Super-resolution land coverage mapping method coupled with large language model knowledge base

Abstract

The invention relates to the technical field of earth space information, in particular to a super-resolution land coverage mapping method coupled with a large language model knowledge base. Comprising the following steps: acquiring a remote sensing image and geoscience text data of a target area and constructing a geoscience knowledge system; mining geoscience knowledge from the geoscience text data based on a large language model, constructing a structured geoscience knowledge base, and further generating a geoscience knowledge graph; obtaining a deep feature graph of the remote sensing image based on a convolutional neural network, and inputting a knowledge graph into a graph attention network to extract a regional geoscience semantic embedding vector; realizing geoscience semantic embedding vector and deep feature map fusion based on a global-local attention fusion mechanism; and performing up-sampling on the fused feature map to obtain a high-resolution depth feature map, and generating a super-resolution land coverage map of the remote sensing image by adopting a nonlinear activation function. According to the method, dynamic modeling and computable expression of regional global knowledge can be realized, and the problems of surface feature boundary fuzziness and category confusion are remarkably improved.

Description

A super-resolution land cover mapping method coupled with a large language model knowledge base Technical Field

[0001] This invention relates to the field of geospatial information technology, and in particular to a super-resolution land cover mapping method coupled with a large language model knowledge base. Background Technology

[0002] With the acceleration of urbanization and the increasing severity of ecological and environmental problems, accurate acquisition and dynamic monitoring of land cover have become crucial foundations for land resource management, environmental assessment, and urban and rural planning. While high-resolution remote sensing imagery can significantly improve the accuracy of land cover identification, it is limited by acquisition costs and coverage area; low-resolution imagery, though advantageous, has limited accuracy in mapping results. Super-resolution mapping techniques (sub-pixel unmixing) can utilize low-resolution imagery to obtain high-resolution land cover mapping results. Traditional sub-pixel resolution techniques (such as Atkinson, 1997) use pixel spatial correlation to infer mixed pixel components, suitable for homogeneous surfaces but prone to boundary blurring and category confusion in heterogeneous regions; statistical learning methods (such as random forests), while capable of fusing multi-source features to improve unmixing accuracy, heavily rely on high-quality training samples and lack adaptability across data sources.

[0003] In recent years, deep learning models (such as convolutional neural networks) have demonstrated significant advantages in super-resolution land cover mapping due to their powerful automatic feature extraction and complex pattern recognition capabilities. However, these models are inherently "data-driven," requiring a large amount of labeled data to ensure performance. In practical applications, their performance often degrades due to imbalanced or scarce samples. Furthermore, their black-box nature means the classification process lacks clear semantic interpretation, failing to provide traceable decision-making basis for actual land management or planning. Therefore, some studies have attempted to introduce domain knowledge to assist classification judgments, but these efforts suffer from high construction costs, poor universality, and difficulties in updating and maintenance, failing to establish a universally effective intelligent classification mechanism. Moreover, geoscientific knowledge is widely embedded in massive amounts of text data, and traditional knowledge bases and methods struggle to mine and integrate this scattered information, resulting in insufficient development and utilization of this geoscientific knowledge. The emergence of large language models offers a new solution for mining geoscientific knowledge and constructing databases. With their powerful natural language understanding and text generation capabilities, large language models can efficiently mine potential geoscientific knowledge from massive amounts of text, providing more comprehensive and accurate geoscientific knowledge support.

[0004] In summary, current super-resolution land cover mapping has the following limitations: 1. Mainstream deep learning algorithms still suffer from black-box characteristics, resulting in a lack of interpretability in the land cover distribution results and making it difficult to reveal the causal relationship between input features and output categories; 2. Existing algorithms couple geoscientific knowledge by fusing feature data reflecting geoscientific characteristics at the feature level and network level, lacking modeling of global geoscientific information within a region; 3. Although large language models have the potential for text parsing, a deep adaptation mechanism with geoscientific ontology has not yet been established, and their application in geoscientific knowledge mining is still in the exploratory stage. Summary of the Invention

[0005] This invention aims to at least improve one of the technical problems existing in the prior art. To this end, this invention proposes a super-resolution land cover mapping method coupled with a large language model knowledge base. Specifically, it includes proposing a large language model knowledge extraction and adaptation mechanism oriented towards geoscientific scenarios, which can automatically extract geoscientific element attributes, feature descriptions, and environmental conditions from unstructured documents and other texts, thereby effectively compensating for the semantic deficiencies of relying solely on single image inputs; by constructing a regional multi-scale geoscientific knowledge graph, dynamic modeling and computable expression of regional multi-dimensional geoscientific knowledge are realized; this invention also proposes a global-local attention fusion module, which can adaptively fuse local image features extracted by convolutional neural networks with regional overall geoscientific semantics aggregated by graph attention networks, enabling the model to make judgments by comprehensively considering global semantics and local features, thereby improving the problems of blurred land cover boundaries and category confusion.

[0006] The technical solution of this invention is as follows: A super-resolution land cover mapping method coupled with a large language model knowledge base, comprising: S1: acquiring remote sensing images and geoscientific text data of a target area, and constructing a multi-level geoscientific knowledge system according to the mapping theme, wherein the geoscientific knowledge system includes corresponding geoscientific element indicators; S2: extracting geoscientific knowledge from the geoscientific text data based on the constructed multi-level geoscientific knowledge system using large language model technology, constructing a geoscientific knowledge base according to the organizational framework of the multi-level geoscientific knowledge system, and further acquiring multi-category geoscientific element indicators, thereby generating a regional multi-scale geoscientific knowledge map; S3: extracting remote sensing data based on a convolutional neural network. The deep feature map of the image is generated, and the regional multi-scale geoscience knowledge graph is input into the graph attention network to generate regional geoscience semantic embedding vectors; S4: The position and channel weights of the regional geoscience semantic embedding vectors are calculated based on the global-local attention mechanism, and the regional geoscience semantic embedding vectors are weighted and modulated using the weights, and fused with the deep feature map of the remote sensing image to form a fused abstract feature map; S5: The fused abstract feature map is upsampled to the target spatial resolution, and a high-resolution depth feature map is obtained based on depth convolution operation, and a nonlinear activation function is used to generate sub-pixel classification results to generate a super-resolution land cover map of the remote sensing image.

