Deep learning geological map intelligent prediction method based on multi-modal data fusion

Abstract

The invention is suitable for the technical field of multi-modal data fusion, and provides a multi-modal data fusion deep learning geological map intelligent prediction method, which comprises the following steps: acquiring multi-modal geological data, constructing a heterogeneous multi-modal feature extraction network, and respectively extracting high-dimensional feature representation of each modal data by adopting a special encoder; designing a multi-modal dynamic fusion module based on a multi-head attention mechanism; establishing a deep learning architecture of a geological knowledge guiding mechanism, and embedding geological rules and priori knowledge into a model training process through a constraint loss function; implementing a multi-task collaborative learning framework, and synchronously completing lithology classification, structure identification and mineralization potential prediction tasks; outputting reliability evaluation of a prediction result by adopting an uncertainty quantification system of a Bayesian deep learning framework; deploying an intelligent active learning system; according to the multi-modal dynamic fusion method based on the multi-head attention mechanism, the problem of depth feature interaction and complementation among heterogeneous geological data is effectively solved.

Description

Technical Field

[0001] This invention relates to the field of multimodal data fusion technology, and more specifically, to a deep learning-based intelligent prediction method for geological maps based on multimodal data fusion. Background Technology

[0002] Geological maps, as fundamental maps representing geological structures, lithological distribution, and mineral resources, are of great value for resource exploration, engineering construction, and social development. Traditional geological mapping methods mainly rely on manual field surveys and manual map compilation, which have inherent drawbacks such as low efficiency, high cost, long cycle time, and strong subjectivity. With the development of Earth observation technology, the ability to acquire multi-source geological data has been significantly improved, and multimodal geological data, including multispectral remote sensing, hyperspectral remote sensing, synthetic aperture radar, gravity, magnetics, geochemistry, and borehole data, are becoming increasingly abundant.

[0003] The development of geological map prediction technology has gone through three main stages. The first stage was based on manual interpretation and relied entirely on expert knowledge. The second stage adopted traditional machine learning methods, achieving semi-automatic prediction, but its ability to fuse and process multimodal data was limited. The third stage began to introduce deep learning technology, using algorithms such as convolutional neural networks to improve feature extraction capabilities, but it still has significant shortcomings in the collaborative utilization of multimodal data. While the existing Chinese patent, "A Multi-Scale Sampling and Input Method for a Deep Learning Recognition Model of Rock Strata" (application number 202010029405.3), solves the multi-scale feature extraction problem, its research on the adaptive fusion mechanism of multimodal data is not in-depth enough.

[0004] Although deep learning technology has made some progress in the field of geosciences, there are still some shortcomings in the intelligent prediction of multimodal geological data. The multimodal fusion strategy is inefficient, and an effective cross-modal feature interaction mechanism has not been established. In addition, the model has weak generalization ability and is difficult to adapt to different geological background regions. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies: although deep learning technology has made some progress in the field of geosciences, it still suffers from low efficiency of multimodal fusion strategies and failure to establish an effective cross-modal feature interaction mechanism in intelligent prediction of multimodal geological data; and weak model generalization ability, making it difficult to adapt to different geological background regions. Therefore, this invention proposes a deep learning-based intelligent prediction method for geological maps based on multimodal data fusion.

[0006] The specific technical solution is: a deep learning-based intelligent prediction method for geological maps based on multimodal data fusion, including the following steps:

[0007] Step 1: Acquire multimodal geological data and perform spatial registration, standardization, and missing value processing on the multimodal geological data; the multimodal geological data includes hyperspectral and multispectral remote sensing data, gravity, magnetic, electrical, seismic and geophysical data, trace and rare earth element geochemical data, digital terrain and geomorphological data, and geological knowledge map data;

[0008] Step 2: Construct a heterogeneous multimodal feature extraction network and use a dedicated encoder to extract high-dimensional feature representations of each modality of data. The heterogeneous multimodal feature extraction network uses a visual Transformer architecture to extract global semantic features of remote sensing images, a three-dimensional convolutional neural network to process geophysical field data, a graph neural network to process geochemical sampling point data, and knowledge graph embedding technology to process structured geological knowledge.

