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

Abstract

The present application relates to a kind of green prediction method and device of mineral based on cross-domain self-adaptation and multimodal fusion, belong to mineral resources exploration and remote sensing technical field.Multiple-source data is obtained, input transfer learning module learns transferable model parameters, the data obtained is fused through modal gating network, then through graph construction model constructs line mapping, and is enhanced using graph convolution network;After enhancement, multimodal data is input into the Transformer encoder of embedded graph convolution network, extract the spatial topology structure and spectral feature of mining area, obtain the potential probability graph of ore body through the prediction head prediction;Environmental risk assessment model is built, " maximum of ore body prediction accuracy " and " minimum of environmental risk " are realized through multi-objective optimization, and the accuracy of ore body prediction is verified through drilling activity.This method not only improves the regional generalization ability of model, but also realizes green mineral exploration and the interpretability of result, 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:

[0003] (1) Strong regional dependence, lack of cross-domain adaptation ability:

[0004] 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, etc. between 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;

[0005] (2) Lack of model interpretability, difficult to provide intuitive basis for exploration decision-making:

[0006] 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 region 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;

[0007] (3) Ignoring environmental factors, difficult to achieve green prospecting:

[0008] 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, and vegetation destruction. 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

[0009] The present application aims 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.

[0010] The technical scheme adopted by the present application is:

[0011] The mineral green prediction method based on cross-domain adaptation and multi-modal fusion comprises the following steps:

[0012] S1. Obtain multi-source satellite remote sensing data, unmanned aerial vehicle low-altitude high-resolution data and geochemical sample data, and preprocess the data;

[0013] 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;

[0014] 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;

[0015] S4. Input the multi-modal fusion data into a graph construction model to construct line mapping, and enhance the spatial information by using a graph convolution network;

[0016] 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 ore body probability map;

[0017] S6. Calculate the ore body prediction accuracy through the prediction accuracy of the potential ore body area, construct an environmental risk assessment model based on various environmental indicators for the environmental risks that may be caused by mineral exploitation, establish two objective functions of ore body prediction accuracy and environmental risk, and realize the calculation of "maximum ore body prediction accuracy" and "minimum environmental risk" through a multi-objective optimization algorithm;

[0018] S7. Verify the accuracy of the 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 through the Q function to adjust and optimize the ore body prediction model after each drilling feedback, and obtain the final prediction model.

[0019] In the 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 capability of the model, and the main task predictor performs ore body potential prediction using the shared features; 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, and the transferable model parameters are subsequently transferred to the Transformer encoder in the target region for initialization.

[0020] The modal gating network in step S3 has the following specific calculation process:

[0021] Let the input modal set be where x i represents the i-th modal, M is the number of modes,

[0022] Calculate the weight for each modal through the modal gating network (Gating Network):

[0023] ,

[0024] 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,

[0025] The final multi-modal fusion result is:

[0026] .

[0027] 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.

[0028] In 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:

[0029] ,

[0030] where S(x) is the saliency map representing the key regions that the model focuses on, F(x) represents the predicted output of the model, which is a probability value, and y is the true label.

[0031] In step S5, the Transformer encoder embedded in the graph convolutional network processes the multi-modal data and spatial topology of the mining area by combining the graph convolutional network (GCN) and the Transformer self-attention mechanism. First, the GCN constructs a graph structure according to the geological structure information of the mining area (such as veins and fault zones), represents each data point as a node in the graph, and performs spatial feature aggregation. Subsequently, the Transformer encoder receives the node features processed by the GCN, captures global dependencies and long-range information using the self-attention mechanism, and enhances the connection between multi-modal data. In this way, the model can integrate information from different modalities in terms of spatial and spectral features, ultimately generating a potential ore body prediction map or a metallogenic probability map of the mining area. Meanwhile, a cross-modal consistency loss function is introduced to ensure the logical unity of different modalities in space and spectrum.

[0032] The calculation of the cross-modal consistency loss function is as follows, where the predicted output of different modalities at the same position is: where represents the prediction function corresponding to the i-th modality,

[0033] The cross-modal consistency loss function is defined as:

[0034] ,

[0035] where is the number of modalities, representing different types of data sources or different input modalities, and i and j are the corresponding modalities.

