Mineral green prediction method and device based on cross-domain self-adaption and multi-modal fusion

Abstract

The invention relates to a mineral green prediction method and device based on cross-domain self-adaption and multi-modal fusion, and belongs to the technical field of mineral resource exploration and remote sensing. The method comprises the following steps: acquiring multi-source data, inputting the multi-source data into a transfer learning module to learn transferable model parameters, fusing the acquired data through a modal gating network, then constructing a line map through a map construction model, and enhancing the line map by using a map convolutional network; the enhanced multi-modal data is input into a Transformer encoder of an embedded graph convolutional network, a mining area space topological structure and spectral features are extracted, and an ore forming body potential probability graph is obtained through prediction of a prediction head; and an environmental risk assessment model is constructed, ore body prediction accuracy maximization and environmental risk minimization are realized through multi-objective optimization, and the accuracy of ore body prediction is verified through drilling activities. The method not only improves the regional generalization ability of the model, but also realizes the interpretability of green mineral exploration and results, and improves the precision, efficiency and sustainability of mineral exploration.

Description

TECHNICAL FIELD

[0001] The present application relates to a mineral green prediction method and device based on cross-domain adaptation and multi-modal fusion, belonging to the field of mineral resource exploration and remote sensing technology. BACKGROUND

[0002] Traditional mineral resource exploration relies on the experience of geologists and manual data analysis, which is limited in accuracy and high in time cost. With the rapid development of remote sensing technology and artificial intelligence, the prospecting method based on multi-source remote sensing data has gradually become an important means in the field of mineral exploration. In recent years, the prospecting method based on remote sensing technology has achieved certain results in the positioning and prediction of mineral resources by using satellite remote sensing data, unmanned aerial remote sensing data, geochemical data and other multi-source information. However, the existing technology still faces the following key problems: (1) Strong regional dependence, lack of cross-domain adaptation ability: Most of the current mineral exploration models rely on samples trained in a specific region, which leads to the lack of cross-regional generalization ability of the model. Due to the significant differences in geological structure, topographic features, remote sensing data acquisition conditions and other factors in different mining areas, these models often cannot maintain high prediction accuracy when migrated to new mining areas. Therefore, the existing method can only be used in a specific region, limiting its promotion and application in other regions; (2) Lack of model interpretability, difficult to provide intuitive basis for exploration decision-making: With the introduction of deep learning methods, the accuracy of remote sensing prospecting has been significantly improved, but most deep learning models still have the "black box" problem. Even if the model shows high accuracy in the training process, its decision-making process still lacks sufficient transparency and cannot explain why a certain area is determined as a potential mining area. This lack of interpretability makes it difficult for geologists to trust the results of the model, limiting its widespread application in actual exploration; (3) Ignoring environmental factors, difficult to achieve green prospecting: Existing ore body prediction methods generally focus on ore body distribution and alteration information extraction, and almost no sufficient consideration is given to environmental risks in the process of mineral exploitation, such as water pollution, soil degradation, vegetation destruction, etc. This one-sided prediction method leads to the fact that while obtaining high-precision ore body prediction, mineral exploration activities may cause serious impact on the ecological environment, which cannot meet the current requirements of green mining and sustainable development. SUMMARY

[0003] The purpose of the present application is to overcome the above-mentioned deficiencies and provide a mineral green prediction method based on cross-domain adaptation and multi-modal fusion, which not only improves the regional generalization ability of the model, but also realizes the green mineral exploration and the interpretability of the model results, greatly improving the precision, efficiency and sustainability of mineral exploration.

[0004] The technical scheme adopted by the present application is: The mineral green prediction method based on cross-domain adaptation and multi-modal fusion comprises the following steps: S1. Obtain multi-source satellite remote sensing data, unmanned aerial vehicle low-altitude high-resolution data and geochemical sample data, and preprocess the data; S2. Input the obtained data including source domain data and target domain data into a transfer learning module through adversarial training, learn transferable model parameters, and initialize the transferable model parameters in the Transformer encoder of the target region; S3. Input the obtained data into a modal gating network to calculate the weight of each modality, and obtain multi-modal fusion data through weighted fusion; S4. Input the multi-modal fusion data into a graph construction model to construct line mapping, and enhance spatial information using a graph convolution network; S5. Input the enhanced multi-modal data into the Transformer encoder of the embedded graph convolution network, extract the spatial topological structure and spectral features of the mining area, introduce a cross-modal consistency loss function to ensure the logical unity of different modalities in spatial and spectral features, input the extracted features into a prediction head to obtain a potential probability map of ore bodies; S6. Calculate the accuracy of ore body prediction through the prediction accuracy of the potential area of ore bodies, construct an environmental risk assessment model based on various environmental indicators, establish two objective functions of ore body prediction accuracy and environmental risk, and realize the calculation of "maximizing the accuracy of ore body prediction" and "minimizing the environmental risk" through a multi-objective optimization algorithm; S7. Verify the accuracy of ore body prediction through drilling activities, calculate the Q function by introducing deep Q learning through drilling feedback data, and adjust the neural network weight of the ore body prediction model using the Q function after each drilling feedback to adjust and optimize itself, and obtain the final prediction model.

