Lithology identification method and feature prediction network construction method

Abstract

The invention provides a lithology identification method and a feature prediction network construction method, and relates to the technical field of lithology intelligent identification, and the lithology identification method comprises the steps: obtaining multi-source remote sensing data of a lithology classification region, and carrying out the feature extraction of the multi-source remote sensing data, and obtaining a lithology fusion feature representation; traversing geological units in each image in the multi-source remote sensing data through the category co-occurrence matrix to obtain a lithology co-occurrence matrix; inputting the lithologic fusion feature representation and the lithologic co-occurrence matrix into a graph convolutional network, and performing global context modeling to obtain global-local lithologic feature representation of the lithologic classification region; and inputting the global-local lithology feature representation into a feature prediction network for prediction to obtain a final lithology identification result of the lithology classification region. According to the method, the heterogeneity problem of multi-source remote sensing data is solved, the advantages of each data source are fully utilized, global and local feature information is combined, and the lithology identification precision is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithology intelligent identification, and in particular to a lithology identification method and a feature prediction network construction method. Background Art

[0002] The identification of lithology plays a very important role in geology, resource exploration, geotechnical exploration, rock mechanics and engineering. Generally, multi-source remote sensing data is used to identify the lithology in a certain area, and the accuracy of lithology identification has a great impact on the underground lithology distribution, exploration of minerals, oil and other resources.

[0003] In the prior art, when processing multi-source remote sensing data, different remote sensing data sources have different characteristics and resolutions, which leads to heterogeneity problems in feature extraction and fusion. For example, hyperspectral data has rich spectral information, while multispectral data has high spatial resolution. Therefore, it is difficult to fully utilize the advantages of each data source in feature extraction of multi-source remote sensing data, which leads to deviations in feature extraction and makes it impossible to ensure the accuracy of lithology identification. Summary of the invention

[0004] The problem solved by the present invention is how to improve the accuracy of lithology identification.

[0005] In order to solve the above problems, the present invention provides a lithology identification method and a feature prediction network construction method.

[0006] In a first aspect, the present invention provides a lithology identification method, comprising:

[0007] Acquire multi-source remote sensing data of a lithology classification area, and perform feature extraction on the multi-source remote sensing data to obtain a lithology fusion feature representation;

[0008] Through the category co-occurrence matrix, the geological units in each image in the multi-source remote sensing data are traversed to obtain a lithology co-occurrence matrix;

[0009] Inputting the lithology fusion feature representation and the lithology co-occurrence matrix into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area;

[0010] The global-local lithology feature representation is input into a feature prediction network for prediction to obtain a final lithology identification result of the lithology classification area.

[0011] Optionally, extracting features from the multi-source remote sensing data to obtain lithology fusion feature representation includes:

[0012] The multi-source remote sensing data is cropped according to a preset size to obtain a plurality of image data;

[0013] Through the transformer network, the image data is linearly embedded to obtain the hierarchical features corresponding to each level in the transformer network;

[0014] All the hierarchical features are fused to obtain the lithology fusion feature representation.

[0015] Optionally, the transformer network includes a block embedding layer and a linear connection layer, and the image data is linearly embedded through the transformer network to obtain hierarchical features corresponding to each level in the transformer network, including:

[0016] Inputting the image data into the block embedding layer of the transformer network, and decomposing the image data into a plurality of sequence blocks through the block embedding layer;

[0017] Linearly embedding the sequence blocks through the linear connection layer of the transformer network to obtain multiple feature vectors;

[0018] The feature vector is sequentially input into multiple levels of the transformer network to obtain the level features corresponding to each level.

[0019] Optionally, the traversal of the geological units in each image in the multi-source remote sensing data by using the category co-occurrence matrix to obtain a lithology co-occurrence matrix includes:

[0020] An initial adjacency matrix is ​​used as the category co-occurrence matrix, wherein all elements of the initial adjacency matrix are 0, and the size of the initial adjacency matrix is ​​N×N, where N is the number of types of different lithologies;

[0021] Traversing each image in the multi-source remote sensing data, and updating the element value of the corresponding position of the initial adjacency matrix according to the number of geological units in the image;

[0022] After traversing all the images, the updated initial adjacency matrix is ​​used as the lithology co-occurrence matrix.

[0023] Optionally, updating the element values ​​of corresponding positions of the initial adjacency matrix according to the number of geological units in the image includes:

[0024] Determining the number of geological units contained in the image according to the category label vectors of the geological units in the image;

[0025] According to the number of the geological units, the elements of the image at corresponding positions in the initial adjacency matrix are increased by corresponding values.

[0026] Optionally, the step of inputting the lithology fusion feature representation and the lithology co-occurrence matrix into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area includes:

[0027] representing the lithology fusion feature as an initial feature of the geological unit;

[0028] Using the lithology co-occurrence matrix as the adjacency matrix of the graph convolutional network;

[0029] Inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling;

[0030] Wherein, inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling includes:

[0031] Performing convolution operations according to the adjacency matrix and the initial features through each convolution layer in the graph convolution network in turn to update the feature representation of the geological unit; wherein the output of each convolution layer serves as the input of the next convolution layer;

[0032] According to the output of the last convolutional layer of the graph convolutional network, a global-local lithology feature representation of the lithology classification area is obtained.

[0033] Optionally, the inputting the global-local lithology feature representation into the feature prediction network for prediction to obtain the final lithology identification result of the lithology classification area includes:

[0034] inputting the global-local lithology feature representation into the feature prediction network;

[0035] By means of the feature prediction network, the global-local lithological feature is compared with the preset lithological feature to obtain a comparison result;

[0036] The final lithology identification result is obtained according to the comparison result.

[0037] The lithology identification method described in the present invention obtains multi-source remote sensing data of a lithology classification area, wherein these data sources have heterogeneity problems due to their different characteristics and resolutions. In order to solve this problem, the present invention uses a feature extraction structure to extract features from multi-source remote sensing data with different resolutions to obtain a lithology fusion feature representation, thereby fully utilizing the advantages of different data sources in the feature extraction process. In addition, based on feature extraction, the present invention traverses the geological units in each image in the multi-source remote sensing data through a category co-occurrence matrix to obtain a lithology co-occurrence matrix, which helps to integrate features from different data sources and enhance the contextual information of features through co-occurrence relationships, thereby solving the heterogeneity problem. In response to the heterogeneity problem of multi-source remote sensing data in the prior art, the present invention fully utilizes information from different data sources, provides a rich feature basis for lithology identification, and solves the fusion problem caused by differences in characteristics and resolutions of different remote sensing data sources; ensures that features extracted from multiple data sources can be effectively integrated, providing a rich feature basis for subsequent lithology identification. Next, by counting the co-occurrence relationships of different lithology categories, important contextual information is provided for lithology identification. Then, the local features are combined with the global contextual information to enhance the recognition ability of the model. The extracted features and co-occurrence relationships are transformed into deeper feature representations by using a graph convolutional network, combined with lithology fusion features and lithology co-occurrence relationships, to achieve the integration of global and local contextual information of lithology features, and improve the accuracy and robustness of lithology identification. Finally, these features are analyzed through a feature prediction network, and the lithology identification results are finally output. The present invention solves the heterogeneity problem of multi-source remote sensing data, fully utilizes the advantages of each data source, combines global and local feature information, and significantly improves the accuracy of lithology identification.