[0009] Step 3: Design a multimodal dynamic fusion module based on a multi-head attention mechanism to achieve deep interaction and adaptive fusion of cross-modal features;

[0010] Step 4: Establish a deep learning architecture for a geological knowledge-guided mechanism that includes stratigraphic sequence constraints, lithofacies assemblage constraints, tectonic ore-controlling constraints, and multi-source anomaly synergistic constraints. Geological rules and prior knowledge are embedded into the model training process through constraint loss functions. Stratigraphic sequence constraints are used to ensure that stratigraphic contact relationships conform to geological laws; lithofacies assemblage constraints are used to ensure the rationality of spatial combinations of specific lithofacies; tectonic ore-controlling constraints are used to ensure that prediction results conform to regional metallogenic laws; and multi-source anomaly synergistic constraints are used to ensure the consistency of geophysical-geochemical-remote sensing anomaly responses.

[0011] Step 5: Implement a multi-task collaborative learning framework to simultaneously complete lithological classification, structural identification, and mineralization potential prediction tasks;

[0012] Step 6: Using a Bayesian deep learning framework, the uncertainty quantification system outputs a reliability assessment of the prediction results.

[0013] Step 7: Deploy an intelligent active learning system to dynamically guide field validation and model iteration optimization based on prediction uncertainties.

[0014] A further technical solution of the present invention, in step one, includes multimodal geological data comprising:

[0015] Hyperspectral and multispectral remote sensing data are used to extract surface lithological spectral characteristics and alteration information;

[0016] Gravity, magnetoelectric, seismic, and geophysical data are used to infer deep geological structures and rock mass distribution;

[0017] Trace and rare earth element geochemical data are used to analyze elemental distribution and mineralization anomalies.

[0018] Digital terrain and geomorphological data are used to extract structural and topographic features;

[0019] Geological knowledge map data, including stratigraphic sequences, tectonic frameworks, and mineralization patterns.

[0020] A further technical solution of the present invention is that, in step two, the dedicated encoder is a visual encoder based on SwinTransformer; for remote sensing image data, a visual encoder based on SwinTransformer is used to capture multi-scale spatial-spectral features through a hierarchical shift window mechanism; the dedicated encoder divides the hyperspectral image into non-overlapping patches, projects them into a high-dimensional space through a linear embedding layer, and performs self-attention calculations in multiple Transformer layers; the dedicated encoder also integrates a spectral attention module to adaptively weight the importance of different spectral bands to effectively highlight feature bands related to lithology and alteration.

[0021] In a further technical solution of the present invention, in step three, the multimodal dynamic fusion module includes:

[0022] A cross-modal attention mechanism is used to calculate the correlation weights between features from different modalities.

[0023] Gated feature fusion unit dynamically adjusts the contribution of each modal feature;

[0024] A multi-scale feature alignment module addresses spatial resolution differences in multi-source data.

[0025] The characteristic distillation mechanism enables the extraction and transfer of important features.

[0026] A further technical solution of the present invention, in step five, implementing the multi-task collaborative learning framework includes:

[0027] A shared feature encoding layer is used to extract common feature representations from multimodal data;

[0028] Task-specific decoders are designed to be optimized specifically for different prediction tasks;

[0029] A knowledge transfer mechanism is used to enable knowledge sharing between different tasks;

[0030] Dynamic weight adjusters are used to balance the training process of multi-task learning.

[0031] In a further technical solution of the present invention, in step five, the multi-task collaborative learning framework is a shared feature encoding layer, which extracts common feature representations with rich geological significance through a multimodal fusion module; the shared feature encoding layer contains comprehensive information related to lithology, structure and mineralization, providing basic feature support for upper-level tasks; above the shared layer, a dedicated decoder branch is designed for each specific task.