[0036] In step S6, the classification accuracy is selected to calculate the prediction accuracy of the potential ore body area, and the calculation formula is as follows:

[0037] ,

[0038] where is the true positive, is the true negative, is the false positive, is the false negative;

[0039] The maximum accuracy of ore body prediction is:

[0040] ,

[0041] where is the position of the potential ore body area predicted by the model, i.e., the prediction result; ​The data is labeled for the actual ore body, i.e., the true label.

[0042] The environmental risk assessment model in step S6 is:

[0043] ,

[0044] wherein is the weight of different environmental factors, is the environmental risk score of the i th region, and N is the number of regions.

[0045] The multi-objective optimization algorithm adopts a combination of a Pareto optimal solution set and an NSGA-II multi-objective optimization algorithm.

[0046] In step S7, the Q function expression is as follows:

[0047] ,

[0048] wherein represents an expected value, represents the average value of all possible rewards under a given state and action, represents a future time step or a termination time, is a discount factor, usually between 0 and 1, representing the influence of future rewards, R t represents a reward function, represents an initial state s , i.e., starting from state s , represents an initial action a , i.e., taking action s from state a。

[0049] The reward function R is defined as:

[0050] ,

[0051] wherein is the improvement of model prediction accuracy; is the environmental risk introduced in the prediction process, used to minimize environmental impact; w1 and w2 are weight coefficients;

[0052] In each round of ore body prediction, the model calculates the Q function value of each action through a deep neural network, and selects the action with the maximum Q function value;

[0053] The update formula of the Q function is as follows:

[0054] ,

[0055] wherein, Q(s, a) is the Q-value of the current state-action pair, representing the expected long-term return of the action in the current state, Rt+1is the reward function at time t+1, representing the environment feedback obtained after performing action at state , is the discount factor, measuring the importance of future rewards, represents the maximum value among the Q-values of all possible actions a' in the next state S t+1 , is the learning rate, controlling the speed of Q-value update.

[0056] The mineral green prediction device based on cross-domain self-adaption and multi-modal fusion comprises a data acquisition module, which is used for acquiring multi-source satellite remote sensing data, unmanned aerial vehicle low-altitude high-resolution data and geochemical sample data, and pre-processing the data;

[0057] A transfer learning module learns model parameters that are transferable from a source domain through adversarial training, and initializes these transferable model parameters in a Transformer encoder of a target region;

[0058] A multi-modal fusion module inputs the obtained data into a modal gating network to calculate the weight of each modality, and obtains multi-modal fusion data through weighted fusion;

[0059] A graph construction module extracts features of each modality to construct a graph structure, constructs a line graph, and enhances spatial information by using a graph convolution network;

[0060] A spatial and spectral feature extraction module is used for extracting spatial topological structure and spectral features of a mining area through a Transformer encoder of an embedded graph convolution network, and calculating through a cross-modal consistency loss function to ensure the logical unity of different modalities in spatial and spectral features;

[0061] A prediction module obtains a mineralization probability through a prediction head;

[0062] A prediction accuracy calculation module is used for calculating a mineral body prediction accuracy through the prediction accuracy of a mineral body potential area;

[0063] An environmental risk assessment module is used for constructing an environmental risk assessment model based on various environmental indicators for the environmental risks that may be caused by mineral exploitation;

[0064] A multi-objective optimization module is used for establishing two objective functions of mineral body prediction accuracy and environmental risk, and realizing the calculation of "maximizing the mineral body prediction accuracy" and "minimizing the environmental risk" through a multi-objective optimization algorithm;

[0065] A drilling feedback module verifies the accuracy of the ore body prediction through drilling activities and provides drilling feedback data;

[0066] A reinforcement learning and optimization updating module adjusts the neural network weights by introducing a deep Q-learning calculation Q function to adjust and optimize the ore body prediction model after each drilling feedback.

