A method and system for constructing a lithology commonality joint feature classification model

Abstract

The present invention provides a method and system for constructing a lithology commonality joint feature classification model, and relates to the field of image processing technology. The method comprises: extracting local-global features of lithology features of remote sensing image data to obtain lithology feature representation of remote sensing image data; optimizing single image contrast loss according to lithology feature representation by contrast learning; obtaining cross-image contrast loss according to the same lithology feature representation of multi-source image data, and constructing a classifier to determine the classification loss; training an initial classification model according to the single image contrast loss, cross-image contrast loss and classification loss to obtain a lithology commonality joint feature classification model. The present invention utilizes the idea of ​​contrast learning to obtain cross-image contrast loss according to multi-source data, so that in the feature space, features of different lithologies are dispersed and features of the same lithology are aggregated, thereby improving the representation capability of features, and utilizing lithology diversity to improve the perception capability of lithology features.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a method and system for constructing a lithology commonality joint feature classification model. Background Art

[0002] In remote sensing images, high-level abstract semantic features such as lithology are manifested as spatial aggregation features at a certain scale, and the spatial distribution has a certain continuity. In addition, the color, texture, morphology and other features of lithology presented in multi-source data are different, and different lithologies have diverse characteristics and similar morphologies, resulting in the same lithology showing different characteristics in different images, making lithology classification more complicated.

[0003] In the prior art, lithology classification models generally use a single data source to describe lithology characteristics. However, the amount of information provided by a single data source to comprehensively describe lithology characteristics is relatively small. At the same time, since the same lithology may show different characteristics in different images, the judgment of lithology characteristics is ambiguous, making the judgment results of lithology characteristics relatively single and more prone to misjudgment of lithology characteristics. Summary of the invention

[0004] The problem solved by the present invention is how to improve the perception capability of the lithology classification model on lithology characteristics.

[0005] In order to solve the above problems, the present invention provides a method for constructing a lithology commonality joint feature classification model, comprising:

[0006] By performing local-global feature extraction on the lithological features of the remote sensing image data, a lithological feature representation of the remote sensing image data is obtained;

[0007] By contrast learning optimization, a single image contrast loss is obtained according to the lithological feature representation;

[0008] According to the same lithological feature representation of multi-source image data, a cross-image comparison loss is obtained, and a classifier is constructed to determine the classification loss;

[0009] The initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model.

[0010] Optionally, the extracting local-global features of the lithological features of the remote sensing image data to obtain the lithological feature representation of the remote sensing image data includes:

[0011] Extracting features from the remote sensing image data using a convolution kernel of a preset size to obtain a multi-scale feature map;

[0012] Slicing the multi-scale feature map to divide the remote sensing image data into multiple feature maps;

[0013] The feature map is sequentially passed through the three layers of transformer modules for feature extraction to obtain the lithological feature representation.

[0014] Optionally, the step of extracting features from the feature map through three layers of transformer modules in sequence to obtain the lithological feature representation includes:

[0015] Through the attention mechanism, the feature map input into the transformer module is linearly processed to obtain a query vector, a key-value vector, and a value vector of the feature map;

[0016] The query vector, the key-value vector and the value vector are normalized by a normalization function to obtain the output of the transformer module, and the lithology feature representation is obtained according to the output.

[0017] Optionally, the optimization through contrast learning, obtaining a single image contrast loss according to the lithological feature representation, includes:

[0018] Obtaining, through a deep encoder network, a plurality of similar lithological feature representations corresponding to each of the remote sensing image data according to the plurality of remote sensing image data;

[0019] Mapping the similar lithology feature representation and the lithology feature representation to the same feature hypersphere through a feature projection network, and determining the image similarity corresponding to each of the remote sensing image data;

[0020] Wherein, for one of the remote sensing image data, the image similarity includes: cross-image similarity and single-image similarity, the cross-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the similar lithological feature representations corresponding to any other remote sensing image data, and the single-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the lithological feature representation corresponding to the remote sensing image data;

[0021] The single image comparison loss corresponding to each remote sensing image data is determined according to the image similarity through a loss function.

[0022] Optionally, determining the single image contrast loss corresponding to each remote sensing image data respectively according to the image similarity through a loss function includes:

[0023] For one of the remote sensing image data, determining a logarithmic value of a single image similarity corresponding to the remote sensing image data;

[0024] Dividing the logarithm of the single image similarity by the sum of all the cross-image similarities corresponding to the remote sensing image data, as the loss value of the similar lithological feature representation;

[0025] The loss values ​​represented by all the similar lithological features are averaged to obtain the single image contrast loss.