[0032] In a further technical solution of the present invention, in step six, the uncertainty quantification system employs:

[0033] The Monte Carlo Dropout method calculates the prediction variance through multiple samplings;

[0034] Deep ensemble learning strategies fuse the prediction results of multiple independent models;

[0035] A Bayesian neural network framework provides probabilistic prediction outputs;

[0036] An uncertainty visualization tool that generates a spatial distribution map of prediction confidence.

[0037] In a further technical solution of the present invention, in step seven, the intelligent active learning system includes:

[0038] Based on an uncertainty-based sampling strategy, the sample point with the most information is automatically identified.

[0039] An optimal verification point planning algorithm is used to generate the optimal path for field verification.

[0040] An online incremental learning mechanism integrates newly collected validation data in real time;

[0041] An adaptive model optimizer continuously improves prediction accuracy and reliability.

[0042] Compared with existing technologies, the deep learning-based intelligent prediction method for geological maps based on multimodal data fusion of this invention can achieve the following:

[0043] 1. In designing a multimodal dynamic fusion module based on a multi-head attention mechanism to achieve deep interaction and adaptive fusion of cross-modal features, the multimodal dynamic fusion method based on the multi-head attention mechanism effectively solves the problem of deep feature interaction and complementarity between heterogeneous geological data; a deep learning architecture with a geological knowledge guidance mechanism is designed, embedding a prior geological knowledge system into the model training process to ensure the geological rationality of the prediction results; a multi-task collaborative learning architecture is established to realize collaborative prediction and knowledge sharing of multiple geological elements; an uncertainty quantification system based on Bayesian deep learning is developed to provide reliability assessment for prediction results; and an intelligent active learning loop system is constructed to realize autonomous optimization and continuous improvement of model performance.

[0044] 2. The output of the multimodal dynamic fusion module is a multimodal fusion feature that has undergone deep interaction and adaptive weighting. It not only retains the unique information of each modality, but also captures the complementary relationship between modalities, providing rich and accurate feature representations for subsequent geological prediction tasks. This design significantly improves the model's ability to understand complex geological phenomena, especially showing good robustness when dealing with inconsistencies in multi-source data or missing data.

[0045] 3. The multi-task collaborative learning framework not only improves the prediction accuracy of each individual task but also ensures the consistency of prediction results in a geological sense. For example, predicted mineralization prospective areas often coincide with specific lithological combinations and tectonic locations. This consistency greatly enhances the geological rationality and practical value of the prediction results, providing more reliable technical support for mineral exploration and geological research. Attached Figure Description

[0046] Figure 1 This is a flowchart of the deep learning-based intelligent prediction method for geological maps based on multimodal data fusion, as described in this invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0048] In embodiments of the present invention, such as Figure 1 As shown: A deep learning-based intelligent prediction method for geological maps based on multimodal data fusion includes the following steps:

[0049] Step 1: Acquire multimodal geological data and perform spatial registration, standardization, and missing value processing on the multimodal geological data; the multimodal geological data includes hyperspectral and multispectral remote sensing data, gravity, magnetic, electrical, seismic and geophysical data, trace and rare earth element geochemical data, digital terrain and geomorphological data, and geological knowledge map data;

[0050] Step 2: Construct a heterogeneous multimodal feature extraction network and use a dedicated encoder to extract high-dimensional feature representations of each modality of data. The heterogeneous multimodal feature extraction network uses a visual Transformer architecture to extract global semantic features of remote sensing images, a three-dimensional convolutional neural network to process geophysical field data, a graph neural network to process geochemical sampling point data, and knowledge graph embedding technology to process structured geological knowledge.

