Rock and mineral classification method based on three-dimensional convolution residual network

Abstract

This invention discloses a rock and mineral classification method based on a three-dimensional convolutional residual network, relating to the field of geological exploration technology. The method includes the following steps: Step S1: Acquire hyperspectral image data of the rock and minerals, wherein the hyperspectral image data includes spatial and spectral information; Step S2: Preprocess the hyperspectral image data to obtain a normalized three-dimensional data matrix; Step S3: Construct a three-dimensional convolutional residual network model, wherein the network model includes an input layer, multiple three-dimensional convolutional residual blocks, a feature fusion module, and a classification output layer. This invention achieves end-to-end automated processing, eliminating the need for manual feature engineering design, reducing reliance on experience, adapting to practical engineering applications, and its modular architecture also possesses strong transferability and scalability, adapting to different data and extended scenarios, balancing classification accuracy and practicality.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, specifically to a rock and mineral classification method based on a three-dimensional convolutional residual network. Background Technology

[0002] Rock and mineral classification is a core foundational technology in geological exploration, resource development, and environmental monitoring. Its accuracy and efficiency directly impact the depth of geological research, the accuracy of resource exploration, and the reliability of engineering applications. With the rapid development of hyperspectral remote sensing technology, rock and mineral hyperspectral images, with their advantages of multi-band and high spectral resolution, can simultaneously carry spatial (e.g., texture, morphology) and spectral (e.g., characteristic absorption peaks of mineral composition) information. They have become the core data source for rock and mineral classification, significantly improving the automation level and comprehensiveness of information acquisition compared to traditional methods such as manual visual identification and chemical analysis.

[0003] Currently, rock and mineral classification methods based on hyperspectral images are mainly divided into two categories: traditional machine learning methods and deep learning methods. Traditional machine learning methods include maximum likelihood estimation, random forests, spectral angle mapping (SAM), and support vector machines (SVM). These methods typically require manual extraction of spectral or spatial features of rocks and minerals before inputting them into a classifier. However, manual feature extraction is not only time-consuming and labor-intensive, requiring a high level of expertise from operators, but also struggles to capture complex, deep features in hyperspectral images of rocks and minerals. This is especially true when similar mineral compositions lead to spectral curve confusion, significantly reducing classification accuracy. Furthermore, traditional methods often fail to effectively integrate spatial and spectral information, neglecting the inherent correlation between the two. In complex field environments, they are susceptible to external interference such as ground cover and pixel mixing, further reducing classification accuracy and robustness. Their classification accuracy typically falls short of 90%, failing to meet the practical requirements for high-precision rock and mineral classification. To address this, we propose a rock and mineral classification method based on a three-dimensional convolutional residual network. Summary of the Invention

[0004] To address the aforementioned technical problems, a rock and mineral classification method based on a three-dimensional convolutional residual network is provided. This technical solution solves the problem of failing to effectively integrate spatial and spectral information.

[0005] To achieve the above objectives, the technical solution adopted in this invention is: a rock and mineral classification method based on a three-dimensional convolutional residual network, comprising the following steps: Step S1: Acquire hyperspectral image data of rocks and minerals, wherein the hyperspectral image data includes spatial dimension information and spectral dimension information; Step S2: Preprocess the hyperspectral image data to obtain a normalized three-dimensional data matrix; Step S3: Construct a three-dimensional convolutional residual network model, which includes an input layer, multiple three-dimensional convolutional residual blocks, a feature fusion module, and a classification output layer; Step S4: Input the preprocessed three-dimensional data matrix into the three-dimensional convolutional residual network model. The three-dimensional convolutional kernel in the three-dimensional convolutional residual block simultaneously extracts spatial and spectral features. The residual connection is used to enhance feature transfer. The extracted spatial and spectral features are fused using the feature fusion module to obtain spatial-spectral fused features. Step S5: Based on the spatial-spectral fusion features, output the rock and mineral classification results through the classification output layer.

[0006] Preferably, the hyperspectral image data of rocks and minerals in step S1 includes a three-dimensional cube with spatial and spectral dimensions. The spatial distribution and continuous spectral fingerprint of rocks and minerals are acquired simultaneously in the 400-2500nm band by using a hyperspectral imager mounted on satellite, airborne, and ground platforms.