[0026] Optionally, obtaining the cross-image contrast loss according to the same lithological feature representation of multi-source image data includes:

[0027] According to the multi-source image data, a plurality of other lithological feature representations in the multi-source image data that are the same as the lithological feature in the remote sensing image data are obtained, and each of the other lithological feature representations in the multi-source image data is mapped to the same feature space as the lithological feature representation in the remote sensing image data;

[0028] Performing feature fusion according to the lithological feature representation of the remote sensing image data and the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation;

[0029] According to the cross-image lithological feature representation, the cross-image contrast loss is obtained.

[0030] Optionally, the step of performing feature fusion on the lithological feature representation according to the remote sensing image data and the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation includes:

[0031] Normalizing the other lithological feature representations of the remote sensing image data to convert the other lithological feature representations into unit vectors;

[0032] The unit vector corresponding to each of the other lithological feature representations is mapped onto a hypersphere of a preset radius to obtain the cross-image lithological feature representation.

[0033] Optionally, constructing a classifier and determining a classification loss includes:

[0034] A classifier constructed based on the lithology feature representation and a plurality of other lithology feature representations of the multi-source image data, obtaining the probability of the lithology feature in each lithology category;

[0035] The classification loss is determined according to the probability through a cross entropy loss function.

[0036] Optionally, the initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model, including:

[0037] The single image contrast loss, the cross-image contrast loss and the classification loss are weighted by using preset weight coefficients corresponding to the single image contrast loss, the cross-image contrast loss and the classification loss, respectively, to obtain a joint loss of the lithological feature;

[0038] The initial classification model is trained according to the joint loss, and the trained initial classification model is used as the lithology commonality joint feature classification model.

[0039] The present invention also provides a lithology commonality joint feature classification model construction system, comprising a memory and a processor;

[0040] The memory is used to store computer programs;

[0041] The processor is used to implement the method for constructing a lithology commonality joint feature classification model as described in any one of the above items when executing the computer program.

[0042] The method and system for constructing a lithology commonality joint feature classification model of the present invention extracts local-global features of the lithology features of remote sensing image data to obtain lithology feature representations of remote sensing image data, so that the model can simultaneously capture local details and global context information in the remote sensing image, thereby obtaining a richer and more comprehensive lithology feature representation. Then, through contrast learning optimization, a single image contrast loss is obtained according to the lithology feature representation, and the single image contrast loss emphasizes the feature dispersion of different lithologies in the same image, thereby distinguishing the subtle differences between similar lithology categories and improving the classification accuracy of the lithology classification model. In addition, multi-source image data is combined, and through a feature combination method, data from different images are integrated to extract consistent lithology feature representations, and cross-image contrast loss is obtained, and the cross-image contrast loss emphasizes the consistency of features of the same lithology in different images, so that the features of different lithologies in the feature space are dispersed, and the features of the same lithology are aggregated, thereby overcoming the limitations or deviations that may exist in a single data source, improving the comprehensiveness and reliability of lithology features, and also helping to improve the perception of lithology features. By combining lithology feature representation with multi-source image data to build a classifier and determine the classification loss, the lithology classification model is guided to learn the category information of lithology features. On this basis, the initial model is trained by combining single image contrast loss, cross-image contrast loss and classification loss. The resulting lithology commonality joint feature classification model can fuse lithology features in image data from different sources with different lithology features in the same image data, and use the diversity of lithology to improve the perception of lithology features. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart of a method for constructing a lithology commonality joint feature classification model according to an embodiment of the present invention;

[0044] Figure 2 It is one of the schematic diagrams of the local-global feature extraction process in the method for constructing a lithology commonality joint feature classification model in another embodiment of the present invention;

[0045] Figure 3 This is a second schematic diagram of a local-global feature extraction process in a method for constructing a lithology commonality joint feature classification model in another embodiment of the present invention;

[0046] Figure 4 It is a flowchart of a method for constructing a lithology commonality joint feature classification model in another embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] Combination Figure 1 As shown, the present invention provides a method for constructing a lithology commonality joint feature classification model, comprising:

[0049] The lithological feature representation of the remote sensing image data is obtained by performing local-global feature extraction on the lithological feature of the remote sensing image data.