[0005] In the above method, the transfer learning module in step S2 includes a feature extractor, the feature extractor is connected with a main task predictor and a domain classifier designed in parallel through a gradient reversal layer (GRL), the feature extractor extracts preliminary features from source domain data, the domain classifier forces the feature extractor to generate shared features effective for both the source domain and the target domain through adversarial training, thereby enhancing the cross-domain generalization ability of the model, and the main task predictor uses the shared features to perform ore body potential prediction; in this way, the transfer learning module learns transferable model parameters, including the convolutional layer weights of the feature extractor, the adversarial training parameters of the domain classifier, and the weights of the main task classification head, which are then transferred to the Transformer encoder in the target area for initialization.

[0006] The modal gating network in step S3 has the following specific calculation process: Let the input modal set be where x i represents the i-th modal, M is the number of modes, Calculate the weight for each modal through the modal gating network (Gating Network): , where, is an environmental factor vector, is a gating function, the input is the modal feature x i and the environmental condition , and the output is the importance score of the modal, The final multi-modal fusion result is: .

[0007] In step S4, the graph construction model first extracts features from multi-source data, uses these features to construct node features of the graph, and then defines nodes and edges of the graph based on geological structure information of the mining area. Each sampling point or spatial unit of the mining area will be a node, and the geological relationship of the mining area will construct edges in the graph through an adjacency matrix to form a line graph.

[0008] In the above step S5, the Grad-CAM explainability module is introduced to generate a saliency map and serve as a regularization constraint to guide model learning for model training. In the training process, Grad-CAM is used to generate a saliency map S(x), which is used as a regularization term: , where S(x) is the saliency map, representing the key area that the model focuses on; F(x) represents the prediction output of the model, which is a probability value, and y is the true label.

[0009] In step S5, the Transformer encoder, embedded with a graph convolutional network (GCN), processes the multimodal data and spatial topology of the mining area by combining the GCN and the Transformer self-attention mechanism. First, the GCN constructs a graph structure based on the geological structure information of the mining area (such as veins and fault zones), representing each data point as a node in the graph and aggregating spatial features. Then, the Transformer encoder receives the node features processed by the GCN and uses the self-attention mechanism to capture global dependencies and long-range information, enhancing the connections between multimodal data. In this way, the model can integrate information from different modalities in both spatial and spectral features, ultimately generating a ore body potential prediction map or a mineralization probability map of the mining area. Simultaneously, a cross-modal consistency loss function is introduced to ensure logical consistency between different modalities in both spatial and spectral dimensions.

[0010] The calculation of the cross-modal consistency loss function assumes that the prediction outputs of different modes at the same location are: in Indicates the first Prediction function corresponding to each mode The cross-modal consistency loss function is defined as: , in The number of modalities represents different types of data sources or different input modalities, with i and j being the corresponding modalities.

[0011] In step S6, the classification accuracy is selected to calculate the prediction accuracy of the ore body potential area. The calculation formula is as follows: , in For a real example, As a true counterexample, This is a false positive example. This is a false counterexample; Maximizing the accuracy of ore body prediction: , in This refers to the location of the ore body potential zone predicted by the model, i.e., the prediction result; This refers to the actual data labeled for the ore body, i.e., the real labels.

[0012] The environmental risk assessment model in step S6 is as follows: , in These are the weights of different environmental factors. Let N be the environmental risk score for the i-th region, and N be the number of regions.

[0013] The multi-objective optimization algorithm described above combines Pareto optimal solution set and NSGA-II multi-objective optimization algorithm.

[0014] In step S7, the expression for the Q function is as follows: , in Expected value represents the average of all possible rewards given a state and action. Indicates the future time step or the end time. R is a discount factor, typically between 0 and 1, representing the influence of future rewards. t Represents the reward function, Indicates the initial state is s That is, from the state s start, Indicates the initial action is a That is, from the state s Take action a。

[0015] The reward function R is defined as: , in To improve the accuracy of model predictions; The environmental risks introduced during the prediction process are used to minimize environmental impact; w1 and w2 are weighting coefficients. In each round of ore body prediction, the model calculates the Q-function value of each action using a deep neural network and selects the action with the largest Q-function value. The update formula for the Q function is as follows: , in, Q is the Q-value of the current state-action pair, representing the expected long-term reward of the action in the current state. It is the reward function at time t+1, representing the state. Next action The environmental feedback received afterwards It's a discount factor that measures the importance of future rewards. Indicates the next state S t+1 In the middle, select the maximum value of the Q-values ​​corresponding to all possible actions a′. It is the learning rate, which controls the speed at which the Q-value is updated.