[0038] In a second aspect, the present invention provides a method for constructing a feature prediction network, which is applied to the above-mentioned lithology identification method, comprising:

[0039] Through the multimodal contrast learning module, constructing a plurality of positive sample pairs and negative sample pairs of preset geological units, wherein the positive sample pairs are sample pairs of the same preset geological unit; and the negative sample pairs are sample pairs of different preset geological units;

[0040] According to the positive sample pair and the negative sample pair, the similarity between the preset geological units is obtained, wherein the similarity between the preset geological units includes the similarity of the positive sample pair and the similarity of the negative sample pair;

[0041] According to the similarity of the positive sample pair and the similarity of the negative sample pair, the initial fully connected neural network is trained by a back propagation algorithm to obtain the feature prediction network.

[0042] Optionally, the training of the initial fully connected neural network by a back propagation algorithm according to the similarity of the positive sample pair and the similarity of the negative sample pair to obtain the feature prediction network includes:

[0043] Establishing a first similarity threshold and a second similarity threshold according to the similarity of the positive sample pair and the similarity of the negative sample pair, wherein the first similarity threshold is greater than the second similarity threshold;

[0044] Training the initial fully connected neural network according to the first similarity threshold and the second similarity threshold;

[0045] When the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and at the same time, the similarity of the negative sample pair is less than the second similarity threshold, the fully connected neural network is used as the feature prediction network.

[0046] Optionally, when the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is less than the second similarity threshold, using the fully connected neural network as the feature prediction network includes:

[0047] Continuously adjusting the parameters of the initial fully connected neural network by comparing the similarity of the positive sample pair with the first similarity threshold and the similarity of the negative sample pair with the second similarity threshold;

[0048] until the similarity of the positive sample pair is continuously greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is continuously less than the second similarity threshold;

[0049] The adjusted initial fully connected neural network is used as the feature prediction network.

[0050] The feature prediction network construction method described in the present invention can effectively distinguish the characteristics of different geological units and enhance the model's ability to identify lithological characteristics by constructing positive sample pairs and negative sample pairs through a multimodal contrast learning module; for example, the construction of positive sample pairs enables the model to learn similar geological unit characteristics, while the construction of negative sample pairs enables the model to identify and distinguish different geological unit characteristics. Using the similarity between positive and negative sample pairs, the method trains the initial fully connected neural network through a back propagation algorithm, optimizes the feature representation of the geological unit, and enables the feature prediction network to more accurately capture the subtle differences in lithology. Moreover, during the training process, the feature prediction network learns the similarities and differences between the preset geological units, thereby enhancing the generalization ability of the model in practical applications and improving the accuracy and robustness of lithology identification; and the trained feature prediction network can accurately predict the lithology of unknown geological units. In summary, the feature prediction network construction method of the present invention enhances feature distinguishability, optimizes feature representation, improves the generalization ability and prediction accuracy of the model through contrast learning, and provides a powerful tool for lithology identification.

[0051] Furthermore, the construction of the feature prediction network plays a vital role in the lithology identification method. Specifically, the multimodal contrast learning module enhances the model's ability to aggregate the same lithology feature and disperse different lithology features, improves feature distinguishability, and utilizes the global context modeling function of the graph convolutional network to enable the feature prediction network to comprehensively consider the local characteristics and global spatial relationships of the geological unit, thereby more accurately capturing the local and global characteristics of the lithology. In addition, the feature prediction network improves the accuracy and robustness of lithology identification by optimizing model parameters, while improving the efficiency of automated processing of large-scale remote sensing data. The adaptability and generalization ability of the network enable it to cope with the analysis needs of complex geological structures, providing an efficient and accurate automated solution for lithology identification, significantly improving the accuracy of lithology identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 is a flow chart of a lithology identification method according to an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of a multi-level feature fusion module in a Transformer-based feature extraction structure according to an embodiment of the present invention;

[0054] Figure 3 It is an overall structural diagram of the feature extraction structure based on Transformer in an embodiment of the present invention;

[0055] Figure 4 It is a schematic diagram of the multi-source remote sensing data fusion and lithology identification processing flow of an embodiment of the present invention;

[0056] Figure 5 The present invention is a flowchart of a method for constructing a feature prediction network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. Although certain embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be interpreted as being limited to the embodiments described herein. On the contrary, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are only for exemplary purposes and are not intended to limit the scope of protection of the present invention.

[0058] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. In addition, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0059] The term "including" and its variations used in this document are open inclusions, that is, "including but not limited to"; the term "based on" means "based at least in part on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one other embodiment"; the term "some embodiments" means "at least some embodiments"; the term "optionally" means "optional embodiments". The relevant definitions of other terms will be given in the following description. It should be noted that the concepts of "first", "second", etc. mentioned in the present invention are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0060] It should be noted that the modifications of "one" and "plurality" mentioned in the present invention are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise clearly indicated in the context, it should be understood as "one or more".

[0061] The names of the messages or information exchanged between multiple devices in the embodiments of the present invention are only used for illustrative purposes, and are not used to limit the scope of these messages or information.

[0062] Combination Figure 1 As shown, the present invention provides a lithology identification method, comprising:

[0063] Multi-source remote sensing data of a lithology classification area is acquired, and feature extraction is performed on the multi-source remote sensing data to obtain a lithology fusion feature representation.

[0064] Specifically, multi-source remote sensing data of the lithology classification area are obtained. These data include but are not limited to high-resolution optical images, SAR images, and DEM images, which provide information on surface and underground structures from different dimensions. By performing feature extraction on these multi-source remote sensing data, the information from different data sources is effectively integrated to obtain a fusion feature representation of lithology. This step is the basis, which ensures that the features extracted from multiple data sources can be effectively integrated, providing a rich feature basis for subsequent lithology identification.

[0065] The geological units in each image in the multi-source remote sensing data are traversed through the category co-occurrence matrix to obtain a lithology co-occurrence matrix.

[0066] Specifically, the geological units of each image in the multi-source remote sensing data are traversed through the category co-occurrence matrix to obtain the lithology co-occurrence matrix. The category co-occurrence matrix is ​​a statistical tool used to capture the co-occurrence relationship between different lithology categories, that is, the frequency of their co-occurrence in the image sample set; thus, it helps to reveal the correlation between lithologies, provides important contextual information for lithology identification, and enhances the richness and accuracy of feature representation.

[0067] The lithology fusion feature representation and the lithology co-occurrence matrix are input into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area.

[0068] Specifically, the graph convolutional network uses the lithology co-occurrence matrix as the adjacency relationship and combines it with the lithology fusion feature representation to achieve the integration of global and local contextual information of lithology characteristics. By aggregating the information of node neighbors and updating the feature representation of the node, the model can capture the local and global characteristics of the lithology, thereby improving the accuracy and robustness of lithology identification.

[0069] The global-local lithology feature representation is input into a feature prediction network for prediction to obtain a final lithology identification result of the lithology classification area.

[0070] Specifically, the feature prediction network can learn the similarities and differences between preset geological units, and can train the initial fully connected neural network through, for example, a back-propagation algorithm, and ultimately determine it as a feature prediction network, thereby being able to accurately predict the lithology of unknown geological units.