[0067] The beneficial effects of the present application are:

[0068] 1. Improve the accuracy and generalization ability of ore body prediction. The present application introduces cross-domain adaptive modeling and transfer learning technology, so that the model can be transferred across regions, overcoming the problem of strong regional dependence in traditional prospecting methods. Even in the case of insufficient samples, the model can maintain high-precision ore body prediction between different mining areas, greatly enhancing the generalization ability and cross-regional adaptability of the model;

[0069] 2. Realize green mining, taking into account ore body prediction and environmental protection. The present application uses a multi-objective optimization algorithm to optimize the accuracy of ore body prediction and minimize environmental risk as dual objectives. During the optimization process, the model not only provides an accurate ore potential distribution map, but also assesses environmental sensitivity risk, helping to achieve efficient resource utilization while reducing ecological damage, in line with current green mining and sustainable development policy requirements;

[0070] 3. Enhance model interpretability and improve geologist trust. The present application combines multi-modal data and graph structure enhanced interpretable dynamic multi-modal Transformer model (referred to as GEDMT model) in the process of ore body prediction, providing intuitive prediction basis through attention mechanism and Grad-CAM interpretability method, helping relevant personnel understand the model's decision-making process. This high interpretability allows exploration personnel to trust the model results and make more scientific and reasonable exploration decisions based on model output;

[0071] 4. Real-time feedback and adaptive optimization, continuously improving prediction accuracy. The present application realizes the closed-loop mechanism of "remote sensing prediction - field verification - model optimization" through real-time fusion of drilling feedback data and dynamic optimization of reinforcement learning. Each drilling result feedback will prompt the model to adjust and optimize itself, gradually improving the 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;

[0072] 5. The present application can quickly realize ore body prediction and effectively screen out high-potential mining areas, significantly improving the efficiency of exploration, by combining cross-domain adaptive modeling and multi-modal data fusion. At the same time, it reduces manual operation and misjudgment in traditional exploration methods, reduces exploration cost, and is especially suitable for large-scale mining exploration and efficient exploration in resource scarce areas. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 The prediction process of the multi-modal-structure joint modeling GEDMT model of the present application;

[0074] Figure 2 The network architecture diagram of the method model of the present application;

[0075] Figure 3 The flowchart of the method of the present application. DETAILED DESCRIPTION

[0076] The following will be further illustrated with specific embodiments.

[0077] Example 1: A green prediction method for mineral resources based on cross-domain adaptive modeling and multi-modal fusion, comprising the following steps:

[0078] S1. Obtain multi-source satellite remote sensing data, unmanned aerial low-altitude high-resolution data and geochemical sample data, and pre-process the data:

[0079] Obtain optical satellite remote sensing data for mineral identification and preliminary positioning of large-scale mining areas. Collect local high-resolution data of the mining area by unmanned aerial vehicle carrying high-resolution imaging equipment, including micro features, vein trends and geological structures of the mining area. Obtain SAR data for penetrating clouds and vegetation layers to obtain topographic and structural features of the mining area. Collect geochemical sample data, including mineral composition, soil chemical composition, heavy metals, etc. Collect environmental monitoring data, such as water quality, NDVI, soil moisture and other environmental parameters for subsequent environmental risk assessment.

[0080] After data collection, standardization preprocessing is required: including geometric correction, atmospheric correction, radiation correction and spatial registration. The preprocessed data has cross-scale consistency and high-quality feature expression, which can be used as a unified basis for subsequent model input.

[0081] S2. Input the obtained data including source domain data and target domain data into the transfer learning module through adversarial training, learn transferable model parameters, and initialize these transferable model parameters into the target area's Transformer encoder:

[0082] The acquired data includes not only source domain data but also target domain data. First, the source domain and target domain data are input into a feature extractor together for feature extraction. Then, the features are passed to a main task predictor and a domain classifier through an extreme inversion layer, where the main task predictor and the domain classifier work 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 domain and the target domain, while the main task predictor uses these features for ore body potential prediction. In this way, the feature extractor and the 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, which will be initialized in the Transformer encoder of the target region.

[0083] This modeling method not only improves the adaptability of the ore prospecting model, but also reduces the dependence on sample data for ore prospecting in new regions, making the method have stronger practical promotion value.

[0084] S3. The obtained data is input into a modal gating network to calculate the weight of each modality, and the multi-modal fusion data is obtained by weighted fusion:

[0085] A modal gating network is set in the input link to adaptively adjust the weight of each modality according to data quality and environmental conditions, and to shield inefficient modalities when necessary. The specific principle is as follows:

[0086] Let the input modal set be where x i represents the i-th modality, and M is the number of modalities.

[0087] Through a modal gating network (Gating Network), the weight of each modality is calculated:

[0088] ,

[0089] 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 modality,

[0090] The final multi-modal fusion input is:

[0091] .