[0016] The mineral green prediction device based on cross-domain adaptation and multimodal fusion includes a data acquisition module for acquiring multi-source satellite remote sensing data, UAV low-altitude high-resolution data and geochemical sample data, and preprocessing the data. The transfer learning module learns transferable model parameters from the source domain through adversarial training, and then transfers these transferable model parameters to the Transformer encoder in the target region for initialization. The multimodal fusion module inputs the acquired data into the modal gating network to calculate the weight of each modality, and obtains multimodal fused data through weighted fusion. The graph construction module extracts features from each modality to build a graph structure, constructs a line graph, and uses a graph convolutional network to enhance spatial information. The spatial and spectral feature extraction module is used to extract the spatial topology and spectral features of the mining area from multimodal data through the Transformer encoder embedded in the graph convolutional network, and to ensure the logical consistency of different modes in spatial and spectral features by calculating the cross-modal consistency loss function. The prediction module uses a prediction head to predict the probability of mineralization. The prediction accuracy calculation module is used to calculate the prediction accuracy of the ore body based on the prediction accuracy of the ore body potential zone. The environmental risk assessment module is used to construct environmental risk assessment models based on various environmental indicators to address the environmental risks that may arise from mineral mining. The multi-objective optimization module is used to establish two objective functions: ore body prediction accuracy and environmental risk. The multi-objective optimization algorithm is used to calculate "maximizing ore body prediction accuracy" and "minimizing environmental risk". The drilling feedback module verifies the accuracy of ore body predictions through drilling activities and provides drilling feedback data. The reinforcement learning and optimization update module introduces deep Q-learning to calculate the Q function. After each drilling feedback, the ore body prediction model uses the Q function to adjust the neural network weights for self-adjustment and optimization.

[0017] The beneficial effects of this invention are: 1. To improve the accuracy and generalization ability of orebody prediction, this invention introduces cross-domain adaptive modeling and transfer learning techniques, enabling the model to migrate across regions and overcoming the problem of strong regional dependence in traditional mineral exploration methods. Even with insufficient samples, the model can still maintain high-precision orebody prediction across different mining areas, greatly enhancing the model's generalization ability and cross-regional adaptability; 2. To achieve green mining while balancing ore body prediction and environmental protection, this invention employs a multi-objective optimization algorithm to optimize both the accuracy of ore body prediction and the minimization of environmental risks. During the optimization process, the model not only provides accurate ore body potential distribution maps but also assesses environmental sensitivity risks, helping mineral exploration activities achieve efficient resource utilization while reducing ecological damage, thus meeting current requirements for green mining. 3. Enhancing model interpretability and increasing the trust of geological experts: This invention combines multimodal data and a graph-enhanced interpretable dynamic multimodal Transformer model (GEDMT model for short) in the ore body prediction process. Through attention mechanisms and the Grad-CAM interpretability method, it provides intuitive prediction basis, helping relevant personnel understand the model's decision-making process. This high interpretability enables exploration personnel to trust the model results and make more scientific and reasonable exploration decisions based on the model output. 4. Real-time feedback and adaptive optimization continuously improve prediction accuracy. This invention achieves a closed-loop mechanism of "remote sensing prediction - field verification - model optimization" through real-time fusion of drilling feedback data and dynamic optimization using reinforcement learning. Each feedback from drilling results prompts the model to self-adjust and optimize, gradually improving prediction accuracy and adapting to different mining conditions in each actual exploration. This continuous learning and optimization capability ensures the long-term effectiveness and reliability of the model in practical applications. 5. Improve mineral exploration efficiency and reduce exploration costs. By combining cross-domain adaptive modeling and multimodal data fusion, this invention can quickly predict ore bodies and effectively screen out high-potential mining areas, significantly improving exploration efficiency. At the same time, it reduces manual operation and misjudgments in traditional exploration methods, lowering exploration costs, and is particularly suitable for large-scale mining area exploration and efficient exploration in resource-scarce regions. Attached Figure Description

[0018] Figure 1 This is the prediction process of the multimodal-construction joint modeling GEDMT model of the present invention; Figure 2 This is a network architecture diagram of the method model of the present invention; Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0019] The following description, in conjunction with specific embodiments, provides further details.

[0020] Example 1: A green prediction method for mineral resources based on cross-domain adaptation and multimodal fusion, comprising the following steps: S1. Acquire multi-source satellite remote sensing data, UAV low-altitude high-resolution data, and geochemical sample data, and preprocess the data: Acquire optical satellite remote sensing data for large-scale mineral identification and preliminary ore body location in the mining area. Use UAVs equipped with high-resolution imaging equipment to collect local high-resolution data of the mining area, including its microscopic features, vein orientation, and geological structure. Acquire SAR data to penetrate cloud and vegetation layers to obtain the topographic and structural features of the mining area. Collect geochemical sample data, including mineral composition, soil chemical composition, and heavy metals. Collect environmental monitoring data, including water quality, NDVI, and soil moisture, for subsequent environmental risk assessment.

[0021] After data acquisition, standardized preprocessing is required, including geometric correction, atmospheric correction, radiometric correction, and spatial registration. Preprocessed data possesses both cross-scale consistency and high-quality feature representation, serving as a unified basis for subsequent model input.

