Remote sensing interpretation agent system based on large language model

Through a remote sensing interpretation agent system based on a large language model, combined with LoRA fine-tuning and retrieval enhancement generation technology, the problem of inefficient multi-source multi-modal interpretation of remote sensing data is solved, and efficient and accurate processing of multiple remote sensing data and The decision-making process is generated, which improves the applicability of the system and self-learning ability.

CN119005242APending Publication Date: 2024-11-22SHANGHAI UNIV

Patent Information

Application Number
CN202410864445.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-30
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

The existing remote sensing data interpretation methods are inefficient and have low interpretation accuracy when facing multi-source and multi-modal data. The traditional methods need to be optimized for specific application scenarios or single tasks, which increases development and maintenance costs and limits The flexibility and wide application of remote sensing technology.

Method used

Develop a remote sensing interpretation agent system based on large language models, and realize adaptive processing by collecting and decoupling various remote sensing proprietary models. Multimodal remote sensing data and unified solution to various interpretation task requirements.

Benefits of technology

It realizes efficient and accurate interpretation of multiple remote sensing data types, reduces the complex process of building and training models for each data type, improves the applicability and scalability of the system, and can dynamically generate decision-making processes and continuously use user feedback. Optimization improves the overall performance and user satisfaction of the system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119005242A_ABST
    Figure CN119005242A_ABST
Patent Text Reader

Abstract

The invention discloses a remote sensing interpretation agent system based on a large language model, which is characterized in that various remote sensing proprietary models, such as hyperspectral classification, change detection, target detection and scene classification, are integrated and decoupled into an image coding module and a task-related decoding module, and flexible combination is realized through interface packaging; therefore, the large language model can process different remote sensing interpretation tasks. A structured output mechanism is introduced, a retrieval enhancement generation technology (RAG) is further introduced in the system, an external vector database is established, and marked decision process examples and decision processes with good user evaluation are stored. The invention not only provides a universal agent suitable for various remote sensing interpretation scenes, but also enriches methods and applications of remote sensing interpretation tasks, can adaptively meet user requirements according to multi-modal data input by a user, has wide application prospects and practical value, and is suitable for popularization and application. And important technical support is brought to research and practice in the remote sensing field.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of remote sensing intelligent interpretation, and in particular to the development of a remote sensing interpretation intelligent agent system based on a large language model. Background Art

[0002] Remote sensing technology has a wide range of applications in earth observation, environmental monitoring, agricultural management, urban planning and other fields. Remote sensing data usually includes multispectral, hyperspectral, radar and other data forms, which are obtained through satellites, drones and other platforms and can provide detailed information about the earth's surface. However, with the rapid development of remote sensing technology and the diversification of data acquisition methods, the amount and complexity of remote sensing data are increasing. For different data types and different types of interpretation tasks, various independent and redundant methods are often required to deal with various application scenarios in a targeted manner. Therefore, it is of great practical value to develop an intelligent system that integrates remote sensing interpretation methods to efficiently and accurately interpret various data types and uniformly handle multiple application scenarios.

[0003] Existing remote sensing data interpretation methods mainly rely on expert experience and rule-based algorithms. These methods have certain advantages when dealing with simple, single-type data, but they are prone to inefficiency and low interpretation accuracy when facing multi-source, multi-modal data. In addition, traditional methods are often optimized for specific application scenarios or single tasks. Whenever new data sources or new application requirements are introduced, it is usually necessary to rebuild the model and perform a lot of training and tuning work. Although deep learning technology has achieved remarkable results in the field of remote sensing interpretation, remote sensing large models have been widely studied and their overall performance has been improved. They have the ability to process multimodal data and solve various downstream tasks, but the performance of these large remote sensing models directly applied to various types of downstream tasks is not even as efficient as some proprietary models, and often requires separate fine-tuning to meet the requirements of the task. This not only increases the development and maintenance costs, but also limits the flexibility and wide application of remote sensing technology. Therefore, it is particularly necessary to build an intelligent agent system that can be uniformly applied to various application scenarios and can efficiently and accurately interpret and fuse various data types.