[0092] This mechanism can dynamically adjust the weight. When the optical image is affected by clouds, the system automatically increases the weight of the SAR modality; when the geochemical anomaly is significant, the influence of the geochemical modality is preferentially increased. This effectively avoids the interference of invalid modalities and improves the robustness and computational efficiency of the model.

[0093] S4. Constructing a graph model for multi-modal fusion data input and enhancing spatial information using graph convolution network:

[0094] First, features are extracted from multi-source data, such as satellite images, UAV images, and geochemical data, which will be 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 will be a node, and the geological relationship of the mining area will be built into the edges of the graph through the adjacency matrix. In this way, the topological structure of the graph can reflect the different geological features and spatial relationships in the mining area. Graph convolution network (GCN) will propagate information on the structure of the graph, aggregate information from adjacent nodes, and enhance the model's perception of the geological space of the mining area to support subsequent prediction.

[0095] S5. Enhancing the multi-modal data input to the Transformer encoder embedded with graph convolution network, extracting the spatial topology and spectral features of the mining area, introducing a cross-modal consistency loss function to ensure the logical unity of different modalities in spatial and spectral features, and extracting features to input the prediction head to obtain the potential probability map of the ore body:

[0096] In the Transformer encoder, the graph convolution network (GCN Layer) is embedded to depict the spatial topology of the mining area, such as fault networks and vein trends. Specifically, the graph convolution network (GCN) captures the spatial relationships between nodes in the mining area by converting geological structure information (such as fault zones and vein trends) into graph nodes and edges. Each node represents a geological unit or sampling point in the mining area, and the edges represent the spatial association between different units. GCN aggregates information from adjacent nodes to extract features with spatial dependencies and combines them with spectral features from other modalities. Then, the Transformer encoder further integrates these spatial and spectral features through self-attention mechanisms, allowing the model to consider both "spectral features + geological structure" and improve the consistency of the prediction results with actual geological rules. The specific principle is to embed the GCN layer in the Transformer encoder to enhance the model's ability to model the spatial topology and geological features of the mining area, ensuring effective integration of geological structure and spectral features during prediction. The specific principle is as follows:

[0097] Let the spatial topology of the mining area be represented as a graph G=(V,E), where: V V is the set of nodes, representing spatial points or grid cells; E E is the set of edges, representing the connection relationship of geological structure,

[0098] The update formula of the graph convolution network (GCN) is:

[0099] ,

[0100] wherein is the adjacency matrix A + identity matrix I, which is used to represent the connection between nodes: each spatial unit / sampling point obtained by dividing the mining area is regarded as a node, if two nodes have a connection in the geological structure, the corresponding position of the adjacency matrix is recorded as 1, indicating that they are connected with each other; if there is no such direct connection between the two in terms of geology or space, it is recorded as 0. That is, the adjacency matrix is a matrix representation of "which areas are geologically connected to each other" in the mining area, which is used to depict the spatial topological structure such as fault, vein trend, etc. is the degree matrix of , is the node representation of the l-th layer, is the weight matrix of the l-th layer, and σ is the activation function.

[0101] A cross-modal consistency loss function is introduced to ensure the logical unity of different modalities in spatial and spectral features. The principle is as follows:

[0102] Let the prediction output of different modalities at the same position be: wherein represents the prediction function corresponding to the i-th modality, The cross-modal consistency loss function is defined as:

[0103]

[0104] ,

[0105] wherein is the number of modalities, indicating different types of data sources or different input modalities. i, j are the corresponding modalities.

[0106] When the SAR identifies the fault structure, the spectral features should correspondingly show alteration anomalies, otherwise the model will reduce its prediction reliability. This mechanism can reduce the misjudgment caused by "false anomalies", and significantly improve the reliability of the prediction results.

[0107] In the model training process, the Grad-CAM explainability method is integrated to generate saliency maps, which are used as regularization constraints to guide model learning. In the training process, Grad-CAM is used to generate saliency map S(x), which is used as a regularization term:

[0108] ,

[0109] wherein S(x) is the saliency map, indicating the key area that the model focuses on; F(x) is the model prediction output y is the true label;

[0110] ​The design not only ensures that the model can explain the prediction basis in the reasoning stage, but also strengthens the focus on key spectral bands and structural features in the training stage. Through the above mechanism, the GEDMT model can not only output ore body potential prediction results, but also provide intuitive explanation basis. This way not only improves prediction accuracy, but also enhances the transparency and credibility of the results, providing a verifiable and explainable reference.

