A Mineral Prospect Prediction Method Based on Multimodal Autoencoder

Abstract

This invention discloses a method for predicting mineral potential areas based on a multimodal autoencoder, addressing the problems of poor statistical models and single-modal recognition capabilities. The method includes: acquiring multimodal data of the metallogenic belt to be predicted and preprocessing the multimodal data; constructing a multimodal deep autoencoder; the deep autoencoder comprising an encoder part and a reconstructor part connected in sequence; acquiring the preprocessed multimodal data and performing modal feature encoding and weighted fusion processing to obtain standardized fusion features; training the multimodal deep autoencoder using the standardized fusion features; setting a dynamic threshold for anomaly detection and target area determination, and using a spatial clustering algorithm to divide potential areas; evaluating the potential areas to obtain the mineral potential areas. Through the above scheme, this invention has advantages such as simple logic and high accuracy and reliability, and has high practical and promotional value in the fields of mineral resource exploration and geological big data analysis technology.

Description

Technical Field

[0001] This invention relates to the field of mineral resource exploration and geological big data analysis technology, and in particular to a method for predicting prospective mineral areas based on a multimodal autoencoder. Background Technology

[0002] Predicting potential mineral deposits requires comprehensive information from multiple sources, including geology, geochemistry, and geophysics. Traditional methods (such as weighted evidence, logistic regression, and multivariate statistical discriminant analysis) rely on expert experience to assign weights or select variables, making them susceptible to subjective bias. Furthermore, these models typically assume linear relationships between input variables, while actual mineralization processes are often highly nonlinear and multi-scale coupled. The limited expressive power of traditional statistical models leads to insufficient prediction accuracy and generalization ability.

[0003] Currently, deep autoencoders (DAEs) possess unsupervised feature extraction and automatic dimensionality reduction capabilities. By reconstructing input data and minimizing reconstruction errors, they can automatically capture and compress features, demonstrating advantages in tasks such as anomaly detection and rare mineralization identification. However, existing DAE-based methods often rely on single-modal data (such as geochemical data alone), making it difficult to comprehensively consider geological structures, geophysical fields, and remote sensing features. This results in an inability to fully characterize complex mineralization processes, and predictive reliability remains insufficient. For example, Zhang Xin's paper, "Research on Prediction of Metallogenic Prospects Based on Deep Learning," combines data augmentation, residual structures, graph neural networks, and Transformers to construct graph data from geochemical element data. Based on a residual graph neural network model, it performs feature extraction and classification on the data.

[0004] Therefore, there is an urgent need to propose a logically simple, accurate, and reliable method for predicting mineral potential areas based on multimodal autoencoders. This method constructs feature encoders for different modalities, mapping multi-source data from geology, geophysics, geochemistry, and remote sensing to a shared latent space, achieving multimodal fusion. Then, combined with the nonlinear feature extraction capabilities of deep autoencoders, it effectively identifies rare mineralization anomalies, improving the model's adaptability and prediction accuracy in complex geological contexts. This invention utilizes reconstructed input data and minimizes reconstruction errors. The model can also perform anomaly detection, reducing reliance on human experience and improving the objectivity of prediction results, providing a more scientific means for the quantitative identification of mineralized target areas. Summary of the Invention

[0005] To address the above problems, the present invention aims to provide a method for predicting mineral prospecting areas based on a multimodal autoencoder. The technical solution adopted by the present invention is as follows:

[0006] A method for predicting mineral prospecting areas based on a multimodal autoencoder includes the following steps:

[0007] Step S1: Obtain multimodal data of the metallogenic belt to be predicted, and preprocess the multimodal data; the multimodal data includes geological data, geophysical data, geochemical data, and remote sensing information;

[0008] Step S2: Construct a multimodal deep autoencoder; the deep autoencoder includes an encoder part and a reconstructor part connected in sequence; the encoder part includes a single-modal feature encoder, a multimodal fusion layer and a main encoder connected in sequence; the reconstructor part includes a reconstructor layer, an anomaly detection layer and a clustering and visualization layer connected in sequence; the main encoder is connected to the reconstructor layer;

[0009] Step S3: Obtain the preprocessed multimodal data, and use modal feature encoding and weighted fusion processing to obtain standardized fusion features;

[0010] Step S4: Train the multimodal deep autoencoder using standardized fusion features;

[0011] Step S5: Set a dynamic threshold, perform anomaly detection and target area determination, and use a spatial clustering algorithm to divide the potential area;

[0012] Step S6: Evaluate the potential area to obtain the mineral prospective area.