[0004] Based on the above practical needs, it is required to build an intelligent agent system that can adaptively match the corresponding proprietary model to process remote sensing images in combination with specific task requirements. In addition, the system must be able to dynamically process different modal data input by the user according to the input situation to better meet the user's needs. In recent years, with the continuous improvement of computing power, the number of network parameters of deep learning models has also continued to increase, and the emergence of large models has been fully reflected in large language models. For example, ChatGPT can already answer questions input by users very reasonably. Therefore, the excellent analytical ability of the large language model provides the possibility for the realization of the intelligent agent, and can be well used as the "brain" responsible for decision-making in the intelligent agent. After the "brain" makes a decision, the intelligent agent system can parse according to the structured output of the large language model, and accurately call various proprietary models to interpret the remote sensing images specified by the user to complete the user's specified needs. With the help of a large model, building an intelligent agent system that can be applied to various task scenarios and uniformly solve the needs of various remote sensing interpretation tasks is the core innovation pursued by the present invention. Because the intelligent agent can rationally decouple task processes, the system can easily fuse multimodal data and can update and improve the system's interpretation capabilities for various scenarios at any time as the proprietary model of remote sensing interpretation is iterated. Summary of the invention

[0005] In order to effectively deal with various application scenarios of remote sensing interpretation and the high cost of joint processing of multimodal remote sensing data, the present invention proposes a remote sensing interpretation intelligent agent system based on a large language model to solve the problems pointed out in the background technology.

[0006] To achieve the above purpose, the technical solution of the present invention is implemented in the following steps:

[0007] Step 1: Collect and decouple various existing remote sensing proprietary models of various modal types, including but not limited to hyperspectral classification, hyperspectral change detection, high-resolution remote sensing image target detection, scene classification and other task scenarios. By decoupling the network into an image encoding module and a decoding module that is closely related to the task and encapsulating the interface, the large language model can reasonably combine different modules to realize various remote sensing interpretation application scenarios.

[0008] Step 2: Build the structured output of the large language model so that the output of the large language model can be parsed by the system.

[0009] The specific prompt template details are as follows:

[0010] system message:

[0011] a) You are an expert in remote sensing interpretation. You must always make decisions independently for given task requirements without seeking help from users, and use your advantages as a large language model. Pursue simple and efficient strategies.

[0012] b) Description of restrictions:

[0013] i. Only the actions listed below can be used;

[0014] ii. The actions in the three different action pools of pre-action, mid-action, and post-action can only be used as components of their corresponding actions, and the change of action identity is not allowed;

[0015] iii. Each step in the action flow list must consist of three action components: pre-action, mid-action, and post-action. If one of the components is not needed, the action of that component should be set to "None";

[0016] iv. The last step of the action flow list must consist of three "Finish" to tell the system that the task is finished;

[0017] v. The image to be processed is represented by ```Input-Image-{serial number}``` in the structured output;

[0018] vi. You can only act proactively, and you need to take this into account when planning your actions;

[0019] vii. You cannot interact with physical objects. If it is absolutely necessary to complete the task or goal, you must ask the user to complete it for you, and directly output the termination action while giving a reminder to avoid wasting time and energy.

[0020] c)Action description:

[0021] i. Pre-action action pool: (mainly encoders for encoding remote sensing data of various modalities. The action description can indicate the type of remote sensing interpretation problem solved before the encoder is decoupled to better help the large language model make decisions)

[0022] 1.{<action 1 name (must correspond to the interface name)>:<action 1 description>,

[0023] 'args':[{'name':<parameter 1 of action 1>,'description':<parameter description>,'type':<parameter type>},

[0024] {'name':<parameter 2 of action 1>,'description':<parameter description>,'type':<parameter type>},

[0025] … ]

[0027] }

[0028] 2.……

[0029] 3.{'None':'This action component is not required for the current step','args':'None'}

[0030] 4.{'Finish':'End the action flow identification, output the results and ask the user to evaluate the results',

[0031] 'args':[{'name':'answer','description':'Final output'}]

[0032] ii. Mid-action action pool: (mainly based on the post-encoding feature processing operation module, including complex actions such as prior feature enhancement and network retrieval, as well as simple data operation actions such as feature difference and normalization. The specific form is the same as the pre-action action pool)

[0033] iii. Post-action action pool: (mainly task-related decoders)

[0034] d) Resource Description:

[0035] i.pre-action contains remote sensing image encoders of various modalities;

[0036] ii. mid-action contains various feature operation methods;

[0037] iii. post-action contains various decoders that complete the goal or its sub-goals;

[0038] iv. You are a large language model that has been trained on a large number of decision-making processes, and use this knowledge to avoid unnecessary process decisions as much as possible.