[0050] Specifically, the lithological characteristics of the remote sensing image data are extracted locally and globally to obtain the lithological characteristics representation of the remote sensing image data. In the preferred embodiment of the present invention, the lithological characteristics of the remote sensing image data are extracted locally and globally to obtain the lithological characteristics representation of the remote sensing image data. Figure 4 As shown in the figure, the transformer model is used to extract spatial information and low-level feature information through convolutions of different sizes, pooling is used to reduce the size of the feature map, and then a segmentation operation is performed based on convolution to obtain the lithological feature representation of the remote sensing image data.

[0051] Through contrast learning optimization, a single image contrast loss is obtained according to the lithological feature representation.

[0052] Specifically, through contrast learning optimization, the single image contrast loss is obtained according to the lithological feature representation. This step applies the feature dispersion method, and the feature representation of the same source lithological images of different categories can be extracted respectively through the deep encoder network. In the preferred embodiment of the present invention, the deep encoder network is a transformer model based on the convolution segmentation strategy, combined with Figure 4As shown in the figure, the feature projection network is used to transform the feature representation into the same feature hypersphere, calculate the image similarity, and apply the contrast learning idea to construct the loss. The network is constrained to obtain the lithological feature representation of the image, which promotes the dispersion of features of similar lithologies and obtains the single image contrast loss.

[0053] The cross-image comparison loss is obtained according to the same lithological feature representation of multi-source image data, and a classifier is constructed to determine the classification loss.

[0054] Specifically, according to the lithology feature representation and multi-source image data, the cross-image contrast loss is obtained, wherein this step effectively fuses the multi-source image data with the remote sensing image data. In a preferred embodiment of the present invention, a feature combination method can be used to extract the feature representation of the same category of lithology from different sources, and combined with Figure 4 As shown, the feature representation is transformed into the same feature space through three fully connected layers to obtain the same feature space representation. The consistent feature representation of the same lithology in multi-source image data is extracted by constructing contrast loss, and the cross-image contrast loss is obtained. The classifier is constructed by lithology feature representation and the classification loss is determined. The same lithology feature representation in multi-source image data is used to fuse and construct a classifier to obtain the classification loss.

[0055] The initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model.

[0056] Specifically, by combining single-image contrast loss, cross-image contrast loss and classification loss to train the model, the model can simultaneously optimize the discrimination of lithological features within a single image, the consistency between different images and the classification accuracy, thereby obtaining a better feature representation.

[0057] The method for constructing a lithology commonality joint feature classification model of the present invention obtains the lithology feature representation of the remote sensing image data by performing local-global feature extraction on the lithology features of the remote sensing image data, so that the model can simultaneously capture the local details and global context information in the remote sensing image, thereby obtaining a richer and more comprehensive lithology feature representation. Then, through contrast learning optimization, a single image contrast loss is obtained according to the lithology feature representation, and the single image contrast loss is used to emphasize the feature dispersion of different lithologies in the same image, so as to distinguish the subtle differences between similar lithology categories and improve the classification accuracy of the lithology classification model. In addition, multi-source image data is combined, and through a feature joint method, data from different images are integrated to extract a consistent lithology feature representation, and a cross-image contrast loss is obtained, and the cross-image contrast loss is used to emphasize the consistency of the features of the same lithology in different images, so that the features of different lithologies in the feature space are dispersed, and the features of the same lithology are aggregated, thereby overcoming the limitations or deviations that may exist in a single data source, improving the comprehensiveness and reliability of lithology features, and also helping to improve the perception of lithology features. By combining lithology feature representation with multi-source image data to build a classifier and determine the classification loss, the lithology classification model is guided to learn the category information of lithology features. On this basis, the initial model is trained by combining single image contrast loss, cross-image contrast loss and classification loss. The resulting lithology commonality joint feature classification model can fuse lithology features in image data from different sources with different lithology features in the same image data, and use the diversity of lithology to improve the perception of lithology features.

[0058] In the embodiment of the present invention, the step of extracting local-global features of the lithological features of the remote sensing image data to obtain the lithological feature representation of the remote sensing image data includes:

[0059] Extracting features from the remote sensing image data using a convolution kernel of a preset size to obtain a multi-scale feature map;

[0060] Slicing the multi-scale feature map to divide the remote sensing image data into multiple feature maps;

[0061] The feature map is sequentially passed through the three layers of transformer modules for feature extraction to obtain the lithological feature representation.