[0071] The lithology identification method described in the present invention obtains multi-source remote sensing data of a lithology classification area, wherein these data sources have heterogeneity problems due to their different characteristics and resolutions. In order to solve this problem, the present invention uses a feature extraction structure to extract features from multi-source remote sensing data with different resolutions to obtain a lithology fusion feature representation, thereby fully utilizing the advantages of different data sources in the feature extraction process. In addition, based on feature extraction, the present invention traverses the geological units in each image in the multi-source remote sensing data through a category co-occurrence matrix to obtain a lithology co-occurrence matrix, which helps to integrate features from different data sources and enhance the contextual information of features through co-occurrence relationships, thereby solving the heterogeneity problem. In response to the heterogeneity problem of multi-source remote sensing data in the prior art, the present invention fully utilizes information from different data sources, provides a rich feature basis for lithology identification, and solves the fusion problem caused by differences in characteristics and resolutions of different remote sensing data sources; ensures that features extracted from multiple data sources can be effectively integrated, providing a rich feature basis for subsequent lithology identification. Next, by counting the co-occurrence relationships of different lithology categories, important contextual information is provided for lithology identification. Then, the local features are combined with the global contextual information to enhance the recognition ability of the model. The extracted features and co-occurrence relationships are transformed into deeper feature representations by using a graph convolutional network, combined with lithology fusion features and lithology co-occurrence relationships, to achieve the integration of global and local contextual information of lithology features, and improve the accuracy and robustness of lithology identification. Finally, these features are analyzed through a feature prediction network, and the lithology identification results are finally output. The present invention solves the heterogeneity problem of multi-source remote sensing data, fully utilizes the advantages of each data source, combines global and local feature information, and significantly improves the accuracy of lithology identification.

[0072] Optionally, extracting features from the multi-source remote sensing data to obtain lithology fusion feature representation includes:

[0073] The multi-source remote sensing data is cropped according to a preset size to obtain a plurality of image data;

[0074] Through the transformer network, the image data is linearly embedded to obtain the hierarchical features corresponding to each level in the transformer network;

[0075] All the hierarchical features are fused to obtain the lithology fusion feature representation.

[0076] Specifically, multi-source remote sensing data, such as high-resolution optical images, SAR images, and DEM images, are cropped according to preset sizes to obtain multiple image data. This cropping process ensures the consistency and comparability of the data and provides standardized input for subsequent feature extraction. Then, these image data are input into the Transformer-based network for linear embedding processing; in the multiple levels of the Transformer network, each level will output corresponding hierarchical features, which contain semantic information at different levels from local to global; finally, by fusion of all hierarchical features and integration of information at different levels, a comprehensive global feature vector containing all input data is generated, namely, the lithology fusion feature representation. This feature representation not only contains the global information and local detail information of the image, but also enhances the model's recognition ability of different parts through the collaborative work of the feature grouping fusion module and the multi-level feature fusion module, providing a comprehensive feature basis for lithology recognition.

[0077] For example, in combination Figure 2 As shown in the figure, the fully connected layers are used to map features to a new space for subsequent processing or classification tasks; the output of the fully connected layer is fused with another feature stream through multiplication, and the fused features are enhanced through a channel shuffle structure to enhance the expressiveness of the features, and the information exchange between different features is promoted by disrupting and reorganizing the channels; the channel shuffled features are combined with another feature stream through residual connections; this structure helps the network learn more effective feature representations while avoiding the problem of gradient disappearance in deep networks. At the same time, multiple Transformer blocks or other types of neural network layers in the network are used to gradually extract and process features. At some levels, features are grouped and processed for further refinement or specific types of processing. Finally, the features processed by multiple layers are integrated and the final lithology fusion feature representation is output for lithology identification or other related tasks. This example processes and integrates features from multi-source remote sensing data through multiple fully connected layers, feature fusion, channel shuffle, and residual connections to improve the accuracy and efficiency of lithology identification.

[0078] In this optional embodiment, firstly, the multi-source remote sensing data processed by preset size clipping ensures the standardization and consistency of the data, providing high-quality input for feature extraction. Secondly, the multi-level feature extraction and linear embedding of the Transformer network can effectively capture multi-level semantic information from local to global, enhancing the expressiveness of features. In addition, through feature fusion technology, features at different levels are integrated to generate a comprehensive lithology fusion feature representation, which not only improves the richness of features, but also enhances the model's ability to recognize lithology.

[0079] Optionally, the transformer network includes a block embedding layer and a linear connection layer, and the image data is linearly embedded through the transformer network to obtain hierarchical features corresponding to each level in the transformer network, including:

[0080] Inputting the image data into the block embedding layer of the transformer network, and decomposing the image data into a plurality of sequence blocks through the block embedding layer;

[0081] Linearly embedding the sequence blocks through the linear connection layer of the transformer network to obtain multiple feature vectors;

[0082] The feature vector is sequentially input into multiple levels of the transformer network to obtain the level features corresponding to each level.

[0083] Specifically, first, the image data is input into the block embedding layer of the Transformer network, which decomposes the entire image data into multiple sequence blocks, each of which contains local image information. Then, these sequence blocks are linearly embedded through the linear connection layer of the Transformer network, mapping the sequence blocks to a high-dimensional space, thereby obtaining multiple feature vectors, which are then sequentially input into multiple layers of the Transformer network. Each layer processes these feature vectors accordingly, extracts features at different levels, and finally obtains the hierarchical features corresponding to each layer. This process not only retains the spatial structural information of the image data, but also enhances the expressiveness of the features through the self-attention mechanism of the Transformer network, providing a rich feature basis for subsequent lithology identification.

[0084] In a preferred embodiment of the present invention, the multi-source remote sensing data is cropped according to a preset size to obtain multiple image data, which include high-resolution optical images, SAR images and DEM images of the area. The multi-source data includes two branches, one branch uses 10m optical images, SAR images and 30DEM images, and the other branch uses 2m optical images, SAR images and 10mDEM images, which are used as inputs of the global local transform network. The input image is first decomposed into sequence blocks, and then input into the linear connection layer and linearly embedded; then, through multiple levels of Transformer, feature representations of different levels are obtained; and the feature representations of different levels are input into the multi-level feature fusion module to obtain global information of different levels.

[0085] The basic Transformer network first receives multi-source images and performs Patch Embedding processing, and then linearly embeds the processed sequence blocks.

[0086] Linear embedding can be expressed as:

[0087] E = W·X+b;

[0088] Among them, E is the embedded feature, W is the weight matrix, X is the input multi-source image data, and b is the bias term. The linearly embedded image is input into multiple Transformer blocks, each of which includes a self-attention mechanism and a feedforward network.

[0089] The self-attention mechanism can be expressed as:

[0090]

[0091] Among them, Q represents the query matrix, which is used to perform dot product operations with the key matrix in the attention mechanism to calculate the attention weights; K represents the key matrix, which performs dot product operations with the query matrix, and the result is used to calculate the attention weights. V represents the value matrix, which is multiplied with the weights to generate the output after obtaining the attention weights; Attention (Q, K, V) attention mechanism is used to calculate the relationship between query, key and value, d k is the dimension of the key vector; k T represents the transpose of the key vector; softmax represents the softmax function, which is used to convert the dot product result into a probability distribution, that is, the attention weight; v represents the value vector (Valuevector), which is multiplied by the attention weight obtained by the softmax function to obtain the weighted output.