[0039] e) Decision output format description:

[0040] i. You should only respond in json format, the response format is as follows:

[0041] {

[0042] 'action':[{

[0043] 'pre-action':{'name':'action name','args':{'argument name':'argumentinput'}},

[0044] 'mid-action':{'name':'action name','args':{'argument name':'argumentinput'}},

[0045] 'post-action':{'name':'action name','args':{'argument name':'argumentinput'}}

[0046] },

[0047] {

[0048] 'pre-action':{'name':'action name','args':{'argument name':'argumentinput'}},

[0049] 'mid-action':{'name':'action name','args':{'argument name':'argumentinput'}},

[0050] 'post-action':{'name':'action name','args':{'argument name':'argumentinput'}}

[0051] },

[0052] …

[0053] ],

[0054]

[0055] ii. Make sure the response result can be parsed by python json.loads.

[0056] In step 3, for some classic and representative remote sensing interpretation task scenarios, a certain amount of task requirements and decision-making process pairs are collected and labeled, and converted into the corresponding structured output form in step 2, thereby forming input-output pairs that can be used for large language model training.

[0057] Step 4: Use LoRA fine-tuning technology to add a learnable low-rank incremental matrix to the weights of the large language model. Fine-tune this additional small number of parameters on the training samples annotated in step 3 while retaining the original pre-trained weights. Then, while retaining the rich prior general knowledge of the large language model, inject the unique knowledge in the remote sensing field into the large language model to improve the performance of the large language model in remote sensing interpretation. The specific steps are as follows:

[0058] Step 4-1 introduces an additional weight increment in the Transformer module of the large language model to replace the update of the model weight, that is:

[0059]

[0060] in Represents the pre-trained weight matrix.

[0061] Step 4-2, further fine-tune the large language model on the ('target requirement', 'decision process json') training sample pairs annotated in step 2 to improve the model's understanding ability for remote sensing interpretation tasks.

[0062] Step 5: In order to enable the intelligent system to continuously improve and learn, the system introduces the Retrieval-Augmented Generation (RAG) technology. By building an external vector database and storing the decision-making process examples that we have already labeled and the decision-making processes that have been well received by users in practice, the next time a user demand comes, the retrieval results of the vector database can be used as a contextual reference for the large language model, thereby helping the large language model to make decisions that are more in line with user needs and more efficient. Specifically:

[0063] When executing the retrieval-augmented generation (RAG) technique, the agent system first needs to parse the user's input to extract key demand information and context. Then, the system searches the external vector database to find the decision process examples that best match the current user's needs and the decision process that users have praised. The specific steps are as follows:

[0064] Step 5-1, vectorize user input: Convert the user input requirements and context information into vector representation. This step uses the encoding capability of the large language model to embed text information into a high-dimensional vector space.

[0065] Step 5-2, search for similar vectors: Use vector similarity calculation to search for the vector closest to the user input vector in the external vector database. This step can use a variety of similarity measurement methods, such as cosine similarity, Euclidean distance, etc., to ensure the accuracy of the search results.

[0066] Step 5-3, extract relevant decision processes: Based on the retrieved similar vectors, extract corresponding decision process examples and decision processes with user reviews. These decision processes have been proven to have high accuracy and effectiveness in practical applications.

[0067] Step 5-4, build context reference: The retrieved decision process examples are used as context references and input into the large language model. In this way, the large language model can combine the experience of historical decision processes to generate interpretation solutions for current needs more accurately and efficiently.

[0068] Step 5-5, Generate decision process: Based on context reference and user's specific needs, the large language model generates a preliminary interpretation decision process. This process includes detailed steps from data preprocessing, model selection to result interpretation, aiming to meet the user's specific needs.

