Geochemical Anomaly Identification Method Using Unsupervised Spatial-Spectral Dual-Branch Deep Learning Model

Abstract

This invention provides an unsupervised spatial-spectral dual-branch deep learning model for geochemical anomaly identification. This model, based on generative adversarial networks (GANs), mainly consists of a spatial branch and a spectral branch. The spatial branch effectively captures geochemical spatial patterns and extracts spatial relationships between adjacent pixels using a GAN based on graph convolution and inverse graph convolution operations. Specifically, inverse graph convolution restores the original features of the data by using the difference between linear and graph convolutional layers. The spectral branch, composed of a long short-term memory network, characterizes geochemical element combinations. This invention simultaneously considers the spatial and spectral features of geochemical anomalies, optimizing the spatial generator, spatial discriminator, spectral generator, and spectral discriminator through adversarial learning. This trains the model and delineates geochemical anomalies, improving the accuracy and effectiveness of deep learning models for geochemical anomaly identification.

Description

Technical Field

[0001] This invention relates to the field of Earth science technology, specifically to a method for geochemical anomaly identification using an unsupervised spatial-spectral dual-branch deep learning model. Background Technology

[0002] The spatial distribution patterns of geochemical elements are influenced by a variety of factors, including geological background, mineralization intensity, and weathering degree. Furthermore, due to the complexity of geological systems and the diversity of mineralization processes, geochemical models are often complex and varied. Big data and deep learning provide new research paradigms and effective tools for the analysis of exploration geochemical data, effectively mining complex, nonlinear geoscientific spatial data and exhibiting strong classification and predictive capabilities.

[0003] Generative Adversarial Networks (GANs) consist of a generator (G) and a discriminator (D), which are trained simultaneously in an adversarial manner. The generator takes random noise as input and synthesizes pseudo-samples by learning the data distribution, aiming to generate samples indistinguishable from real data. D is a binary classifier used to distinguish between real data samples and pseudo-samples generated by the generator, aiming to correctly classify real and fake samples. G attempts to generate real samples indistinguishable from noisy data, while D attempts to distinguish between real and generated samples; this process is achieved through adversarial training. Graph Convolutional Networks (GCNNs) are graph-based neural networks, typically used to extract important spatial features from input data and adjacency matrices, while considering the spatial anisotropy of the data. Long Short-Term Memory (LSTM) networks are a variant of Recurrent Neural Networks (RNNs). While RNNs can receive hidden states from previous time points, LSTMs can solve the long-term dependency problem common in RNNs, effectively conveying and representing information in long-term sequences while avoiding the loss of useful information. Furthermore, LSTMs can also solve the vanishing or exploding gradient problems in RNNs. Compared to traditional RNNs, LSTMs introduce three gates (input gate, forget gate, and output gate) and a long-term memory. These mechanisms enable LSTMs to better handle long-term dependencies in sequences.

[0004] Existing models cannot efficiently identify and extract spatial and spectral distribution patterns of geochemical data simultaneously, nor can they fully consider the relationships between feature sequences. Limited sample size, high-dimensional features, and mixed geochemical information make geochemical anomaly identification a significant challenge. Therefore, developing effective methods to identify and extract spatial distribution patterns of geochemical data is crucial for mineral exploration. Currently, research on constructing unsupervised deep learning model structures that simultaneously consider the spatial and spectral features of geochemical data to improve the accuracy of geochemical anomaly identification is of considerable importance. Summary of the Invention

[0005] To address the limitations of existing models in efficiently identifying and extracting both spatial and spectral distribution patterns from geochemical data, and in fully considering the relationships between feature sequences, this invention provides an unsupervised spatial-spectral dual-branch deep learning model for geochemical anomaly identification. This model utilizes independent spatial and spectral branches to identify the spatial and spectral features of geochemical anomalies. The spatial branch is a generative adversarial network based on graph convolution, while the spectral branch is a generative adversarial network based on long short-term memory (LSTM). This model can simultaneously consider the spatial and spectral features of geochemical anomalies, thereby improving the accuracy of geochemical anomaly identification.

[0006] A geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model is proposed. This method generates new sample data by learning the spatial and spectral features of random input noise through a spatial generator and a spectral generator, respectively. The generator is optimized by minimizing the probability of a discriminator successfully identifying the generated data; and the discriminator is optimized by maximizing the difference between the generated data and the original data. The specific steps are as follows:

[0007] S1. Decompose the original geochemical data into spatial data X and spectral data R, and construct random noise data with a normal distribution that has the same structure as the original geochemical data as the input data of the generative adversarial network.