[0022] S2. Input the acquired data, including source domain data and target domain data, into the transfer learning module. Through adversarial training, learn transferable model parameters, and then transfer these transferable model parameters to the Transformer encoder of the target region for initialization. The acquired data includes not only source domain data but also target domain data. First, both source and target domain data are fed into a feature extractor for feature extraction. These features are then passed through an extreme inversion layer to the main task predictor and the domain classifier, which operate in parallel. The domain classifier forces the feature extractor to learn region-independent shared features through adversarial training, ensuring that the feature extractor can extract common features applicable to both the source and target domains. Simultaneously, the main task predictor utilizes these features for ore body potential prediction. In this way, the feature extractor and domain classifier jointly learn transferable model parameters, including the convolutional layer weights of the feature extractor and the adversarial training parameters of the domain classifier. These parameters are initialized in the Transformer encoder for the target region.

[0023] This modeling approach not only improves the adaptability of mineral exploration models but also reduces the dependence on sample data for mineral exploration in new areas, making the method more valuable for practical application.

[0024] S3. Input the obtained data into a modal gating network to calculate the weight of each modality, and then perform weighted fusion to obtain multimodal fused data: A modal gating network is set up at the input stage to adaptively adjust the weights of each modality based on data quality and environmental conditions, and to mask inefficient modalities when necessary. The specific principle is as follows: Let the set of input modes be Where x i Let M represent the i-th mode, and M be the number of modes.

[0025] Weights are calculated for each modality using a modality-gating network. , in, For environmental factor vectors, This is a gating function, with the modal feature x as input. i and environmental conditions The output is the importance score for that modality. The final multimodal fusion input is: .

[0026] This mechanism can dynamically adjust weights. When optical imagery is affected by clouds, the system automatically increases the weight of the SAR mode; when geochemical anomalies are significant, the influence of the geochemical mode is preferentially increased. This effectively avoids interference from invalid modes and improves the robustness and computational efficiency of the model.

[0027] S4. Input multimodal fusion data into the graph to construct a model, build lines to create the graph, and use graph convolutional networks to enhance spatial information: First, features are extracted from multi-source data, such as satellite imagery, UAV images, and geochemical data. These features are used to construct the node features of the graph. Then, based on the geological structure information of the mining area, the nodes and edges of the graph are defined. Each sampling point or spatial unit in the mining area serves as a node, and the geological relationships within the mining area are constructed into edges in the graph through an adjacency matrix. In this way, the topological structure of the graph can reflect the different geological features and spatial relationships within the mining area. Graph Convolutional Networks (GCNs) propagate information across the graph structure, aggregating information from neighboring nodes to enhance the model's ability to perceive the geological space of the mining area, thus supporting subsequent predictions.

[0028] S5. The enhanced multimodal data is input into the Transformer encoder of the embedded graph convolutional network to extract the spatial topology and spectral features of the mining area. A cross-modal consistency loss function is introduced to ensure the logical consistency of different modes in spatial and spectral features. The extracted features are then input into the prediction head to predict the probability map of ore-forming potential. A Graph Convolutional Network (GCN) layer is embedded in the Transformer encoder to characterize the spatial topology of the mining area, including fault networks and vein strikes. Specifically, the GCN transforms the geological structural information of the mining area (such as fault zones and vein strikes) into nodes and edges of a graph, capturing the spatial relationships between nodes within the mining area. Each node represents a geological unit or sampling point in the mining area, and edges represent the spatial associations between different units. The GCN aggregates information from neighboring nodes, extracts spatially dependent features, and combines them with spectral features from other modalities. Then, the Transformer encoder further fuses these spatial and spectral features through a self-attention mechanism, enabling the model to simultaneously consider "spectral features + geological structure," thereby improving the consistency between prediction results and actual geological patterns. The specific principle is as follows: By embedding a GCN layer in the Transformer encoder, the model's ability to model the spatial topology and geological features of the mining area is enhanced, ensuring that geological structure and spectral features are effectively fused during the prediction process. Let the spatial topology of the mining area be represented by graph G=(V,E), where: V A set of nodes, representing spatial points or grid cells; E Boundary sets represent the connections between geological structures. The update formula for Graph Convolutional Networks (GCNs) is: , in Let A be the adjacency matrix and I be the identity matrix. The adjacency matrix is ​​used to represent the connection relationships between nodes: each spatial unit / sampling point obtained from the division of the mining area is regarded as a node. If two nodes are geologically connected, they are marked as 1 in the corresponding position of the adjacency matrix, indicating that they are connected to each other; if there is no such direct geological or spatial connection between them, they are marked as 0. In other words, the adjacency matrix is ​​a matrix representation of "which areas in the mining area are geologically connected to each other", used to characterize the spatial topology such as faults and vein trends. for The degree matrix, Let the nodes of the l-th layer be represented as follows: Let σ be the weight matrix of the l-th layer, and σ be the activation function.

[0029] A cross-modal consistency loss function is introduced to ensure logical consistency in spatial and spectral characteristics across different modes. The principle is as follows: Suppose the prediction outputs of different modes at the same location are: in Indicates the first Prediction function corresponding to each mode The cross-modal consistency loss function is defined as: , in The modality represents the number of modalities, indicating different types of data sources or different input modalities. i and j represent the corresponding modalities.

[0030] When SAR identifies a fracture structure, its spectral features should correspondingly show alteration anomalies; otherwise, the model's prediction reliability will decrease. This mechanism can reduce misjudgments caused by "false anomalies" and significantly improve the reliability of prediction results.

