Lithology scene classification method and system and storage medium

Abstract

The invention provides a lithologic scene classification method and system and a storage medium, and relates to the technical field of image processing, and the method comprises the steps: carrying out the feature extraction of a plurality of remote sensing images of a detection region, and obtaining a plurality of remote sensing image features, the remote sensing image features comprising spatial features and spectral features; performing adaptive learning on the spatial features and the spectral features to generate an adjacent matrix; updating the remote sensing image features through a graph convolutional network according to information between nodes in the adjacent matrix to obtain global-local features corresponding to the remote sensing image features; performing comparative learning on the global-local features through a comparative learning model to obtain optimized final features; and carrying out feature fusion and classification processing on the final features to obtain a lithologic scene classification result of the detection area. According to the method, the lithology categories with close feature distribution in the remote sensing image are effectively classified, and the classification precision of the lithology categories is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a lithology scene classification method, system and storage medium. Background Art

[0002] Lithology scene classification has important application value in many fields such as geological survey, mineral resource exploration and environmental protection. Traditional lithology classification mainly relies on field sampling and manual interpretation, which is not only time-consuming and labor-intensive, but also limited by environmental factors such as geography and weather, making it difficult to achieve large-scale and refined lithology information acquisition.

[0003] In the prior art, with the development of remote sensing technology, remote sensing images are generally used for lithology classification. However, due to the limitations of its resolution and spectral characteristics, a single remote sensing image is difficult to provide comprehensive lithology information. At the same time, since the performance differences of different lithologies on a single spectral image are not obvious, the lithology scene features extracted from a single remote sensing image are relatively similar, which easily leads to confusion of lithology categories and errors in lithology scene classification. Summary of the invention

[0004] The problem solved by the present invention is how to classify lithology categories with similar feature distribution in remote sensing images.

[0005] In order to solve the above problems, the present invention provides a lithology scene classification method, system and storage medium.

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

[0007] Extracting features from multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features;

[0008] Adaptively learning the spatial features and the spectral features to generate an adjacency matrix;

[0009] By using a graph convolutional network, the remote sensing image features are updated according to information between nodes in the adjacency matrix to obtain global-local features corresponding to the remote sensing image features;

[0010] By using a contrastive learning model, the global-local features are contrastively learned to obtain optimized final features;

[0011] The final features are subjected to feature fusion and classification processing to obtain the lithology scene classification result of the detection area.

[0012] Optionally, the extracting features of multiple remote sensing images of the detection area to obtain multiple remote sensing image features includes:

[0013] Preprocessing the multiple remote sensing images to obtain multiple remote sensing images with a unified spatial coverage range;

[0014] The multiple remote sensing images with unified spatial coverage are input into a convolutional neural network for feature extraction to obtain the spatial features and the spectral features of the multiple remote sensing images, and the spatial features and the spectral features are used as the remote sensing image features.

[0015] Optionally, the adaptively learning the spatial features and the spectral features to generate an adjacency matrix includes:

[0016] Resizing the feature graph of the spatial feature and the feature graph of the spectral feature to obtain the feature graph of the spatial feature and the feature graph of the spectral feature that match the size of the initial matrix;

[0017] Determining the nodes of the initial matrix according to the pixel positions in the feature graph of the spatial feature and the feature graph of the spectral feature, wherein each pixel position corresponds to one node;

[0018] The adjacency matrix is ​​obtained by adaptively learning the feature graphs of the spatial features and the spectral features and updating the nodes of the initial matrix.

[0019] Optionally, updating the remote sensing image features according to information between nodes in the adjacency matrix through a graph convolutional network to obtain global-local features corresponding to the remote sensing image features includes:

[0020] Inputting the adjacency matrix and the remote sensing image features into the graph convolutional network, and updating the remote sensing image features through the convolutional layer of the graph convolutional network;

[0021] wherein, the output features of each of the convolutional layers are aggregated according to the information between the nodes in the adjacency matrix to obtain the local features of the remote sensing image features;

[0022] Obtaining the global features of the remote sensing image features according to the output of the last convolutional layer of the graph convolutional network;

[0023] By fusing the local features with the global features, global-local features corresponding to the remote sensing image features are obtained.

[0024] Optionally, performing comparative learning on the global-local features through a comparative learning model to obtain optimized final features includes:

[0025] Inputting the global-local features into the contrastive learning model, and determining the similarity between each feature in the global-local features and other features;

[0026] By comparing the loss function, all the features of the global-local features are divided according to the similarity, and the features are divided into similar features and heterogeneous features;

[0027] The global-local features are optimized according to the homogeneous features and the heterogeneous features to obtain the optimized final features.

