A geospatial analysis agent system and method based on large models

Abstract

The application discloses a kind of geospatial analysis intelligent agent systems and methods based on large model, it is related to geographic information science and artificial intelligence technical field.The geospatial analysis intelligent agent system based on large model provided in the present application includes: planning agent, for geospatial data and natural language spatial analysis task instruction, through semantic understanding, complexity evaluation, decomposition and generation structured multi-step action plan;Execution agent includes multiple professional categories of professional execution agent;Review agent includes multiple professional categories of review agent, and review agent and professional execution agent one-to-one correspond;Wherein, planning agent, execution agent and review agent are all agents, agent is based on language model, pre-setting for language model understanding and calling role information, work flow and tool set;Therefore, it solves the problem that traditional geospatial analysis intelligent agent technology based on large model has high threshold and low degree of automation.

Description

Technical Field

[0001] This invention relates to the field of geospatial information processing technology, and in particular to a geospatial analysis intelligent agent system and method based on a large model. Background Technology

[0002] Traditional Geographic Information System (GIS) spatial analysis is the mainstream technical means for solving geospatial problems. Its standard workflow usually relies on professional GIS software (such as ArcGIS, QGIS, etc.) and is manually operated by technical personnel with professional knowledge.

[0003] The GIS spatial analysis toolchain has extremely high professional barriers and complex operating procedures. Users not only need a deep understanding of geographic information science theory, but also must be proficient in operating specific GIS software. This makes it difficult for researchers and policymakers without a GIS background to directly use spatial analysis tools to solve their practical business problems, greatly limiting the popularization and application of spatial analysis technology. Summary of the Invention

[0004] The purpose of this invention is to provide a geospatial analysis intelligent agent system and method based on a large model.

[0005] Firstly, this application provides a geospatial analysis intelligent agent system based on a large model. The system includes: a planning agent, used to decompose and generate a structured multi-step action plan from geospatial data and natural language spatial analysis task instructions through semantic understanding and complexity assessment; an execution agent, including multiple specialized execution agents, used to execute multiple tasks and obtain multiple task results based on the multi-step action plan; and a review agent, including multiple specialized review agents, with each review agent corresponding one-to-one with a specialized execution agent. The review agent is used to: review the results of multiple tasks; if the review fails, the review result is returned to the execution agent to continue updating the task results; wherein, the planning agent, execution agent, and review agent are all agents, and the agents are defined by language... Based on the language model, pre-set role information, workflow, and toolset for the language model to understand and invoke; the agent's working process includes: selecting target tools from the toolset based on natural language function descriptions of natural language instructions and toolset; the agent converts natural language instructions into structured computational tasks according to the parameter specifications of the selected target tools, and obtains the execution results of the computational tasks, including execution status and returned data; the toolset includes a data understanding toolset; the data understanding toolset integrates a geospatial data processing library to parse and extract metadata information from geospatial data, which includes one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial extent, number of bands and resolution of raster data.

[0006] In some embodiments, the toolset further includes a system interaction toolset and a retrieval toolset; wherein the system interaction tools are used to provide controlled interfaces for the agent to interact with the underlying operating system and file system, and the types of controlled interfaces include, but are not limited to: reading, writing and deleting files, creating and querying directories, and executing external code scripts; the retrieval toolset is used to provide the agent with an interface to access external knowledge bases.

[0007] In some embodiments, the executing agent includes a manager agent, which is used to: receive and manage the multi-step action plan, generate a task sequence according to the multi-step action plan, and schedule and assign tasks in the task sequence to corresponding professional executing agents; the professional executing agents include an analyst agent, which is used to process and analyze geospatial data, generate and execute geospatial analysis code by calling tools, and generate analysis results; the professional executing agents include a cartographer agent, which is used to visualize the analysis results and generate visualization results including maps and charts; the professional executing agents include a reporter agent, which is used to integrate the analysis process, analysis results, and visualization results to generate a structured analysis report.

[0008] In some embodiments, the reviewing agent includes a code reviewer agent for reviewing geospatial analysis code generated by the analyst agent; an image reviewer agent for reviewing visualization results; and a report reviewer agent for reviewing analysis reports. If the review result indicates failure, the manager agent re-issues the task to the appropriate professional execution agent for modification based on the feedback generated by the reviewing agent.