[0051] Step 3: Design a multimodal dynamic fusion module based on a multi-head attention mechanism to achieve deep interaction and adaptive fusion of cross-modal features;

[0052] Step 4: Establish a deep learning architecture for a geological knowledge-guided mechanism that includes stratigraphic sequence constraints, lithofacies assemblage constraints, tectonic ore-controlling constraints, and multi-source anomaly synergistic constraints. Geological rules and prior knowledge are embedded into the model training process through constraint loss functions. Stratigraphic sequence constraints are used to ensure that stratigraphic contact relationships conform to geological laws; lithofacies assemblage constraints are used to ensure the rationality of spatial combinations of specific lithofacies; tectonic ore-controlling constraints are used to ensure that prediction results conform to regional metallogenic laws; and multi-source anomaly synergistic constraints are used to ensure the consistency of geophysical-geochemical-remote sensing anomaly responses.

[0053] Step 5: Implement a multi-task collaborative learning framework to simultaneously complete lithological classification, structural identification, and mineralization potential prediction tasks;

[0054] Step 6: Using a Bayesian deep learning framework, the uncertainty quantification system outputs a reliability assessment of the prediction results.

[0055] Step 7: Deploy an intelligent active learning system to dynamically guide field validation and model iteration optimization based on prediction uncertainties.

[0056] In designing a multimodal dynamic fusion module based on a multi-head attention mechanism to achieve deep interaction and adaptive fusion of cross-modal features, the multimodal dynamic fusion method based on the multi-head attention mechanism effectively solves the problem of deep feature interaction and complementarity between heterogeneous geological data. A deep learning architecture with a geological knowledge guidance mechanism is designed to embed a prior geological knowledge system into the model training process to ensure the geological rationality of the prediction results. A multi-task collaborative learning architecture is established to realize collaborative prediction and knowledge sharing of multiple geological elements. An uncertainty quantification system based on Bayesian deep learning is developed to provide reliability assessment for the prediction results. An intelligent active learning loop system is constructed to realize autonomous optimization and continuous improvement of model performance.

[0057] In another embodiment of the present invention, in step one, the multimodal geological data includes:

[0058] Hyperspectral and multispectral remote sensing data are used to extract surface lithological spectral characteristics and alteration information;

[0059] Gravity, magnetoelectric, seismic, and geophysical data are used to infer deep geological structures and rock mass distribution;

[0060] Trace and rare earth element geochemical data are used to analyze elemental distribution and mineralization anomalies.

[0061] Digital terrain and geomorphological data are used to extract structural and topographic features;

[0062] Geological knowledge map data, including stratigraphic sequences, tectonic frameworks, and mineralization patterns.

[0063] In another embodiment of the present invention, in step two, the dedicated encoder is a visual encoder based on the Swin Transformer. For remote sensing image data, a visual encoder based on the Swin Transformer is used to capture multi-scale spatial-spectral features through a hierarchical shift window mechanism. The dedicated encoder divides the hyperspectral image into non-overlapping patches, projects them into a high-dimensional space through a linear embedding layer, and performs self-attention calculations in multiple Transformer layers. The dedicated encoder also integrates a spectral attention module to adaptively weight the importance of different spectral bands to effectively highlight feature bands related to lithology and alteration.

[0064] For geophysical field data, a 3D convolutional neural network encoder is used to process volumetric data such as gravity, magnetism, electricity, and seismicity. This encoder employs anisotropic convolutional kernels, using a large receptive field in the horizontal direction to capture regional field features, and fine convolutions in the vertical direction to characterize the vertical variations of field values. Residual connections and dense connections are also incorporated into the network to ensure smooth gradient propagation, facilitating the training of deep networks. Furthermore, the encoder includes a multi-scale feature pyramid, enabling the simultaneous extraction of local anomaly features and regional field background features.

[0065] For geochemical sampling data, a graph neural network-based encoder is used to process irregularly distributed sampling points. First, a spatial adjacency graph of the sampling points is constructed, where nodes contain elemental abundance features, and edge weights are determined by both spatial distance and geological background. Then, neighborhood information is iteratively aggregated through graph convolutional layers while preserving the spatial distribution characteristics of the sampling points. The encoder also includes attention pooling layers to adaptively highlight the importance of anomalous sampling points.