[0062] In this embodiment, the remote sensing image data is feature extracted by a convolution kernel of a preset size to obtain a multi-scale feature map, wherein, Figure 2As shown in the figure, the transformer feature extraction structure based on the convolutional segmentation strategy in this study first obtains a multi-scale feature map through the convolutional segmentation strategy, then segments the multi-scale feature map, and inputs the feature map blocks generated after segmentation into the subsequent transformer model. This strategy uses convolutional neural networks to increase the ability to extract low-level feature information, thereby reducing the size of the feature map, reducing the operation parameters, and including spatial information and low-level features. Figure 3 As shown in the figure, after passing through three layers of transformer modules, the stacking of three layers of transformer modules can further deepen the fusion and representation of features, so that the model can understand and express the lithological characteristics more carefully. Among them, the transformer module is composed of a multi-layer perceptron and multi-head attention, which is used to extract global features.

[0063] The method for constructing a lithology commonality joint feature classification model of the present invention can simultaneously extract features of remote sensing image data at different scales through convolution operations of preset sizes, which helps to capture the performance of lithology features at different resolutions and increase the richness of features. After feature extraction, the size of the feature map is reduced through pooling operations, which effectively reduces the operation parameters in subsequent processing, thereby reducing the computational complexity and resource consumption of the model. The Transformer module, especially the multi-head attention mechanism therein, is used to capture long-distance dependencies in remote sensing image data and extract global features.

[0064] In the embodiment of the present invention, the feature map is sequentially subjected to feature extraction through three layers of transformer modules to obtain the lithology feature representation, including:

[0065] Through the attention mechanism, the feature map input into the transformer module is linearly processed to obtain a query vector, a key-value vector, and a value vector of the feature map;

[0066] The query vector, the key-value vector and the value vector are normalized by a normalization function to obtain the output of the transformer module, and the lithology feature representation is obtained according to the output.

[0067] In this embodiment, the residual idea is adopted to connect the output features of the transformer module to achieve effective fusion of shallow features and deep features. The attention mechanism formula needs to be used for connection, where the attention mechanism formula is as follows:

[0068]

[0069] Among them, Q, K, V, are the input query vector, key value vector and value vector respectively, dk is the dimension of the query vector, key-value vector, and value vector. Softmax means normalizing the weights so that the sum of the query vector, key-value vector, and value vector is equal to 1.

[0070] The method for constructing a lithology commonality joint feature classification model of the present invention adopts the residual idea to connect the output features of the Transformer module with the input features, thereby achieving effective fusion of shallow and deep features.

[0071] In the embodiment of the present invention, the single image contrast loss is obtained according to the lithological feature representation through contrast learning optimization, including:

[0072] Obtaining, through a deep encoder network, a plurality of similar lithological feature representations corresponding to each of the remote sensing image data according to the plurality of remote sensing image data;

[0073] Mapping the similar lithology feature representation and the lithology feature representation to the same feature hypersphere through a feature projection network, and determining the image similarity corresponding to each of the remote sensing image data;

[0074] Wherein, for one of the remote sensing image data, the image similarity includes: cross-image similarity and single-image similarity, the cross-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the similar lithological feature representations corresponding to any other remote sensing image data, and the single-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the lithological feature representation corresponding to the remote sensing image data;

[0075] The single image comparison loss corresponding to each remote sensing image data is determined according to the image similarity through a loss function.