[0111] S6. Calculate the ore body prediction accuracy by the prediction accuracy of the ore body potential area, construct an environmental risk assessment model based on various environmental indicators for the environmental risks that may be caused by mineral exploitation, establish two objective functions of ore body prediction accuracy and environmental risk, and realize the calculation of "maximizing ore body prediction accuracy" and "minimizing environmental risk" through multi-objective optimization algorithm:

[0112] (1) Objective function construction

[0113] In the prospecting task, the ore body prediction accuracy is one of the key indicators to measure the performance of the model, which aims to evaluate the recognition accuracy of the model in the prediction of the location and range of the real ore body in the ore body potential area.

[0114] The calculation of the objective function can measure the performance of the model through classification accuracy, recall rate, F1 score or mean square error, etc. In this invention, the classification accuracy (Accuracy) is selected as the objective function to calculate the prediction accuracy of the ore body potential area. The calculation formula is as follows:

[0115] ,

[0116] Among them, is the true positive, is the true negative, is the false positive, is the false negative.

[0117] Maximize the ore body prediction accuracy,

[0118] ,

[0119] Among them, is the position of the ore body potential area predicted by the model (prediction result), i.e. the ore body prediction label output by the model; is the actual ore body annotation data (true label), i.e. the ore body area annotated by experts.

[0120] By optimizing the above objective function L mining, improve the accuracy of the model in ore body prediction, and ensure that the model can accurately identify the location and range of potential mining areas. Specifically, the optimization process will improve the prediction accuracy of ore potential areas, reduce false positives and false negatives, and make the mining area prediction more accurate. By continuously optimizing the accuracy objective function, the model can improve its generalization ability in unknown areas and effectively identify ore potential areas in different regions.

[0121] Environmental risk minimization aims to minimize the impact of mineral exploitation on the environment, especially the potential threats to water sources, soil, vegetation, and other ecosystems. By analyzing remote sensing data, geochemical data, and environmental monitoring data, the model can assess the environmental sensitivity of different mining areas and quantify the environmental risks that may be caused by mineral exploitation. Based on environmental indicators such as vegetation index (NDVI) and water quality index, an environmental risk assessment model is constructed:

[0122] ,

[0123] where is the weight of different environmental factors, is the environmental risk score of the ith region, and N is the number of regions. The optimization goal of this objective function is to reduce the damage to the ecological environment and ensure the sustainability of resource development.

[0124] (2) Dual-objective optimization algorithm

[0125] To balance the accuracy of ore body prediction and environmental risk, the invention uses the Pareto optimal solution set (Pareto Optimal Solution Set) to solve the dual-objective optimization problem. In multi-objective optimization, the Pareto frontier refers to the solution set that cannot be further improved without sacrificing another objective. The specific implementation steps are as follows:

[0126] ① Weight adjustment: Adjust the weights λ1 and λ2 of the two objective functions to control the balance between ore body prediction accuracy and environmental risk:

[0127] ,

[0128] where λ1 and λ2 represent the weight coefficients of ore body prediction and environmental protection, respectively, and satisfy λ1+λ2=1.

[0129] ② Pareto optimization process: Use multi-objective optimization algorithms such as particle swarm optimization (PSO) or genetic algorithm (GA) to find a set of Pareto optimal solutions in the objective function space, representing the best balance point between ore body prediction and environmental protection under given constraints. Each solution corresponds to a different mining area and environmental impact assessment, and the final model will output a set of ore potential areas and environmental risk sensitivity maps.

[0130] ③Non-dominated Sorting Genetic Algorithm II (NSGA-II) algorithm is introduced to ensure the balance of different solutions in the optimization process. The algorithm evaluates the superiority of solutions through non-dominated sorting and crowding distance calculation. Specifically, NSGA-II performs non-dominated sorting according to the relative superiority of solutions, and selects the optimal solution by calculating the crowding distance of each solution (i.e. the distance in the objective space from other solutions). In this process, the Pareto optimization algorithm and the NSGA-II algorithm are used side by side, the Pareto optimization algorithm is used to generate a Pareto front solution set for the problem, which represents different trade-offs between ore body prediction accuracy and environmental risk, and the NSGA-II further optimizes on this basis, by calculating the crowding distance between solutions, to ensure the diversity of the solution set and avoid selecting too similar solutions. Through this method, an effective trade-off between maximizing ore body prediction accuracy and minimizing environmental risk can be achieved, so as to select a set of optimal solutions for practical application. Finally, based on the evaluation results of the Pareto front and NSGA-II, combined with the crowding distance calculation, an optimal balance point is determined, i.e. to minimize environmental risk while ensuring accuracy, to meet the needs of practical application.