[0007] In one possible technical solution, S2 further includes: S21: segmenting the geoscientific text data to be extracted, and preserving paragraph boundaries and context markers during segmentation; S22: constructing corresponding prompt word templates for different types of geoscientific knowledge based on the constructed geoscientific knowledge system; S23: integrating the large language model interface and prompt word template chain based on the LangChain framework to achieve unified invocation and organization of the extraction process; S24: during execution, using the key parameters in the geoscientific text data to be extracted and the prompt word templates as input. The results are passed to the large language model, and multiple rounds of calls are executed. The extraction results from the multiple rounds of calls are compared and merged for consistency, and the extraction results are evaluated and scored to select the best extraction result; S25: The extraction results returned by the large language model are indexed according to element category, indicator type, spatial location and time information to obtain a structured geoscience knowledge base; S26: Based on the constructed structured geoscience knowledge base, multi-scale and multi-category geoscience element indicators are obtained, and regional multi-scale geoscience knowledge graphs are constructed according to administrative divisions and spatial relationships to realize the expression of multi-dimensional geoscience knowledge in multi-scale regions.

[0008] In one possible technical solution, S22 further includes: S221, prompt word requirements analysis and thought chain construction: based on the constructed geoscientific knowledge system and corresponding geoscientific element indicators, the overall extraction task is decomposed into several sub-tasks, forming a chain-like execution sequence, and generating a description of the logical steps required for the prompt words, specifically expressed as: In the formula, To extract the complete task set, For chained execution of subtasks; S222, construct a unified prompt word template: clarify the model's extraction target and output format, including the fixed fields and format of the model output, and embed geoscientific feature classifications in the prompt words. The output can be represented as a set of fields: In the formula To extract the complete set of fields, S223: The logical steps described in the thought chain constructed in S221 are embedded into the prompt words in the form of phased instructions; including identifying the spatiotemporal information involved in the text, the category of geoscientific elements, the matching geoscientific indicators, numerical and unit extraction, and structured output; S224: A self-checking and scoring process is added to the prompt words, where the self-checking includes: whether there are any missing elements, missing units, or indicators that do not match the elements, and the extraction results are scored according to the check results, and the position of the corresponding input text is marked, where the extraction results are introduced into a scoring function. : In the formula, Indicates integrity check, Indicates consistency. Indicating accuracy, , , S225: Introduce background information and typical extraction examples related to the study area into the prompt words, thereby improving the model's adaptability to specific regional features and specific extraction tasks.

[0009] In one possible technical solution, further, in S3, the deep feature map of the remote sensing image is obtained based on the convolutional neural network, specifically including: S311, shallow feature extraction of the image: the shallow features of the remote sensing image are extracted through the convolutional layer, specifically represented as: In the formula, This represents the convolution operation. For remote sensing images, S312, Deep Feature Extraction of Image: A convolutional layer is used to further extract deep features from the remote sensing image. The shallow features are standardized using batch normalization. Activation functions implement feature mapping Then, for feature mapping Perform the same convolution and normalization operations, introducing a residual structure in this operation, and finally obtain the image features. : In the formula, This represents the convolution operation. This indicates Batch Normalization. S313, Activation function layer; Deep feature downsampling: for image features Perform max pooling and record the pooling result and the max pooling position index: In the formula, This indicates a max pooling operation. This represents the position index corresponding to max pooling. This represents the deep feature map after the first downsampling; S314, multi-level feature extraction and downsampling: repeating the processes of S312 and S313 four times to form a multi-level convolutional pooling structure; in the... In this iteration, the input is the deep feature map output from the previous pooling layer. Deep features are obtained through convolution, standardization, activation, and downsampling: In the formula, Indicates the first The position index corresponding to the second max pooling. Indicates the first Deep feature map after subsampling.

[0010] In one possible technical solution, further, in S3, feature modeling and information propagation are performed on the regional multi-scale geoscience knowledge graph based on graph attention networks to extract regional geoscience semantic embedding vectors. Specifically, this includes: S321, constructing a node feature matrix: based on the administrative division nodes in the regional multi-scale geoscience knowledge graph, a node feature matrix is ​​constructed. The node feature matrix can be represented as: Where o is the number of nodes and d is the feature dimension.

[0011] S322, Construction and Weighting of Multi-Relation Adjacency Matrix: A weighted adjacency matrix is ​​constructed based on different relation types in the regional multi-scale geoscientific knowledge graph. The edge relation weights are determined by spatial adjacency and hierarchical subordination. Adjacency relations are denoted as... Subordination is denoted as For stable training calculations, the self-loop is denoted as... ,in, It is an adjacency matrix. For the identity matrix; S323, calculate the multi-relation multi-head attention layer: assuming the network has a total of Layer, number Layer usage One attention head; the output dimension of this layer is set to... any node Its neighboring nodes The attention coefficient is calculated as follows: in, express Layer The linear transformation matrix of each attention head. No. Layer Attention vectors of attention heads, for Layer node characteristics, For activation function, Indicates the first The attention weights are calculated for any two connected nodes in the layer; these attention weights are then normalized using Softmax to obtain the final attention weights. The current node's features are updated as follows: in, For nodes Neighboring nodes that have an adjacency or dependency relationship The set that is formed This represents the normalized attention weights. express Layer nodes of The node vectors obtained from multiple attention heads are concatenated or averaged to obtain the node representation matrix for that layer. S324, Inter-layer propagation and update: After each layer of propagation, a new node representation matrix is ​​obtained. When the graph attention network propagates to the first... After layers, the final node embedding matrix is ​​obtained: in, For the number of nodes, Indicates the first The feature dimension of the layer.

[0012] In one possible technical solution, further, in S4, the deep features of remote sensing images and the regional geoscientific semantic embedding vectors are fused based on a global-local attention mechanism. Specifically, this includes: fusing the regional geoscientific semantic embedding vectors... Deep feature maps of remote sensing images By location index Mapping function Generate global geoscientific semantic vectors for the corresponding locations in the image. The specific process is as follows: Based on global-local attention fusion mechanism Attention weights for image features are adaptively generated from geoscientific semantic embedding vectors. The global-local attention fusion mechanism employs a parallel spatial channel attention mechanism to adaptively adjust global semantic features; for the generated global geoscientific semantic vector... Apply two consecutive 1×1 convolutions Then apply the Sigmoid activation function. To generate an independent spatial selection mask. The specific process is as follows: Simultaneously, global average pooling is applied. To compress all spatial information in each channel; through a fully connected layer A nonlinear transformation is applied to this global feature to obtain a more compact global feature. Design a learnable vector Multiply with global features, then use The function generates independent channel selection masks. The process can be represented as follows: Channel selection mask With the corresponding spatial selection mask The final weight matrix is ​​obtained by matrix multiplication. This matrix integrates importance information from both spatial and channel dimensions, and the process can be represented as follows: Will pass Normalization yields attention weights Regional knowledge embedding vector With attention weight Multiply the features and add them to the deep feature map of the same type to obtain the fused feature map. The specific process is as follows: Features fused from fourth downsampled image features and geoscientific semantic embedding vectors Furthermore, the Atrous spatial pyramid pooling operation is employed to enhance the model's ability to extract features from small ground features. The specific process is as follows: In the formula, This indicates the image pyramid pooling operation. This indicates the expansion coefficient used in Atrous spatial pyramid pooling. This is the output of Atrous spatial pyramid pooling.

