between the two values ​​indicates the direction of geological learning of the model in the previous training iteration. The difference between the two values ​​describes the rate of change of the model's geological cognition in continuous learning. When the two values ​​tend to be consistent, it means that the model has learned the mineralization control law of shallow low-temperature hydrothermal deposits and achieved geological consistency convergence. Represents a small smoothing constant;

[0066] Using gradient oscillation threshold To evaluate the learning performance of the model, ,like The model is in the geological learning phase, at which point the gradient oscillation amplitude is relatively large, and the geological learning rate is set to... The loss curve fluctuates significantly, indicating that the model's understanding of geological patterns is still unstable and it is in a rapid exploration phase. Therefore, we should strengthen the geological coupling weights, enable the geological feedback reinforcement update mechanism, and appropriately extend the number of training rounds for geological samples to promote the model's full learning of the geological feature space.

[0067] like The model enters the geological equilibrium stage, where its learning of geological features gradually stabilizes, the relationship between geology and data constraints reaches equilibrium, and the model begins to gradually form a coupled geological understanding, capable of capturing the intrinsic laws between faulting, alteration, and geochemical processes. The geological learning rate is then adjusted to... Fixed geological coupling weights and enabled the geological balance monitoring module.

[0068] like The model enters the geological convergence stage. At this point, the gradient change is minimal, the loss function tends to stabilize, and the model has achieved geologically consistent convergence. The spatial distribution of the output results exhibits continuity and geological interpretability, and the geological learning rate decreases to [value missing]. The geological coupling weights are frozen to stop updates, the model structure and parameters are fixed, and the final geological identification results and confidence layers are output to ensure the stability of the model in a geological sense and the reliability of predictions.

[0069] It should also be noted that by establishing a geological consistency loss mechanism, spatial constraints are applied to the attention response and alteration labels, enabling collaborative optimization of the model at both the geological spatial and semantic classification levels. This mechanism constructs a composite loss structure through nonlinear amplification terms, logarithmic penalty terms, and smoothing boundary constraint terms, which participate in model training together with classification loss and regularization terms. By weighting and adjusting the proportions of each loss term, the model can simultaneously ensure the spatial continuity of geological structures and the accuracy of classification predictions during the optimization process, thus effectively overcoming the problem of correct classification but fragmented geology in traditional algorithms. The nonlinear amplification term enhances the sensitive response to highly disparate regions, the logarithmic penalty term suppresses abnormal gradient fluctuations, and the smoothing constraint term maintains the continuous transition of alteration boundaries. The model ultimately achieves consistent convergence in geological space and stable output of prediction results, which not only improves the geological credibility of the identification results of epithermal deposits but also provides a more interpretable intelligent analysis tool for the study of mineralization regularities.

[0070] Example 2, an embodiment of the present invention, provides a modeling system for shallow low-temperature hydrothermal deposits based on shortwave infrared, including a multi-source feature mapping module, a geological prior guidance module, and a geological coupling optimization module.

[0071] The multi-source feature mapping module is used to perform weighted normalization processing on multi-source heterogeneous geological data. Through a nonlinear fusion coding algorithm with weighted coefficients, the multi-source heterogeneous geological data is mapped to a unified feature space to generate a fused feature vector. The geological prior guidance module is used to construct a self-attention mechanism for geological structure perception. Based on the key, query, and value matrices, a geological prior matrix reflecting the distribution of topographic gradients and geochemical anomalies is introduced. During the attention calculation process, the local geological coupling weight parameters are dynamically adjusted. The geological coupling optimization module is used to establish a geological consistency loss mechanism. The attention response and alteration label are spatially constrained. A composite loss structure is constructed through nonlinear amplification terms, logarithmic penalty terms, and smooth boundary constraint terms. The geological consistency loss mechanism, classification loss, and regularization terms are jointly optimized.

Claims

1. A method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared radiation, characterized in that, include: The multi-source heterogeneous geological data is weighted and normalized, and then mapped to a unified feature space by a nonlinear fusion coding algorithm with weighted coefficients to generate a fusion feature vector. A self-attention mechanism for geological structure perception is constructed. Based on the key, query, and value matrices, a geological prior matrix reflecting the distribution of topographic gradient and geochemical anomalies is introduced. Local geological coupling weight parameters are dynamically adjusted during the attention calculation process. A geological consistency loss mechanism is established, which spatially constrains the attention response and alteration label. A composite loss structure is formed by nonlinear amplification term, logarithmic penalty term and smooth boundary constraint term. The geological consistency loss mechanism, classification loss and regularization term are jointly used in the optimization.

2. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 1, characterized in that: The weighted normalization process for multi-source heterogeneous geological data includes... The spectral reflectance data is normalized by band, and the gray values ​​of different image sources are in a uniform range after normalization. Geochemical element concentration data are standardized based on the anomaly sensitivity of each element, and the standardized results maintain a consistent scale on the numerical scale. The terrain gradient data is smoothed based on the slope change rate. The smoothing results converge in high slope areas to reduce numerical offset. After normalization, each type of data corresponds to a weighting coefficient for spectral features, a weighting coefficient for geochemical features, and a weighting coefficient for topographic features. These are determined through preset or self-learning methods to adjust the proportional relationship of different data sources in the fusion feature calculation. Before weighted fusion, the remote sensing spectral data, geochemical data and topographic data are spatially registered. A unified geographic projection coordinate system is adopted, and all types of data are resampled to the same grid resolution so that the corresponding pixel positions correspond one-to-one in space. After weighting, the matrix is ​​superimposed to obtain the preliminary fusion feature matrix, and the obtained matrix is ​​used as the input for the nonlinear fusion coding step.

3. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 2, characterized in that: The generation of the fused feature vector includes inputting weighted and normalized spectral data, geochemical data, and topographic data into a fusion function, obtaining multi-source input feature data through linear superposition, transforming the multi-source input feature data using a nonlinear activation unit, wherein the nonlinear activation unit adopts a hyperbolic tangent function, and the fused signal is limited to a finite interval under the action of the activation function. The dynamic weights of input features are automatically adjusted based on the standard deviation changes of each type of input feature. When the standard deviation fluctuation of an input feature exceeds the standard deviation fluctuation threshold, the balance adjustment weight factor of the corresponding input feature automatically decreases. The fusion encoding output is a multi-dimensional feature vector, with each component corresponding to the response value after combining the input features. The parameters in the fusion process are updated through backpropagation during the training phase.

4. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 3, characterized in that: The self-attention mechanism for constructing geological structure perception includes, Within the geological prior attention model, a key matrix, query matrix, and value matrix are established to calculate the feature correlation between spatial units. A geological prior matrix is ​​generated through the joint calculation of topographic gradient and geochemical anomaly degree to characterize the spatial differences in geological structure. When performing attention calculation, the geological prior matrix and the attention similarity matrix are superimposed according to the geological coupling weight to form a comprehensive attention score; The attention weights are dynamically adjusted based on the relative deviation between local values ​​and global statistical features in the geological prior matrix.

5. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 4, characterized in that: The dynamic adjustment of local geological coupling weight parameters during attention calculation includes, When performing attention-weighted operations, the local difference between the input features and the geological prior matrix is ​​monitored, and the local difference is used as the basis for adaptive adjustment of the local geological coupling weight parameters. When the local difference value is greater than the set geological difference sensitivity threshold, the value of the local geological coupling weight parameter is increased within the corresponding spatial unit. During the training phase, the local geological coupling weight parameters are continuously updated to determine whether the local geological coupling weight parameters converge to a stable interval according to the optimization criteria during multiple iterations. Through a dynamic adjustment mechanism, the attention matrix in the geological prior attention model is adapted to the spatial distribution characteristics of geological priors in each round of calculation, forming a weighted matrix structure that is updated in real time with changes in geological conditions.

6. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 5, characterized in that: The establishment of the geological consistency loss mechanism includes, In the multimodal geological feature recognition model, the spatial difference between the attention response distribution output by the model and the alteration label is calculated, and a loss function is constructed based on the spatial difference. The loss function consists of three types of sub-terms, including a nonlinear amplification term, a logarithmic penalty term, and a smoothing boundary constraint term; The three types of sub-items are combined into a composite loss structure by weighting. During training, either fixed weight or adaptive weight mode can be used for calculation. The fixed weight is a preset constant, while the adaptive weight is dynamically adjusted according to the convergence rate of each sub-item loss. The multimodal geological feature recognition model updates parameters based on the convergence of the composite loss function, and the composite loss structure performs a complete backpropagation process in each iteration.

7. The method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in claim 6, characterized in that: The optimization process, which integrates the geological consistency loss mechanism with classification loss and regularization terms, includes... During the training phase of the multimodal geological feature recognition model, the geological consistency loss function, the sample classification loss function, and the regular smoothing function are proportionally superimposed to form a comprehensive optimization objective. In each training iteration, the weighted sum of the three sub-items is minimized simultaneously. The training control unit dynamically adjusts the weights of each item according to the rate of change of loss. When the rate of decrease of geological consistency loss is lower than that of classification loss, the geological consistency loss weight is automatically increased to balance the convergence speed. When the gradient oscillation exceeds the preset gradient oscillation threshold, the weight of the regularization term is automatically increased. The joint optimization mechanism performs gradient normalization in each iteration, and all parameters of the multimodal geological feature recognition model are updated synchronously within the unified optimization framework.

8. A modeling system for shallow low-temperature hydrothermal deposits based on shortwave infrared radiation, employing the modeling method for shallow low-temperature hydrothermal deposits based on shortwave infrared radiation as described in any one of claims 1 to 7, characterized in that: It includes a multi-source feature mapping module, a geological prior guidance module, and a geological coupling optimization module; The multi-source feature mapping module is used to perform weighted normalization processing on multi-source heterogeneous geological data. Through a nonlinear fusion coding algorithm with weighted coefficients, the multi-source heterogeneous geological data is mapped to a unified feature space to generate a fusion feature vector. The geological prior guidance module is used to construct a self-attention mechanism for geological structure perception. Based on the key, query and value matrix, a geological prior matrix reflecting the distribution of topographic gradient and geochemical anomaly is introduced, and the local geological coupling weight parameters are dynamically adjusted during the attention calculation process. The geological coupling optimization module is used to establish a geological consistency loss mechanism, which spatially constrains the attention response and alteration label, and constructs a composite loss structure through nonlinear amplification terms, logarithmic penalty terms, and smooth boundary constraint terms. The geological consistency loss mechanism, classification loss, and regularization terms are jointly optimized.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method for modeling shallow low-temperature hydrothermal deposits based on shortwave infrared as described in any one of claims 1 to 7.

Images