[0008] S2. Input random noise data into the constructed spatial generator and spectrum generator to generate spatial data. Spectral data

[0009] S3, Generate spatial data Generated spectral data The original spatial data x and the original spectral data R are input into the constructed spatial discriminator and spectral discriminator to obtain high-level spatial features and high-level spectral features.

[0010] S4. Construct a fusion network H to fuse the spatial high-level features and spectral high-level features output by the spatial discriminator and the spectral discriminator;

[0011] S5. Design loss functions to train the constructed generator, discriminator, and fusion network. The generator includes a spatial generator and a spectral generator, and the discriminator includes a spatial discriminator and a spectral discriminator, until Nash equilibrium is reached, that is, the loss functions of the generator and discriminator tend to stabilize. After training, the final unsupervised spatial-spectral dual-branch deep learning model is obtained, and this model is used for geochemical anomaly identification.

[0012] The beneficial effects of the technical solution provided by this invention are as follows: This invention proposes an unsupervised spatial-spectral dual-branch deep learning model for geochemical anomaly identification. First, the original geochemical data is decomposed into spatial data X and spectral data R. Random noise data with a normal distribution and the same structure as the original geochemical data is constructed as input data for a generative adversarial network. Then, the random noise data is input into the spatial generator and the spectral generator to generate spatial data. Spectral data Spatial data Spectral data The original spatial data X and original spectral data R are input into the spatial discriminator and spectral discriminator to obtain high-level spatial and spectral features. A fusion network H is constructed to fuse the high-level spatial and spectral features. Finally, the generator, discriminator, and fusion network are trained using a loss function, resulting in a final unsupervised spatial-spectral dual-branch deep learning model. This model can simultaneously consider both spatial and spectral geochemical features in areas with few known mineral deposits to identify geochemical anomalies, thus improving the accuracy of geochemical anomaly identification. Attached Figure Description

[0013] Figure 1 A flowchart of the geochemical anomaly identification method using an unsupervised spatial-spectral dual-branch deep learning model provided by the present invention;

[0014] Figure 2 This is a structural diagram of the geochemical anomaly identification method based on the unsupervised spatial-spectral dual-branch deep learning model provided by the present invention.

[0015] Figure 3 This is a schematic diagram of the inverse graph convolution operation in this invention;

[0016] Figure 4 This is a schematic diagram illustrating the working principle of the Long Short-Term Memory (LSTM) network in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0018] Please refer to Figure 1-2This invention provides an unsupervised geochemical anomaly identification method using a spatial-spectral dual-branch deep learning model. The method constructs spatial and spectral branches based on a generative adversarial network (GAN). The spatial branch effectively captures geochemical spatial patterns and extracts spatial relationships between adjacent pixels using a GAN based on graph convolution operations; inverse graph convolution restores the original features of the data through the difference between linear and graph convolutional layers. The spectral branch consists of a long short-term memory (LSTM) network to study the geochemical element combinations within a single pixel, thus simultaneously considering the spatial and spectral features of geochemical anomalies. Adversarial learning is used to optimize the spatial generator, spectral generator, spatial discriminator, and spectral discriminator, thereby training the model and identifying geochemical anomalies, improving the accuracy and effectiveness of the deep learning model in geochemical anomaly identification. The specific steps are as follows:

[0019] S1. Decompose the original geochemical data into original spatial data X and original spectral data R, and construct random noise data with the same structure as the original geochemical data (mean 0, standard deviation 1) as the input data of the generator. The random noise data is decomposed into spatial noise data x and spectral noise data r, and x and r are respectively used as the spatial generator G. S Spectrum Generator G U The input data. Taking a certain study area as an example, the original data has 15,450 data points, each with 39 features, generating random noise data with a normal distribution (mean of 0 and standard deviation of 1) of 15,450×39.