[0028] Optionally, the optimizing the global-local features according to the homogeneous features and the heterogeneous features to obtain the optimized final features includes:

[0029] Aggregating the similar features to obtain aggregated features;

[0030] Performing dispersion processing on the heterogeneous features to obtain dispersed features;

[0031] The optimized final feature is obtained according to the aggregation feature and the dispersion feature.

[0032] Optionally, the multiple remote sensing images include ZY1 images, Landsat images and GF5 images;

[0033] The performing feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area includes:

[0034] By channel stitching, the final features of the Landsat image and the GF5 image are fused to obtain stitching features;

[0035] Determine the classification probability of each lithology category according to the splicing features and the final features of the ZY1 image through soft classification;

[0036] Determining weight values ​​of the splicing features and the final features of the ZY1 image according to the classification probability;

[0037] The classification probability of the spliced ​​feature and the classification probability of the final feature of the ZY1 image are weighted and fused according to the weight value, so as to obtain the lithology classification label of each pixel in the multi-remote sensing image, and the lithology classification label is used as the lithology scene classification result

[0038] Optionally, it also includes:

[0039] Establishing the contrastive learning model, which specifically includes:

[0040] Obtain lithology characteristic data sets;

[0041] constructing positive sample pairs according to the same lithology labels in the lithology feature data set, and constructing negative sample pairs according to different lithology labels in the lithology feature data set;

[0042] According to the positive sample pairs and the negative sample pairs, the initial learning model is trained to maximize the similarity of positive samples and to minimize the similarity of negative samples;

[0043] The trained initial learning model is used as the comparison learning model.

[0044] In a second aspect, the present invention provides a lithology scene classification system, comprising:

[0045] A feature extraction unit is used to extract features from multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features;

[0046] An adaptive learning unit, used for adaptively learning the spatial features and the spectral features to generate an adjacency matrix;

[0047] A graph convolution unit, used to update the remote sensing image features according to the information between the nodes in the adjacency matrix through a graph convolution network, so as to obtain the global-local features corresponding to the remote sensing image features;

[0048] A graph contrast learning unit, used to perform contrast learning on the global-local features through a contrast learning model to obtain optimized final features;

[0049] The fusion and classification unit is used to perform feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area.

[0050] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the above-mentioned lithology scene classification method is implemented.

[0051] The lithology scene classification method, system and storage medium of the present invention capture the spatial and spectral features of each image by extracting features from multi-source remote sensing images. Then, an adjacency matrix is ​​generated using an adaptive learning method, and the extracted spatial features and spectral features are adaptively learned to generate an adjacency matrix. The adjacency matrix describes the relationship between features and is used in a graph convolutional network to simulate the connection strength between features, providing structured information for subsequent feature updates. Then, the remote sensing image features are updated according to the adjacency matrix through a graph convolutional network, and feature representations containing global and local information are extracted, so that the model can capture richer contextual information, thereby enhancing the ability to distinguish lithology categories. Further, these global-local features are optimized by contrasting learning models, enhancing the model's ability to distinguish similar lithology features, and improving the classification accuracy of lithology categories close to feature distribution. Finally, the optimized final features are subjected to feature fusion and classification processing to obtain the final lithology scene classification results. The present invention provides rich initial information for the model through feature extraction, and uses the adjacency matrix generated by adaptive learning to provide structured relationship information for the graph convolutional network. The graph convolutional network and contrastive learning model further enhance the feature differentiation capability. Finally, accurate lithology scene classification results are obtained through feature fusion and classification processing, which effectively classifies lithology categories with similar feature distribution in remote sensing images and improves the classification accuracy of lithology categories. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 Schematic diagram of the process of the lithology scene classification method in an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of lithology scene classification in an embodiment of the present invention;

[0054] Figure 3 Schematic diagram of the structure of the lithology scene classification system in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] 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.

[0056] 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.

[0057] 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.

[0058] 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".

[0059] 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.

[0060] In view of the problems existing in the above-mentioned related technologies, this embodiment provides a lithology scene classification method, system and storage medium.

[0061] Combination Figure 1 As shown, a lithology scene classification method provided by an embodiment of the present invention includes:

[0062] Feature extraction is performed on multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features.

[0063] Specifically, we first extract features from multi-source remote sensing images of the detection area, which involves extracting key information from images of different resolutions and spectral ranges. Spatial features can be extracted from high-resolution ZY1 images, including but not limited to detailed information such as image texture, shape, and edges. At the same time, spectral features are also extracted from Landsat and GF5 images, which reflect the reflectivity differences of different lithologies in different spectral bands. These spatial and spectral features together constitute the initial feature representation of the remote sensing image, providing rich information for subsequent classification.