[0013] In step S1, the preprocessing of multimodal data includes the following steps:

[0014] Step S101, Data Collection: Collect multi-source data for a metallogenic belt, including geological, geophysical, geochemical, and remote sensing information. Geological data comes from a 1:200,000 scale geological map, including lithology, fault structures, and the distribution of intrusive intrusive bodies. Geophysical data comes from airborne magnetic surveys with a resolution of 2 km; geochemical data comes from the National Geochemical Survey Program, containing the concentrations of 39 elements. Remote sensing data comes from Landsat 8 imagery, extracting features such as iron staining index and vegetation index.

[0015] Step S102: Project the multimodal data to the same coordinate system: Project the different data sources mentioned above to the same coordinate system (such as WGS84 / UTM), and use GIS tools to convert the vector data and raster data into a unified 1km×1km grid format.

[0016] Step S103: Under the same coordinate system, geological data, geophysical data, geochemical data, and remote sensing information are subjected to z-score normalization, followed by removal of outliers and missing values. Incomplete samples are discarded or missing features are filled using nearest neighbor interpolation. The normalization formula is as follows:

[0017]

[0018] in, This represents the data after standardization. This refers to data before it has been annotated (such as geological data, geophysical data, geochemical data, and remote sensing information). The mean of the feature; The standard deviation of a feature.

[0019] Step S104: Concatenate the standardized data into a multimodal feature matrix X. The expression for this multimodal feature matrix is:

[0020]

[0021] in, Represents a submatrix of geological features; Represents the geophysical characteristic submatrix; Represents the geochemical characteristic submatrix; This represents the remote sensing feature submatrix.

[0022] In step S2, the deep autoencoder includes an encoder section and a reconstructor section connected in sequence. The encoder section includes a single-modal feature encoder, a multimodal fusion layer, and a main encoder connected in sequence, while the reconstructor section includes a reconstructor layer, an anomaly detection layer, and a clustering and visualization layer connected in sequence; the main encoder is connected to the reconstructor layer.

[0023] Here, the encoder adopts a layer-by-layer dimensionality reduction structure: 128-64-32-16. The input is the modality fusion feature h, and the output is the encoded low-dimensional latent representation z, whose expression is:

[0024]

[0025] Where Z represents the low-dimensional latent representation obtained after encoding; This represents the main encoder mapping function; This represents the encoder weight matrix, used to linearly transform the input features into latent features; This represents the bias vector, used for offset compensation before nonlinear activation.

[0026] In addition, the reconstructor layer adopts symmetric dimensionality increase: 16-32-64-128, and its expression is:

[0027]

[0028] in, This indicates that the input is the unified latent feature after fusion. The reconstructed feature vector; This represents the decoder mapping function, used to restore the low-dimensional latent feature z to the original feature space; This represents the weight matrix of the decoder, used for feature upscaling; This represents the bias vector of the decoder.

[0029] In addition, the reconstructor layer calculates and optimizes the reconstruction error, and includes a loss function. Its expression is:

[0030]

[0031] Where N represents the total number of training samples; Represents the ith modality fusion feature; Represents the ith modal fusion feature The corresponding reconstruction results.

[0032] Introducing sparsity constraints, its expression is:

[0033]

[0034] Where λ represents the regularization coefficient, which controls the degree of influence of the sparse term on the total loss.

[0035] Here, the Adam optimizer is used with a learning rate of 0.001 and a batch size of 64 for 300 rounds of iterative training.

[0036] In this technique, after training is complete, some parameters of the encoder are frozen for subsequent potential region identification. The reconstructor is only used to calculate the reconstruction error during training and does not participate in prediction during the inference phase.

[0037] Step S301, in this embodiment, single-modal feature encoding uses a three-layer fully connected neural network connected sequentially for each modality m, and employs layer-by-layer dimensionality reduction to extract modal features. Its expression is:

[0038]

[0039] Where σ(·) represents the ReLU function; This represents the neural network weight matrix corresponding to the m-th mode; This represents the input feature matrix for the m-th mode; This represents the bias term of the m-th mode, used to introduce a nonlinear offset; the value of m ranges from [1,2,3,4].

[0040] Step S302: In each modal encoder, the number of nodes is reduced layer by layer (e.g., 128→64→32), mapping the high-dimensional features to low-dimensional latent representations. Each modal encoder is trained separately to obtain stable feature representations.

[0041] Step S303: The multimodal fusion layer employs weighted fusion, combining the latent features of the four modalities in a weighted manner, expressed as follows:

[0042]

[0043] in, The unified latent feature after fusion is represented by the encoder output feature vector of the m-th modality. The weighted sum; Represents the weighting coefficients for the m-th modal feature; This represents the encoder output feature vector for the m-th mode.