[0013] In one possible technical solution, further, in S5, the fused abstract feature map is upsampled to the target spatial resolution, and a nonlinear activation function is used to generate sub-pixel classification results. Specifically, this includes: merging the feature map obtained in step S4... Upsampling is performed to obtain feature maps with higher spatial resolution at a specific scale, followed by further convolutional and pooling operations, ultimately... The activation function is used to obtain the land cover classification results for each sub-pixel. The specific steps are as follows: in, For upsampling operation, This represents the upsampling amplification factor. This represents the convolution operation. For the activation function layer, This indicates the execution of multiple identical module steps, and is related to downsampling. Maintain consistency. Feature map representing the target scale. For activation function, This represents the classification result for each sub-pixel of the feature map.

[0014] The super-resolution land cover mapping method based on the coupled large language model knowledge base of the present invention has the following advantages compared with the prior art: (1) The present invention breaks through the limitations of single modal input and shallow knowledge utilization in the existing methods, and proposes a multimodal information fusion mechanism driven by a large language model knowledge base. By mining and organizing geoscience knowledge in multi-source texts through a large language model and embedding it into a deep learning network, the joint modeling of global geoscience semantics and local image features is realized, thereby effectively improving the accuracy of land use classification and enhancing the interpretability and cross-domain adaptability of the results.

[0015] (2) In view of the problem that the geoscience knowledge in the massive text is complex and traditional natural language processing methods are difficult to mine and integrate, this invention proposes a knowledge extraction and adaptation mechanism for a large language model oriented to geoscience scenarios, which integrates multi-source geoscience texts, remote sensing interpretation rules and geospatial association knowledge, and provides rich geoscience prior knowledge for remote sensing intelligent land use classification.

[0016] (3) In view of the current problem that a single image as model input lacks the ability to model the overall geoscientific semantics of a region, this invention proposes a global-local attention fusion mechanism. This module can adaptively guide the global geoscientific semantics aggregated by the graph attention network and the cross-scale attention interaction of the local image features extracted by the CNN, so that the model can make a judgment by combining global semantics and local features, thereby significantly improving the classification accuracy and mapping effect.

[0017] (4) The method of the present invention can obtain high-resolution and high-precision land cover products. The framework is compatible with multiple types of geoscience data in super-resolution land cover mapping and has good scalability. It provides a new path for the innovation of remote sensing information extraction technology and helps to promote the application of remote sensing in resource management and ecological assessment.

[0018] (5) This invention can promote the standardized construction of geoscience knowledge bases and open up a new path for the vertical application of large language models in remote sensing mapping and earth system science, thereby promoting the level of intelligence and automation in earth system science research.

[0019] A super-resolution land cover mapping system coupled with a large language model knowledge base includes: a multi-source data acquisition module for acquiring remote sensing images and geoscientific text data of a target area, constructing a geoscientific knowledge system based on the mapping theme, wherein the structure of the geoscientific knowledge system includes corresponding geoscientific element indicators; a geoscientific knowledge mining module for extracting geoscientific knowledge from the geoscientific text data based on the constructed knowledge system and large language model technology, constructing a geoscientific knowledge base according to the organizational framework of the multi-level geoscientific knowledge system, further acquiring multi-category geoscientific element indicators, thereby generating a multi-scale geoscientific knowledge map of the region; and a modal feature extraction module for extracting depth features from remote sensing images based on convolutional neural networks. The system generates a regional geoscientific semantic embedding vector by inputting a multi-scale regional geoscientific knowledge graph into a graph attention network. A modal feature fusion module calculates the position and channel weights of the regional geoscientific semantic embedding vector based on a global-local attention mechanism, and uses these weights to perform weighted modulation on the vector. This is then fused with the deep feature map of the remote sensing image to form a fused abstract feature map. A super-resolution feature classification module upsamples the fused abstract feature map to the target spatial resolution, obtains a high-resolution depth feature map based on depth convolution operations, and uses a non-linear activation function to generate sub-pixel classification results, thus generating a super-resolution land cover map of the remote sensing image.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 is a flowchart of a super-resolution land cover mapping method coupled with a large language model knowledge base in an embodiment of the present invention; Figure 2 is a flowchart of the construction of a large language model knowledge base for geoscientific applications in an embodiment of the present invention; Figure 3 is a flowchart of the prompt word engineering template technology for geoscientific applications in an embodiment of the present invention; Figure 4 is a structural diagram of the modal feature extraction, modal feature fusion module and the super-resolution feature classification module in an embodiment of the present invention; Figure 5 is a structural diagram of the global-local attention fusion mechanism in the modal feature fusion module in an embodiment of the present invention; Figure 6 is a flowchart of super-resolution land cover mapping coupled with a large language model geoscientific knowledge base in an embodiment of the present invention; Figure 7 is a structural block diagram of a super-resolution land cover mapping method coupled with a large language model knowledge base in an embodiment of the present invention.

[0023] Figure labels: Multi-source data acquisition module 100, geoscience knowledge mining module 200, modal feature extraction module 300, modal feature fusion module 400, super-resolution feature classification module 500. Detailed Implementation

[0024] The embodiments of the present invention are described in detail below. The embodiments described with reference to the accompanying drawings are exemplary. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0025] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0027] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects and not to describe a particular order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, it may include a series of steps or units, or optionally, steps or units not listed, or other steps or units inherent to these processes, methods, products, or devices.

[0028] The accompanying drawings show only the portions relevant to this application, not all of them. Before discussing exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe operations (or steps) as sequential processes, many of these operations may be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations may be rearranged. The process may be terminated when its operation is completed, but may also have additional steps not included in the drawings. The process may correspond to a method, function, procedure, subroutine, subprogram, etc.