[0069] Step 5-6, user feedback: present the generated decision process to the user, allowing the user to review and provide feedback. The user can make changes to each step of the process or confirm its accuracy.

[0070] Steps 5-7, optimize and update the database: adjust and optimize the generated decision process based on user feedback, and store the process examples confirmed and optimized by users into the external vector database to further improve the system's retrieval and generation capabilities.

[0071] The technical implementation principle of the present invention: the system utilizes the excellent problem analysis and decision-making capabilities of the large language model, and combines the large model field technologies such as retrieval-augmented generation technology (RAG) and LoRA (Low-Rank Adaptation of Large Language Models) fine-tuning, and accesses various existing remote sensing proprietary models to achieve adaptive processing of multimodal remote sensing data, and uniformly solve various remote sensing interpretation scenarios. When building the system, it is first necessary to collect various types of remote sensing proprietary models, and decouple the encoder and decoder of the model to obtain encoders of various modes and decoders of various interpretation tasks. Then, by collecting some task requirements in representative application scenarios and manually annotating their corresponding standardized processes. By using the annotated task requirements and decision-making processes, the large language model (such as the LLaMA model) is combined with the LoRA strategy to fine-tune the large language model to improve the understanding ability of the large language model in the remote sensing task scenario. In order to enable the large language model to continuously optimize itself and enrich its knowledge reserve when it is actually deployed, the system builds an external vector database by introducing retrieval enhancement generation technology, and uses manually annotated task requirements and decision-making process pairs as the basic information of the external vector library. When the agent is actually deployed, the external vector database is enriched by the user's satisfaction with the results to achieve self-learning and improvement of the agent.

[0072] Beneficial effects: Through the above steps, the intelligent system can not only dynamically generate efficient and accurate remote sensing interpretation decision-making processes according to user needs, but also gradually improve the overall performance and user satisfaction of the system by continuously accumulating and learning user feedback. In this way, the system can always maintain high efficiency and accuracy when processing diverse remote sensing data and complex interpretation tasks, meeting the needs of various application scenarios. The specific effects are as follows:

[0073] 1. By introducing the large language model and LoRA fine-tuning technology, the remote sensing interpretation intelligent agent system of the present invention can efficiently and accurately process remote sensing data of multiple modalities according to user needs and inputs, avoiding the complex process of building and training models for each data type in traditional methods.

[0074] 2. The present invention provides a unified interpretation framework that can be applied to a variety of remote sensing interpretation tasks and application scenarios. Through modular design, the system can flexibly combine different model modules to achieve the interpretation of different types of remote sensing data, greatly improving the applicability and scalability of the system.

[0075] 3. By leveraging the excellent analysis and decision-making capabilities of the large language model and the retrieval-augmented generation technology (RAG), the system can dynamically generate decision-making processes based on user needs, and continuously optimize and improve through user feedback1, thereby enhancing the system's intelligence and self-learning capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 It is a schematic diagram of the overall module structure of the present invention.

[0077] Figure 2 It is a flow chart of the RAG part of an embodiment of the present invention. DETAILED DESCRIPTION

[0078] The following is a detailed description of an embodiment of the present invention in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0079] Reference Figure 1 As shown in the figure, the overall implementation steps of the remote sensing interpretation agent system are as follows:

[0080] Step 1: The user inputs the requirement text and image data as the input information of the system. The requirement text contains the user's specific requirements and interpretation tasks.

[0081] Step 2: The system compares and searches the requirements input by the user with the information in the external vector database, retrieves the relevant case information as additional context information and sends it to the large language model together with the requirements.

[0082] Step 3: The large language model makes decisions based on the input information according to the requirements of the system prompt and generates structured output. The structured output consists of two parts: action and observation. The action part describes in detail the specific steps such as pre-action, mid-action and post-action. The observation part includes information such as plan and reason.

[0083] Step 4: The system parses the structured output of the large language model and submits the parsed action part to the action execution module. The action execution module calls the corresponding interface to perform specific operations based on the parsing results.

[0084] Step 5: After the system completes all actions, it returns the interpretation results and the analysis of the "observation" part to the user. The user reviews and gives feedback on the results. The system further enriches the external vector database based on the user's feedback information, and then optimizes and adjusts the interpretation process.