[0020] S2, Using graph convolutional layers f G and inverse graph convolutional layer f T Construct space generator G S In this model, the graph convolutional layer extracts latent high-level features z from spatially noisy data, and the inverse graph convolutional layer maps these high-level features back to the original input space. The inverse graph convolutional layer uses linear layers (i.e.,...) Figure 3 (projection layer in) f L (·) and graph convolutional layers (i.e.) Figure 3 (GCN layer in) f G (·) Restore the original noise features to obtain the synthesized spatial data. Figure 3 In the middle, z i h represents the latent high-level features extracted by the graph convolutional layer. i N represents the output of the projection layer. i This represents the neighborhood information aggregated by node i after passing through the GCN layer. A spectrum generator G is constructed using long short-term memory network layers. UThe first long short-term memory (LSTM) network layer extracts latent high-level features from the spectral noise data, and the second LSM network layer maps these high-level features back to the original input space to obtain the synthesized spectral data. The generator includes the space generator G. S Spectrum Generator G U This is used to generate pseudo-samples. The spatial and spectral data obtained after inputting the noise generated in S1 into the generator are collectively referred to as synthetic data, which are the pseudo-samples. Taking a certain study area as an example, the generator is equipped with two graph convolutional layers with 16 and 2 output channels respectively, and two inverse graph convolutional layers, such as... Figure 3 As shown, the number of output channels are 16 and 39 respectively, thus completing the generation of spatial data; the hidden layers of the first long short-term memory network layer are set to 16, and the number of layers is set to 4, and the second long short-term memory network layer ( Figure 4 The hidden layer is set to 2, and the number of layers is set to 4; Figure 4 In LSTM, a single LSTM unit includes a forget gate, an input gate, and an output gate. t Indicates the input of the current LSTM unit; h t This represents the hidden layer output of the current LSTM unit; This represents the long-term memory of the current LSTM cell; f t The computational operation representing the forget gate determines which information needs to be discarded; C t-1 This represents the node state of the previous LSTM unit (i.e., the final output, long-term memory); i t =σ(W i [h t-1 ,a t ]+b i ) represents important information output by the input gate; h t-1 This represents the output of the hidden layer of the previous LSTM unit; This represents the long-term memory of the input gate output; o t The output gate computation operation determines the output of the LSTM unit; tanh(·) represents the activation function. In the above description, W... i and W c b represents the trainable weight parameters in the corresponding operation. i and b c This represents the trainable bias parameter in the corresponding operation.

[0021] In other words, in S1 of this invention, spatial noise data x and spectral noise data r are constructed, and in S2, x and r are input into the generator to obtain spatial data. Spectral data Spatial data Spectral data That is, the generated pseudo data, the original spatial data X and x and Having the same structure, the original spectral data R and r are... They have the same structure.

[0022] S3. Construct a spatial discriminant D using graph convolutional layers. S A spectral discriminator D is constructed using long short-term memory network layers. U The generated spatial data were processed separately. Spectral data Feature extraction is performed on the original spatial data X and the original spectral data R to facilitate high-level feature fusion. The spatial discriminator is configured with two graph convolutional layers, with 8 and 2 output channels respectively, and the spectral discriminator is configured with a long short-term memory network layer, with 2 hidden layers and a total of 2 layers.

[0023] S4. Construct a fusion network H to fuse the spatial and spectral high-level features output by the discriminator. The resulting data undergoes a fusion operation after passing through the discriminator. The raw data is fused after passing through the discriminator to become H[D] S (X),D U (R)], and classify the fused features, outputting the discriminator's judgment result, i.e., whether it is real data or synthetic data. Representing spatial data, This refers to spectral data, which is generated data (i.e., pseudo data) that has the same structure as the original noise data.

[0024] S5. Design two cross-entropy loss functions to train the generator and discriminator separately. The generator aims to generate pseudo-samples that the discriminator cannot recognize, i.e., minimize the probability of the discriminator correctly classifying the sample. The data to be generated is to be classified as 1; the discriminator aims to distinguish between real and fake samples, minimizing the probability of misclassification. The process involves identifying real data as 1 and generated data as 0 until Nash equilibrium is reached, meaning the loss functions of the generator and discriminator stabilize. The generator and discriminator are trained simultaneously, allowing the model to adjust trainable parameters during backpropagation, improving the learning ability of the spatial and spectral branches to data patterns and enhancing the accuracy of geochemical anomaly identification. After multiple training iterations, the loss functions of the generator and discriminator stabilize, reaching Nash equilibrium, completing the training of the unsupervised spatial-spectral dual-branch deep learning model. The final unsupervised spatial-spectral dual-branch deep learning model is used for geochemical anomaly identification. Since the generator uses noise input to simulate the geochemical distribution patterns of the study area to assist in training the discriminator, and only real geochemical data is needed for anomaly delineation, the final geochemical anomaly identification uses only the discriminator and the fusion model H, expressed as: y = softamx(H[DS (X),D U (R)]), where y represents the geochemical anomaly output by the model, and softmax(·) is the activation function; when y is 0, it indicates that there is no geochemical anomaly, and when y is greater than 0, it indicates that there is a geochemical anomaly. The closer y is to 1, the higher the level of the identified geochemical anomaly, and vice versa. By inputting the actual acquired geochemical data into the final unsupervised spatial-spectral dual-branch deep learning model, geochemical anomalies and their degree can be identified.