[0092] The feedforward network can be expressed as:

[0093] FFN(x)=max(0,xW1+b1)W2+b2;

[0094] Among them, FFN(x) represents the function of the feedforward network, which accepts input x and produces output; x represents the input vector, that is, the input of the feedforward network; W1 represents the weight matrix of the first layer; W2 represents the weight matrix of the second layer; b1 represents the bias term of the first layer, which is added to the result of multiplying with the weight matrix W1 to adjust the output; b2 represents the bias term of the second layer, which is added to the result of multiplying with the weight matrix W2 to adjust the output; max(0,·) represents the ReLU activation function (Rectified LinearUnit), which sets all negative values ​​to 0 and keeps positive values ​​unchanged, which is used to introduce nonlinearity.

[0095] The multi-level feature fusion module is used to extract and aggregate global features of the image. The multi-level feature fusion module collects features from different layers of Transformer blocks. These features contain global semantic information at different levels. The features at different levels are reduced in dimension through a fully connected layer. The fully connected layer maps the fused features to a new output space, which can be expressed as:

[0096] Y=W f ·F+b f ;

[0097] Among them, Y is the new feature after processing by the fully connected layer, W f is the weight matrix, F is the output features of the last few layers of the Transformer network, and b f It is a bias term. It reduces the dimensional features through the fully connected layer and performs feature selection at the same time, selects the most effective feature information, captures the implicit patterns and relationships in the data, and then aggregates the three features processed by the fully connected layer, integrates the information of different layers, and generates a comprehensive global feature vector containing all the input data; at the same time, the category labels of different levels are reshaped into feature maps of the same size, and the feature maps are placed in a channel shuffle structure and spliced.

[0098] Among them, it should be noted that the core idea of ​​the channel shuffle operation is to rearrange the channels of the feature map to promote information fusion between different channels. Channel shuffle disrupts the order of channels so that the features of each group can be dispersed into different groups, thereby enhancing the information flow between feature maps. The channel shuffle operation can be implemented through the following steps: First, the channel dimension of the input feature map is reshaped into two dimensions: the number of groups and the number of channels in each group. Next, the reshaped feature map is transposed to permute the two dimensions of the group convolution. Finally, the transposed feature map is flattened to complete the channel shuffle and obtain the fused feature. The fused feature is multiplied with the global feature vector to obtain a new comprehensive global feature vector, which enhances the recognition ability of the model. It is connected with the feature vector after channel shuffle splicing through a residual connection to finally realize image feature extraction.

[0099] For example, in combination Figure 3 As shown in the figure, the black square on the left side represents multi-source remote sensing image data. These data sources are diverse and may include optical images of different resolutions, SAR images, and digital elevation models (DEMs). Multi-source image data first passes through a linear embedding layer to convert the image data into a format suitable for Transformer network processing; linear embedding involves adjusting the data dimension or applying a weight matrix to facilitate subsequent feature extraction. The linearly embedded image data is fed into multiple Transformer blocks; each Transformer block includes a self-attention mechanism and a feedforward network, which can capture local and global features in the image data. Multi-level feature extraction combined Figure 3 As shown in the figure, multiple stacked rectangles are shown, representing different levels in the Transformer network; each level extracts features from the image data to generate hierarchical features; these features contain different levels of information from basic to advanced. After multi-level feature extraction, the feature grouping and fusion module groups and fuses features at different levels to enhance the distinguishability of features and integrate richer information. Finally, the grouped and fused features are processed through a series of operations (fully connected layers, channel shuffling, etc.) to finally output the lithology fusion feature representation. These fusion feature representations can comprehensively reflect the global information and local details of the lithology, providing key information for subsequent lithology identification. In summary, this example shows the process from the input of multi-source remote sensing data to the output of the final lithology fusion feature representation. The entire process aims to obtain a rich feature set for lithology classification through the powerful feature extraction capability of the Transformer network.

[0100] In this optional embodiment, first, the introduction of the block embedding layer enables the network to effectively process high-resolution image data. By decomposing the image data into smaller sequence blocks, the network can capture finer local features. Secondly, the linear embedding capability of the linear connection layer enables the network to effectively map these local features to a high-dimensional space, thereby enhancing the distinctiveness of the features. In addition, the multi-level structure of the Transformer network enables the model to extract and integrate features layer by layer, thereby obtaining deeper global features. This embodiment not only improves the efficiency and accuracy of feature extraction, but also enhances the model's ability to capture key information in image data through the self-attention mechanism, ultimately significantly improving the accuracy of lithology identification.

[0101] Optionally, the traversal of the geological units in each image in the multi-source remote sensing data by using the category co-occurrence matrix to obtain a lithology co-occurrence matrix includes:

[0102] An initial adjacency matrix is ​​used as the category co-occurrence matrix, wherein all elements of the initial adjacency matrix are 0, and the size of the initial adjacency matrix is ​​N×N, where N is the number of types of different lithologies;

[0103] Traversing each image in the multi-source remote sensing data, and updating the element value of the corresponding position of the initial adjacency matrix according to the number of geological units in the image;

[0104] After traversing all the images, the updated initial adjacency matrix is ​​used as the lithology co-occurrence matrix.

[0105] Specifically, first, an adjacency matrix of size N×N is initialized as a category co-occurrence matrix, where N represents the number of different lithology types. In the initial stage, all elements are set to 0, indicating that the co-occurrence relationship of any lithology has not been counted. Subsequently, each image in the multi-source remote sensing data is traversed, and the element values ​​of the corresponding lithology types in the initial adjacency matrix are updated according to the actual distribution of geological units in the image, in order to count the co-occurrence frequency of each lithology in the image sample set. For example, if two lithologies co-occur in a certain image, the value of the corresponding position in the adjacency matrix is ​​increased by 1, and the process continues until all images have been traversed; finally, the updated initial adjacency matrix, that is, the lithology co-occurrence matrix, reflects the co-occurrence relationship between different lithologies, providing important contextual information for lithology identification.

[0106] In a preferred embodiment of the present invention, geological units represent nodes, edges represent the relationship between geological units, and the adjacency matrix is ​​an N×N matrix, in which the element A[i, j] represents the connection strength between node i and node j, and N is the number of all lithological types in the region. A lithological co-occurrence matrix is ​​constructed based on the spatial proximity relationship of geological units. If two geological units are adjacent in space, the value of the corresponding position in the adjacency matrix is ​​1, otherwise it is 0. For example, assuming that there are N different categories in the region, C1, C2, ..., C N Denotes, C1, C2, ..., C N Represents the labels of different categories; given an image sample set, which contains M images, M represents the total number of images in the image sample set. For each image i, a binary vector t i =(t 1i , t 2i , ..., t Ni ) represents its category label, t i For each image i, this is a binary vector representing its category label. 1i , t 2i , ..., t Ni Represents a binary vector t i The components of , indicating whether image i belongs to category C1, C2, ..., C N , if it belongs to, then t ji =1, otherwise t ji = 0. Then the category co-occurrence matrix can be expressed as an N×N matrix CCM, where CCM ij Indicates category C i and Category C j The frequency or number of times it occurs at the same time. CCM ij It can be obtained by counting the co-occurrence of corresponding category labels in all images.