[0066] In processing prior geological knowledge, a knowledge graph embedding encoder is used to transform structured geological knowledge into numerical vectors. Based on the TransR model, this encoder maps geological entities and relationships to different vector spaces, capturing complex geological patterns through vector operations. Specifically, the encoder also integrates a rule constraint module to ensure that the embedded representations satisfy geological logical constraints.

[0067] The output features of all dedicated encoders undergo layer normalization and feature standardization before being projected onto a unified feature space, preparing for subsequent multimodal fusion. This heterogeneous encoder design fully considers the characteristics of each modality's data, providing a high-quality feature representation foundation for subsequent cross-modal fusion.

[0068] In another embodiment of the present invention, in step three, the multimodal dynamic fusion module includes:

[0069] A cross-modal attention mechanism is used to calculate the correlation weights between features from different modalities.

[0070] Gated feature fusion unit dynamically adjusts the contribution of each modal feature;

[0071] A multi-scale feature alignment module addresses spatial resolution differences in multi-source data.

[0072] The characteristic distillation mechanism enables the extraction and transfer of important features.

[0073] In another embodiment of the present invention, in step three, the feature representations extracted by each modality encoder are projected onto a unified feature space, and a multi-head attention mechanism is used to compute the intermodal correlations in different representation subspaces in parallel. For each attention head, the module generates a set of query vectors, key vectors from the target modality, and value vectors from the source modality. By calculating the similarity scores between the query and the key, an attention weight matrix is ​​generated. This design allows the model to simultaneously focus on multiple dependencies between different modality features, such as the correlation between remote sensing spectral features and geochemical element content, or the correspondence between geophysical field features and geological structural models.

[0074] The multimodal dynamic fusion module also introduces a gated fusion mechanism, which dynamically adjusts the contribution of each modal feature through learnable gate weights. The gated fusion mechanism automatically adjusts the proportion of each feature in the fused features based on the importance and reliability of the current feature, thereby enhancing its influence when important features are prominent and suppressing its contribution when feature noise is high. This mechanism is particularly suitable for dealing with common noise and incompleteness problems in geological data.

[0075] The multimodal dynamic fusion module also includes a feature refinement layer, which maintains training stability through residual connections and layer normalization techniques, and uses a feedforward neural network to enhance the expressive power of features. The multimodal dynamic fusion module adopts cross-modal attention distillation technology to transfer important features from information-rich modalities to information-scarce modalities, thereby improving the overall feature quality.

[0076] The output of the multimodal dynamic fusion module is multimodal fusion features that have undergone deep interaction and adaptive weighting. This not only preserves the unique information of each modality but also captures the complementary relationships between modalities, providing rich and accurate feature representations for subsequent geological prediction tasks. This design significantly improves the model's ability to understand complex geological phenomena, especially demonstrating good robustness when dealing with inconsistencies in multi-source data or missing data.

[0077] In another embodiment of the present invention, in step five, implementing the multi-task collaborative learning framework includes:

[0078] A shared feature encoding layer is used to extract common feature representations from multimodal data;

[0079] Task-specific decoders are designed to be optimized specifically for different prediction tasks;

[0080] A knowledge transfer mechanism is used to enable knowledge sharing between different tasks;

[0081] Dynamic weight adjusters are used to balance the training process of multi-task learning.

[0082] In another embodiment of the present invention, in step five, the multi-task collaborative learning framework is a shared feature encoding layer, which extracts common feature representations with rich geological significance through a multimodal fusion module. The shared feature encoding layer contains comprehensive information related to lithology, structure, and mineralization, providing basic feature support for upper-level tasks. Above the shared layer, a dedicated decoder branch is designed for each specific task: for the lithology classification task, a spatial pyramid pooling decoder based on an attention mechanism is used, which can capture lithology distribution features at different scales; for the structure recognition task, an edge-aware decoder is used, which focuses on highlighting the boundary features of linear and ring structures; for the mineralization potential prediction task, a multi-scale feature fusion decoder is designed, which comprehensively considers ore-controlling factors and mineralization display features.