[0131] (3) Result output

[0132] The dual-objective optimization method of the present application will eventually output two key results:

[0133] Ore body potential distribution map: Through the 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;

[0134] Environmental sensitivity risk map: Through the environmental risk assessment model, the ecological risk of each mining area is displayed, identifying ecologically fragile and sensitive areas, and pointing out the possible environmental impact of mining activities;

[0135] Through the above optimization, the present application can achieve an efficient balance between prospecting accuracy and ecological protection, ensuring the sustainable development of mineral resources while minimizing environmental damage.

[0136] S7. The accuracy of the ore body prediction is verified by drilling activities, and the drilling feedback data is introduced into the depth Q learning to calculate the Q function. The ore body prediction model adjusts the neural network weights after each drilling feedback to adjust and optimize itself, and finally obtains the prediction model:

[0137] In the actual mineral resources exploration process, the mining environment and geological conditions have great uncertainty, therefore, the ore body prediction model needs to have dynamic adaptability to continuously improve the prediction accuracy and effectively respond to new data and actual exploration results feedback. The application realizes the dynamic optimization and iterative update of the model by introducing the reinforcement learning (RL) mechanism, forming a closed-loop mechanism of "remote sensing prediction - field verification - model optimization". The specific steps are as follows:

[0138] (1) Fusion of drilling feedback data

[0139] In the process of mineral exploration, drilling activities are usually used to verify the accuracy of ore body prediction. After each drilling, some geological samples and data will be obtained. These drilling feedback data will be input into the model as actual labels, thus forming a data closed loop.

[0140] (2) Dynamic optimization of reinforcement learning

[0141] The application adopts deep Q learning (DQN) for reinforcement learning optimization. DQN combines Q-learning and deep neural network (DNN), which uses neural network to approximate Q function, so that the model can process high-dimensional and complex remote sensing data and geological data, and adjust its prediction strategy in real time.

[0142] The reward function of reinforcement learning is used to guide the decision of the model, ensuring that the environmental impact is minimized while optimizing the ore body prediction. The reward function provides immediate feedback, measuring the effect of each action. The Q function (quality function) is used to estimate the long-term return of taking a certain action in a certain state in reinforcement learning. Q function evaluates the long-term value of each decision by accumulating the immediate feedback provided by the reward function, guiding the model to choose the optimal strategy. Specifically,

[0143] The Q function expression is as follows:

[0144] ,

[0145] Where represents the expected value, which represents the average value of all possible rewards under the given state and action. represents the future time step or termination time. is the discount factor, usually between 0 and 1, representing the influence of future rewards. A smaller will make the model focus more on short-term rewards, and a larger will make the model focus more on long-term rewards. represents the initial state s, i.e. starting from state s. represents the initial action a, i.e. taking action a from state s.

[0146] The reward function is defined as:

[0147] ,

[0148] 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.

[0149] The update formula for the Q function is as follows:

[0150] ,

[0151] 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.

[0152] 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.

[0153] 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.

[0154] (3) Establishment of a closed-loop mechanism

[0155] 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:

[0156] Remote sensing prediction → Drilling verification → Data feedback → Reinforcement learning optimization → Remote sensing prediction

[0157] 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.

[0158] (4) Application effects

[0159] ① Continuous optimization and adaptive improvement:

[0160] By introducing deep Q-learning (DQN), the ore body prediction model can adjust and optimize itself after each drilling feedback. With the addition of new data each time, the model can maintain a better balance between ore body prediction and environmental protection, improving prediction accuracy;

[0161] ② Improve ore body prediction accuracy and environmental protection:

[0162] Dynamic optimization of reinforcement learning enables the model to adaptively adjust its prediction strategy while reducing environmental risks associated with mineral extraction. Through continuous feedback updates, the model continuously learns to achieve the best balance between ore body prediction accuracy and environmental protection effect;

[0163] ③ Closed-loop feedback mechanism:

[0164] Through the closed-loop mechanism of "remote sensing prediction - field verification - model optimization", the model can gradually improve in actual exploration, and each feedback data helps the model to learn and optimize itself in real time. This mechanism ensures more accurate ore body prediction and the ability to adapt to geological and environmental changes in different mining areas.