[0064] In a preferred embodiment of the present invention, three remote sensing images with different resolutions and spectral characteristics, ZY1, Landsat and GF5, are used. ZY1 images provide rich spatial details with their 2.5-meter high resolution, while Landsat and GF5 images provide spectral information with their 30-meter resolution. Convolutional neural networks (such as ResNet50) are used to extract the spatial and spectral features of these images, which include lithology boundaries and structural features, as well as subtle differences in the spectrum of different lithologies.

[0065] Adaptively learn the spatial features and the spectral features to generate an adjacency matrix.

[0066] Specifically, the extracted spatial and spectral features are adaptively learned to generate an adjacency matrix. This step is achieved through a neural network, which learns how to construct an adjacency matrix based on the spatial and spectral similarities between features. This adjacency matrix not only reflects the connection strength between features, but can also be dynamically adjusted to better capture the complex relationship between lithological features. In this way, a graph convolutional network that can capture the structured information of image features is constructed.

[0067] Through the graph convolutional network, the remote sensing image features are updated according to the information between the nodes in the adjacency matrix to obtain the global-local features corresponding to the remote sensing image features.

[0068] Specifically, the graph convolutional network is used to update the remote sensing image features according to the node information in the adjacency matrix. In this process, the feature representation of each node (i.e., each pixel or image area) is aggregated and updated according to the features of its neighboring nodes, thereby extracting a feature representation containing global and local information. This feature update mechanism enables the model to capture local details and global structures in the image data, enhancing the ability to distinguish lithology categories.

[0069] The global-local features are subjected to contrastive learning through a contrastive learning model to obtain optimized final features.

[0070] Specifically, the global-local features are further optimized by contrastive learning models. In this step, the model learns how to enhance the similarity of similar lithological features and reduce the similarity of heterogeneous lithological features by contrastive learning loss functions. In a preferred embodiment of the present invention, the model constructs positive and negative sample pairs, where the positive sample pairs refer to features with the same lithological label, and the negative sample pairs refer to features with different lithological labels. In this way, the model can more accurately distinguish lithological categories with similar feature distributions. In a preferred embodiment of the present invention, optimization is performed by maximizing the similarity of positive sample features and minimizing the similarity of negative sample features, thereby improving the ability to distinguish lithological categories.

[0071] The final features are subjected to feature fusion and classification processing to obtain the lithology scene classification result of the detection area.

[0072] Specifically, the optimized final features are finally subjected to feature fusion and classification processing. The remote sensing image features from different sources are weighted and fused, and a classifier, such as a softmax classifier, is used to classify the fused features. In this way, the lithology scene classification results of the detection area can be obtained. Feature fusion not only integrates the advantages of multi-source data, but also improves the accuracy and robustness of classification.

[0073] In summary, combined with Figure 2 As shown in the figure, it starts with multiple input layers, which may represent different remote sensing image data sources, such as ZY1 image, Landsat image and GF5 image. Each input layer is subjected to feature extraction through a series of convolutional layers to capture the spatial and spectral characteristics of the image. Subsequently, the extracted features may be fed into the graph convolutional network layers, which update the features through the adjacency matrix generated by adaptive learning to obtain global-local features. The contrastive learning model is then used to perform graph comparison to further optimize the global-local features, and the aggregation of similar lithological features and the dispersion of heterogeneous lithological features are enhanced through the contrastive learning mechanism. Finally, the steps of feature fusion and classification processing are performed, in which the features of Landsat image and GF5 image are fused through channel splicing, that is, the feature fusion module, and then soft-classified with the features of ZY1 image to obtain the classification probability. These probabilities are then weighted fused according to the weight values ​​to obtain the final lithological classification label, which integrates all processed features and gives the classification result, and finally back-propagated through the fusion loss. Back-propagation starts from the loss layer and calculates the gradient of the loss function relative to all weights in the network. These gradients indicate how to adjust the weights to reduce the value of the loss function.

[0074] The lithology scene classification method of the present invention captures the spatial and spectral features of each image by extracting features from multi-source remote sensing images. Then, an adjacency matrix is ​​generated using an adaptive learning method, and the extracted spatial features and spectral features are adaptively learned to generate an adjacency matrix. The adjacency matrix describes the relationship between features and is used in a graph convolutional network to simulate the connection strength between features, providing structured information for subsequent feature updates. Then, the remote sensing image features are updated according to the adjacency matrix through a graph convolutional network, and feature representations containing global and local information are extracted, so that the model can capture richer contextual information, thereby enhancing the ability to distinguish lithology categories. Further, these global-local features are optimized by contrasting learning models, enhancing the model's ability to distinguish similar lithology features, and improving the classification accuracy of lithology categories close to feature distribution. Finally, the optimized final features are subjected to feature fusion and classification processing to obtain the final lithology scene classification results. The present invention provides rich initial information for the model through feature extraction, and uses the adjacency matrix generated by adaptive learning to provide structured relationship information for the graph convolutional network. The graph convolutional network and contrastive learning model further enhance the feature differentiation capability. Finally, accurate lithology scene classification results are obtained through feature fusion and classification processing, which effectively classifies lithology categories with similar feature distribution in remote sensing images and improves the classification accuracy of lithology categories.