[0085] Reference Figure 2 As shown in FIG. 1 , an example of the implementation steps of the retrieval enhancement generation technology (RAG) in the remote sensing interpretation agent system is as follows:

[0086] S1: The construction of the external vector database is composed of a large number of reliable target decision pairs. For example, the user's requirement is "please perform change detection on the two input hyperspectral images", and the model's structured output decision is:

[0087] {

[0088] 'action':['Hyperspectral Encoder 1', 'Feature Map Interpolation', 'Change Detection Decoder 1'],

[0089] 'observation':{

[0090] 'First, the two spectral images are encoded by the hyperspectral encoder 1, and then the difference is input into the change detection decoder to generate the change detection result.

[0091] Since change detection needs to detect the area that has changed between two images, the difference needs to be processed after encoding to highlight the features of the changed part, and finally the change detection decoder is used to achieve the goal'

[0092] }

[0093] }

[0094] A large number of reliable cases are represented into vector form through a text encoder to form a vector database.

[0095] S2: The user inputs a requirement, such as "please perform change detection on two high-resolution images". The system encodes the user's input requirement into a vector representation using the same text encoder as the external vector database.

[0096] S3: Calculate the similarity between the vector representation of the user's needs and the example vectors in the vector database to retrieve the Top K example vectors that are most similar to the current needs.

[0097] S4: Extract the data corresponding to similar vectors, and then pass the search results together with user needs as input information to the large language model.

[0098] S5: The large language model combines the retrieved similar data and the user’s current specific needs to generate an interpretation decision process.

[0099] Through the above steps, the system can not only dynamically generate efficient and accurate remote sensing interpretation decision-making processes, but also continuously improve the overall performance and user satisfaction of the system through user feedback, meeting the needs of diverse remote sensing data and complex interpretation tasks. The present invention introduces a structured output mechanism to ensure that the output of the large language model can be accurately parsed and applied by the system. In view of the classic remote sensing interpretation task requirements and decision-making processes, the system collects and annotates rich training data, converts it into a structured output format, and provides the necessary input and output pairs for the training of the large language model. In order to combine the professional knowledge in the remote sensing field with the large language model, this technology adopts the LoRA fine-tuning technology, which effectively improves the performance and adaptability of the model in remote sensing interpretation tasks while retaining the general knowledge of the original large language model by fine-tuning these learnable low-rank incremental matrices on the annotated training samples. In addition, the system also introduces the retrieval enhanced generation technology (RAG), establishes an external vector database, and stores annotated decision process examples and decision processes with good user evaluation. When the user needs it, the system provides context references for the large language model based on the database retrieval results, helping the model to generate more practical and efficient remote sensing interpretation decisions.

[0100] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A remote sensing interpretation agent system based on a large language model, characterized in that: The steps include: Step 1: Collect and decouple various remote sensing proprietary models, and decouple the network into image encoding modules and task-related decoding modules for interface encapsulation; Step 2: Build a system message to create a structured output so that the output of the large language model can be parsed by the system; Step 3: For classic remote sensing interpretation tasks, collect and annotate a certain amount of task requirements and decision-making processes, and convert them into the structured output format in step 2 to provide input and output pairs for large language model training; Step 4: Use LoRA fine-tuning technology to add a learnable low-rank incremental matrix to the weight of the large language model, and fine-tune these additional parameters on the training samples annotated in step 3; Step 5: Introduce the retrieval-augmented generation technology (RAG), build an external vector database and store annotated decision process examples and decision processes with good user evaluation.

2. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 1, various existing remote sensing proprietary models of various modal types are collected and decoupled, including but not limited to hyperspectral classification, hyperspectral change detection, high-resolution remote sensing image target detection and scene classification and other task scenarios. By decoupling the network into an image encoding module and a decoding module that is closely related to the task and performing interface encapsulation, the large language model can reasonably combine different modules to realize various remote sensing interpretation application scenarios, thereby preparing for the intelligent system.

3. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 2, a structured output of a large language model is constructed so that the output of the large language model can be parsed by the system.

4. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 3, for some classic and representative remote sensing interpretation task scenarios, a certain amount of task requirements and decision-making process pairs are collected and labeled, and converted into the corresponding structured output form in step 2, thereby forming input-output pairs that can be used for large language model training.

5. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 4, the LoRA fine-tuning technique is used to add a learnable low-rank incremental matrix to the weights of the large language model.

6. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 4, the original pre-trained weights are retained by fine-tuning the additionally introduced small number of parameters on the training samples annotated in step 3, and then the unique knowledge in the remote sensing field is injected into the large language model while retaining the rich prior general knowledge of the large language model, so as to improve the performance of the large language model in remote sensing interpretation.

7. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: When step 4 is implemented, it specifically includes the following steps: Step 4-1 introduces an additional weight increment in the Transformer module of the large language model to replace the update of the model weight, that is: in Represents the pre-trained weight matrix; Step 4-2, further fine-tune the large language model on the ('target requirement', 'decision process json') training sample pairs annotated in step 2 to improve the model's understanding ability for remote sensing interpretation tasks.

8. The remote sensing interpretation intelligent agent system based on a large language model as claimed in claim 1, characterized in that: In step 5, in order to enable the intelligent agent system to have the ability to continuously improve and learn, the system introduces the Retrieval-Augmented Generation (RAG) technology.

9. A remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 5, we build an external vector database and store the decision process examples we have annotated and the decision processes that have been well received by users in practice. This way, the next time a user demand comes, we can use the search results of the vector database as a context reference for the large language model, thereby helping the large language model make decisions that are more in line with user needs and more efficient.

10. The remote sensing interpretation agent system based on a large language model as claimed in claim 1, characterized in that: In step 5, when the agent system performs the retrieval augmentation generation technology (RAG), it first needs to parse the user's input to extract key demand information and context. Then, the system searches the external vector database to find the decision process examples that best match the current user's needs and the decision process that is well received by users. Specifically, the steps include: Step 5-1, vectorize user input: convert the user input requirements and context information into vector representation. This step uses the encoding ability of the large language model to embed text information into a high-dimensional vector space. Step 5-2, retrieve similar vectors: use vector similarity calculation to search for the vector closest to the user input vector in the external vector database. This step can use a variety of similarity measurement methods, such as cosine similarity, Euclidean distance, etc., to ensure the accuracy of the search results; Step 5-3, extract relevant decision processes: Based on the retrieved similar vectors, extract corresponding decision process examples and decision processes with user reviews. These decision processes have been proven to have high accuracy and effectiveness in practical applications. Step 5-4, construct context reference: The retrieved decision process examples are used as context references and input into the large language model. In this way, the large language model can combine the experience of historical decision processes to generate interpretation solutions for current needs more accurately and efficiently. Step 5-5, generate decision process: Based on context reference and specific needs of users, the large language model generates a preliminary interpretation decision process, which includes detailed steps from data preprocessing, model selection to result interpretation, aiming to meet specific needs of users; Step 5-6, user feedback phase: present the generated decision process to the user, allowing the user to review and provide feedback. The user can make changes to each step of the process or confirm its accuracy; Steps 5-7, optimize and update the database: adjust and optimize the generated decision process based on user feedback, and store the process examples confirmed and optimized by users into the external vector database to further improve the system's retrieval and generation capabilities.

Citation Information

Patent Citations

  • Large model retrieval enhancement generation method and device

    CN117891838A

  • RAG knowledge question-answering method and device based on fusion vector and keyword retrieval

    CN117951274A

  • Visual big language model construction method for multi-source ship remote sensing image interpretation

    CN117994677A

  • Text information extraction method based on large language model and efficient parameter fine tuning

    CN118132674A

  • Diversity-enhanced text retrieval enhancement generation method and system

    CN118260406A

Cited By

  • Financial service processing method and device, storage medium and electronic equipment

    CN119831728A

  • Remote sensing intelligent interpretation proxy method and device based on multi-agent cooperation and medium

    CN120164094A

  • Remote sensing intelligent interpretation system based on large language model

    CN120375201A

  • Through-the-wall radar high-resolution imaging method based on large language model

    CN120802255A

  • Wall-penetrating radar high-resolution imaging method based on large language model

    CN120802255B