[0083] The multi-task collaborative learning framework introduces a dynamic task weight adjustment mechanism. By analyzing the learning difficulty and training status of each task, it adaptively adjusts the weight ratio of each item in the loss function. In the early stages of training, more emphasis is placed on the basic lithological classification task; as training progresses, the weights of structural identification and mineralization prediction tasks are gradually increased. This dynamic adjustment strategy ensures the collaborative optimization of multiple tasks and avoids the problem of certain tasks dominating the training process.

[0084] The multi-task collaborative learning framework also employs knowledge distillation techniques, enabling different tasks to learn from and promote each other. For example, the mineralization prediction task can learn fracture network features from the structural identification task, while the lithological classification task can draw on the altered rock identification experience from the mineralization prediction task. This cross-task knowledge transfer significantly improves the performance of each task, especially in scenarios with limited training samples.

[0085] The multi-task collaborative learning framework not only improves the prediction accuracy of each individual task but also ensures the consistency of the prediction results in a geological sense. For example, the predicted mineralization prospective areas often coincide with specific lithological combinations and tectonic locations. This consistency greatly enhances the geological rationality and practical value of the prediction results, providing more reliable technical support for mineral exploration and geological research.

[0086] In another embodiment of the present invention, in step six, the uncertainty quantification system employs:

[0087] The Monte Carlo Dropout method calculates the prediction variance through multiple samplings;

[0088] Deep ensemble learning strategies fuse the prediction results of multiple independent models;

[0089] A Bayesian neural network framework provides probabilistic prediction outputs;

[0090] An uncertainty visualization tool that generates a spatial distribution map of prediction confidence.

[0091] In another embodiment of the present invention, in step seven, the intelligent active learning system includes:

[0092] Based on an uncertainty-based sampling strategy, the sample point with the most information is automatically identified.

[0093] An optimal verification point planning algorithm is used to generate the optimal path for field verification.

[0094] An online incremental learning mechanism integrates newly collected validation data in real time;

[0095] An adaptive model optimizer continuously improves prediction accuracy and reliability.

[0096] In another embodiment of the present invention, in step seven, the intelligent active learning system includes an uncertainty quantification module. The uncertainty quantification module uses a combination of Monte Carlo Dropout and deep ensemble learning to generate a confidence assessment for each prediction point. For lithological classification tasks, it calculates the entropy value of the predicted probability of the category. For mineralization prediction tasks, it evaluates the variance and confidence interval of the predicted value. These uncertainty indicators together constitute a "knowledge gap" distribution map, clearly identifying the areas where the model most needs to be validated.

[0097] The intelligent active learning system employs a multi-objective optimization algorithm to design a field validation scheme. The field validation scheme comprehensively considers the area with the highest uncertainty, which is the area that provides the greatest benefit for model improvement. The field validation scheme also comprehensively considers accessibility and safety, selecting areas with convenient transportation and good working conditions. The field validation scheme also includes the cost-effectiveness of the validation work, aiming to solve the most uncertainties with the fewest validation points possible. The system will generate an optimal validation route map and sampling point layout scheme, and provide them to the field geological work team.

[0098] New data obtained from field validation is immediately incorporated into the online learning process. The system employs an incremental learning algorithm to rapidly integrate new validation data without disrupting the existing model knowledge. In particular, the system distinguishes between "confirmatory" samples (validation results consistent with predictions) and "corrective" samples (validation results inconsistent with predictions), assigning higher learning weights to the latter to accelerate the correction of model error patterns.

[0099] After each model update, the system reassesses the predictive uncertainty across the entire study area and plans the focus of the next round of validation. This iterative process continues until the predictive uncertainty is reduced to a predetermined threshold, or budget and time constraints are met. The system also includes a performance monitoring module that tracks and records the model's performance improvement on various tasks after each iteration, providing a quantitative basis for decision-making.