[0029] The terms “component,” “module,” “system,” “unit,” etc., used in this specification are used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution. For example, a unit can be, but is not limited to, a process running on a processor, a processor, an object, an executable file, a thread of execution, a program, and / or distributed between two or more computers. Furthermore, these units can be executed from various computer-readable media on which various data structures are stored. Units can communicate, for example, via local and / or remote processes based on signals having one or more data packets (e.g., data from a second unit interacting with another unit between a local system, a distributed system, and / or a network; for example, the Internet interacting with other systems via signals).

[0030] Example 1, as shown in Figure 1, provides a super-resolution land cover mapping method coupled with a large language model knowledge base. The method includes: S1: acquiring remote sensing images and massive amounts of geoscientific text data of the target area to construct a multi-level, multi-category, and multi-theme geoscientific knowledge system. This geoscientific knowledge system includes corresponding geoscientific element indicators. Specifically, S1 includes: S11: acquiring geoscientific text data based on multi-source data platforms (such as Wikipedia, statistical yearbooks, open access platforms, etc.) according to the application theme and scenario area, and performing preprocessing such as filtering on the geoscientific text data; acquiring remote sensing images of the corresponding target area and performing preprocessing on the remote sensing image data; S12: constructing a multi-level geoscientific knowledge system based on the geoscientific text data. This geoscientific knowledge system mainly includes five categories: topography, soil, climate, hydrology, and biology. Each category includes three levels of subcategories, specifically corresponding geoscientific element indicators, such as climate-climate elements-temperature-average temperature.

[0031] S2: Geological knowledge is extracted from geoscientific text data using a large language model prompt word engineering approach. A structured geoscientific knowledge base is constructed according to a multi-level geoscientific knowledge system framework, thereby generating a regional multi-scale geoscientific knowledge graph. Specifically, this includes: S21: Segmenting the extracted geoscientific text data to meet the input token constraints of the large language model, while preserving paragraph boundaries and context markers during segmentation to ensure semantic integrity; S22: Based on the constructed multi-level geoscientific knowledge system, constructing corresponding prompt word templates for different types of geoscientific knowledge; S23: Integrating the large language model interface and prompt word template chain based on the LangChain framework for unified invocation and organization of the extraction process; S24: During execution, the data to be extracted... Key parameters from geoscientific text data and prompt word templates are taken as input and passed to a large language model. Multiple rounds of calls are executed, and the extraction results from these rounds are compared and merged for consistency. The extraction results are then evaluated and scored, and the best extraction result is selected. S25: The extraction results returned by the large language model are written into a database table, and indexes are built according to element category, indicator type, spatial location, and temporal information to obtain a structured geoscientific knowledge base, which supports subsequent multi-dimensional retrieval and knowledge graph construction. S26: For the constructed structured geoscientific knowledge base, a regional multi-scale geoscientific knowledge graph with spatial hierarchical relationships is generated based on administrative divisions, including multi-level administrative divisions such as "province-city-district-township," while maintaining the spatial relationships between elements at different levels. Finally, the regional multi-scale geoscientific knowledge graph is stored and visualized using Neo4j software.

[0032] S3: Deep feature maps of remote sensing images are extracted based on convolutional neural networks, and regional multi-scale geoscientific knowledge graphs are input into a graph attention network to generate regional geoscientific semantic embedding vectors, achieving dual-modal feature representation of knowledge and images; among which, remote sensing images... Inputting a convolutional neural network to obtain deep feature maps of an image The main implementation uses multiple convolutional layers, activation functions, and pooling layers. The specific process is as follows: S311, shallow feature extraction of the image: The shallow features of the remote sensing image are extracted through the convolutional layer, specifically represented as follows: In the formula, This represents the convolution operation. For remote sensing images, S312, Deep Feature Extraction of Image: A convolutional layer is used to further extract deep features from the remote sensing image. The shallow features are standardized using batch normalization. Activation functions implement feature mapping Then, for feature mapping Perform the same convolution and normalization operations, introducing a residual structure in this operation, and finally obtain the image features. : In the formula, This represents the convolution operation. This indicates Batch Normalization. S313, Activation function layer; Deep feature downsampling: for image features Perform max pooling and record the pooling result and the max pooling position index: In the formula, This indicates a max pooling operation. This represents the position index corresponding to max pooling. This represents the deep feature map after the first downsampling; S314, multi-level feature extraction and downsampling: repeating the processes of S312 and S313 four times to form a multi-level convolutional pooling structure; in the... In this iteration, the input is the deep feature map output from the previous pooling layer. Deep features are obtained through convolution, standardization, activation, and downsampling: In the formula, Indicates the first The position index corresponding to the second max pooling. Indicates the first Deep feature map after subsampling.

[0033] Regional geoscientific semantic embedding vectors are extracted based on graph attention networks and regional multi-scale geoscientific knowledge graphs. The input to the graph attention network includes the construction of the node feature matrix. and adjacency matrix Node feature matrix The adjacency matrix represents the attribute information of each administrative division node. This represents the adjacency relationships between nodes. The model output is an embedding vector of regional geoscientific knowledge. The specific process includes: S321, constructing the node feature matrix: based on the administrative division nodes in the regional multi-scale geoscience knowledge graph, constructing the node feature matrix. The node feature matrix can be represented as: Where o is the number of nodes and d is the feature dimension.

[0034] S322, Construction and Weighting of Multi-Relation Adjacency Matrix: A weighted adjacency matrix set is constructed based on different relation types in the regional multi-scale geoscientific knowledge graph. The edge relation weights are determined by spatial adjacency and hierarchical subordination. Adjacency relations are denoted as... Subordination is denoted as For stable training calculations, the self-loop is denoted as... ,in, It is an adjacency matrix. For the identity matrix; S323, calculate the multi-relation multi-head attention layer: assuming the network has a total of Layer, number Layer usage One attention head; the output dimension of this layer is set to... any node Its neighboring nodes The attention coefficient is calculated as follows: in, express Layer The linear transformation matrix of each attention head. No. Layer Attention vectors of attention heads, for Layer node characteristics, For activation function, Indicates the first The attention weights are calculated for any two connected nodes in the layer; these attention weights are then normalized using Softmax to obtain the final attention weights. The current node's features are updated as follows: in, For nodes Neighboring nodes that have an adjacency or dependency relationship The set that is formed This represents the normalized attention weights. express Layer nodes of The node vectors obtained from multiple attention heads are concatenated or averaged to obtain the node representation matrix for that layer. S324, Inter-layer propagation and update: After each layer of propagation, a new node representation matrix is ​​obtained. When the graph attention network propagates to the first... After layers, the final node embedding matrix is ​​obtained: in, For the number of nodes, Indicates the first The feature dimension of the layer.