[0007] Preferably, step S2 involves preprocessing the hyperspectral image data, including: performing radiometric and atmospheric correction on the hyperspectral image data; resampling the corrected hyperspectral image data to reduce spectral dimensional redundancy; and normalizing the resampled data using a max-min normalization method to map the data values ​​to a preset numerical range.

[0008] Preferably, the three-dimensional convolutional residual block in step S3 includes a three-dimensional convolutional layer, a batch normalization layer, an activation layer, and a residual branch; the three-dimensional convolutional layer performs convolution operations along the spatial and spectral dimensions, and the convolutional kernel size is K×K×K, where K is a positive integer greater than or equal to 3; the residual branch is configured to directly pass the input features to the output end and add them element-wise with the features processed by the three-dimensional convolution.

[0009] Preferably, the enhancement of feature transfer using residual connections in step S4 specifically involves: The preprocessed 3D data matrix is ​​input into the 3D convolution residual block, and after 3D convolution, normalization and nonlinear activation operations, the nonlinear mapping intermediate feature map of the main branch is obtained. Construct a residual shortcut branch and use identity mapping to preserve the original input features of the three-dimensional convolutional residual block. When the number of channels and size of the input features do not match those of the main branch output features, 1×1×1 three-dimensional convolution is used to align the dimensions and sizes. The deep spatial-spectral joint features of the main branch are added element by element to the original shallow features of the shortcut branch to transfer residual information across layers. The fused features after residual addition are subjected to nonlinear activation screening to output enhanced features; based on the multi-layer cascaded three-dimensional convolutional residual blocks, the feature transmission paths of shallow and deep layers are transferred and reused.

[0010] Preferably, the feature fusion module in step S4 adaptively fuses spatial features and spectral features through a channel attention mechanism; the channel attention mechanism includes: global average pooling operation, global max pooling operation, shared fully connected layer and sigmoid activation function; the shared fully connected layer is a feature vector that compresses the channel dimension and outputs a channel attention weight vector.

[0011] Preferably, the specific process of the feature fusion module is as follows: Dynamic weights are assigned to spatial and spectral features; the weights are adaptively adjusted based on the importance of the two types of features, and the feature weight coefficients are calculated by sharing a fully connected layer and a sigmoid activation function. Based on the assigned dynamic weights, a weighted fusion operation is performed on spatial and spectral features to achieve deep fusion of the two types of features. The weighted fusion features are enhanced to strengthen the effective fusion features and suppress invalid noise, and the final spatial-spectral fusion features are output.

[0012] Preferably, the classification output layer in step S5 includes: The spatial-spectral fusion features are input into the classification output layer, and then global average pooling is used to reduce the dimensionality to obtain a fixed-dimensional global fusion feature vector. The global fusion feature vector is mapped to a low-dimensional feature vector corresponding to the number of rock and mineral categories; the low-dimensional feature vector is input into the softmax activation function and converted into the probability value of each rock and mineral category. Set a probability threshold, and select the rock and mineral category with the highest probability value that is not lower than the threshold as the final classification result.

[0013] Preferably, the threshold is set based on the rock and mineral classification accuracy requirements and the distribution characteristics of the sample data.

[0014] Preferably, the rock and mineral classification results output in step S5 include: Obtain the classification probabilities of each category output by the classification output layer; The category corresponding to the highest probability is determined as the target rock and mineral category; The output includes rock and mineral classification results with category labels and confidence levels.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a 3D convolutional kernel to simultaneously extract joint features from the spatial and spectral dimensions of rocks and minerals, avoiding the shortcomings of traditional methods that separate the two types of features and improving feature representation capabilities. The introduction of residual connections effectively solves the gradient vanishing problem in deep networks, reduces feature propagation loss, accelerates training convergence, and enhances model generalization stability. Through a feature fusion module, the two types of features are deeply complementary, accurately distinguishing rocks and minerals with similar appearances but subtle spectral differences, reducing misclassification and omission rates. Normalization preprocessing eliminates external interference, improving the model's adaptability in complex scenarios. The entire process is automated end-to-end, eliminating the need for manual feature engineering, reducing reliance on experience, and adapting to practical engineering applications. The modular architecture also possesses strong transferability and scalability, adapting to different data and extended scenarios, balancing classification accuracy and practicality. Attached Figure Description

[0016] Figure 1 This is a flowchart of the steps of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, this invention discloses a rock and mineral classification method based on a three-dimensional convolutional residual network. This method achieves accurate and efficient classification of rocks and minerals by simultaneously mining the spatial and spectral features of rock and mineral hyperspectral images and combining residual connection and channel attention feature fusion technology.