[0076] In this embodiment, through contrastive learning optimization, the Transformer module is used to obtain the single image contrast loss according to the lithology feature representation. Since lithology is diverse and complex, different rock types may have similar characteristics. Therefore, through contrastive learning, similar lithology categories are distinguished, and the feature differences and similarities between lithologies are learned to improve the accuracy of lithology classification. In a preferred embodiment of the present invention, a feature dispersion method is adopted to obtain a single image contrast loss. Its purpose is to promote the dispersion of features of similar lithologies. Category information is needed to determine whether lithology images belong to the same class. The basis of contrastive learning has changed from "whether it comes from the same image" to "whether it belongs to the same class." Combined with Figure 4As shown, the feature projection network is used to transform the feature representation into the same feature hypersphere to obtain the same feature space representation, and then the image similarity is calculated. Among them, it is worth mentioning that the similar lithology feature representation is the feature representation that belongs to the same category of features as the lithology feature representation. Image similarity is divided into two types: cross-image similarity and single-image similarity. Cross-image similarity measures the similarity between different remote sensing image data. If the lithology feature representations of two image data are close to each other on the hypersphere, then their cross-image similarity is high. Single-image similarity measures the similarity between different lithology feature representations within the same remote sensing image data, which helps to identify the consistency and difference within the image. The loss function is designed to optimize the network parameters so that similar lithology feature representations are closer on the hypersphere, while dissimilar feature representations are farther away, thereby improving the quality of feature representation. On this basis, the loss function is used to calculate the corresponding loss. Among them, the first step is to select a sample i, where sample i is one of the similar lithology feature representations, and for each sample i in the data set, calculate its feature similarity with all other samples. The second step is to calculate the similarity between sample i and all other samples k, where other samples k represent other similar lithological features. The third step is to calculate the feature dot product similarity between each sample j (same as sample i) and sample i, and scale it by the temperature parameter τ. Finally, the single image comparison loss is calculated by the loss function. Among them, the sample is Figure 4 Samples 1 to n in the optical image.

[0077] The method for constructing a lithology commonality joint feature classification model of the present invention obtains a single image contrast loss according to the lithology feature representation through a contrastive learning optimization method, thereby achieving the purpose of improving the accuracy of lithology classification. Among them, the core advantage of contrastive learning is that it can distinguish and learn the characteristic differences and similarities between lithologies when the lithologies are diverse and complex.

[0078] In the embodiment of the present invention, determining the single image contrast loss corresponding to each remote sensing image data according to the image similarity through a loss function includes:

[0079] For one of the remote sensing image data, determining a logarithmic value of a single image similarity corresponding to the remote sensing image data;

[0080] Dividing the logarithm of the single image similarity by the sum of all the cross-image similarities corresponding to the remote sensing image data, as the loss value of the similar lithological feature representation;

[0081] The loss values ​​represented by all the similar lithological features are averaged to obtain the single image contrast loss.

[0082] In this embodiment, the single image contrast loss is calculated by the loss function, and the loss value needs to be calculated for each sample i. The loss value is calculated by selecting a sample j of the same type as i, calculating the logarithm of the dot product similarity of its feature vector, and dividing it by the sum of the dot product similarities of other samples. Finally, the loss values ​​of all samples are averaged.

[0083] Among them, the loss function is:

[0084]

[0085] in, is the sum of the single image contrast losses of all samples, is the single image contrast loss of sample i. k is the lithology characteristic vector of the kth sample. j The feature vector of the jth sample that belongs to the same class as the ith sample. N is the total number of samples. Represents the category label of the i-th sample. is the number of samples that belong to the same class as the i-th sample. i is the feature vector of the i-th sample. τ is the temperature parameter used to adjust the smoothness of the similarity distribution. i≠j Represents the indicator function, which is 1 when i is not equal to j, otherwise it is 0. Represents the indicator function, which is used to ensure category consistency. If the category of the i-th sample is 1 if the value is 0, otherwise it is 0.

[0086] The method for constructing a lithology commonality joint feature classification model of the present invention effectively quantifies the similarity between similar lithology feature representations and promotes the distinction of similar lithology features in feature space by using a loss function to calculate the single image contrast loss. The loss function uses logarithmic values ​​to enhance the model's sensitivity to subtle differences, and at the same time, by dividing the logarithmic value by the sum of the similarities of similar samples, it balances the contribution of different samples to the loss value, so that the loss function more fairly reflects the similarity distribution of each sample.

[0087] In the embodiment of the present invention, obtaining the cross-image contrast loss according to the same lithological feature representation of multi-source image data includes:

[0088] According to the multi-source image data, a plurality of other lithological feature representations in the multi-source image data that are the same as the lithological feature in the remote sensing image data are obtained, and each of the other lithological feature representations in the multi-source image data is mapped to the same feature space as the lithological feature representation in the remote sensing image data;

[0089] Performing feature fusion according to the lithological feature representation of the remote sensing image data and the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation;

[0090] According to the cross-image lithological feature representation, the cross-image contrast loss is obtained.