[0075] Optionally, the extracting features of multiple remote sensing images of the detection area to obtain multiple remote sensing image features includes:

[0076] Preprocessing the multiple remote sensing images to obtain multiple remote sensing images with a unified spatial coverage range;

[0077] The multiple remote sensing images with unified spatial coverage are input into a convolutional neural network for feature extraction to obtain the spatial features and the spectral features of the multiple remote sensing images, and the spatial features and the spectral features are used as the remote sensing image features.

[0078] Specifically, feature extraction is performed on multiple remote sensing images of the detection area. In a preferred embodiment of the present invention, the preprocessing of ZY1, Landsat and GF5 images is first involved to ensure that all remote sensing images cover the same actual geographical range, which can be 600 meters × 600 meters, so as to achieve consistency in data spatial coverage. In the preprocessing stage, the GF5 image will undergo dimensionality reduction processing to adapt to the input format of the network. Subsequently, these preprocessed images are input into a convolutional neural network, such as ResNet50, to extract spatial and spectral features of multiple remote sensing images, wherein the high resolution of the ZY1 image enables it to have high accuracy in extracting lithology boundaries and structural features, while the Landsat image and GF5 image focus on extracting spectral features, which helps to capture the subtle differences in the spectrum of different lithologies. These extracted features, including the high-resolution spatial features of the ZY1 image and the spectral features of the Landsat and GF5 images, together constitute the remote sensing image features, providing rich input data for subsequent graph convolutional networks and contrastive learning models.

[0079] In the embodiment of the present invention, the spatial consistency of multi-source image data is ensured through preprocessing, precise spatial details are provided by using high-resolution remote sensing images, and the spectral information of spectral remote sensing images is used to enhance the model's ability to distinguish lithology in the spectrum. This joint analysis of multi-source data gives full play to the advantages of different images in spatial resolution and spectral resolution, and realizes the fine classification and accurate identification of lithology scenes.

[0080] Optionally, the adaptively learning the spatial features and the spectral features to generate an adjacency matrix includes:

[0081] Resizing the feature graph of the spatial feature and the feature graph of the spectral feature to obtain the feature graph of the spatial feature and the feature graph of the spectral feature that match the size of the initial matrix;

[0082] Determining the nodes of the initial matrix according to the pixel positions in the feature graph of the spatial feature and the feature graph of the spectral feature, wherein each pixel position corresponds to one node;

[0083] The adjacency matrix is ​​obtained by adaptively learning the feature graphs of the spatial features and the spectral features and updating the nodes of the initial matrix.

[0084] Specifically, in order to ensure the effectiveness of images with different resolutions and spectral features, the process of adaptively learning spatial features and spectral features to generate an adjacency matrix begins with resizing the feature map. In a preferred embodiment of the present invention, the original size of the ZY1 image feature map is 240×240 pixels, while the feature map size of the Landsat image and the GF5 image is 20×20 pixels. In order to make these feature maps adapt to the input of the graph convolutional network, all remote sensing images need to be adjusted to match the size of the initial adjacency matrix through multi-layer convolution and pooling operations, for example, to a size of 16×16. After adjustment, each pixel position is regarded as a node in the initial adjacency matrix, so that each node corresponds to a local feature of the lithological area. Subsequently, through an adaptive learning mechanism, a neural network is used to learn these spatial and spectral feature maps to generate an adjacency matrix. In this process, the nodes of the adjacency matrix are updated according to the spatial and spectral information in the feature graph to determine the strength of the relationship between the features, so that the graph convolutional network can more effectively capture the correlation in the image data. This is achieved by constructing an adaptively learned adjacency matrix, dynamically adjusting the connection strength between nodes during the training process, and optimizing the expressiveness of the features. In a preferred embodiment of the present invention, in order to dynamically adjust the adjacency matrix during the training process so that it can adaptively learn the optimal connection relationship between different lithological features, a learnable parameter matrix θ is introduced. This parameter matrix will be combined with the initial adjacency matrix to obtain an updateable adjacency matrix. The specific process is as follows:

[0085] A * =A+θ;

[0086] Among them, A is the initial adjacency matrix; θ is the learnable parameter matrix; A * is an adaptive adjacency matrix.

[0087] During model training, the value of θ is continuously adjusted based on the feedback of the loss function, thereby changing the connection strength between nodes in the adjacency matrix.

[0088] In the embodiment of the present invention, through resizing and adaptive learning, the generated adjacency matrix can accurately capture the relationship between spatial and spectral features in remote sensing images. This method not only improves the expressiveness of features, but also enhances the model's ability to model complex relationships between lithological features.