[0031] The Grad-CAM interpretability method is integrated during model training to generate a saliency map, which is then used as a regularization constraint to guide model learning. During training, Grad-CAM is used to generate a saliency map S(x), which is then used as a regularization term. , Where S(x) is the saliency map, representing the key regions of interest to the model; F(x) is the model's predicted output; and y is the ground truth label. This design not only ensures that the model can explain its predictions during the inference phase, but also strengthens the focus on key spectral bands and structural features during the training phase. Through this mechanism, the GEDMT model can not only output ore body potential prediction results, but also provide intuitive explanations. This approach improves prediction accuracy, enhances the transparency and credibility of the results, and provides verifiable and interpretable references.

[0032] S6. Calculate the ore body prediction accuracy based on the prediction accuracy of the ore body potential area. Addressing the potential environmental risks from mineral mining, construct an environmental risk assessment model based on various environmental indicators. Establish two objective functions: ore body prediction accuracy and environmental risk. Calculate the "maximization of ore body prediction accuracy" and the "minimization of environmental risk" using a multi-objective optimization algorithm. (1) Construction of the objective function In mineral exploration tasks, the accuracy of ore body prediction is one of the key indicators for measuring model performance, aiming to evaluate the model's accuracy in identifying the location and extent of the real ore body in ore body potential prediction.

[0033] The objective function can be calculated, and model performance can be measured using metrics such as classification accuracy, recall, F1 score, or mean squared error. In this invention, classification accuracy is chosen as the objective function to calculate the prediction accuracy of the ore body potential area. The calculation formula is as follows: , in For a real example, As a true counterexample, This is a false positive example. This is a false counterexample.

[0034] Maximize the accuracy of ore body prediction. , in The location of the ore body potential area predicted by the model (prediction result), i.e., the ore body prediction label output by the model; This refers to the actual ore body labeling data (real labels), i.e., the ore body areas labeled by experts.

[0035] By optimizing the above objective function L mining This improves the accuracy of the model in orebody prediction, ensuring that the model can accurately identify the location and extent of potential mining areas. Specifically, the optimization process enhances the prediction accuracy of orebody potential areas, reduces false positives and false negatives, and makes mining area predictions more precise. By continuously optimizing the accuracy objective function, the model can improve its generalization ability in unknown areas and effectively identify orebody potential areas in different regions.

[0036] Environmental risk minimization aims to minimize the environmental impact of mining, particularly the potential threats to ecosystems such as water sources, soil, and vegetation. Through comprehensive analysis of remote sensing data, geochemical data, and environmental monitoring data, models can assess the environmental sensitivity of different mining areas and quantify the potential environmental risks of mining. An environmental risk assessment model is constructed based on environmental indicators such as the National Distance Vibration Index (NDVI) and water quality index. , in This represents the weights of different environmental factors. Let N be the number of regions and N be the environmental risk score for the i-th region. The optimization objective of this objective function is to reduce potential damage to the ecological environment and ensure the sustainability of resource development.

[0037] (2) Bi-objective optimization algorithm To balance orebody prediction accuracy with environmental risk, this invention employs a Pareto optimal solution set to solve a bi-objective optimization problem. In multi-objective optimization, the Pareto front refers to the set of solutions for which no further improvement can be achieved without sacrificing the other objective. The specific implementation steps are as follows: ① Weight Adjustment: By adjusting the weights λ1 and λ2 of the two objective functions, the balance between ore body prediction accuracy and environmental risk is controlled. , Where λ1 and λ2 represent the weighting coefficients for ore body prediction and environmental protection, respectively, satisfying λ1+λ2=1.

[0038] ②Pareto optimization process: Using multi-objective optimization algorithms such as Particle Swarm Optimization (PSO) or Genetic Algorithm (GA), a set of Pareto optimal solutions is found in the objective function space, representing the best balance between ore body prediction and environmental protection under given constraints. Each solution corresponds to a different mining area and environmental impact assessment. The final model will output a set of ore body potential areas and environmental risk sensitivity maps.

[0039] ③ Non-dominated Sorting and Crowding Calculation: To ensure a balance between different solutions during optimization, this invention introduces the NSGA-II (Non-dominated Sorting Genetic Algorithm II) algorithm. This algorithm evaluates the superiority of solutions through non-dominated sorting and crowding calculation. Specifically, NSGA-II performs non-dominated sorting according to the relative superiority of solutions and selects the optimal solution by calculating the crowding of each solution (i.e., its distance from other solutions in the target space). In this process, the Pareto optimization algorithm and the NSGA-II algorithm are used in parallel. The Pareto optimization algorithm generates a Pareto front solution set for the problem, representing different trade-offs between ore body prediction accuracy and environmental risk. NSGA-II further optimizes this set by calculating the crowding of solutions to ensure the diversity of the solution set and avoid selecting overly similar solutions. This method achieves an effective trade-off between maximizing ore body prediction accuracy and minimizing environmental risk, thereby selecting a set of optimal solutions for practical application. Ultimately, based on the evaluation results of the Pareto Frontier and NSGA-II, and combined with congestion calculations, an optimal balance point was determined, which minimizes environmental risks while ensuring accuracy and meeting the needs of practical applications.