[0091] In this embodiment, combined with Figure 4 As shown in the multi-source sample feature fusion in , the lithological feature representation is feature fused with other lithological feature representations with the same lithological feature in the multi-source image data to obtain the cross-image contrast loss. Usually, a feature combination method is used to obtain other lithological feature representations with the same lithological feature in the multi-source image data according to the multi-source image data, and other lithological feature representations are mapped to the same feature space as the lithological feature representation, that is, feature representations of the same category of lithology from different sources are extracted, and the feature projection network is used to transform the feature representation to the same feature space, and the lithological feature representation is feature fused with other lithological feature representations to obtain a cross-image lithological feature representation, and then the cross-image contrast loss is obtained according to the cross-image lithological feature representation, which can be calculated by the corresponding loss function. The following loss function is generally used for training:

[0092]

[0093] in, is the sum of the cross-image contrast losses of all samples, is the cross-image contrast loss of sample i. τ is a constant greater than 0, z k is the lithology characteristic vector of the kth sample. j The feature vector of the jth sample that belongs to the same class as the i-th sample. exp(z i ·z j / τ) is the dot product of the feature vectors of the i-th sample and the j-th sample, and the similarity is calculated using an exponential function. Calculate the sum of the dot products of the i-th sample and all other similar lithology feature vectors. k indicates the index used to traverse the sample set.

[0094] The method for constructing a lithology commonality joint feature classification model of the present invention realizes the integration and optimization of the same lithology features in multi-source image data by obtaining a method for cross-image contrast loss through feature fusion. Through precise feature extraction technology, feature representations matching the target lithology features are identified and extracted from multi-source image data, and mapped to a unified feature space, ensuring that similar lithologies in different data sources can be compared and learned under a common framework. Then, the target lithology features are combined with other lithology feature representations through feature fusion technology to generate cross-image lithology feature representations, thereby enhancing the expressiveness of features and improving the recognition accuracy and robustness of the model for lithology categories.

[0095] In the embodiment of the present invention, the lithological feature representation according to the remote sensing image data is subjected to feature fusion with the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation, including:

[0096] Normalizing the other lithological feature representations of the remote sensing image data to convert the other lithological feature representations into unit vectors;

[0097] The unit vector corresponding to each of the other lithological feature representations is mapped onto a hypersphere of a preset radius to obtain the cross-image lithological feature representation.

[0098] In this embodiment, other lithologic feature representations are normalized to obtain unit vectors corresponding to other lithologic feature representations, and each unit vector corresponding to other lithologic feature representations is mapped to a hypersphere of a preset radius to obtain a cross-image lithologic feature representation. In a preferred embodiment of the present invention, it is assumed that a round contains N images, which are used as subsequent network inputs for training. After calculation by the encoder network, N images will generate N lithologic feature vectors, and each lithologic feature vector is normalized to become a unit vector. In this way, each lithologic feature vector falls on a hypersphere with a radius of 1.

[0099] The obtained feature vector is: feature = {z 1 ,z 2 ,…,z N}; For any image i, there is an image j in the remaining N-1 images, and images i and j come from the same image. Because they come from the same image, the feature vectors of images i and j are as close as possible; except for images i and j, for the remaining N-2 images, because they come from different images from image i, their feature vectors are as far away from the feature vector of image i as possible.

[0100] The method for constructing a lithology commonality joint feature classification model of the present invention obtains a cross-image lithology feature representation by fusing lithology feature representation with other lithology feature representations, thereby effectively enhancing the model's ability to characterize lithology features.

[0101] In the embodiment of the present invention, the step of constructing a classifier and determining the classification loss includes:

[0102] A classifier constructed based on the lithology feature representation and a plurality of other lithology feature representations of the multi-source image data, obtaining the probability of the lithology feature in each lithology category;

[0103] The classification loss is determined according to the probability through a cross entropy loss function.

[0104] In this embodiment, combined with Figure 4 As shown in FIG. 1 , the classifier constructed according to the lithological feature representation obtains the probability of the lithological feature in each lithological category through the fully connected layer, and the classification loss is determined according to the probability through the cross entropy loss function. The classification loss is used to guide the network to learn the lithological features. The cross entropy loss function is:

[0105]

[0106] Where p = [p 0 ,…,p C-1 ] is a distribution of the probability of a lithology feature in each lithology category, and each element p i Indicates the probability that the sample belongs to the i-th category; y = [y 0 ,…,y C-1 ] is the onehot representation of the sample label. When the sample belongs to the i-th category, y i =1, otherwise y i =0; c is the lithology label.