[0019] Example 1 Step S1: Acquire hyperspectral image data of rocks and minerals In this embodiment, the hyperspectral image data of rocks and minerals is three-dimensional cubic data containing spatial and spectral dimensions. Specifically, it is acquired through the coordinated use of satellite, airborne, and ground-based acquisition platforms, equipped with a hyperspectral imager, simultaneously collecting spatial distribution information and continuous spectral fingerprint information of rock and mineral samples within the visible-near-infrared band of 400-2500 nm, forming the original hyperspectral image data of rocks and minerals. The spatial dimension information reflects the texture structure, morphological distribution, and spatial positional relationships of the rocks and minerals, while the spectral dimension information reflects the mineral composition characteristics. Together, they constitute the core discrimination criteria for rock and mineral classification, ensuring the comprehensiveness of subsequent feature extraction.

[0020] Example 2 Step S2: Preprocess the hyperspectral image data to obtain a normalized three-dimensional data matrix; In order to eliminate external interference, reduce data redundancy, and improve the training efficiency and classification accuracy of subsequent network models, this embodiment performs multi-step preprocessing on the original hyperspectral image data obtained in step S1. First, radiometric correction and atmospheric correction are performed on the original hyperspectral image data in sequence. Radiometric correction is used to eliminate the radiation distortion caused by the hyperspectral imager’s own response differences and changes in light intensity. Atmospheric correction is used to remove the interference of atmospheric scattering, absorption and other factors on spectral information and restore the true spectral characteristics of rocks and minerals. The corrected hyperspectral image data is subjected to spectral resampling. By merging adjacent similar bands and filtering key feature bands, the redundancy of spectral dimensions is reduced, the amount of data computation is reduced, and the core spectral information required for rock and mineral classification is preserved. The resampled data is standardized using the max-min normalization method, which maps all data values ​​to a preset range of [0,1] to eliminate the problem of inconsistent dimensions of data in different bands. This provides standardized data for the input of the subsequent three-dimensional convolutional residual network model, improves the convergence speed and generalization ability of the model, and finally obtains the normalized three-dimensional data matrix.

[0021] Example 3 Step S3: Construct a 3D convolutional residual network model The 3D convolutional residual network model constructed in this embodiment adopts a modular design, and the whole includes an input layer, multiple 3D convolutional residual blocks, a feature fusion module, and a classification output layer. The specific structure and parameters of each module are set as follows: The input layer is used to receive the normalized three-dimensional data matrix obtained in step S2. Its input dimension matches the spatial dimension and spectral dimension of the three-dimensional data matrix to ensure that the data can be completely input into the subsequent network modules. The 3D convolutional residual block, as the core feature extraction unit of the network, includes a 3D convolutional layer, a batch normalization layer, an activation layer, and a residual branch. The 3D convolutional layer performs convolution operations simultaneously along the spatial and spectral dimensions, with a kernel size of K×K×K, where K is 3 (which can be adjusted to a positive integer greater than or equal to 3 based on the actual resolution of the rock and mineral data). This is used to simultaneously extract spatial and spectral features. The batch normalization layer standardizes the convolutional feature maps, accelerating network training convergence and suppressing overfitting. The activation layer uses the ReLU activation function to achieve non-linear feature mapping and enhance the model's feature representation ability. The residual branch is configured to directly pass the input features to the output, adding them element-wise with the features processed by 3D convolution, batch normalization, and activation, thus achieving cross-layer feature transfer. The feature fusion module is used to deeply fuse the spatial and spectral features extracted from multiple 3D convolutional residual blocks. It adopts a channel attention mechanism to achieve adaptive fusion and improve the discriminativeness of the fused features. The classification output layer is used to map the fused features into rock and mineral classification results, thereby achieving accurate identification of rock and mineral categories.