[0044] Step S304: Normalize the fused h once (min-max scaling or layer normalization) to ensure that the features input to the autoencoder are within a similar scale range.

[0045] In step S5, the multimodal deep autoencoder is trained using standardized fusion features, and the reconstruction error is calculated using the trained multimodal deep autoencoder, expressed as follows:

[0046]

[0047] in, Represents the sample of the i-th grid. The reconstruction error.

[0048] In addition, this technology uses a preset dynamic threshold for anomaly detection and target area determination, and employs a spatial clustering algorithm for potential area division, including:

[0049] Based on the distribution of reconstruction errors of samples corresponding to standardized fusion features, the average reconstruction error of training samples is obtained. and reconstruction error standard value ; Reconstruct the average error using training samples and reconstruction error standard value Dynamic threshold for anomaly detection The adjustment is made, and its expression is:

[0050]

[0051] like Then the sample of the i-th grid If anomalies are identified as potential mining anomalies, spatial clustering methods (such as DBSCAN) are used to merge adjacent anomaly rasters into potential areas, generating potential area boundary markers that overlap with known deposits and newly discovered anomaly areas.

[0052] In step S6, the potential area is evaluated to obtain a mineral prospective area, including the following steps:

[0053] Step S601: Convert the reconstruction error into a potential score of 0 to 1 and draw a grid heatmap.

[0054] Step S602: Based on the sorting of reconstruction errors, target area identification and labeling are performed on the grid heatmap.

[0055] Step S603: Use the AUC index evaluation model to evaluate the index in order to obtain the prospective mineral area.

[0056] Step S604: Geochemical sampling is performed on several newly discovered anomalous target areas to confirm whether there are significant metal element anomalies.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] This invention employs a weighted fusion mechanism, utilizing learnable weights normalized by softmax or gated attention fusion. Traditional autoencoders often fuse features from different modalities through simple concatenation or averaging, easily leading to weight imbalances between modalities. This invention achieves joint modeling of data from different sources through a modal feature encoder, overcoming the limitation of traditional single-modal methods in comprehensively depicting complex mineralization backgrounds. It ensures the adaptability and interpretability of each modal feature during fusion, avoiding the "dominance effect" of a single modality on the overall feature representation.

[0059] This invention utilizes the multi-layer nonlinear mapping and reconstruction mechanism of deep autoencoders to extract key information hidden in high-dimensional data without complex manual feature engineering, significantly reducing reliance on expert experience.

[0060] This invention achieves accurate identification of rare mineralization anomalies through quantitative analysis of reconstruction errors, effectively distinguishes mineralization signals from background noise, and has a prediction accuracy significantly higher than that of traditional statistical models.

[0061] This invention, after the encoder outputs the latent space, not only utilizes it for anomaly detection but also proposes a strategy to freeze encoder parameters after training for direct extraction of latent features in subsequent inference stages, thereby improving the model's stability and generalization in practical applications. Simultaneously, this invention introduces sparse constraints (such as regularization and KL divergence sparsity) to enhance the discriminative power of the latent representation. This design differs from the traditional autoencoder optimization method that relies solely on mean squared error.

[0062] This invention proposes a dynamic threshold setting method based on the reconstruction error distribution, and combines an improved method of controlling the false positive rate using quantile thresholds or ROC curves, thereby enhancing the robustness and flexibility of anomaly detection.

[0063] This technology, based on an autoencoder, further combines the DBSCAN spatial clustering method to aggregate anomalous rasters, forming a mineral prospect prediction map. This invention employs a holistic workflow of "model output—spatial clustering—map preparation," which differs from traditional autoencoder tasks that only focus on numerical anomaly detection, thus enhancing the model's usability and engineering value in geological exploration.

[0064] In summary, this invention has the advantages of simple logic and high accuracy and reliability, and has high practical and promotional value in the fields of mineral resource exploration and geological big data analysis technology. Attached Figure Description

[0065] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope of protection. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a logic flowchart of the present invention.

[0067] Figure 2 This is a schematic diagram illustrating the anomaly detection and mineral prospect prediction of the present invention.

[0068] Figure 3 This is the result of target region identification and annotation in this invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0070] In this embodiment, the term "and / or" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0071] The terms "first" and "second," etc., used in the specification and claims of this embodiment are used to distinguish different objects, not to describe a specific order of objects. For example, "first target object" and "second target object," etc., are used to distinguish different target objects, not to describe a specific order of target objects.