[0107] The method for constructing a lithology commonality joint feature classification model of the present invention constructs a classifier through lithology feature representation and determines the classification loss, thereby effectively improving the accuracy and efficiency of lithology identification.

[0108] In the embodiment of the present invention, the initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model, including:

[0109] The single image contrast loss, the cross-image contrast loss and the classification loss are weighted by using preset weight coefficients corresponding to the single image contrast loss, the cross-image contrast loss and the classification loss, respectively, to obtain a joint loss of the lithological feature;

[0110] The initial classification model is trained according to the joint loss, and the trained initial classification model is used as the lithology commonality joint feature classification model.

[0111] In this embodiment, combined with Figure 4 As shown, by using the preset weight coefficients corresponding to the single image contrast loss, the cross-image contrast loss and the classification loss, weighted processing is performed according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain the joint loss, and the initial classification model is trained based on the joint loss. In a preferred embodiment of the present invention, a network total loss function is used to weight the single image contrast loss, the cross-image contrast loss and the classification loss, wherein the network total loss function is:

[0112]

[0113] Among them, α, β, and γ are weight coefficients obtained based on experience. is the single image contrast loss, is the cross-image contrast loss, is the classification loss, For joint losses.

[0114] The method for constructing a lithology commonality joint feature classification model of the present invention obtains the classification results of the lithology characteristics of remote sensing image data through weighted processing of single image contrast loss, cross-image contrast loss and classification loss, and combines the advantages of different loss functions to achieve more accurate lithology characteristic recognition. Single image contrast loss focuses on the distinction of lithology characteristics within the same image, which helps to capture the lithology details and differences within the image; cross-image contrast loss focuses on the consistency of the same lithology characteristics between different images, which enhances the model's ability to generalize lithology characteristics; and classification loss directly optimizes the prediction of lithology categories, improving the accuracy of classification.

[0115] The present invention also provides a lithology commonality joint feature classification model construction system, comprising a memory and a processor;

[0116] The memory is used to store computer programs;

[0117] The processor is used to implement the method for constructing a lithology commonality joint feature classification model as described in any one of the above items when executing the computer program.

[0118] The lithology commonality joint feature classification model construction system of the present invention obtains the lithology feature representation of the remote sensing image data by performing local-global feature extraction on the lithology features of the remote sensing image data, so that the model can simultaneously capture the local details and global context information in the remote sensing image, thereby obtaining a richer and more comprehensive lithology feature representation. Then, through contrast learning optimization, a single image contrast loss is obtained according to the lithology feature representation, and the single image contrast loss is used to emphasize the feature dispersion of different lithologies in the same image, so as to distinguish the subtle differences between similar lithology categories and improve the classification accuracy of the lithology classification model. In addition, multi-source image data is combined, and through the feature combination method, data from different images are integrated to extract consistent lithology feature representations, and cross-image contrast loss is obtained. The cross-image contrast loss is used to emphasize the consistency of the features of the same lithology in different images, so that the features of different lithologies in the feature space are dispersed, and the features of the same lithology are aggregated, thereby overcoming the limitations or deviations that may exist in a single data source, improving the comprehensiveness and reliability of lithology features, and also helping to improve the perception of lithology features. By combining lithology feature representation with multi-source image data to build a classifier and determine the classification loss, the lithology classification model is guided to learn the category information of lithology features. On this basis, the initial model is trained by combining single image contrast loss, cross-image contrast loss and classification loss. The resulting lithology commonality joint feature classification model can fuse lithology features in image data from different sources with different lithology features in the same image data, and use the diversity of lithology to improve the perception of lithology features.