[0035] S4: Calculate the channel and location weights of the regional geoscientific semantic embedding vector based on a global-local attention fusion mechanism, and then use these weights to perform weighted modulation on the regional geoscientific semantic embedding vector. This weighted vector is then fused with the deep feature map of the remote sensing image to obtain the fused abstract feature map. Specifically, this involves: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] Deep feature maps of remote sensing images By location index Mapping function Generate global geoscientific semantic vectors for the corresponding locations in the image. The specific process is as follows: Based on global-local attention fusion mechanism Attention weights for image features are adaptively generated from geoscientific semantic embedding vectors. The global-local attention fusion mechanism employs a parallel spatial channel attention mechanism to adaptively adjust global semantic features; for the generated global geoscientific semantic vector... Apply two consecutive 1×1 convolutions Then apply the Sigmoid activation function. To generate an independent spatial selection mask. The specific process is as follows: Simultaneously, global average pooling is applied. To compress all spatial information in each channel; through a fully connected layer A nonlinear transformation is applied to this global feature to obtain a more compact global feature. Design a learnable vector Multiply with global features, then use The function generates independent channel selection masks. The process can be represented as follows: Channel selection mask With the corresponding spatial selection mask The final weight matrix is ​​obtained by matrix multiplication. This matrix integrates importance information from both spatial and channel dimensions, and the process can be represented as follows: Will pass Normalization yields attention weights Regional knowledge embedding vector With attention weight Multiply the features and add them to the deep feature map of the same type to obtain the fused feature map. The specific process is as follows: Features fused from fourth downsampled image features and geoscientific semantic embedding vectors Furthermore, the Atrous spatial pyramid pooling operation is employed to enhance the model's ability to extract features from small ground features. The specific process is as follows: In the formula, This indicates the image pyramid pooling operation. This indicates the expansion coefficient used in Atrous spatial pyramid pooling. This is the output of Atrous spatial pyramid pooling.

[0036] S5: Upsample the fused abstract feature map to the target spatial resolution, perform depthwise convolution to obtain a high-resolution depth feature map, and use a non-linear activation function to generate sub-pixel classification results, thus generating a super-resolution land cover map of the remote sensing image. The process is as follows: The fused feature map obtained in S4... Upsampling is performed to obtain feature maps with higher spatial resolution at a specific scale, followed by further convolutional and pooling operations, ultimately... The activation function is used to obtain the land cover classification results for each sub-pixel. The specific steps are as follows: in, For upsampling operation, This represents the upsampling amplification factor. This represents the convolution operation. For the activation function layer, This indicates the execution of multiple identical module steps, and is related to downsampling. Maintain consistency. Feature map representing the target scale. For activation function, This represents the classification result for each sub-pixel of the feature map.

[0037] In one embodiment, the regional multi-scale geoscience knowledge graph is modeled using only a single-scale geoscience knowledge graph. This implementation is a special case of the application of regional multi-scale geoscience knowledge graphs.

[0038] It should be noted that, in this embodiment, to demonstrate the construction of corresponding prompt word templates for different types of geoscientific knowledge, Figure 3 shows the technical flowchart of the prompt word engineering template for geoscientific scenarios. First, based on the geoscientific knowledge system, the extraction task is broken down into sub-tasks such as element identification, indicator matching, numerical / unit extraction, and spatiotemporal positioning, and a thought chain logic is constructed. Then, prompt word templates are designed, output fields are defined, and knowledge classification and indicator systems are embedded. In the execution phase, the model is gradually guided to complete element extraction and structured output through phased instructions. Afterward, a self-checking and scoring mechanism is added to verify the completeness and consistency of the extraction results and mark the original text position. Finally, combined with the research area and specific task, contextual information is injected into the prompt words to achieve the adaptation and optimization of the results to different regional characteristics.

[0039] In one embodiment, step S22: the technical process of the prompt word engineering template for geoscientific scenarios mainly includes the following steps: S221, prompt word requirements analysis and thought chain construction: based on the constructed geoscientific knowledge system and corresponding geoscientific element indicators, the overall extraction task is decomposed into several sub-tasks, forming a chain-like execution sequence, and generating the logical step description required for the prompt words, specifically expressed as: In the formula, To extract the complete task set, For chained execution of subtasks; S222, construct a unified prompt word template: clarify the model's extraction target and output format, including the fixed fields and format of the model output, and embed geoscientific feature classifications in the prompt words. The output can be represented as a set of fields: In the formula To extract the complete set of fields, For each extracted element content; S223, embed the thought chain steps constructed in S221 into the prompt words in the form of phased instructions: including identifying the spatiotemporal information involved in the text, geoscientific element categories, matching geoscientific indicators, numerical and unit extraction, and structured output; S224: add a self-checking and scoring step to the prompt words, where the self-checking includes: whether there are any missing elements, missing units, or indicators that do not match the elements, and score the extraction results based on the check results, and mark the position of the corresponding input text, where the extraction results are introduced into a scoring function. : In the formula, Indicates integrity check, Indicates consistency. Indicating accuracy, , , S225: Introduce background information and typical extraction examples related to the study area into the prompt words, thereby improving the model's adaptability to specific regional features and specific extraction tasks.

[0040] Taking super-resolution land cover mapping by coupling a large language model geoscientific knowledge base with remote sensing imagery as an example, Figure 6 illustrates the overall process proposed in this invention. This process systematically integrates the dual advantages of knowledge-driven and data-driven approaches in its methodological design: on the one hand, relying on the semantic parsing and knowledge extraction capabilities of the large language model, it automatically mines implicit multi-type geoscientific element indicators from massive geoscientific texts and organizes them in the form of a knowledge graph, realizing the structured modeling of unstructured textual knowledge; on the other hand, it deeply integrates the geoscientific semantic features obtained from the large language model with the pixel features of remote sensing imagery, forming a joint representation of global geoscientific semantics and local image features. Specifically, this invention first collects and preprocesses multi-source remote sensing images and geoscientific text data to construct a multi-level, multi-topic knowledge system. Second, based on the large language model prompt word project, it extracts and organizes geoscientific knowledge at different scales according to the knowledge system, constructing a dynamically updatable knowledge graph. Then, it uses a convolutional neural network to extract local spectral and texture features from the remote sensing images, and combines this with a graph attention network to extract the global geoscientific semantics of the knowledge graph. Next, it uses a global-local attention fusion mechanism to fuse image features with global geoscientific semantics, enabling the model to comprehensively analyze global semantics and local features for discrimination. Finally, the fused features are upsampled to the target resolution by a super-resolution classification module to obtain high-precision, high-resolution land cover mapping results.