[0089] Optionally, updating the remote sensing image features according to information between nodes in the adjacency matrix through a graph convolutional network to obtain global-local features corresponding to the remote sensing image features includes:

[0090] Inputting the adjacency matrix and the remote sensing image features into the graph convolutional network, and updating the remote sensing image features through the convolutional layer of the graph convolutional network;

[0091] wherein, the output features of each of the convolutional layers are aggregated according to the information between the nodes in the adjacency matrix to obtain the local features of the remote sensing image features;

[0092] Obtaining the global features of the remote sensing image features according to the output of the last convolutional layer of the graph convolutional network;

[0093] By fusing the local features with the global features, global-local features corresponding to the remote sensing image features are obtained.

[0094] Specifically, the process of updating remote sensing image features through graph convolutional networks to obtain global-local features first involves inputting the adjacency matrix and remote sensing image features into the graph convolutional network. The graph convolutional network can capture the structured information of image features and generate an adjacency matrix through adaptive learning to achieve modeling of complex relationships between samples. The convolutional layer of the graph convolutional network updates the remote sensing image features, where the information between nodes in the adjacency matrix is ​​used to perform feature aggregation on the output features of each convolutional layer to obtain the local features of the remote sensing image features. The global relationship between lithologies is captured by layer-by-layer aggregation while retaining the local information of the image features. Subsequently, the global features of the remote sensing image features are obtained based on the output of the last convolutional layer of the graph convolutional network. Finally, the global-local features corresponding to the remote sensing image features are obtained by fusing the local features with the global features.

[0095] In the embodiment of the present invention, information from adjacent nodes is integrated through feature aggregation to enhance the recognition ability of local structural features. At the same time, by extracting global features, the global patterns and trends in the image data can be understood. Finally, the fusion of global-local features provides a comprehensive feature representation that not only contains rich local details, but also incorporates global contextual information. By utilizing the complementarity of multi-source remote sensing image data, high-precision classification of complex lithology scenes can be achieved through graph convolutional networks and feature fusion strategies.

[0096] Optionally, performing comparative learning on the global-local features through a comparative learning model to obtain optimized final features includes:

[0097] Inputting the global-local features into the contrastive learning model, and determining the similarity between each feature in the global-local features and other features;

[0098] By comparing the loss function, all the features of the global-local features are divided according to the similarity, and the features are divided into similar features and heterogeneous features;

[0099] The global-local features are optimized according to the homogeneous features and the heterogeneous features to obtain the optimized final features.

[0100] Specifically, the process of optimizing the global-local features through the contrastive learning model first involves inputting these features into the contrastive learning model, wherein the core of the contrastive learning module is to enhance the feature representation through the contrastive learning mechanism, so that the similar lithological features are more aggregated and the different lithological features are more dispersed. In this step, the similarity between each feature and other features in the global-local features is determined. In the preferred embodiment of the present invention, samples with the same lithological label are regarded as positive samples, and samples with different lithological labels are regarded as negative samples. Then, all features of the global-local features are divided according to the similarity through the contrastive loss function, and the features are divided into similar features and heterogeneous features. In this process, the contrastive loss is optimized by maximizing the similarity of positive sample features and minimizing the similarity of negative sample features. Finally, the global-local features are optimized according to the similar features and heterogeneous features to obtain the optimized final features, which can effectively promote the aggregation of similar lithological features, while ensuring the dispersion of different lithological features, and improving the distinguishing ability of the model. In the preferred embodiment of the present invention, the main purpose of the contrastive learning model is to further enhance the similarity of similar lithological features and reduce the similarity of heterogeneous features through the contrastive loss function, thereby optimizing the feature representation generated by the graph convolutional network.

[0101] The contrast loss function is:

[0102]

[0103] Where: d i,j represents the characteristic distance between node i and node j; y i,j =1 indicates a positive sample pair, y i,j = 0 indicates a negative sample pair; m is a threshold hyperparameter used to control the distance range of negative samples. By minimizing the contrast loss, the feature distance of the positive sample pair can be reduced, while the feature distance of the negative sample pair can be increased.

[0104] In the embodiment of the present invention, the contrastive learning model significantly improves the model's ability to distinguish lithological features by optimizing global-local features, so that the model can more accurately distinguish lithological categories with similar feature distributions, because by maximizing the similarity of similar lithological samples and minimizing the similarity of different lithological samples, the model learns a more robust feature representation. This not only enhances the model's ability to aggregate similar lithological features, but also enhances its ability to distinguish heterogeneous lithological features, thereby achieving higher classification accuracy and better generalization performance in lithological scene classification tasks.