[0040] (3) Result output The dual-objective optimization method of this invention will ultimately output two key results: Ore body potential distribution map: By using an optimized ore body prediction function, the ore body potential score of different regions is obtained, showing the spatial distribution and scale of potential mining areas; Environmental Sensitivity Risk Map: Through an environmental risk assessment model, the ecological risks of each mining area are displayed, ecologically vulnerable and sensitive areas are identified, and the environmental impacts that mining activities may cause are indicated. Through the above optimizations, this invention can achieve an efficient balance between mineral exploration accuracy and ecological protection, ensuring the sustainable development of mineral resources while minimizing environmental damage.

[0041] S7. Verify the accuracy of the ore body prediction through drilling activities. Calculate the Q-function using deep Q-learning based on the drilling feedback data. After each drilling feedback, the ore body prediction model uses the Q-function to adjust the neural network weights for self-adjustment and optimization, resulting in the final prediction model. In actual mineral resource exploration, the environment and geological conditions of mining areas are highly uncertain. Therefore, orebody prediction models need to be dynamically adaptable to continuously improve prediction accuracy and effectively respond to feedback from new data and actual exploration results. This invention introduces a reinforcement learning (RL) mechanism to achieve dynamic optimization and iterative updates of the model, forming a closed-loop mechanism of "remote sensing prediction - field verification - model optimization". The specific steps are as follows: (1) Fusion of drilling feedback data In mineral exploration, drilling activities are typically used to verify the accuracy of ore body predictions. Each drilling operation yields geological samples and data. This drilling feedback data is then input into the model as actual labels, thus forming a data loop.

[0042] (2) Dynamic optimization of reinforcement learning This invention employs Deep Q-learning (DQN) for reinforcement learning optimization. DQN combines Q-learning and deep neural networks (DNN), using neural networks to approximate the Q-function, enabling the model to handle high-dimensional and complex remote sensing and geological data, and adjust its prediction strategy in real time.

[0043] The reward function in reinforcement learning guides the model's decisions, ensuring that environmental impact is minimized while optimizing ore body prediction. The reward function provides immediate feedback, measuring the effect of each action. The Q-function (quality function), on the other hand, is used in reinforcement learning to estimate the long-term reward of taking a particular action in a given state. The Q-function evaluates the long-term value of each decision by accumulating the immediate feedback provided by the reward function, guiding the model to select the optimal strategy. Specifically... The expression for the Q function is as follows: , in Expected value represents the average of all possible rewards given a state and action. It indicates the future time step or the end point. This is a discount factor, typically between 0 and 1, representing the influence of future rewards. Smaller ones... This will cause the model to focus more on short-term rewards, larger This will make the model focus more on long-term rewards. This indicates that the initial state is s, meaning we start from state s. This indicates that the initial action is 'a', meaning that action 'a' is taken from state 's'.

[0044] The reward function is defined as: , in To improve the accuracy of model predictions; The environmental risks introduced during the prediction process are used to minimize environmental impact; w1 and w2 are weighting coefficients.

[0045] The update formula for the Q function is as follows: , in, Q is the Q-value of the current state-action pair, representing the expected long-term reward of the action in the current state. It is the reward function at time t+1, representing the state. Next action The environmental feedback obtained afterwards. It is a discount factor that measures the importance of future rewards. Indicates the next state S t+1 In the middle, select the maximum value of the Q value corresponding to all possible actions a′. It is the learning rate, which controls the speed at which the Q-value is updated.

[0046] In each round of ore body prediction, the model calculates the Q-value of each action using a deep neural network and selects the action with the highest Q-value.

[0047] After each drilling feedback, the model updates the weights in the neural network and gradually optimizes the Q function, enabling the model to continuously improve the accuracy of ore body prediction and the effectiveness of environmental protection under different mining area environmental conditions.

[0048] (3) Establishment of a closed-loop mechanism In each round of orebody prediction, the model makes preliminary predictions about the mining area based on remote sensing data and identifies potential orebody zones. Subsequent drilling activities provide real-world validation for the model, comparing the remote sensing prediction results with drilling feedback data to assess the model's prediction errors. This process allows the model to continuously learn and adjust, forming the following closed-loop mechanism: Remote sensing prediction → Drilling verification → Data feedback → Reinforcement learning optimization → Remote sensing prediction This process is a continuous cycle. After each remote sensing prediction and drilling verification, the model adjusts its prediction strategy based on feedback data in order to better identify ore bodies and maximize the balance between environmental protection and resource utilization.

[0049] (4) Application effect ① Continuous optimization and adaptive improvement: By introducing Deep Q-Learning (DQN), the orebody prediction model can self-adjust and optimize after each drilling feedback. The addition of new data each time drives the model to maintain a better balance between orebody prediction and environmental protection, thereby improving prediction accuracy. ② Improve the accuracy of ore body prediction and protect the environment: The dynamic optimization of reinforcement learning enables the model to adaptively adjust its prediction strategy while reducing the environmental risks of mineral extraction. Through continuous feedback updates, the model learns continuously, achieving the optimal balance between ore body prediction accuracy and environmental protection effectiveness. ③ Closed-loop feedback mechanism: Through a closed-loop mechanism of "remote sensing prediction - field verification - model optimization," the model can be gradually improved during actual exploration. Each addition of feedback data helps the model learn and optimize in real time. This mechanism ensures more accurate ore body prediction and adapts to geological and environmental changes in different mining areas.