[0100] Through this intelligent active learning mechanism, the system significantly improves the efficiency and focus of geological surveys, concentrates limited fieldwork resources on the most critical areas, and achieves a rapid improvement in model prediction accuracy and geological understanding, providing strong technical support for mineral resource evaluation and geological scientific research.

[0101] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning geological map intelligent prediction method of multi-modal data fusion, characterized in that, The method comprises the following steps: Step one, obtaining multi-modal geological data and performing spatial registration, standardization and missing value processing on the multi-modal geological data; the multi-modal geological data comprises hyperspectral and multispectral remote sensing data, gravity-magnetic-electric-seismic geophysical data, trace and rare earth element geochemical data, digital terrain and geomorphology data, and geological knowledge graph data; Step two, constructing a heterogeneous multi-modal feature extraction network, and extracting high-dimensional feature representations of each modal data by using a special encoder; wherein the heterogeneous multi-modal feature extraction network adopts a visual Transformer architecture to extract global semantic features of remote sensing images, adopts a three-dimensional convolutional neural network to process geophysical field data, adopts a graph neural network to process geochemical sampling point data, and adopts a knowledge graph embedding technology to process structured geological knowledge; Step three, designing a multi-modal dynamic fusion module based on a multi-head attention mechanism to realize deep interaction and adaptive fusion of cross-modal features; Step four, establishing a deep learning architecture of a geological knowledge guiding mechanism including stratum sequence constraints, lithofacies combination constraints, structure ore-controlling constraints and multi-source anomaly collaborative constraints, and embedding geological rules and prior knowledge into the model training process through a constraint loss function; the stratum sequence constraints are used to ensure that the stratum contact relationship conforms to the geological law; the lithofacies combination constraints are used to ensure the rationality of the spatial combination of specific lithofacies; the structure ore-controlling constraints are used to make the prediction results conform to the regional metallogenic regularity; and the multi-source anomaly collaborative constraints are used to ensure the consistency of geophysical-geochemical-remote sensing anomaly responses; Step five, implementing a multi-task collaborative learning framework to simultaneously complete the tasks of lithology classification, structure identification and metallogenic potential prediction; Step six, using an uncertainty quantification system of a Bayesian deep learning framework to output reliability evaluation of the prediction results; Step seven, deploying an intelligent active learning system to dynamically guide field verification and model iterative optimization according to the prediction uncertainty.

2. The deep learning geologic map intelligent prediction method of multi-modal data fusion according to claim 1, characterized in that, In the step one, the multi-modal geological data comprises: Hyperspectral and multispectral remote sensing data for extracting surface lithology spectral features and alteration information; Gravity-magnetic-electric-seismic geophysical data for inferring deep geological structure and rock mass distribution; Trace and rare earth element geochemical data for analyzing element distribution and mineralization anomaly regularity; Digital terrain and geomorphology data for extracting structure and terrain features; Geological knowledge graph data including stratum sequence, structure framework and metallogenic regularity knowledge.

3. The deep learning geologic map intelligent prediction method of multi-modal data fusion according to claim 1, characterized in that, In the step two, the special encoder is a Swin Transformer-based visual encoder; for remote sensing image data, the Swin Transformer-based visual encoder is used to capture multi-scale spatial-spectral features through a hierarchical shift window mechanism; The special encoder divides the hyperspectral image into non-overlapping patches, projects them to a high-dimensional space through a linear embedding layer, and performs self-attention calculation in multiple Transformer layers; the special encoder also integrates a spectral attention module to adaptively weight the importance of different spectral bands, which is used to effectively highlight the feature bands related to lithology and alteration.

4. The deep learning geologic map intelligent prediction method of multi-modal data fusion according to claim 1 or 2 or 3, characterized in that, In the step three, the multi-modal dynamic fusion module comprises: Cross-modal attention mechanism to calculate the correlation weight between different modalities; Gating feature fusion unit to dynamically adjust the contribution of each modality feature; Multi-scale feature alignment module to solve the spatial resolution difference of multi-source data; Feature distillation mechanism to extract and transfer important features.