[0041] This invention discloses a super-resolution land cover mapping method coupled with a large language model knowledge base. This method overcomes the limitations of existing methods, which rely on single-modal input and shallow knowledge utilization, by proposing a multi-modal information fusion mechanism driven by a large language model knowledge base. By mining and organizing geoscientific knowledge from multi-source texts using a large language model and embedding it into a deep learning network, it achieves joint modeling of global geoscientific semantics and local image features, thereby effectively improving the accuracy of land use classification and enhancing the interpretability and cross-domain adaptability of the results. Addressing the problem that traditional natural language processing methods struggle to mine and integrate the complex geoscientific knowledge in massive amounts of text, this invention proposes a large language model knowledge extraction and adaptation mechanism oriented towards geoscientific scenarios. This mechanism integrates multi-source geoscientific texts, remote sensing interpretation rules, and geospatial correlation knowledge, providing rich prior geoscientific knowledge for remote sensing intelligent land use classification. To address the problem that current models using single images as input lack comprehensive regional geoscientific semantic modeling, this invention proposes a global-local attention fusion mechanism. This module adaptively guides the cross-scale attention interaction between global geoscientific semantics aggregated by a graph attention network and local image features extracted by a CNN, enabling the model to comprehensively consider both global semantics and local features for discrimination, thereby significantly improving classification accuracy and mapping quality. This method can acquire high-resolution, high-precision land cover products. The framework is compatible with multi-class and multimodal geoscientific data in super-resolution land cover mapping, possesses good scalability, provides a new path for innovation in remote sensing information extraction technology, and helps promote the application of remote sensing in resource management and ecological assessment.

[0042] Example 2 This example provides a super-resolution land cover mapping system coupled with a large language model knowledge base, as shown in Figure 7, for implementing the mapping method of the above example. It includes: a multi-source data acquisition module 100, used to acquire remote sensing images of the target area and massive geoscientific text data to construct a geoscientific knowledge system covering multiple levels, categories, and themes. The geoscientific knowledge system includes corresponding geoscientific element indicators. The multi-source data acquisition module 100 also includes operations such as preprocessing the acquired remote sensing images and geoscientific text data.

[0043] The Geoscience Knowledge Mining Module 200 is used to extract geoscience knowledge from geoscience text data based on the constructed geoscience knowledge system and the prompt word engineering of the large language model. It constructs a structured geoscience knowledge base according to the organizational framework of the geoscience knowledge system, and then generates a regional multi-scale geoscience knowledge graph. Figure 2 shows the flowchart of the construction of the large language model geoscience knowledge base, which mainly includes three parts: a text data processing module, a geoscience knowledge extraction module, and a geoscience knowledge database module. The text data processing module mainly segments the text data into blocks according to the geoscience knowledge system and application themes to meet the input of the large language model. The geoscience knowledge extraction module is responsible for extracting geoscience knowledge from the processed text data according to the knowledge system. In this example, firstly, the API interface of the large language model (such as the GPT series) is obtained; then, the extraction task and objective are determined through the geoscience knowledge system; prompt words and extraction templates are designed according to the extraction objectives; further, prompt word templates are constructed based on the prompt words and extraction templates; finally, a LangChain task chain is created based on LangChain combined with the large language model API and prompt word templates to complete the extraction task of regional geoscience text information. The geoscience knowledge database module organizes and standardizes extracted geoscience knowledge according to regional relationships using knowledge graph technology. In this example, geoscience knowledge is organized using administrative divisions and spatial relationships to construct a four-level regional multi-scale geoscience knowledge graph of province-city-district-township. Finally, Neo4j software is used to store and visualize the regional multi-scale geoscience knowledge graph.

[0044] The modal feature extraction module 300 is used to input the preprocessed remote sensing image into a convolutional neural network, obtain the deep feature map of the remote sensing image based on the convolutional neural network, and extract the regional geoscientific semantic embedding vector from the regional multi-scale geoscience knowledge graph based on the graph attention network, so as to realize the dual-modal feature representation of knowledge and image. The deep feature map includes spectral, spatial and texture features, etc.

[0045] The modal feature fusion module 400 is used to calculate the position and channel weights of regional geoscientific semantic embedding vectors based on the global-local attention fusion module. It then uses these weights to perform weighted modulation on the regional geoscientific semantic embedding vectors and fuses them across scales with the deep feature map of the remote sensing image to obtain an abstract feature map that integrates geoscientific knowledge and image features. Figure 5 shows the structure of the global-local attention fusion mechanism in the modal feature fusion module. First, the geoscientific semantic embedding vectors are mapped to global geoscientific semantic vectors corresponding to the image features using position indices. Then, attention weights are calculated from both the spatial and channel dimensions to characterize the differences in importance between different spatial positions and semantic channels. Finally, the spatial and channel weights are jointly modeled to form a comprehensive attention weight, and this weight is used to perform weighted modulation and fusion of the geoscientific semantic embeddings and remote sensing image features to obtain the fused abstract feature map.

[0046] The super-resolution feature classification module 500 is used to upsample the fused abstract feature map to the target spatial resolution, obtain a high-resolution depth feature map based on depth convolution operation, and generate sub-pixel classification results using a nonlinear activation function to generate a super-resolution land cover map of the remote sensing image.

[0047] Figure 4 shows a schematic diagram of the modal feature extraction, modal feature fusion, and super-resolution feature classification modules. This part is the main component of remote sensing intelligent land cover mapping coupled with a large language model and geoscientific knowledge base. The modal feature extraction module is used to extract deep features from remote sensing images and geoscientific knowledge. The modal feature fusion module calculates the relationship weights between image features and geoscientific knowledge embedding vectors through an attention mechanism to obtain a weighted knowledge embedding vector, which is finally superimposed with the deep feature map of the image and input into subsequent network layers. The super-resolution feature classification module is used to further convolution and pool the knowledge-enhanced feature map to obtain the feature-classified result, i.e., the land cover mapping result.

[0048] The super-resolution land cover mapping system coupled with a large language model knowledge base in this application embodiment can be a device, or a component, integrated circuit, or chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. For example, mobile electronic devices can be mobile phones, tablets, laptops, PDAs, ultra-mobile personal computers (UMPCs), netbooks, or personal digital assistants (PDAs), etc., while non-mobile electronic devices can be servers, network-attached storage (NAS), personal computers (PCs), etc. This application embodiment does not specifically limit the specific implementation.

[0049] The super-resolution land cover mapping system coupled with a large language model knowledge base in this application embodiment can be a device with an operating system. This operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit it.