[0105] Optionally, the optimizing the global-local features according to the homogeneous features and the heterogeneous features to obtain the optimized final features includes:

[0106] Aggregating the similar features to obtain aggregated features;

[0107] Performing dispersion processing on the heterogeneous features to obtain dispersed features;

[0108] The optimized final feature is obtained according to the aggregation feature and the dispersion feature.

[0109] Specifically, the process of optimizing global-local features involves aggregation and dispersion processing in the contrastive learning model. In this process, the same type of features are first aggregated, which means that the features with the same lithology label are regarded as positive samples, and the similarity between these features is enhanced through the contrastive learning mechanism, so that the features of the same lithology are more closely clustered together in the feature space. Next, the heterogeneous features are dispersed, that is, the features with different lithology labels are regarded as negative samples, and the similarity between these features is reduced through contrastive learning, so that the features of different lithologies are more dispersed in the feature space. In a preferred embodiment of the present invention, the optimization is performed by maximizing the similarity of positive sample features and minimizing the similarity of negative sample features through the contrast loss function, thereby promoting the aggregation of similar lithology features while ensuring the dispersion of different lithology features. Finally, the optimized final features are obtained based on the aggregated features and the dispersed features.

[0110] In the embodiment of the present invention, the ability to distinguish lithological features is significantly improved by contrasting the aggregation and dispersion processing of the learning model. The model can more accurately distinguish lithological categories with similar feature distributions, and the model learns more robust feature representations by enhancing the similarity of similar lithological samples and reducing the similarity of different lithological samples. This not only enhances the aggregation ability of similar lithological features, but also enhances the ability to distinguish heterogeneous lithological features, thereby achieving higher classification accuracy and better generalization performance in lithological scene classification tasks.

[0111] Optionally, the multiple remote sensing images include ZY1 images, Landsat images and GF5 images;

[0112] The performing feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area includes:

[0113] By channel stitching, the final features of the Landsat image and the GF5 image are fused to obtain stitching features;

[0114] Determine the classification probability of each lithology category according to the splicing features and the final features of the ZY1 image through soft classification;

[0115] Determining weight values ​​of the splicing features and the final features of the ZY1 image according to the classification probability;

[0116] The classification probability of the spliced ​​feature and the classification probability of the final feature of the ZY1 image are weighted and fused according to the weight value, so as to obtain the lithology classification label of each pixel in the multi-remote sensing image, and the lithology classification label is used as the lithology scene classification result

[0117] Specifically, the process of feature fusion and classification processing first involves fusing the final features of Landsat images and GF5 images through channel stitching to form a stitching feature. The complementarity of the two images in spectral information is utilized, and the spectral advantages of remote sensing images are integrated through channel stitching operations to enhance the complementarity of multi-source data. Then, the classification probability of each lithology category is determined by soft classification, combining the stitching features and the final features of the ZY1 image. Among them, the lithology soft classification label refers to the probability distribution of each category obtained by model prediction in the lithology classification task, rather than a definite and unique category label. It reflects the possibility (probability value) that each pixel or sample belongs to each lithology category, rather than a single lithology category directly given in the hard classification. In this process, the high-resolution features of the ZY1 image provide precise spatial details, while the fusion features of the Landsat and GF5 images provide rich spectral information. Subsequently, the weight values ​​of the stitching features and the ZY1 image are determined according to the classification probability, and this weight value reflects the contribution of different image features in the classification. Finally, the classification probability of the stitching features and the classification probability of the ZY1 image features are weighted and fused according to the weight value to obtain the lithology classification label of each pixel, balancing the feature contributions of different images so that the final result has both the accuracy of spatial details and the distinguishability of spectral features.

[0118] In the embodiment of the present invention, the data of multi-source remote sensing images are effectively integrated through feature fusion and classification processing, and the accuracy and reliability of lithology scene classification are improved. The feature fusion strategy makes full use of the complementary information of different images, such as the spatial details of ZY1 images and the spectral information of Landsat and GF5 images. By weighted fusion of the classification probabilities of different images, the model can balance the contribution of various features, thereby capturing subtle differences in the spectrum while maintaining high-resolution spatial information. Not only does it improve the accuracy of classification, but it also enhances the model's ability to recognize complex geological structures, making the lithology scene classification results more accurate and comprehensive.

[0119] Optionally, it also includes:

[0120] Establishing the contrastive learning model, which specifically includes:

[0121] Obtain lithology characteristic data sets;

[0122] constructing positive sample pairs according to the same lithology labels in the lithology feature data set, and constructing negative sample pairs according to different lithology labels in the lithology feature data set;

[0123] According to the positive sample pairs and the negative sample pairs, the initial learning model is trained to maximize the similarity of positive samples and to minimize the similarity of negative samples;

[0124] The trained initial learning model is used as the comparison learning model.