[0107] In this optional embodiment, through the above process, the co-occurrence relationship between different lithologies can be systematically captured and quantified, providing rich contextual information for lithology identification. By counting the co-occurrence frequency of each lithology in the image sample set, the potential correlation between lithologies can be revealed, which is crucial for understanding geological structures and distribution patterns. Secondly, the construction of the lithology co-occurrence matrix provides adjacency relationships for the graph convolutional network, allowing the network to more accurately simulate the interactions and dependencies between geological units; it not only improves the accuracy of lithology identification, but also enhances the model's ability to interpret geological data.

[0108] Optionally, updating the element values ​​of corresponding positions of the initial adjacency matrix according to the number of geological units in the image includes:

[0109] Determining the number of geological units contained in the image according to the category label vectors of the geological units in the image;

[0110] According to the number of the geological units, the elements of the image at corresponding positions in the initial adjacency matrix are increased by corresponding values.

[0111] Specifically, first, for each image in the multi-source remote sensing data, the number of geological units contained in the image is determined based on its category label vector, where the category label vector is a binary vector in which each element corresponds to a specific lithology category. If the image contains the lithology, the corresponding element value is 1, otherwise it is 0. Next, based on this category label vector, the number of times each lithology appears in the image is counted. Then, based on these statistical data, the corresponding lithology category position in the initial adjacency matrix is ​​increased by a numerical value. Specifically, if a certain lithology exists in the image, the corresponding lithology category position in the adjacency matrix is ​​increased by 1 to record the number of co-occurrences of each lithology in the image. This process is repeated for all images until all images have been processed, thereby obtaining the final lithology co-occurrence matrix.

[0112] In a preferred embodiment of the present invention, an N×N matrix of all zeros is initialized; each image i in the sample set is traversed, and the category label vector t of image i is i , for every pair of categories C k and C j , if t ji =1 and t ki =1, that is, image i belongs to category C k and C j , then increase CCM ij After traversing all images, the obtained matrix CCM is the lithology co-occurrence matrix, and the number of statistics in it is converted into probability.

[0113] In this optional embodiment, by accurately counting the number of geological units in each image and updating the initial adjacency matrix, the co-occurrence relationship between different lithologies can be accurately captured, thereby providing more accurate contextual information for lithology identification. In addition, the lithology co-occurrence matrix of this embodiment can comprehensively record the co-occurrence frequency of lithologies, provide rich data support for subsequent graph convolutional networks, and enhance the model's ability to understand the complex relationship between geological units. At the same time, the accurate lithology co-occurrence matrix as the input of the graph convolutional network helps to optimize network performance and improve the accuracy and robustness of lithology identification, especially when dealing with areas with complex geological structures.

[0114] Optionally, the step of inputting the lithology fusion feature representation and the lithology co-occurrence matrix into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area includes:

[0115] representing the lithology fusion feature as an initial feature of the geological unit;

[0116] Using the lithology co-occurrence matrix as the adjacency matrix of the graph convolutional network;

[0117] Inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling;

[0118] Wherein, inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling includes:

[0119] Performing convolution operations according to the adjacency matrix and the initial features through each convolution layer in the graph convolution network in turn to update the feature representation of the geological unit; wherein the output of each convolution layer serves as the input of the next convolution layer;

[0120] According to the output of the last convolutional layer of the graph convolutional network, a global-local lithology feature representation of the lithology classification area is obtained.

[0121] Specifically, first, the lithology fusion feature representation, that is, the features of multi-source remote sensing data extracted by the Transformer network, is used as the initial features of the geological unit; these initial features contain rich local and global information, providing a detailed description of the geological unit. Then, the co-occurrence relationship between different lithologies is recorded through the lithology co-occurrence matrix, and used as the adjacency matrix of the graph convolutional network to define the connection strength between geological units. After the initial features and the adjacency matrix are input into the graph convolutional network, they are processed through each convolutional layer in the network, and the adjacency matrix and the initial features are used for convolution operations to update the feature representation of the geological unit; this process is carried out layer by layer, and the output of each layer becomes the input of the next layer, thereby realizing the layer-by-layer abstraction and integration of features. Finally, according to the output of the last convolutional layer of the graph convolutional network, the global-local lithology feature representation of the lithology classification area is obtained. These feature representations integrate the local features and global context information of the geological unit, providing strong feature support for lithology identification.

[0122] Among them, the graph convolution operation updates the feature representation of the node by aggregating the information of the node neighbors. The graph convolution is to extend the convolution operation from the traditional regular grid (such as image) to the irregular graph structure data. The convolution operation of GCN can be expressed as:

[0123] H (l+1) =σ(D -1 / 2AD -1 / 2 H (l) W (l) );

[0124] Among them, H (l) is the node feature matrix of the lth layer; H (l+1) represents the node feature matrix of the l+1th layer; W (l) is the weight matrix of the lth layer; A is the adjacency matrix; D is the degree matrix; σ is the nonlinear activation function; D -1 / 2 represents the square root of the inverse of the degree matrix D, which is used to normalize the adjacency matrix so that the feature update of each node not only depends on the number of its neighbors, but is also inversely proportional to the number of neighbors; AD -1 / 2 H (l) W (l) Represents the product of the normalized adjacency matrix and the node feature matrix.

[0125] For example, in combination Figure 4 As shown in the figure, two remote sensing data inputs with different resolutions are shown, including 10-meter optical image + radar (SAR) + 30-meter digital elevation model (DEM) as a coarse classification system, and 2-meter optical image + SAR + 10-meter DEM as a fine classification system; these data sources provide multi-dimensional information required for lithology classification. Multi-source remote sensing data is first processed by a Transformer-based feature extraction structure, including multiple self-attention mechanism layers and feature grouping fusion modules, which are used to extract features of remote sensing data with different resolutions. The features extracted from the Transformer network then enter the multi-level feature fusion module to further integrate and refine the feature information to form a lithology fusion feature representation. In addition, based on geological maps and remote sensing images of different resolutions, a coarse classification system and a fine classification system are constructed, and the lithology co-occurrence matrix is ​​calculated. The lithology co-occurrence matrix is ​​used to reflect the co-occurrence relationship between different lithology categories and to generate an adjacency matrix. The lithology fusion feature representation and the adjacency matrix of the lithology co-occurrence relationship are input into the graph convolutional network; the graph convolutional network implements global context modeling through multi-layer convolution operations to extract global-local lithology feature representations. Through the multimodal contrast learning module, positive and negative sample pairs are constructed, and by obtaining the multimodal contrast loss, feature aggregation of the same lithology and feature dispersion of different lithologies are achieved. The multimodal contrast loss is then weighted and superimposed with the segmentation loss to form a fusion loss. Through the back propagation algorithm, the model parameters are optimized to improve the accuracy of lithology identification. Finally, the global-local lithology feature representation is sent to the decoder, which is a fully connected neural network that is used to output the final lithology identification result of the lithology classification area. Among them, it should be noted that Figure 4The arrows in the figure indicate the direction of data flow, where the solid arrows indicate forward propagation and the dotted arrows indicate reverse propagation. The weighted superposition symbol in the figure indicates the combination of loss functions. The entire process reflects the complete processing from raw data to the final lithology identification results.