[0050] Example 2: A mineral green prediction device based on cross-domain adaptation and multimodal fusion, including a data acquisition module for acquiring multi-source satellite remote sensing data, UAV low-altitude high-resolution data and geochemical sample data, and preprocessing the data; The transfer learning module learns transferable model parameters from the source domain through adversarial training, and then transfers these transferable model parameters to the Transformer encoder in the target region for initialization. The multimodal fusion module inputs the acquired data into the modal gating network to calculate the weight of each modality, and obtains multimodal fused data through weighted fusion. The graph construction module extracts features from each modality to build a graph structure, constructs a line graph, and uses a graph convolutional network to enhance spatial information. The spatial and spectral feature extraction module is used to extract the spatial topology and spectral features of the mining area from multimodal data through the Transformer encoder embedded in the graph convolutional network, and to ensure the logical consistency of different modes in spatial and spectral features by calculating the cross-modal consistency loss function. The prediction module uses a prediction head to predict the probability of mineralization. The prediction accuracy calculation module is used to calculate the prediction accuracy of the ore body based on the prediction accuracy of the ore body potential zone. The environmental risk assessment module is used to construct environmental risk assessment models based on various environmental indicators to address the environmental risks that may arise from mineral mining. The multi-objective optimization module is used to establish two objective functions: ore body prediction accuracy and environmental risk. The multi-objective optimization algorithm is used to calculate "maximizing ore body prediction accuracy" and "minimizing environmental risk". The drilling feedback module verifies the accuracy of ore body predictions through drilling activities and provides drilling feedback data. The reinforcement learning and optimization update module introduces deep Q-learning to calculate the Q function. After each drilling feedback, the ore body prediction model uses the Q function to adjust the neural network weights for self-adjustment and optimization.

[0051] The above is a further description of the present invention in conjunction with specific embodiments, and the scope of protection of the present invention is not limited thereto.

Claims

1. A green prediction method for mineral resources based on cross-domain adaptive and multimodal fusion, characterized by: The steps include the following: S1. Acquire multi-source satellite remote sensing data, UAV low-altitude high-resolution data, and geochemical sample data, and preprocess the data; S2. Input the acquired data, including source domain data and target domain data, into the transfer learning module. Through adversarial training, learn transferable model parameters and transfer these transferable model parameters to the Transformer encoder of the target region for initialization. S3. Input the obtained data into the modal gating network to calculate the weight of each modality, and then perform weighted fusion to obtain multimodal fusion data; S4. Input multimodal fusion data into the graph construction model to construct lines and graphs, and use graph convolutional networks to enhance spatial information; S5. The enhanced multimodal data is input into the Transformer encoder of the embedded graph convolutional network to extract the spatial topology and spectral features of the mining area. A cross-modal consistency loss function is introduced to ensure the logical consistency of different modes in spatial and spectral features. The extracted features are input into the prediction head to predict and obtain the ore-forming potential probability map. S6. Calculate the ore body prediction accuracy by predicting the ore body potential area. Based on various environmental indicators, construct an environmental risk assessment model to address the environmental risks that mining may bring. Establish two objective functions: ore body prediction accuracy and environmental risk. Calculate "maximizing ore body prediction accuracy" and "minimizing environmental risk" through a multi-objective optimization algorithm. S7. Verify the accuracy of ore body prediction through drilling activities. Calculate the Q function by introducing deep Q-learning into the drilling feedback data. After each drilling feedback, the ore body prediction model uses the Q function to adjust the neural network weights for self-adjustment and optimization, resulting in the final prediction model.

2. The mineral green prediction method based on cross-domain adaptation and multimodal fusion as described in claim 1, characterized in that, In step S2, the transfer learning module includes a feature extractor. The feature extractor is connected to the parallel-designed main task predictor and domain classifier through an extreme inversion layer. The feature extractor extracts preliminary features from the source domain data. The domain classifier, through adversarial training, forces the feature extractor to generate shared features that are effective for both the source and target domains, thereby enhancing the model's cross-domain generalization ability. The main task predictor then uses these shared features to predict the ore body potential. In this way, the transfer learning module learns transferable model parameters, including the convolutional layer weights of the feature extractor, the adversarial training parameters of the domain classifier, and the weights of the main task classification head. These transferable model parameters are then transferred to the Transformer encoder of the target region for initialization.

3. The mineral green prediction method based on cross-domain adaptation and multimodal fusion as described in claim 1, characterized in that, The modal gating network described in step S3 is calculated as follows: Let the set of input modes be Where x i Let M represent the i-th mode, and M be the number of modes. Weights are calculated for each mode using a modal gating network: ， in, An environmental factor vector. This is a gating function, with the modal feature x as input. i and environmental conditions The output is the importance score for that modality. The final multimodal fusion result is as follows: 。 4. The mineral green prediction method based on cross-domain adaptation and multimodal fusion as described in claim 1, characterized in that, In step S5, the Grad-CAM interpretability module is introduced to generate a saliency map, which is then used as a regularization constraint to guide model learning during training. During training, Grad-CAM is used to generate a saliency map S(x), which is then used as a regularization term. ， Where S(x) is the saliency map, representing the key regions of interest to the model; F(x) represents the model's predicted output, which is a probability value; and y is the true label.