[0125] Specifically, the process of establishing a comparative learning model first involves obtaining a lithological feature data set, which contains lithological features extracted from multi-source remote sensing images. Then, based on these features and the lithological labels corresponding to the lithological features, positive sample pairs and negative sample pairs are constructed, where the positive sample pairs are composed of features with the same lithological labels, and the negative sample pairs are composed of features with different lithological labels. Subsequently, these sample pairs are used to train the initial learning model, and the feature representation ability of the model is optimized by a training strategy that maximizes the similarity of positive samples and minimizes the similarity of negative samples. Finally, the trained initial learning model is used as a comparative learning model for subsequent global-local feature optimization. In a preferred embodiment of the present invention, the comparative learning module promotes the aggregation of similar lithological features through a comparative loss function, while ensuring the dispersion of different lithological features, thereby improving the model's ability to distinguish.

[0126] In the embodiment of the present invention, the training and application of the contrastive learning model significantly improves the model's ability to distinguish and classify lithological features. This method enables the model to more accurately distinguish lithological categories with similar feature distributions, because by maximizing the similarity of similar lithological samples and minimizing the similarity of different lithological samples, the contrastive learning model learns a more robust feature representation.

[0127] Combination Figure 3 As shown, the present invention also provides a lithology scene classification system, comprising:

[0128] A feature extraction unit is used to extract features from multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features;

[0129] An adaptive learning unit, used for adaptively learning the spatial features and the spectral features to generate an adjacency matrix;

[0130] A graph convolution unit, used to update the remote sensing image features according to the information between the nodes in the adjacency matrix through a graph convolution network, so as to obtain the global-local features corresponding to the remote sensing image features;

[0131] A graph contrast learning unit, used to perform contrast learning on the global-local features through a contrast learning model to obtain optimized final features;

[0132] The fusion and classification unit is used to perform feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area.

[0133] The lithology scene classification system of the present invention captures the spatial and spectral features of each image by extracting features from multi-source remote sensing images. Then, an adjacency matrix is ​​generated using an adaptive learning method, and the extracted spatial features and spectral features are adaptively learned to generate an adjacency matrix. The adjacency matrix describes the relationship between features and is used in a graph convolutional network to simulate the connection strength between features, providing structured information for subsequent feature updates. Then, the remote sensing image features are updated according to the adjacency matrix through a graph convolutional network, and feature representations containing global and local information are extracted, so that the model can capture richer contextual information, thereby enhancing the ability to distinguish lithology categories. Further, these global-local features are optimized by contrastive learning models, the model's ability to distinguish similar lithology features is enhanced, and the classification accuracy of lithology categories close to feature distribution is improved. Finally, the optimized final features are subjected to feature fusion and classification processing to obtain the final lithology scene classification results. The present invention provides rich initial information for the model through feature extraction, and uses the adjacency matrix generated by adaptive learning to provide structured relationship information for the graph convolutional network. The graph convolutional network and contrastive learning model further enhance the feature differentiation capability. Finally, accurate lithology scene classification results are obtained through feature fusion and classification processing, which effectively classifies lithology categories with similar feature distribution in remote sensing images and improves the classification accuracy of lithology categories.

[0134] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned lithology scene classification method is implemented.

[0135] The lithology scene classification method, system and storage medium of the present invention capture the spatial and spectral features of each image by extracting features from multi-source remote sensing images. Then, an adjacency matrix is ​​generated using an adaptive learning method, and the extracted spatial features and spectral features are adaptively learned to generate an adjacency matrix. The adjacency matrix describes the relationship between features and is used in a graph convolutional network to simulate the connection strength between features, providing structured information for subsequent feature updates. Then, the remote sensing image features are updated according to the adjacency matrix through a graph convolutional network, and feature representations containing global and local information are extracted, so that the model can capture richer contextual information, thereby enhancing the ability to distinguish lithology categories. Further, these global-local features are optimized by contrasting learning models, enhancing the model's ability to distinguish similar lithology features, and improving the classification accuracy of lithology categories close to feature distribution. Finally, the optimized final features are subjected to feature fusion and classification processing to obtain the final lithology scene classification results. The present invention provides rich initial information for the model through feature extraction, and uses the adjacency matrix generated by adaptive learning to provide structured relationship information for the graph convolutional network. The graph convolutional network and contrastive learning model further enhance the feature differentiation capability. Finally, accurate lithology scene classification results are obtained through feature fusion and classification processing, which effectively classifies lithology categories with similar feature distribution in remote sensing images and improves the classification accuracy of lithology categories.

[0136] 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.

[0137] 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 scene classification method, characterized in that: include: Extracting features from multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features; Adaptively learning the spatial features and the spectral features to generate an adjacency matrix; By using a graph convolutional network, the remote sensing image features are updated according to information between nodes in the adjacency matrix to obtain global-local features corresponding to the remote sensing image features; By using a contrastive learning model, the global-local features are contrastively learned to obtain optimized final features; The final features are subjected to feature fusion and classification processing to obtain the lithology scene classification result of the detection area.