[0022] Example 4 Step S4: Input the 3D data matrix, extract and fuse spatial-spectral features. The normalized three-dimensional data matrix obtained in step S2 is input into the three-dimensional convolutional residual network model constructed in step S3. Features are extracted through three-dimensional convolutional residual blocks, feature transfer is enhanced through residual connections, and feature fusion is achieved through the feature fusion module. The preprocessed 3D data matrix is ​​input into the first 3D convolutional residual block. After convolution by the 3D convolutional layer, normalization by the batch normalization layer, and nonlinear activation by the activation layer, the nonlinear mapping intermediate feature map of the main branch is obtained. Construct a residual shortcut branch and use an identity mapping to preserve the original input features of the 3D convolutional residual block; if the number of channels and size of the input features do not match the output features of the main branch, then use a 1×1×1 3D convolution to align the dimensions and size of the original input features to ensure that the two can be operated on element-wise. The deep spatial-spectral joint features extracted from the main branch are added element-wise to the original shallow features retained by the shortcut branch to achieve cross-layer transfer of residual information and reduce information loss during feature transfer. The fused features after the residuals are added are then filtered again by nonlinear activation through the activation layer to output enhanced features; by cascading multiple three-dimensional convolutional residual blocks, the transmission path of shallow features and deep features is realized and reused, so as to fully explore the fine-grained features of rocks and minerals. The feature fusion module enables adaptive fusion of spatial and spectral features. In this embodiment, the feature fusion module adaptively fuses spatial and spectral features using a channel attention mechanism. This channel attention mechanism includes global average pooling, global max pooling, a shared fully connected layer, and a sigmoid activation function. The shared fully connected layer is used to compress the channel dimension and output a channel attention weight vector. The specific fusion steps are as follows: Dynamic weights are assigned to the spatial and spectral features extracted from multiple 3D convolutional residual blocks. These weights are adaptively adjusted based on the importance of the two types of features in rock and mineral classification. Specifically, the channel dimension of the features is compressed by a shared fully connected layer, and the feature weight coefficients of each channel are calculated by the Sigmoid activation function. Based on the assigned dynamic weight coefficients, a weighted fusion operation is performed on spatial and spectral features to achieve deep coupling of the two types of features, allowing the fused features to simultaneously take into account both the spatial structure information and spectral composition information of rocks and minerals. The weighted fusion features are subjected to nonlinear processing to enhance effective fusion features and suppress invalid noise interference, ultimately outputting highly discriminative spatial-spectral fusion features, providing reliable feature support for subsequent rock and mineral classification.

[0023] Example 5 Step S5: Output the rock and mineral classification results through the classification output layer. Based on the spatial-spectral fusion features obtained in step S4, rock and mineral categories are determined through the classification output layer, and the final classification result is output. The spatial-spectral fusion features are input into the classification output layer. First, a global average pooling operation is used to reduce the dimensionality, resulting in a fixed-dimensional global fusion feature vector. This eliminates spatial dimensional differences and reduces computational cost. The global fusion feature vector is mapped to a low-dimensional feature vector corresponding to the number of rock and mineral categories through a fully connected layer. The dimension of the low-dimensional feature vector is consistent with the preset number of rock and mineral categories. The low-dimensional feature vector is input into the softmax activation function and converted into probability values ​​for each rock and mineral category. The sum of the probability values ​​for all rock and mineral categories is 1. The higher the probability value, the greater the likelihood that the rock or mineral belongs to the corresponding category. A probability threshold is set, which is flexibly set based on the rock and mineral classification accuracy requirements and the distribution characteristics of sample data (in this embodiment, the threshold is preferably 0.7, which can be adjusted according to the actual application scenario). The rock and mineral category with the highest probability value and not lower than the threshold is selected as the final classification result. The classification output layer obtains the classification probabilities of each category, determines the category corresponding to the highest probability as the target rock and mineral category, and finally outputs the rock and mineral classification results containing category labels and corresponding confidence scores (i.e., the highest probability values), making it easy for staff to intuitively view the reliability of the classification.

[0024] Specific Application Case 1 The classification of rock cores from the Bangpu polymetallic deposit in Tibet requires the identification of six target rock minerals.