[0025] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A geochemical anomaly identification method using an unsupervised spatial-spectral dual-branch deep learning model, characterized in that, Includes the following steps: S1. Decompose the raw geochemical data into spatial data. Spectral data We construct normally distributed random noise data with the same structure as the original geochemical data as the input data for the generative adversarial network. S2. Input random noise data into the constructed spatial generator and spectrum generator to generate spatial data. Spectral data ; S3, Generate spatial data , generated spectral data Raw spatial data and raw spectral data The inputs are fed into the constructed spatial discriminator and spectral discriminator to obtain high-level spatial features and high-level spectral features; S4. Building a converged network The spatial high-level features and spectral high-level features output by the spatial discriminator and the spectral discriminator are fused; S5. Design loss functions to train the constructed generator, discriminator, and fusion network. The generator includes a spatial generator and a spectral generator, and the discriminator includes a spatial discriminator and a spectral discriminator, until Nash equilibrium is reached, that is, the loss functions of the generator and discriminator tend to stabilize. After training, the final unsupervised spatial-spectral dual-branch deep learning model is obtained, and this model is used for geochemical anomaly identification. S2 uses graph convolutional layers and inverse graph convolutional layer Build space generator Among them, graph convolutional layers are used to process spatially noisy data. Extracted as potential high-level features Inverse graph convolutional layers are used to map high-level features back to the original input space. Inverse graph convolutional layers use linear layers. And graph convolutional layer By restoring the original noise characteristics, the synthesized spatial data is obtained. , This represents a non-linear activation function designed to extract high-level features by gradually subtracting neighborhood information. The original node information was recovered. A spectrum generator is constructed using long short-term memory network layers. The first long short-term memory network layer is used to process spectral noise data. The high-level features are extracted and mapped back to the original input space by a second long short-term memory network layer, resulting in synthesized spectral data. ; The spatial branch effectively captures geochemical spatial patterns and extracts spatial relationships between adjacent pixels using a generative adversarial network based on graph convolution operations; inverse graph convolution restores the original features of the data by using the difference between linear layers and graph convolution layers; the spectral branch studies the geochemical element combinations within a single pixel.

2. The geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model according to claim 1, characterized in that, In S1, random noise data is decomposed into spatial noise data. Sum-spectrum noise data , and Each as a space generator Spectrum generator Input data.

3. The geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model according to claim 1, characterized in that, S3 uses graph convolutional layers to build a spatial discriminator. A spectral discriminator is constructed using long short-term memory network layers. The generated spatial data were processed separately. Spectral data Raw spatial data and raw spectral data Feature extraction is performed in order to facilitate the fusion of advanced features.

4. The geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model according to claim 3, characterized in that, In S4, the generated data undergoes a fusion operation after passing through the spatial discriminator. The raw data is fused after passing through the spectral discriminator. The fused features are then classified, and the discriminator's judgment result is output, indicating whether it is real data or synthetic data.

5. The geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model according to claim 4, characterized in that, In S5, two cross-entropy loss functions are designed to train the generator and discriminator separately. The generator is used to minimize the probability of the discriminator correctly classifying a character. The generated data will be classified as 1; the discriminator is used to minimize the probability of the discriminator making a wrong judgment. This involves identifying real data as 1 and generated data as 0; until Nash equilibrium is reached, meaning the loss functions of the generator and discriminator tend to stabilize; where, This represents the cross-entropy loss function.

6. The geochemical anomaly identification method based on an unsupervised spatial-spectral dual-branch deep learning model according to claim 4, characterized in that, In S5, the final unsupervised spatial-spectral dual-branch deep learning model is used for anomaly identification. Since the generator uses noise input to simulate the geochemical distribution patterns of the study area to assist in training the discriminator, and only real geochemical data is needed when delineating anomalies, the final geochemical anomaly identification only uses the trained discriminator and the fusion network. , is represented as: ,in, Geochemical anomalies representing the model output, For activation function; when A value of 0 indicates no geochemical anomalies. A value greater than 0 indicates an anomaly in geochemistry. The closer the value is to 1, the higher the level of geochemical anomaly identified.

Images