[0119] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0120] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0121] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for constructing a lithology commonality joint feature classification model, characterized in that: include: By performing local-global feature extraction on the lithological features of the remote sensing image data, a lithological feature representation of the remote sensing image data is obtained; Through contrast learning optimization, a single image contrast loss is obtained according to the lithological feature representation, specifically including: through a deep encoder network, according to a plurality of the remote sensing image data, a plurality of similar lithological feature representations corresponding to each of the remote sensing image data are obtained; through a feature projection network, the similar lithological feature representations and the lithological feature representations are mapped to the same feature hypersphere, and the image similarity corresponding to each of the remote sensing image data is determined; wherein, for one of the remote sensing image data, the image similarity includes: cross-image similarity and single-image similarity, the cross-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the similar lithological feature representations corresponding to any other of the remote sensing image data, and the single-image similarity represents the similarity between any one of the similar lithological feature representations corresponding to the remote sensing image data and the lithological feature representation corresponding to the remote sensing image data; the single image contrast loss corresponding to each of the remote sensing image data is determined according to the image similarity through a loss function; According to the lithological feature representation in the multi-source image data that is the same as the lithological feature in the remote sensing image data, a cross-image comparison loss is obtained, and a classifier is constructed to determine the classification loss; wherein, the cross-image comparison loss is obtained according to the lithological feature representation that is the same as the lithological feature in the multi-source image data, specifically comprising: according to the multi-source image data, a plurality of other lithological feature representations in the multi-source image data that are the same as the lithological feature in the remote sensing image data are obtained, and each of the other lithological feature representations in the multi-source image data is mapped to the same feature space as the lithological feature representation of the remote sensing image data; according to the lithological feature representation of the remote sensing image data, feature fusion is performed with the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation; according to the cross-image lithological feature representation, a cross-image comparison loss is obtained; The initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model.

2. The method for constructing a lithology commonality joint feature classification model according to claim 1, characterized in that: The method of extracting local-global features of the lithological features of the remote sensing image data to obtain the lithological feature representation of the remote sensing image data includes: Extracting features from the remote sensing image data using a convolution kernel of a preset size to obtain a multi-scale feature map; Slicing the multi-scale feature map to divide the remote sensing image data into multiple feature maps; The feature map is sequentially passed through three layers of transformer modules for feature extraction to obtain the lithology feature representation.

3. The method for constructing a lithology commonality joint feature classification model according to claim 2, characterized in that: The feature map is sequentially subjected to feature extraction through three layers of transformer modules to obtain the lithology feature representation, including: Through the attention mechanism, the feature map input into the transformer module is linearly processed to obtain a query vector, a key-value vector, and a value vector of the feature map; The query vector, the key-value vector and the value vector are normalized by a normalization function to obtain the output of the transformer module, and the lithology feature representation is obtained according to the output.

4. The method for constructing a lithology commonality joint feature classification model according to claim 1, characterized in that: Determining the single image contrast loss corresponding to each remote sensing image data according to the image similarity through a loss function includes: For one of the remote sensing image data, determining a logarithmic value of a single image similarity corresponding to the remote sensing image data; Dividing the logarithm of the single image similarity by the sum of all the cross-image similarities corresponding to the remote sensing image data, as the loss value of the similar lithological feature representation; The loss values ​​represented by all the similar lithological features are averaged to obtain the single image contrast loss.

5. The method for constructing a lithology commonality joint feature classification model according to claim 1, characterized in that: The step of fusing the lithological feature representation according to the remote sensing image data with the other lithological feature representations of the multi-source image data to obtain a cross-image lithological feature representation includes: Normalizing the other lithological feature representations of the remote sensing image data to convert the other lithological feature representations into unit vectors; The unit vector corresponding to each of the other lithological feature representations is mapped onto a hypersphere of a preset radius to obtain the cross-image lithological feature representation.

6. The method for constructing a lithology commonality joint feature classification model according to claim 1, characterized in that: The step of constructing a classifier and determining the classification loss includes: A classifier constructed based on the lithology feature representation and a plurality of other lithology feature representations of the multi-source image data, obtaining the probability of the lithology feature in each lithology category; The classification loss is determined according to the probability through a cross entropy loss function.

7. The method for constructing a lithology commonality joint feature classification model according to claim 1, characterized in that: The initial classification model is trained according to the single image contrast loss, the cross-image contrast loss and the classification loss to obtain a lithology commonality joint feature classification model, including: The single image contrast loss, the cross-image contrast loss and the classification loss are weighted by using preset weight coefficients corresponding to the single image contrast loss, the cross-image contrast loss and the classification loss, respectively, to obtain a joint loss of the lithological feature; The initial classification model is trained according to the joint loss, and the trained initial classification model is used as the lithology commonality joint feature classification model.

8. A lithology commonality joint feature classification model construction system, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is used to implement the method for constructing a lithology commonality joint feature classification model as described in any one of claims 1 to 7 when executing the computer program.

Images