5. The mineral green prediction method based on cross-domain adaptation and multimodal fusion according to claim 1, characterized in that, In step S5, the Transformer encoder embedded with a graph convolutional network processes the multimodal data and spatial topology of the mining area by combining the graph convolutional network and the Transformer self-attention mechanism. First, the GCN constructs a graph structure based on the geological structure information of the mining area, representing each data point as a node in the graph and performing spatial feature aggregation. Then, the Transformer encoder receives the node features processed by the GCN and uses the self-attention mechanism to capture global dependencies and long-range information, thereby enhancing the connection between multimodal data.

6. The mineral green prediction method based on cross-domain adaptation and multimodal fusion according to claim 1, characterized in that, The calculation of the cross-modal consistency loss function assumes that the prediction outputs of different modes at the same location are: in Indicates the first Prediction function corresponding to each mode The cross-modal consistency loss function is defined as: ， in The number of modalities represents different types of data sources or different input modalities, with i and j being the corresponding modalities.

7. The mineral green prediction method based on cross-domain adaptation and multimodal fusion according to claim 1, characterized in that, In step S6, the classification accuracy is selected to calculate the prediction accuracy of the ore body potential area. The calculation formula is as follows: ， in For a real example, As a true counterexample, As a false positive example, This is a false counterexample; Maximizing the accuracy of orebody prediction: ， in This refers to the location of the ore body potential zone predicted by the model, i.e., the prediction result; This refers to the actual data labeled for the ore body, i.e., the real labels.

8. The mineral green prediction method based on cross-domain adaptation and multimodal fusion according to claim 1, characterized in that, The environmental risk assessment model in step S6 is as follows: ， in These are the weights of different environmental factors. Let N be the environmental risk score for the i-th region, and N be the number of regions.

9. The mineral green prediction method based on cross-domain adaptive and multimodal fusion according to claim 1, characterized in that, In step S7, the expression for the Q function is as follows: ， in Expected value represents the average of all possible rewards given a state and action. Indicates the future time step or the end time. R is a discount factor, typically between 0 and 1, representing the influence of future rewards. t Let S0 = S represent the initial state as S, i.e., starting from state S, and a0 = a represent the initial action as a, i.e., taking action a from state s. The reward function R is defined as: ， in To improve the accuracy of model predictions; The environmental risks introduced during the prediction process are used to minimize environmental impact; w1 and w2 are weighting coefficients. In each round of ore body prediction, the model calculates the Q-function value of each action using a deep neural network and selects the action with the largest Q-function value. The update formula for the Q function is as follows: ， in, Q is the Q-value of the current state-action pair, representing the expected long-term reward of the action in the current state. It is the reward function at time t+1, representing the state S. t Next, execute action a t The environmental feedback received afterwards It's a discount factor that measures the importance of future rewards. Indicates the next state S t+1 In the middle, select the maximum value of the Q-values ​​corresponding to all possible actions a′. It is the learning rate, which controls the speed at which the Q-value is updated.

10. A mineral resource green prediction device based on cross-domain adaptive and multimodal fusion, characterized in that, It includes a data acquisition module for acquiring multi-source satellite remote sensing data, UAV low-altitude high-resolution data, and geochemical sample data, and for preprocessing the data; The transfer learning module learns transferable model parameters from the source domain through adversarial training, and then transfers these transferable model parameters to the Transformer encoder in the target region for initialization. The multimodal fusion module inputs the acquired data into the modal gating network to calculate the weight of each modality, and obtains multimodal fused data through weighted fusion. The graph construction module extracts features from each modality to build a graph structure, constructs a line graph, and uses a graph convolutional network to enhance spatial information. The spatial and spectral feature extraction module is used to extract the spatial topology and spectral features of the mining area from multimodal data through the Transformer encoder embedded in the graph convolutional network, and to ensure the logical consistency of different modes in spatial and spectral features by calculating the cross-modal consistency loss function. The prediction module uses a prediction head to predict the probability of mineralization. The prediction accuracy calculation module is used to calculate the prediction accuracy of the ore body based on the prediction accuracy of the ore body potential zone. The environmental risk assessment module is used to construct environmental risk assessment models based on various environmental indicators to address the environmental risks that may arise from mineral mining. The multi-objective optimization module is used to establish two objective functions: ore body prediction accuracy and environmental risk. The multi-objective optimization algorithm is used to calculate "maximizing ore body prediction accuracy" and "minimizing environmental risk". The drilling feedback module verifies the accuracy of ore body predictions through drilling activities and provides drilling feedback data. The reinforcement learning and optimization update module introduces deep Q-learning to calculate the Q function. After each drilling feedback, the ore body prediction model uses the Q function to adjust the neural network weights for self-adjustment and optimization.

Images