[0126] In this optional embodiment, by using the lithology fusion feature representation as the initial feature input to the graph convolutional network, the information in the multi-source remote sensing data can be fully utilized, thereby enhancing the expressive power of the geological unit features. And by using the lithology co-occurrence matrix as the adjacency matrix, the co-occurrence relationship between geological units is effectively fused, so that the model can capture the global context information of the geological unit. Through the layer-by-layer abstraction and integration of the graph convolutional network, the accuracy of lithology identification is effectively improved, especially in complex geological structures, different lithology categories can be more accurately distinguished. At the same time, the generation of global-local lithology feature representation takes into account the local characteristics and global context information of the geological unit, and enhances the robustness of the model to outliers and noise. The global context modeling step of this embodiment not only improves the accuracy and robustness of lithology identification, but also provides an effective tool for geological data analysis.

[0127] Optionally, the inputting the global-local lithology feature representation into the feature prediction network for prediction to obtain the final lithology identification result of the lithology classification area includes:

[0128] inputting the global-local lithology feature representation into the feature prediction network;

[0129] By means of the feature prediction network, the global-local lithological feature is compared with the preset lithological feature to obtain a comparison result;

[0130] The final lithology identification result is obtained according to the comparison result.

[0131] Specifically, after obtaining the global-local lithology feature representation, the lithology identification process in this embodiment further uses the feature prediction network to perform the final lithology classification, and inputs these comprehensive feature representations into the feature prediction network, wherein the feature prediction network is a trained neural network that can identify and compare the similarity or difference between the input features and the preset lithology features. Within the network, through a series of calculations and hierarchical processing, the global-local lithology features are compared with the pre-stored or learned lithology feature templates, wherein the comparison process involves the calculation of the similarity of the feature vectors, and the cosine similarity, Euclidean distance and other measurement methods can be used. Finally, based on these comparison results, the network outputs the final lithology identification results, which represent the prediction and determination of the lithology of each geological unit in the lithology classification area. Moreover, this process is automated, can efficiently process a large amount of geological data, and provide accurate lithology classification results.

[0132] In a preferred embodiment of the present invention, the comprehensive feature vector obtained through multi-layer graph convolution can be sent to the decoder, which is a neural network with a fully connected layer. The decoder receives the feature vector and outputs the identification result of the lithology. Assume that we have the following three lithology feature vectors: Granite feature vector: F granite , gneiss eigenvector: F gneiss , limestone feature vector: F limestone , these feature vectors are used to train the decoder, which learns the characteristic patterns of each lithology. In the prediction phase, the decoder receives a new feature vector and compares it with the patterns learned in the training phase, outputting the most likely lithology category. For example, if the new feature vector is most similar to the granite feature vector in the training set, the decoder will predict that the sample is granite.

[0133] In this optional embodiment, the feature prediction network can integrate global and local lithology features, thereby improving the accuracy of lithology identification, especially under complex geological conditions. By comparing with preset lithology features, the network can not only identify known lithologies, but also effectively classify new and unseen lithology samples. In addition, the automated lithology identification process greatly reduces manual intervention and improves the speed and efficiency of data processing, especially when processing large-scale remote sensing data, thereby further reducing the manpower and material costs required in traditional geological exploration, especially when conducting geological surveys in remote or difficult-to-reach areas.

[0134] Combination Figure 5 As shown, the present invention provides a method for constructing a feature prediction network, which is applied to the above-mentioned lithology identification method, comprising:

[0135] Through the multimodal contrast learning module, positive sample pairs and negative sample pairs of multiple preset geological units are constructed, wherein the positive sample pairs are sample pairs of the same preset geological unit; and the negative sample pairs are sample pairs of different preset geological units.

[0136] Specifically, the network is trained to identify the characteristic patterns of different geological units by constructing positive and negative sample pairs using a multimodal contrastive learning module. The positive sample pairs consist of feature vectors representing the same lithology, while the negative sample pairs consist of feature vectors representing different lithologies. In this way, the network can learn the difference between similar and dissimilar features, thereby improving its ability to distinguish different lithologies. Moreover, this process is the basis for feature prediction network training, providing the necessary sample data for subsequent similarity calculation and network optimization.

[0137] The similarity between the preset geological units is obtained according to the positive sample pair and the negative sample pair, wherein the similarity between the preset geological units includes the similarity of the positive sample pair and the similarity of the negative sample pair.

[0138] Specifically, by evaluating the high similarity between positive sample pairs and the low similarity between negative sample pairs, a criterion for measuring the performance of the network is provided, wherein the similarity can be calculated using a variety of methods, such as cosine similarity or Euclidean distance, etc., which will not be repeated here. These metrics can quantify the closeness between feature vectors, thereby providing a quantitative feedback mechanism for network training so that the network can adjust its parameters to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs.

[0139] In a preferred embodiment of the present invention, the global-local lithology feature representation extracted by the graph convolutional network guided by the co-occurrence relationship is used to construct positive and negative sample pairs. When constructing the positive sample pairs, the feature representations representing the same lithology are selected as the positive sample pairs. For example, the features of granite are selected as the positive sample pairs. When constructing the negative sample pairs, the features representing different lithologies are selected as the negative sample pairs. For example, the features of granite and the features of gneiss are selected as the negative sample pairs. For each pair of positive and negative sample pairs, the similarity between them is calculated using cosine similarity or Euclidean distance.

[0140] According to the similarity of the positive sample pair and the similarity of the negative sample pair, the initial fully connected neural network is trained by a back propagation algorithm to obtain the feature prediction network.

[0141] Specifically, based on the calculated similarity, the initial fully connected neural network is trained through the back propagation algorithm to obtain a feature prediction network. In this process, the weights and biases of the network are adjusted according to the gradient of the loss function, where the loss function is usually a combination of contrast loss and segmentation loss; the network can learn how to better distinguish different lithological features and improve its ability to predict lithological classification. Among them, the back propagation algorithm is a supervised learning technique that allows the network to gradually optimize its parameters through an iterative process to achieve better classification performance; ultimately, the fully trained network can accurately predict the lithology of the lithological classification area.

[0142] The feature prediction network construction method described in the present invention can effectively distinguish the characteristics of different geological units and enhance the model's ability to identify lithological characteristics by constructing positive sample pairs and negative sample pairs through a multimodal contrast learning module; for example, the construction of positive sample pairs enables the model to learn similar geological unit characteristics, while the construction of negative sample pairs enables the model to identify and distinguish different geological unit characteristics. Using the similarity between positive and negative sample pairs, the method trains the initial fully connected neural network through a back propagation algorithm, optimizes the feature representation of the geological unit, and enables the feature prediction network to more accurately capture the subtle differences in lithology. Moreover, during the training process, the feature prediction network learns the similarity and difference between the preset geological units, thereby enhancing the generalization ability of the model in practical applications and improving the accuracy and robustness of lithology identification; and the trained feature prediction network can accurately predict the lithology of unknown geological units. In summary, the feature prediction network construction method of the present invention enhances feature distinguishability, optimizes feature representation, improves the generalization ability and prediction accuracy of the model through contrast learning, and provides a powerful tool for lithology identification.