[0165] Example 2: Green mineral prediction device based on cross-domain adaptation and multi-modal fusion, comprising a data acquisition module for acquiring multi-source satellite remote sensing data, unmanned aerial low-altitude high-resolution data and geochemical sample data, and preprocessing the data;

[0166] Transfer learning module, through adversarial training, learn the model parameters that can be transferred from the source domain, and initialize these transferable model parameters in the Transformer encoder of the target region;

[0167] Multi-modal fusion module, input the obtained data into the modal gating network to calculate the weight of each modality, and obtain multi-modal fusion data through weighted fusion;

[0168] Graph construction module, extract features of each modality to construct graph structure, construct line graph, and use graph convolution network to enhance spatial information;

[0169] Spatial and spectral feature extraction module, used to extract the spatial topology structure and spectral features of the mining area through the Transformer encoder embedded in the graph convolution network, and calculate through the cross-modal consistency loss function to ensure the logical unity of different modalities in spatial and spectral features;

[0170] Prediction module, obtain the probability of mineralization through the prediction head;

[0171] a prediction accuracy calculation module configured to calculate the prediction accuracy of the ore body by the prediction accuracy of the potential area of the ore body;

[0172] an environmental risk assessment module configured to construct an environmental risk assessment model based on various environmental indicators for the environmental risk that may be caused by the mineral exploitation;

[0173] a multi-objective optimization module configured to establish two objective functions of the prediction accuracy of the ore body and the environmental risk, and to realize the calculation of "maximization of the prediction accuracy of the ore body" and "minimization of the environmental risk" through a multi-objective optimization algorithm;

[0174] a drilling feedback module configured to verify the accuracy of the prediction of the ore body through drilling activities and to provide drilling feedback data;

[0175] a reinforcement learning and optimization updating module configured to introduce a deep Q learning to calculate a Q function, and to adjust the neural network weight through the Q function to perform self-adjustment and optimization of the ore body prediction model after each drilling feedback.

[0176] The above is a further description of the present application in combination with specific embodiments, and the protection scope of the present application is not limited thereto.

Claims

1. A mineral green prediction method based on cross-domain adaptation and multi-modal fusion, characterized in that, The steps include the following: 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 for adversarial training, the transfer learning module includes a feature extractor, the feature extractor is connected with a main task predictor and a domain classifier designed in parallel through an extreme inversion layer, the feature extractor extracts preliminary features from the 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 capability of the model, and the main task predictor uses the shared features for ore body potential prediction; In this way, the transfer learning module learns transferable model parameters, including the convolution 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 of the target area for initialization; 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 by using a graph convolution network; S5. The enhanced multi-modal data are input into a Transformer encoder embedded with a graph convolution network, the spatial topological structure and spectral features of the mining area are extracted, a cross-modal consistency loss function is introduced to ensure the logical unity of different modalities in spatial and spectral features, and the extracted features are input into a prediction head to obtain a metallogenic potential probability map; the Transformer encoder embedded with a graph convolution network processes the multi-modal data and spatial topological structure of the mining area by combining the graph convolution network and the Transformer self-attention mechanism, first, the GCN constructs a graph structure according to the geological structure information of the mining area, represents each data point as a node in the graph, and performs spatial feature aggregation, then the Transformer encoder receives the node features processed by the GCN, uses the self-attention mechanism to capture global dependency and long-range information, and enhances the connection between multi-modal data; S6. Calculate the ore body prediction accuracy through the prediction accuracy of the ore body potential area, construct an environmental risk assessment model based on various environmental indicators for the environmental risk that may be caused by mining, establish two objective functions of ore body prediction accuracy and environmental risk, and realize the calculation of "maximum ore body prediction accuracy" and "minimum 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 weights of the ore body prediction model through the Q function after each drilling feedback to realize self-adjustment and optimization, and obtain the final prediction model.