2. The lithology scene classification method according to claim 1, characterized in that: The feature extraction of multiple remote sensing images of the detection area is performed to obtain multiple remote sensing image features, including: Preprocessing the multiple remote sensing images to obtain multiple remote sensing images with a unified spatial coverage range; The multiple remote sensing images with unified spatial coverage are input into a convolutional neural network for feature extraction to obtain the spatial features and the spectral features of the multiple remote sensing images, and the spatial features and the spectral features are used as the remote sensing image features.

3. The lithology scene classification method according to claim 1, characterized in that: The adaptive learning of the spatial features and the spectral features to generate an adjacency matrix includes: Resizing the feature graph of the spatial feature and the feature graph of the spectral feature to obtain the feature graph of the spatial feature and the feature graph of the spectral feature that match the size of the initial matrix; Determining the nodes of the initial matrix according to the pixel positions in the feature graph of the spatial feature and the feature graph of the spectral feature, wherein each pixel position corresponds to one node; The adjacency matrix is ​​obtained by adaptively learning the feature graphs of the spatial features and the spectral features and updating the nodes of the initial matrix.

4. The lithology scene classification method according to claim 1, characterized in that: The remote sensing image features are updated according to the information between the nodes in the adjacency matrix through the graph convolution network to obtain the global-local features corresponding to the remote sensing image features, including: Inputting the adjacency matrix and the remote sensing image features into the graph convolutional network, and updating the remote sensing image features through the convolutional layer of the graph convolutional network; wherein, the output features of each of the convolutional layers are aggregated according to the information between the nodes in the adjacency matrix to obtain the local features of the remote sensing image features; Obtaining the global features of the remote sensing image features according to the output of the last convolutional layer of the graph convolutional network; By fusing the local features with the global features, global-local features corresponding to the remote sensing image features are obtained.

5. The lithology scene classification method according to claim 1, characterized in that: The contrast learning model is used to perform contrast learning on the global-local features to obtain the optimized final features, including: Inputting the global-local features into the contrastive learning model, and determining the similarity between each feature in the global-local features and other features; By comparing the loss function, all the features of the global-local features are divided according to the similarity, and the features are divided into similar features and heterogeneous features; The global-local features are optimized according to the homogeneous features and the heterogeneous features to obtain the optimized final features.

6. The lithology scene classification method according to claim 5, characterized in that: The step of optimizing the global-local features according to the homogeneous features and the heterogeneous features to obtain the optimized final features includes: Aggregating the similar features to obtain aggregated features; Performing dispersion processing on the heterogeneous features to obtain dispersed features; The optimized final feature is obtained according to the aggregation feature and the dispersion feature.

7. The lithology scene classification method according to claim 1, characterized in that: The multiple remote sensing images include ZY1 images, Landsat images and GF5 images; The performing feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area includes: By channel stitching, the final features of the Landsat image and the GF5 image are fused to obtain stitching features; Determine the classification probability of each lithology category according to the splicing features and the final features of the ZY1 image through soft classification; Determining weight values ​​of the splicing features and the final features of the ZY1 image according to the classification probability; The classification probability of the spliced ​​feature and the classification probability of the final feature of the ZY1 image are weightedly fused according to the weight value to obtain the lithology classification label of each pixel in the multi-remote sensing image, and the lithology classification label is used as the lithology scene classification result.

8. The lithology scene classification method according to claim 1, characterized in that: Also includes: Establishing the contrastive learning model, which specifically includes: Obtain lithology characteristic data sets; constructing positive sample pairs according to the same lithology labels in the lithology feature data set, and constructing negative sample pairs according to different lithology labels in the lithology feature data set; According to the positive sample pairs and the negative sample pairs, the initial learning model is trained to maximize the similarity of positive samples and to minimize the similarity of negative samples; The trained initial learning model is used as the comparison learning model.

9. A lithology scene classification system, characterized in that: include: A feature extraction unit is used to extract features from multiple remote sensing images of the detection area to obtain multiple remote sensing image features, wherein the remote sensing image features include spatial features and spectral features; An adaptive learning unit, used for adaptively learning the spatial features and the spectral features to generate an adjacency matrix; A graph convolution unit, used to update the remote sensing image features according to the information between the nodes in the adjacency matrix through a graph convolution network, so as to obtain the global-local features corresponding to the remote sensing image features; A graph contrast learning unit, used to perform contrast learning on the global-local features through a contrast learning model to obtain optimized final features; The fusion and classification unit is used to perform feature fusion and classification processing on the final features to obtain the lithology scene classification result of the detection area.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the lithology scene classification method according to any one of claims 1 to 8 is implemented.

Images