[0143] Furthermore, the construction of the feature prediction network plays a vital role in the lithology identification method. Specifically, the multimodal contrast learning module enhances the model's ability to aggregate the same lithology feature and disperse different lithology features, improves feature distinguishability, and utilizes the global context modeling function of the graph convolutional network to enable the feature prediction network to comprehensively consider the local characteristics and global spatial relationships of the geological unit, thereby more accurately capturing the local and global characteristics of the lithology. In addition, the feature prediction network improves the accuracy and robustness of lithology identification by optimizing model parameters, while improving the efficiency of automated processing of large-scale remote sensing data. The adaptability and generalization ability of the network enable it to cope with the analysis needs of complex geological structures, providing an efficient and accurate automated solution for lithology identification, significantly improving the accuracy of lithology identification.

[0144] Optionally, the training of the initial fully connected neural network by a back propagation algorithm according to the similarity of the positive sample pair and the similarity of the negative sample pair to obtain the feature prediction network includes:

[0145] Establishing a first similarity threshold and a second similarity threshold according to the similarity of the positive sample pair and the similarity of the negative sample pair, wherein the first similarity threshold is greater than the second similarity threshold;

[0146] Training the initial fully connected neural network according to the first similarity threshold and the second similarity threshold;

[0147] When the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and at the same time, the similarity of the negative sample pair is less than the second similarity threshold, the fully connected neural network is used as the feature prediction network.

[0148] Specifically, the first and second similarity thresholds are established through similarity to optimize the training process of the initial fully connected neural network. This process involves setting two key thresholds: the first similarity threshold is used for positive sample pairs, requiring the similarity of positive sample pairs to reach or exceed this threshold to ensure that the model can effectively identify the characteristics of the same lithology; the second similarity threshold is used for negative sample pairs, requiring the similarity of negative sample pairs to be lower than this threshold to ensure that the model can distinguish the characteristics of different lithologies. The setting of these two thresholds makes the network training have a clear goal, that is, to maximize the similarity of positive sample pairs and minimize the similarity of negative sample pairs. During the training process, the network parameters are continuously adjusted through the back propagation algorithm until the similarity of the positive sample pairs meets the requirements of the first similarity threshold, and the similarity of the negative sample pairs is lower than the second similarity threshold. At this time, the fully connected neural network is regarded as a trained feature prediction network and is used for lithology identification. Among them, it should be noted that "the similarity of positive sample pairs" and "the similarity of negative sample pairs" and "the similarity between preset geological units" actually describe the same concept, but are explained from different perspectives. The similarity of positive sample pairs and negative sample pairs: are described from the perspective of sample pairs; in contrastive learning, positive sample pairs refer to sample pairs belonging to the same geological unit, while negative sample pairs refer to sample pairs belonging to different geological units. The similarity of positive sample pairs is high, while the similarity of negative sample pairs is low. The similarity between the preset geological units is described from the perspective of the geological units. In the process of constructing the feature prediction network, the similarity between different geological units is calculated by the contrast learning module, and these similarities reflect the proximity of the geological units in the feature space. In this embodiment, the calculation results of the similarity of the positive sample pairs and negative sample pairs constructed by the multimodal contrast learning module are used to measure the similarity between different geological units. These similarity values ​​are used to train the fully connected neural network, and the network parameters are adjusted by the back propagation algorithm to maximize the similarity of the positive sample pairs and minimize the similarity of the negative sample pairs. The training of the network is more accurately controlled by setting two similarity thresholds (the first similarity threshold and the second similarity threshold). When the similarity of the positive sample pairs reaches or exceeds the first similarity threshold, and the similarity of the negative sample pairs is lower than the second similarity threshold, it is considered that the network training has achieved the expected effect, and the fully connected neural network can be used as a feature prediction network. In summary, whether from the perspective of the sample pairs or the perspective of the geological units, the concept of similarity is consistent, and is used to measure the proximity between samples or geological units in the feature space.

[0149] In a preferred embodiment of the present invention, the loss value can be calculated by applying the loss function formula:

[0150] L contrastive =∑ (i，j)∈P max(0, ms(i, j)) 2 +∑ (i，j)∈N s(i, j) 2 ;

[0151] Among them, L contrastive is the calculated loss value, P represents the set of positive sample pairs, N represents the set of negative sample pairs, s(i, j) represents the similarity of sample pair (i, j), m is the threshold of similarity, (i, j)∈P represents that sample pair (i, j) belongs to the set of positive sample pairs P, max(0, ms(i, j)) 2 Represents a part of the contrast loss, which is used to penalize the similarity s(i, j) of the positive sample pair, when s(i, j) is less than the threshold m. If s(i, j) is greater than or equal to m, this item is 0; s(i, j) represents the similarity of the sample pair (i, j), m represents the similarity threshold, (i, j)∈N represents that the sample pair (i, j) belongs to the negative sample pair set N, s(i, j) 2 The other part of the contrast loss is used to penalize the situation where the similarity s(i, j) of the negative sample pair is too large. contrastive Optimize model parameters. During training, the model learns how to increase the similarity of positive sample pairs and reduce the similarity of negative sample pairs.

[0152] Get the segmentation loss L from the decoder seg and contrastive learning loss L contrastive Perform weighted processing to obtain the final total loss L all , the specific calculation formula is as follows:

[0153] L all =L contrastive *a+L seg *b;

[0154] Among them, a and b represent weight coefficients used to balance the influence of contrastive learning loss and segmentation loss in the total loss. Select sample pairs representing the same geological unit from multimodal data as positive sample pairs; through feature comparison, train the model to maximize the similarity of positive sample pairs; select sample pairs representing different geological units from multimodal data as negative sample pairs; through feature comparison, train the model to minimize the similarity of negative sample pairs; calculate the similarity between positive and negative sample pairs based on the features extracted by the graph convolutional network guided by the aforementioned co-occurrence relationship; obtain the fusion loss by weighted superposition of multimodal contrast loss and segmentation loss, and use back propagation to optimize model parameters to improve the model's discrimination ability.

[0155] In this optional embodiment, firstly, by setting a clear similarity threshold, a quantitative standard is provided to evaluate and guide the training of the network, making the training process more scientific and accurate. Secondly, by contrasting the learning strategy, the network's ability to distinguish lithological features is enhanced, and the accuracy of lithological identification is improved. In addition, through the iterative optimization of the back-propagation algorithm, the network can automatically adjust parameters to better adapt to the set threshold, reducing manual intervention and improving training efficiency. Ultimately, a robust feature prediction network is obtained that can not only identify known lithologies, but also adapt to new and unseen lithological samples, and has good generalization ability, thereby improving the work efficiency and accuracy of lithological identification.

[0156] Optionally, when the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is less than the second similarity threshold, using the fully connected neural network as the feature prediction network includes:

[0157] Continuously adjusting the parameters of the initial fully connected neural network by comparing the similarity of the positive sample pair with the first similarity threshold and the similarity of the negative sample pair with the second similarity threshold;

[0158] until the similarity of the positive sample pair is continuously greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is continuously less than the second similarity threshold;

[0159] The adjusted initial fully connected neural network is used as the feature prediction network.