2. The mineral green prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, 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 modality, M is the number of modalities, The modal gating network calculates the weight for each modality as follows: ， wherein, is an environmental factor vector, is a gating function, input is the modal feature x i with environmental conditions , output is the importance score of the modal, The final multi-modal fusion result is: 。 3. The mineral green prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, The Grad-CAM explainability module is introduced in step S5 to generate a saliency map and guide the model learning as a regularization constraint. In the training process, the Grad-CAM generates a saliency map S(x) and uses it as a regularization term: ， where S(x) is the saliency map representing the key areas 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.

4. The mineral green prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, The calculation of the cross-modal consistency loss function is as follows: assuming that the prediction output of different modalities at the same position is: wherein represents the prediction function corresponding to the i-th modality, and represents the prediction function corresponding to the j-th modality. The cross-modal consistency loss function is defined as: ， wherein is the number of modalities, representing different types of data sources or different input modalities, i, j are the corresponding modalities.

5. The green mineral prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, In step S6, the classification accuracy is calculated to determine the prediction accuracy of the ore body potential area, and the calculation formula is as follows: ， wherein is a true case, is a true counter case, is a false positive case, is a false counter case; Maximize the accuracy of ore body prediction: ， wherein is the model predicted ore body potential zone location, i.e. the prediction result; is the actual ore body annotation data, i.e. the true label.

6. The green mineral prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, In step S6, the environmental risk assessment model is: ， wherein is the weight of different environmental factors, is the environmental risk score of the ith region, and N is the number of regions.

7. The green mineral prediction method based on cross-domain adaptation and multi-modal fusion according to claim 1, characterized in that, In step S7, the Q function expression is as follows: ， wherein denotes the expected value, the average value of all possible rewards under a given state and action, denotes the time step or termination time in the future, is a discount factor between 0 and 1, indicating the influence of future rewards, R t is the reward function at time t, S0= S denotes the initial state S, i.e. starting from state S, a0= a denotes the initial action a, i.e. taking action a from state s; The reward function R is defined as: ， wherein for the improvement of the prediction accuracy of the model; for the environmental risk introduced in the prediction process, for minimizing the environmental impact; w1, w2 are weight coefficients; In each round of ore body prediction, the model calculates the Q function value of each action through a deep neural network and selects the action with the maximum Q function value; The update formula of the Q function is as follows: ， where, is the Q-value of the current state-action pair, representing the expected long-term return of the action in the current state, is the reward function at time t+1, representing the environment feedback obtained after performing action a t in state S t , is the discount factor, measuring the importance of future rewards, represents the maximum value among the Q-values corresponding to all possible actions a' in the next state S t+1 , is the learning rate, controlling the speed of Q-value updates.

8. The mineral green prediction device based on cross-domain adaptation and multi-modal fusion, adopting the mineral green prediction method based on cross-domain adaptation and multi-modal fusion according to any one of claims 1-7, characterized in that, It includes a data acquisition module for acquiring multi-source satellite remote sensing data, unmanned aerial low-altitude high-resolution data, and geochemical sample data, and pre-processing the data; The transfer learning module learns the transferable model parameters of the source domain through adversarial training, and initializes these transferable model parameters in the Transformer encoder of the target region; The multi-modal fusion module inputs the obtained data into the modal gating network to calculate the weight of each modality, and obtains the multi-modal fusion data through weighted fusion; The graph construction module extracts the features of each modality to construct the graph structure, constructs the line graph, and enhances the spatial information using the graph convolution network; The spatial and spectral feature extraction module is used to extract the spatial topology structure and spectral features of the mining area through the Transformer encoder embedded in the graph convolution network, and to ensure the logical unity of different modalities in spatial and spectral features through the cross-modal consistency loss function calculation; The prediction module obtains the ore-forming probability through the prediction head; The prediction accuracy calculation module is used to calculate the ore body prediction accuracy through the prediction accuracy of the ore body potential area; The environmental risk assessment module is used to construct an environmental risk assessment model based on various environmental indicators for the environmental risks that may be caused by mineral exploitation; The multi-objective optimization module is used to establish two objective functions of ore body prediction accuracy and environmental risk, and to realize "maximization of ore body prediction accuracy" and "minimization of environmental risk" calculation through multi-objective optimization algorithm; The drilling feedback module verifies the accuracy of the ore body prediction through drilling activities and provides drilling feedback data; The reinforcement learning and optimization update module introduces the deep Q learning to calculate the Q function, and the ore body prediction model adjusts the neural network weights for self-adjustment and optimization after each drilling feedback.

Images