[0025] S1 uses a dual platform of airborne and ground-based systems to acquire 200 sets of hyperspectral three-dimensional cubic data from rock cores in the 400-2500nm band; After radiation and atmospheric correction, S2 resamples the spectrum to 128 bands and maps it to [0,1] using max-min normalization; S3 constructs a model containing 6 three-dimensional convolutional residual blocks with 3×3×3 convolutional kernels, residual branches adapted to channel size, and feature fusion using a channel attention mechanism; S4 transfers features through residual connections and weighted fuses spatial and spectral features; S5 sets a threshold of 0.75 and outputs results including category labels and confidence levels.

[0026] In this case, this method was used to classify 200 sets of core samples. The classification accuracy of the data acquired by the SPECIM hyperspectral imager reached 92.3%, with a Kappa coefficient of 0.88, which is better than the 91.5% of the traditional spectral angle mapping method. The classification accuracy of the data acquired by the domestic core scanner reached 76.5%, with a Kappa coefficient of 0.68, which is 1.5 percentage points higher than the traditional method. The main error was caused by mixed pixel interference due to the lower spatial resolution of the domestic equipment. Six target rocks and minerals were successfully and accurately identified. The wavelength range of the Fe-OH absorption peak of chlorite in borehole ZK0010 was clearly captured as 2252.3~2260nm, and that in borehole ZK0011 was 2251.7~2258.5nm. The high wavelength value accurately indicated the center position of the ore body, providing reliable technical support for mining exploration. At the same time, it verified the adaptability of this method to different types of hyperspectral imaging equipment, which greatly improved the efficiency of rock and mineral classification. Compared with traditional manual classification, the work efficiency was improved by more than 80%.

[0027] Specific Application Case 2 In Fengxian County, Shaanxi Province, altered rock minerals need to be classified into three types: altered rock minerals and non-altered rock minerals.

[0028] S1 uses a combination of satellite and drone to collect 150 sets of data within a 50km² area in the 400-2500nm band; S2 is uniformly resampled to 128 bands to complete correction and normalization; S3 constructs a model with 8 residual blocks, with a 5×5×5 convolution kernel, and strengthens the weights of alteration features; S4 extracts and fuses features to suppress vegetation disturbance; S5 sets a threshold of 0.7 and outputs classification results and distribution maps. After application, the classification accuracy reaches 88.7%, the altered rock mineral identification accuracy is 90.2%, the efficiency is improved by more than 90%, and three prospecting target areas are delineated.

[0029] In this case, the method achieved a rock and mineral classification accuracy of 88.7% within a 50 km² area, with an altered rock and mineral identification accuracy of 90.2%, representing a 12.5 percentage point improvement over the traditional spectral angle mapping method. This effectively solved the problem of difficult identification of altered rock and mineral in vegetated areas. High-precision classification of two key areas could be completed in a single day, increasing work efficiency by over 90% compared to traditional manual exploration and significantly reducing the workload of field exploration. Three prospecting target areas with concentrated distribution of altered rock and mineral were successfully delineated. Combined with the confidence data in the classification results, precise guidance was provided for subsequent field verification and mineral exploration. An integrated "sky-air-ground" rapid identification technology system for altered minerals was constructed, verifying the applicability and superiority of this method in complex terrain and shallowly covered areas. It can be applied on a large scale to green mineral resource exploration in special landscape areas of Northwest China.

[0030] In summary, the two case studies cover two typical scenarios: precise core classification and large-scale regional exploration. Both strictly followed the entire process of the aforementioned rock and mineral classification method and optimized parameter configurations for the actual scenarios, achieving excellent application results. This fully demonstrates that the rock and mineral classification method based on three-dimensional convolutional residual networks has high classification accuracy, efficiency, and scenario adaptability, and can be widely applied in practical engineering fields such as geological exploration, mine surveying, and mineral prospecting.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A rock and mineral classification method based on a three-dimensional convolutional residual network, characterized in that, Includes the following steps: Step S1: Acquire hyperspectral image data of rocks and minerals, wherein the hyperspectral image data includes spatial dimension information and spectral dimension information; Step S2: Preprocess the hyperspectral image data to obtain a normalized three-dimensional data matrix; Step S3: Construct a three-dimensional convolutional residual network model, which includes an input layer, multiple three-dimensional convolutional residual blocks, a feature fusion module, and a classification output layer; Step S4: Input the preprocessed three-dimensional data matrix into the three-dimensional convolutional residual network model. The three-dimensional convolutional kernel in the three-dimensional convolutional residual block simultaneously extracts spatial and spectral features. The residual connection is used to enhance feature transfer. The extracted spatial and spectral features are fused using the feature fusion module to obtain spatial-spectral fused features. Step S5: Based on the spatial-spectral fusion features, output the rock and mineral classification results through the classification output layer.

2. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that: The hyperspectral image data of rocks and minerals mentioned in step S1 includes a three-dimensional cube with spatial and spectral dimensions. The spatial distribution and continuous spectral fingerprint of rocks and minerals are acquired simultaneously in the 400-2500nm band by using a hyperspectral imager mounted on satellite, airborne, and ground platforms.

3. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that, Step S2 involves preprocessing the hyperspectral image data, including: performing radiometric and atmospheric corrections on the hyperspectral image data; resampling the corrected hyperspectral image data to reduce spectral dimensional redundancy; and normalizing the resampled data using a max-min normalization method to map the data values ​​to a preset numerical range.

4. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that: In step S3, the 3D convolutional residual block includes a 3D convolutional layer, a batch normalization layer, an activation layer, and a residual branch. The 3D convolutional layer performs convolution operations along the spatial and spectral dimensions, and the convolutional kernel size is K×K×K, where K is a positive integer greater than or equal to 3. The residual branch is configured to directly pass the input features to the output end and add them element-wise with the features processed by the 3D convolution.

5. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that, The enhancement of feature transfer using residual connections described in step S4 specifically involves: The preprocessed 3D data matrix is ​​input into the 3D convolution residual block, and after 3D convolution, normalization and nonlinear activation operations, the nonlinear mapping intermediate feature map of the main branch is obtained. Construct a residual shortcut branch and use identity mapping to preserve the original input features of the three-dimensional convolutional residual block. When the number of channels and size of the input features do not match those of the main branch output features, 1×1×1 three-dimensional convolution is used to align the dimensions and sizes. The deep spatial-spectral joint features of the main branch are added element by element to the original shallow features of the shortcut branch to transfer residual information across layers. The fused features after residual addition are subjected to nonlinear activation screening to output the enhanced features; Based on the cascaded three-dimensional convolutional residual blocks, the feature transmission paths of shallow and deep layers are transmitted and reused.

6. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that: The feature fusion module in step S4 adaptively fuses spatial features and spectral features through a channel attention mechanism. The channel attention mechanism includes: global average pooling, global max pooling, a shared fully connected layer, and a sigmoid activation function. The shared fully connected layer is a feature vector that compresses the channel dimension and outputs a channel attention weight vector.

7. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 6, characterized in that, The specific process of the feature fusion module is as follows: Dynamic weights are assigned to spatial and spectral features; The weights are adaptively adjusted based on the importance of the two types of features, and the feature weight coefficients are calculated by sharing a fully connected layer and a sigmoid activation function. Based on the assigned dynamic weights, a weighted fusion operation is performed on spatial and spectral features to achieve deep fusion of the two types of features. The weighted fusion features are enhanced to strengthen the effective fusion features and suppress invalid noise, and the final spatial-spectral fusion features are output.

8. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that, The classification output layer in step S5 includes: The spatial-spectral fusion features are input into the classification output layer, and then global average pooling is used to reduce the dimensionality to obtain a fixed-dimensional global fusion feature vector. The global fusion feature vector is mapped to a low-dimensional feature vector corresponding to the number of rock and mineral categories; the low-dimensional feature vector is input into the softmax activation function and converted into the probability value of each rock and mineral category. Set a probability threshold, and select the rock and mineral category with the highest probability value that is not lower than the threshold as the final classification result.

9. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 8, characterized in that: The threshold is set based on the accuracy requirements of rock and mineral classification and the distribution characteristics of sample data.

10. The rock and mineral classification method based on a three-dimensional convolutional residual network according to claim 1, characterized in that, The rock and mineral classification results output in step S5 include: Obtain the classification probabilities of each category output by the classification output layer; The category corresponding to the highest probability is determined as the target rock and mineral category; The output includes rock and mineral classification results with category labels and confidence levels.

Images