[0160] Specifically, when the similarity of the positive sample pairs reaches or exceeds the preset first similarity threshold, and the similarity of the negative sample pairs is lower than the second similarity threshold, the fully connected neural network is confirmed as a valid feature prediction network, and the network parameters are continuously monitored and adjusted by comparing the similarity of the positive and negative sample pairs with their respective thresholds, and fine-tuning the parameters of the initial fully connected neural network using the back propagation algorithm to ensure that the similarity of the positive sample pairs is always higher than the first threshold, and the similarity of the negative sample pairs is always lower than the second threshold. This process will continue until the network can stably meet these conditions in continuous verification, at which point the network is considered to be fully trained and ready for lithology identification tasks.

[0161] In this optional embodiment, by setting and monitoring the first similarity threshold and the second similarity threshold, it is ensured that the network is sensitive and accurate in identifying lithological features, thereby improving the accuracy of lithological identification. In addition, the network parameters are continuously adjusted until specific conditions are met, and the network steps are automatically optimized, which reduces manual intervention and improves the efficiency of the training process. In addition, by continuously training until the similarity of the positive sample pairs is stably higher than the first threshold and the similarity of the negative sample pairs is stably lower than the second threshold, the feature prediction network learns to more effectively distinguish the features of different lithologies, thereby enhancing the model's ability to distinguish. The feature prediction network trained in this embodiment has learned to deal with the similarity differences of various sample pairs during the training process, so that it can perform lithological identification more robustly when facing new and unseen data. The feature prediction network construction method of this embodiment not only improves the accuracy and efficiency of lithological identification, but also enhances the robustness and generalization ability of the model by accurately controlling the similarity threshold and automatic network optimization, providing an effective technical support for lithological identification.

[0162] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc. In the present application, the unit described as a separate component may or may not be physically separated, and the component displayed as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment of the present invention. In addition, each functional unit in each embodiment of the present invention can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0163] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A lithology identification method, characterized in that: include: Acquire multi-source remote sensing data of a lithology classification area, and perform feature extraction on the multi-source remote sensing data to obtain a lithology fusion feature representation; Through the category co-occurrence matrix, the geological units in each image in the multi-source remote sensing data are traversed to obtain a lithology co-occurrence matrix; Inputting the lithology fusion feature representation and the lithology co-occurrence matrix into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area; The global-local lithology feature representation is input into a feature prediction network for prediction to obtain a final lithology identification result of the lithology classification area.

2. The lithology identification method according to claim 1, characterized in that: The extracting features of the multi-source remote sensing data to obtain lithology fusion feature representation includes: The multi-source remote sensing data is cropped according to a preset size to obtain a plurality of image data; Through the transformer network, the image data is linearly embedded to obtain the hierarchical features corresponding to each level in the transformer network; All the hierarchical features are fused to obtain the lithology fusion feature representation.

3. The lithology identification method according to claim 2, characterized in that: The transformer network includes a block embedding layer and a linear connection layer. The image data is linearly embedded through the transformer network to obtain the hierarchical features corresponding to each level in the transformer network, including: Inputting the image data into the block embedding layer of the transformer network, and decomposing the image data into a plurality of sequence blocks through the block embedding layer; Linearly embedding the sequence blocks through the linear connection layer of the transformer network to obtain multiple feature vectors; The feature vector is sequentially input into multiple levels of the transformer network to obtain the level features corresponding to each level.

4. The lithology identification method according to claim 1, characterized in that: The lithology co-occurrence matrix is ​​obtained by traversing the geological units in each image in the multi-source remote sensing data through the category co-occurrence matrix, including: An initial adjacency matrix is ​​used as the category co-occurrence matrix, wherein all elements of the initial adjacency matrix are 0, and the size of the initial adjacency matrix is ​​N×N, where N is the number of types of different lithologies; Traversing each image in the multi-source remote sensing data, and updating the element value of the corresponding position of the initial adjacency matrix according to the number of geological units in the image; After traversing all the images, the updated initial adjacency matrix is ​​used as the lithology co-occurrence matrix.

5. The lithology identification method according to claim 4, characterized in that: The updating of the element values ​​of the corresponding positions of the initial adjacency matrix according to the number of geological units in the image includes: Determining the number of geological units contained in the image according to the category label vectors of the geological units in the image; According to the number of the geological units, the elements of the image at corresponding positions in the initial adjacency matrix are increased by corresponding values.

6. The lithology identification method according to claim 1, characterized in that: The step of inputting the lithology fusion feature representation and the lithology co-occurrence matrix into a graph convolutional network to perform global context modeling to obtain a global-local lithology feature representation of the lithology classification area includes: representing the lithology fusion feature as an initial feature of the geological unit; Using the lithology co-occurrence matrix as the adjacency matrix of the graph convolutional network; Inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling; Wherein, inputting the initial features and the adjacency matrix into the graph convolutional network for global context modeling includes: Performing convolution operations according to the adjacency matrix and the initial features through each convolution layer in the graph convolution network in turn to update the feature representation of the geological unit; wherein the output of each convolution layer serves as the input of the next convolution layer; According to the output of the last convolutional layer of the graph convolutional network, a global-local lithology feature representation of the lithology classification area is obtained.

7. The lithology identification method according to claim 1, characterized in that: The step of inputting the global-local lithology feature representation into the feature prediction network for prediction to obtain the final lithology identification result of the lithology classification area includes: inputting the global-local lithology feature representation into the feature prediction network; By means of the feature prediction network, the global-local lithological feature is compared with the preset lithological feature to obtain a comparison result; The final lithology identification result is obtained according to the comparison result.

8. A method for constructing a feature prediction network, characterized in that: The lithology identification method applied to any one of items 1 to 7 above comprises: Through the multimodal contrast learning module, constructing a plurality of positive sample pairs and negative sample pairs of preset geological units, wherein the positive sample pairs are sample pairs of the same preset geological unit; and the negative sample pairs are sample pairs of different preset geological units; According to the positive sample pair and the negative sample pair, the similarity between the preset geological units is obtained, wherein the similarity between the preset geological units includes the similarity of the positive sample pair and the similarity of the negative sample pair; According to the similarity of the positive sample pair and the similarity of the negative sample pair, the initial fully connected neural network is trained by a back propagation algorithm to obtain the feature prediction network.

9. The method for constructing a feature prediction network according to claim 8, characterized in that: The method of training the initial fully connected neural network by a back propagation algorithm according to the similarity of the positive sample pair and the similarity of the negative sample pair to obtain the feature prediction network includes: Establishing a first similarity threshold and a second similarity threshold according to the similarity of the positive sample pair and the similarity of the negative sample pair, wherein the first similarity threshold is greater than the second similarity threshold; Training the initial fully connected neural network according to the first similarity threshold and the second similarity threshold; When the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and at the same time, the similarity of the negative sample pair is less than the second similarity threshold, the fully connected neural network is used as the feature prediction network.

10. The method for constructing a feature prediction network according to claim 9, characterized in that: When the similarity of the positive sample pair is greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is less than the second similarity threshold, using the fully connected neural network as the feature prediction network includes: Continuously adjusting the parameters of the initial fully connected neural network by comparing the similarity of the positive sample pair with the first similarity threshold and the similarity of the negative sample pair with the second similarity threshold; until the similarity of the positive sample pair is continuously greater than or equal to the first similarity threshold, and the similarity of the negative sample pair is continuously less than the second similarity threshold; The adjusted initial fully connected neural network is used as the feature prediction network.

Images