[0050] The super-resolution land cover mapping system coupled with a large language model knowledge base provided in this application embodiment can realize the various processes of the super-resolution land cover mapping method coupled with a large language model knowledge base as shown in Figure 1. To avoid repetition, these processes will not be described again here.

[0051] The super-resolution land cover mapping method and system based on the coupled large language model knowledge base of the present invention has the following advantages compared with the prior art: (1) The present invention breaks through the limitations of single modal input and shallow knowledge utilization in the existing methods, and proposes a multimodal information fusion mechanism driven by a large language model knowledge base. By mining and organizing geoscience knowledge in multi-source texts through a large language model and embedding it into a deep learning network, the joint modeling of global geoscience semantics and local image features is realized, thereby effectively improving the accuracy of land use classification and enhancing the interpretability and cross-domain adaptability of the results.

[0052] (2) In view of the problem that the geoscience knowledge in the massive text is complex and traditional natural language processing methods are difficult to mine and integrate, this invention proposes a knowledge extraction and adaptation mechanism for a large language model oriented to geoscience scenarios, which integrates multi-source geoscience texts, remote sensing interpretation rules and geospatial association knowledge, and provides rich geoscience prior knowledge for remote sensing intelligent land use classification.

[0053] (3) In view of the current problem that single images as model input lack the ability to model the overall geoscientific semantics of the region, this invention proposes a global-local attention fusion mechanism. This module can adaptively guide the interaction between the global geoscientific semantics aggregated by the graph attention network and the local image features extracted by the CNN across scales, so that the model can make a judgment by combining global semantics and local features, thereby significantly improving the classification accuracy and mapping effect.

[0054] (4) The method of the present invention can obtain high-resolution and high-precision land cover products. The framework is compatible with multiple types of geoscience data in super-resolution land cover mapping and has good scalability. It provides a new path for the innovation of remote sensing information extraction technology and helps to promote the application of remote sensing in resource management and ecological assessment.

[0055] (5) This invention can promote the standardized construction of geoscience knowledge bases and also opens up a new path for the vertical application of large language models in remote sensing mapping and Earth system science, thereby promoting the intelligentization and automation of Earth system science research. Optionally, this application embodiment also provides an electronic device, including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the various processes of the above-described embodiment of a super-resolution land cover mapping method coupled with a large language model knowledge base, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0056] This application also provides a readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described embodiment of a super-resolution land cover mapping method coupled with a large language model knowledge base, and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0057] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0058] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention.

[0059] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0060] Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The reference to "embodiment" herein means that a specific feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily indicate the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0061] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A super-resolution land cover mapping method coupled with a large language model knowledge base, characterized in that, include: S1: Acquire remote sensing images and geoscientific text data of the target area, and construct a multi-level geoscientific knowledge system based on the mapping theme. The geoscientific knowledge system includes corresponding geoscientific element indicators. S2: Based on the constructed multi-level geoscientific knowledge system, geoscientific knowledge is extracted from geoscientific text data using large language model technology. A geoscientific knowledge base is constructed according to the organizational framework of the multi-level geoscientific knowledge system, and multi-category geoscientific element indicators are further obtained to generate a regional multi-scale geoscientific knowledge graph. S3: Deep feature maps of remote sensing images are extracted based on convolutional neural networks, and the regional multi-scale geoscientific knowledge graph is input into a graph attention network to generate regional geoscientific semantic embedding vectors. S4: The position and channel weights of the regional geoscientific semantic embedding vectors are calculated based on a global-local attention mechanism, and the weights are used to perform weighted modulation on the regional geoscientific semantic embedding vectors. The vectors are then fused with the deep feature maps of remote sensing images to form a fused abstract feature map. S5: Upsample the fused abstract feature map to the target spatial resolution, obtain a high-resolution depth feature map based on depth convolution operation, and generate sub-pixel classification results using a nonlinear activation function to generate a super-resolution land cover map of the remote sensing image.

2. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, S2 Includes: S21: Segment the extracted geoscientific text data into blocks, and preserve paragraph boundaries and context markers during segmentation; S22: Based on the constructed geoscience knowledge system, construct corresponding prompt word templates for different types of geoscience knowledge; S23: Based on the LangChain framework, integrate the large language model interface and prompt word template chain to realize unified calling and organize the extraction process; S24: During execution, the key parameters in the geoscientific text data to be extracted and the prompt word template are passed as input to the large language model. Multiple rounds of calls are performed, and the extraction results from the multiple rounds of calls are compared and merged for consistency. The extraction results are evaluated and scored, and the best extraction result is selected. S25: The extraction results returned by the large language model are indexed according to element category, indicator type, spatial location, and temporal information to obtain a structured geoscientific knowledge base. S26: Based on the constructed structured geoscientific knowledge base, multi-scale and multi-category geoscientific element indicators are obtained, and a regional multi-scale geoscientific knowledge graph is constructed according to the spatial relationship of administrative divisions to realize the expression of geoscientific knowledge in a multi-dimensional and multi-scale region.

3. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, S22 includes: S221, Prompt Requirement Analysis and Mind Chain Construction: Based on the constructed geoscientific knowledge system and corresponding geoscientific element indicators, the overall extraction task is decomposed into several sub-tasks, forming a chain-like execution sequence, and generating a description of the logical steps required for the prompts, specifically expressed as: In the formula, To extract the complete task set, For chained execution of subtasks; S222, construct a unified prompt word template: clarify the model's extraction target and output format, including the fixed fields and format of the model output, and embed geoscientific feature classifications in the prompt words. The output can be represented as a set of fields: In the formula To extract the complete set of fields, S223: The logical steps described in the thought chain constructed in S221 are embedded into the prompt words in the form of phased instructions; including identifying the spatiotemporal information involved in the text, the category of geoscientific elements, the matching geoscientific indicators, numerical and unit extraction, and structured output; S224: A self-checking and scoring process is added to the prompt words, where the self-checking includes: whether there are any missing elements, missing units, or indicators that do not match the elements, and the extraction results are scored according to the check results, and the position of the corresponding input text is marked, where the extraction results are introduced into a scoring function. ： In the formula, Indicates integrity check, Indicates consistency. Indicating accuracy, ， ， S225: Introduce background information and typical extraction examples related to the study area into the prompt words, thereby improving the model's adaptability to specific regional features and specific extraction tasks.

4. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, In S3, deep feature maps of remote sensing images are extracted based on convolutional neural networks, specifically including: S311: Shallow feature extraction of images: shallow features of the remote sensing images are extracted through convolutional layers, specifically represented as: In the formula, This represents the convolution operation. For remote sensing images, S312: Deep Feature Extraction of Image: Deep features of the remote sensing image are further extracted using convolutional layers. Shallow features are standardized through batch normalization. Activation functions implement feature mapping Then, for feature mapping Perform the same convolution and normalization operations, introduce residual structures during the normalization operation, and finally obtain the image features. ： In the formula, This represents the convolution operation. This indicates Batch Normalization. S313: Activation function layer; S313: Deep feature downsampling: for image features Perform max pooling and record the pooling result and the max pooling position index: In the formula, This indicates a max pooling operation. This represents the position index corresponding to max pooling. This represents the deep feature map after the first downsampling; S314: Multi-level feature extraction and downsampling: Repeat steps S312 and S313 four times to form a multi-level convolutional pooling structure; in the... In this iteration, the input is the deep feature map output from the previous pooling layer. Deep features are obtained through convolution, standardization, activation, and downsampling: In the formula, Indicates the first The position index corresponding to the second max pooling. Indicates the first Deep feature map after subsampling.

5. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, In S3, feature modeling and information propagation are performed on the regional multi-scale geoscience knowledge graph based on graph attention networks to extract regional geoscience semantic embedding vectors. Specifically, this includes: S321, constructing the node feature matrix: based on the administrative division nodes in the regional multi-scale geoscience knowledge graph, constructing the node feature matrix. The node feature matrix can be represented as: Where o is the number of nodes and d is the feature dimension; S322, Construction and weighting of multi-relation adjacency matrix: A weighted adjacency matrix is ​​constructed based on different relation types in the regional multi-scale geoscientific knowledge graph, where the edge relation weights are determined by spatial adjacency and hierarchical subordination; adjacency relations are denoted as... Subordination is denoted as For stable training calculations, the self-loop is denoted as... ,in, It is an adjacency matrix. For the identity matrix; S323, calculate the multi-relation multi-head attention layer: Assume the graph attention network has a total of Layer, number Layer usage The first point of attention; the second The output dimension of the layer is set to any node Its neighboring nodes The attention coefficient is calculated as follows: in, express Layer The linear transformation matrix of each attention head. The Layer Attention vectors of attention heads, for Layer node characteristics, For activation function, Indicates the first The attention weights are calculated for any two connected nodes in the layer; these attention weights are then normalized using Softmax to obtain the final attention weights. The current node's features are updated as follows: in, For nodes Neighboring nodes that have an adjacency or dependency relationship The set that is formed This represents the normalized attention weights. express Layer nodes of The node vectors obtained from multiple attention heads are concatenated or averaged to obtain the node representation matrix for that layer. S324, Inter-layer propagation and update: After each layer of propagation, a new node representation matrix is ​​obtained. When the network spreads to the first After layers, the final node embedding matrix is ​​obtained: in, For the number of nodes, Indicates the first The feature dimension of the layer.

6. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, S4 includes: Embedding regional geoscience semantics into vectors Deep feature maps of remote sensing images By location index Mapping function Generate global geoscientific semantic vectors for the corresponding locations in the image. The specific process is as follows: Based on global-local attention fusion mechanism Attention weights for image features are adaptively generated from geoscientific semantic embedding vectors. The global-local attention fusion mechanism employs a parallel spatial channel attention mechanism to adaptively adjust global semantic features; for the generated global geoscientific semantic vector... Apply two consecutive 1×1 convolutions Then apply the Sigmoid activation function. To generate an independent spatial selection mask. The specific process is as follows: Simultaneously, global average pooling is applied. To compress all spatial information in each channel; through a fully connected layer A nonlinear transformation is applied to this global feature to obtain a more compact global feature. Design a learnable vector Multiply with global features, then use The function generates independent channel selection masks. The process can be represented as follows: Channel selection mask With the corresponding spatial selection mask The final weight matrix is ​​obtained by matrix multiplication. This matrix integrates importance information from both spatial and channel dimensions, and the process can be represented as follows: Will pass Normalization yields attention weights Regional knowledge embedding vector With attention weight Multiply the features and add them to the deep feature map of the same type to obtain the fused feature map. The specific process is as follows: Features fused from fourth downsampled image features and geoscientific semantic embedding vectors Furthermore, the Atrous spatial pyramid pooling operation is employed to enhance the model's ability to extract features from small ground features. The specific process is as follows: In the formula, This indicates the image pyramid pooling operation. This indicates the expansion coefficient used in Atrous spatial pyramid pooling. This is the output of Atrous spatial pyramid pooling.

7. The super-resolution land cover mapping method with coupled large language model knowledge base according to claim 1, characterized in that, S5 includes The fused feature map obtained in step S4 Upsampling is performed to obtain feature maps with higher spatial resolution at a specific scale, followed by further convolutional and pooling operations, ultimately... The activation function is used to obtain the land cover classification results for each sub-pixel. The specific steps are as follows: in, For upsampling operation, This represents the upsampling amplification factor. This represents the convolution operation. For the activation function layer, This indicates the execution of multiple identical module steps, and is related to downsampling. Maintain consistency. Feature map representing the target scale. For activation function, This represents the classification result for each sub-pixel in the feature map.

8. A super-resolution land cover mapping system coupled with a large language model knowledge base, characterized in that, include: The multi-source data acquisition module is used to acquire remote sensing images and geoscientific text data of the target area, and to construct a geoscientific knowledge system based on the mapping theme. The structure of the geoscientific knowledge system includes corresponding geoscientific element indicators. The geoscience knowledge mining module is used to extract geoscience knowledge from geoscience text data based on the constructed knowledge system and large language model technology, construct a geoscience knowledge base according to the organizational framework of the multi-level geoscience knowledge system, and further obtain multi-category geoscience element indicators to generate a regional multi-scale geoscience knowledge map. The modal feature extraction module is used to extract deep feature maps of remote sensing images based on convolutional neural networks, and input the regional multi-scale geoscience knowledge graph into the graph attention network to generate regional geoscience semantic embedding vectors; The modal feature fusion module is used to calculate the position and channel weights of the regional geoscientific semantic embedding vector based on the global-local attention mechanism, and to use the weights to perform weighted modulation on the regional geoscientific semantic embedding vector, and fuse it with the deep feature map of remote sensing image to form a fused abstract feature map. The super-resolution feature classification module is used to upsample the fused abstract feature map to the target spatial resolution, obtain a high-resolution depth feature map based on depth convolution operation, and generate sub-pixel classification results using a nonlinear activation function to generate a super-resolution land cover map of the remote sensing image.

Images