[0072] In the embodiments of this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design that is described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or design. Specifically, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0073] In the description of the embodiments in this application, unless otherwise stated, "multiple" means two or more. For example, multiple processing units means two or more processing units; multiple systems means two or more systems.

[0074] like Figures 1 to 3 As shown, this embodiment provides a method for predicting mineral prospecting areas based on a multimodal autoencoder, which includes the following steps:

[0075] Step 1: Experimental area and data preparation:

[0076] A typical metallogenic belt in southeastern my country was selected as the study area. The main metallogenic factors in this area are several large iron ore deposits, a certain period of granite bodies and regional faults, which have typical skarn-type iron polymetallic mineralization characteristics.

[0077] Geological data: 1:200,000 scale geological map, including lithology, fault structures, and intrusive rock distribution. Geophysical data: Airborne magnetic survey data (2km resolution), interpolated to a 1km×1km grid using inverse distance weighting (IDW). Geochemical data: Contents of 39 elements (e.g., Fe, Cu, Zn, Pb), from the National Geochemical Survey Plan, with a sampling density of approximately 1 sample / 4km². Remote sensing data: Ferrous Index and NDVI extracted from Landsat 8, used to characterize surface alteration information.

[0078] All the above data were uniformly projected and standardized into 1km×1km raster data, and z-score normalization was applied.

[0079] The second step is model structure and feature fusion:

[0080] Here, an improved multimodal autoencoder (MDAE) is used for latent feature extraction and anomaly detection.

[0081] First, modal feature encoding: For the four types of data—geological, geophysical, geochemical, and remote sensing—a three-layer fully connected encoder is established for each, with the expression as follows:

[0082]

[0083] Where m=1,2,3,4 are modal indices.

[0084] Secondly, the multimodal features are weighted and summed, and the expression is as follows:

[0085]

[0086] in, The unified latent feature after fusion is represented by the encoder output feature vector of the m-th modality. The weighted sum; Represents the weighting coefficients for the m-th modal feature; This represents the encoder output feature vector for the m-th mode.

[0087] The third step is the construction and training of the deep autoencoder:

[0088] The deep autoencoder comprises an encoder section and a reconstructor section connected in sequence. The encoder section includes a single-modal feature encoder, a multimodal fusion layer, and a main encoder connected in sequence, while the reconstructor section includes a reconstructor layer, an anomaly detection layer, and a clustering and visualization layer connected in sequence.

[0089] The encoder part compresses the fused high-dimensional features h into a low-dimensional latent representation z, the expression of which is:

[0090] .

[0091] The reconstructor part reconstructs the latent representation z into an approximation of the original features. It is used to calculate the reconstruction error, and its expression is:

[0092] .

[0093] In addition, the reconstructor layer calculates and optimizes the reconstruction error, and includes a loss function. Its expression is:

[0094] .

[0095] Introducing sparsity constraints, its expression is:

[0096] .

[0097] Here, the optimizer is Adam (learning rate 0.001, batch size 64). Iterations: converged after 300 epochs. After training, the prediction phase only needs to use the encoder part and the reconstruction error for calculation, without relying on the reconstructor output.

[0098] Step 4: Anomaly Detection and Potential Area Determination

[0099] First, calculate the reconstruction error for each grid cell: And set a dynamic threshold: ;like Then the sample of the i-th grid If anomalies are identified as potential mining anomalies, spatial clustering methods (such as DBSCAN) are used to merge adjacent anomaly rasters into potential areas, generating potential area boundary markers that overlap with known deposits and newly discovered anomaly areas.

[0100] In this embodiment, a quantile threshold is used, which can replace... To enhance the robustness of anomaly detection, its expression is:

[0101]

[0102] Where E represents the set of reconstruction errors for all samples; q represents the quantile threshold level. At that time, the threshold The 95th to 99th percentiles of the reconstruction error distribution are used to identify potential anomalies.

[0103] Step 5: Experimental Results and Verification

[0104] Figure 3 Solid circles represent newly discovered potential target areas, while dashed circles indicate areas overlapping with previous target areas. Here, high-potential areas are mainly distributed around the core metallogenic belt of the study area, closely coinciding with known iron ore locations. The model identifies four potential areas, one of which overlaps with a known deposit, and three are newly discovered anomalous target areas. Field validation shows that one of the new target areas exhibits a significant Fe element anomaly. The AUC value is 0.91, an improvement of approximately 12% compared to the traditional weighted evidence method.

[0105] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any changes made based on the design principles of the present invention, or any non-creative modifications made thereon, shall fall within the scope of protection of the present invention.