5. The deep learning geological map intelligent prediction method of multi-modal data fusion according to claim 1 or 2 or 3, characterized in that, In the step three, the features extracted by each modality encoder are projected into a unified feature space, and a multi-head attention mechanism is used to calculate the inter-modal correlation in different representation subspaces in parallel; For each attention head, the module generates a set of query vectors from the target modality, key vectors, and value vectors from the source modality, and generates an attention weight matrix by calculating the similarity score between the query and the key; The multi-modal dynamic fusion module also introduces a gating fusion mechanism, which dynamically adjusts the contribution of each modality feature through a learnable gating weight; The gating fusion mechanism automatically adjusts the proportion of the feature in the fused feature according to the importance and reliability of the current feature, so as to enhance the influence of important features and suppress the contribution of features with large noise; The multi-modal dynamic fusion module also contains a feature refining layer, which maintains training stability through residual connection and layer normalization technology, and enhances the expression ability of features using feedforward neural networks; The multi-modal dynamic fusion module uses cross-modal attention distillation technology to transfer important features from the modality with rich information to the modality with relatively insufficient information, improving the overall feature quality. 6.The deep learning-based intelligent geological map prediction method using multi-modal data fusion according to claim 1, wherein, In the step five, the multi-task collaborative learning framework includes: Shared feature encoding layer for extracting common feature representation of multi-modal data; Task-specific decoder for specialized optimization for different prediction tasks; Knowledge transfer mechanism for knowledge sharing between different tasks; Dynamic weight regulator for balancing the training process of multi-task learning.

7. The deep learning geological map intelligent prediction method of multi-modal data fusion according to claim 1, characterized in that, In the step five, the multi-task collaborative learning framework is a shared feature encoding layer that extracts common feature representation with rich geological significance through a multi-modal fusion module; The shared feature encoding layer contains comprehensive information related to lithology, structure and mineralization, providing basic feature support for upper-layer tasks; Above the shared layer, a dedicated decoder branch is designed for each specific task: The multi-task collaborative learning framework introduces a dynamic task weight adjustment mechanism, which adaptively adjusts the weight proportion of each item in the loss function by analyzing the learning difficulty and training state of each task; The multi-task collaborative learning framework also uses knowledge distillation technology to enable different tasks to learn from and promote each other. 8.The deep learning geologic map intelligent prediction method of multi-modal data fusion according to claim 1, wherein, In the step six, the uncertainty quantification system uses: Monte Carlo Dropout method to calculate the prediction variance by sampling multiple times; Deep ensemble learning strategy to fuse the prediction results of multiple independent models; Bayesian neural network framework to provide probabilistic prediction output; Uncertainty visualization tools to generate prediction confidence spatial distribution maps. 9.The deep learning-based intelligent geological map prediction method using multi-modal data fusion according to claim 1, wherein, In the step seven, the intelligent active learning system includes: Based on the uncertainty of the sampling strategy, the sample points with the largest amount of information are automatically identified; Optimal verification point planning algorithm to generate the optimal path for field verification; Online incremental learning mechanism to integrate new collected validation data in real time; Adaptive model optimizer to continuously improve the prediction accuracy and reliability.

10. The deep learning geological map intelligent prediction method of multi-modal data fusion according to claim 1, characterized in that, In the step seven, the intelligent active learning system includes an uncertainty quantification module, which uses a combination of Monte Carlo Dropout and deep ensemble learning to generate confidence evaluation for each prediction point; for the lithology classification task, the entropy value of the class prediction probability is calculated; for the ore prediction task, the variance and confidence interval of the prediction value are evaluated; The intelligent active learning system uses a multi-objective optimization algorithm to design the field verification scheme; The field verification scheme takes into account the areas with the highest uncertainty, which have the greatest benefit for model improvement; the field verification scheme also considers accessibility and safety, and selects areas with convenient transportation and better working conditions; the field verification scheme also includes the cost-effectiveness of the verification work; The new data obtained from the field verification immediately enters the online learning process.

Images