Claims

1. A method for predicting mineral prospecting areas based on a multimodal autoencoder, characterized in that, Includes the following steps: Acquire multimodal data of the metallogenic belt to be predicted, and preprocess the multimodal data; the multimodal data includes geological data, geophysical data, geochemical data, and remote sensing information; A multimodal deep autoencoder is constructed; the deep autoencoder includes an encoder part and a reconstructor part connected in sequence; the encoder part includes a single-modal feature encoder, a multimodal fusion layer and a main encoder connected in sequence; the reconstructor part includes a reconstructor layer, an anomaly detection layer and a clustering and visualization layer connected in sequence; the main encoder is connected to the reconstructor layer; Preprocessed multimodal data is obtained, and modal feature encoding and weighted fusion are used to obtain standardized fusion features; A multimodal deep autoencoder is trained using standardized fusion features; Preset dynamic thresholds for anomaly detection and target area determination, and use spatial clustering algorithms to divide potential areas; The potential areas are assessed to identify prospective mineral deposit areas; The single-modal feature encoder employs a three-layer fully connected neural network connected sequentially, and uses layer-by-layer dimensionality reduction to extract modal features. Its expression is as follows: Where σ(·) represents the ReLU function; This represents the neural network weight matrix corresponding to the m-th mode; This represents the input feature matrix for the m-th mode; This represents the bias term for the m-th mode; m ranges from [1, 2, 3, 4]. The multimodal fusion layer employs weighted fusion, and its expression is as follows: ,in, This indicates the unified potential characteristics after fusion; Represents the weighting coefficients for the m-th modal feature; This represents the encoder output feature vector for the m-th mode; The main encoder employs a layer-by-layer dimensionality reduction approach, and its expression is as follows: , where Z represents the low-dimensional latent representation obtained after encoding; This represents the main encoder mapping function; Represents the encoder weight matrix; This represents the bias vector, used for offset compensation before nonlinear activation; The reconstructor layer employs symmetric dimensionality increase, and its expression is: ； in, This indicates that the input is the unified latent feature after fusion. The reconstructed feature vector; This represents the decoder mapping function; The weight matrix represents the decoder; This represents the bias vector of the decoder.

2. The mineral prospect prediction method based on a multimodal autoencoder according to claim 1, characterized in that, Preprocessing multimodal data includes the following steps: Project multimodal data onto the same coordinate system; Under the same coordinate system, geological data, geophysical data, geochemical data and remote sensing information are processed using z-score normalization; The standardized data are concatenated into a multimodal feature matrix X, the expression of which is: ,in, Represents a submatrix of geological features; Represents the geophysical characteristic submatrix; Represents the geochemical characteristic submatrix; This represents the remote sensing feature submatrix.

3. The method for predicting mineral prospecting areas based on a multimodal autoencoder according to claim 2, characterized in that, The reconstructor layer calculates and optimizes the reconstruction error; the reconstructor layer includes a loss function. Its expression is: , where N represents the total number of training samples; Represents the ith modality fusion feature; Represents the ith modal fusion feature The corresponding reconstruction results; Introducing sparsity constraints, its expression is: ； Where λ represents the regularization coefficient.

4. The mineral prospect prediction method based on a multimodal autoencoder according to claim 3, characterized in that, A multimodal deep autoencoder is trained using standardized fusion features, and the reconstruction error is calculated using the trained multimodal deep autoencoder. The expression for this error is as follows: ； in, Represents the sample of the i-th grid. The reconstruction error.

5. The mineral prospect prediction method based on a multimodal autoencoder according to claim 4, characterized in that, A preset dynamic threshold is used for anomaly detection and target region determination. A spatial clustering algorithm is employed for potential region delineation, including: Based on the distribution of reconstruction errors of samples corresponding to standardized fusion features, the average reconstruction error of training samples is obtained. and reconstruction error standard value ; Reconstruct the average error using training samples and reconstruction error standard value Dynamic threshold for anomaly detection The adjustment is made, and its expression is: ； like Then the sample of the i-th grid The anomalies identified as potential mining anomalies are then merged using a spatial clustering algorithm to obtain potential areas.

6. The method for predicting mineral prospective areas based on a multimodal autoencoder according to claim 5, characterized in that, The assessment of potential areas to identify mineral prospective areas includes the following steps: The reconstruction error was converted into a potential score of 0 to 1, and a grid heatmap was plotted. Based on the sorting of reconstruction errors, target areas are identified and labeled in the grid heatmap; The AUC index evaluation model is used to evaluate the indicators in order to obtain the prospective mineral areas.

Images