[0009] In some embodiments, the planning agent is used to: invoke the semantic understanding capabilities of the language model to analyze the natural language spatial analysis task instructions, invoke data understanding tools to extract metadata from geospatial data, and generate the multi-step action plan.

[0010] In some embodiments, the professional execution agent is used to: invoke a data understanding tool to obtain data attributes of the input data; invoke a retrieval tool to obtain technical specifications or function usage; the professional execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result; in response to the execution result indicating successful task execution, submits the generated execution result to the corresponding review agent for review; wherein, the computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment, and after execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result; wherein, the execution result includes the execution status and returned data.

[0011] In some embodiments, the system includes an agent registration portal and a tool registration portal; the agent registration portal is used to: input the role information, workflow, and list of callable tools for a new role; the tool registration portal is used to: encapsulate custom functions or APIs of external services into new tools; the new tools have tool identifiers, natural language text describing the tool's functions, and parameter specifications defining the input and output.

[0012] Secondly, this application provides a geospatial analysis method based on a large model. The method includes: receiving a natural language spatial analysis task instruction input by a user in natural language form; acquiring one or more geospatial data sets; using a planning agent to decompose and generate a structured multi-step action plan from the geospatial data and the natural language spatial analysis task instruction through semantic understanding and complexity assessment; using an execution agent comprising multiple professional execution agents to execute multiple tasks based on the multi-step action plan and obtain multiple task results; using a review agent comprising multiple professional review agents to review the multiple task results; if the review fails, the review result is returned to the execution agent to continue updating the task results; wherein the planning agent, execution agent, and review agent are all agents. The intelligent agent is based on a language model, with pre-set role information, workflow, and tool set for the language model to understand and invoke. The working process of the intelligent agent includes: selecting a target tool from the tool set based on natural language function descriptions of natural language instructions and the tool set; converting natural language instructions into structured computational tasks according to the parameter specifications of the selected target tool, and obtaining the execution results of the computational tasks, whereby the execution results include execution status and returned data; the tool set includes a data understanding tool set; the data understanding tool set integrates a geospatial data processing library, used to parse and extract metadata information of geospatial data, the metadata including one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial extent, number of bands and resolution of raster data.

[0013] In some embodiments, the executing agent includes a manager agent, which receives and manages the multi-step action plan, generates a task sequence based on the multi-step action plan, and schedules and assigns tasks in the task sequence to corresponding professional executing agents; the professional executing agents include an analyst agent, which processes and analyzes spatial data, generates and executes geospatial analysis code by calling tools, and generates analysis results; the professional executing agents include a cartographer agent, which visualizes the analysis results and generates visualization results including maps and charts; and the professional executing agents include a reporter agent, which integrates the analysis process, analysis results, and visualization results to generate a structured analysis report.

[0014] In some embodiments, the professional execution agent is used to: invoke a data understanding tool to obtain data attributes of the input data; invoke a retrieval tool to obtain technical specifications or function usage; the professional execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result; in response to the execution result indicating successful task execution, submits the generated execution result to the corresponding review agent for review; wherein, the computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment, and after execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result; wherein, the execution result includes the execution status and returned data.

[0015] The embodiments of the geospatial analysis intelligent agent system and method based on large models provided in this application offer a hierarchical, multi-agent collaborative architecture by planning intelligent agents, executing intelligent agents, reviewing intelligent agents, and a set of tools to support intelligent agents. It also introduces a self-verifying quality control mechanism to solve the problems of high technical threshold, low degree of automation, poor flexibility, and difficulty in guaranteeing the reliability of results in traditional geospatial analysis intelligent agents based on large models. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0017] Figure 1 A schematic diagram of the structure of a geospatial analysis intelligent agent system based on a large model is shown.

[0018] Figure 2 An exemplary operational flowchart of a geospatial analysis intelligent agent system based on a large model is shown.

[0019] Figure 3 An exemplary flowchart of the planning layer is shown.

[0020] Figure 4 An exemplary flowchart of the execution layer is shown.

[0021] Figure 5 An exemplary flowchart of the review layer is shown.

[0022] Figure 6 An exemplary flowchart of a geospatial analysis method based on a large model is shown. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0025] In one or more embodiments of this application, a multi-agent large language model system, GeoAgent, is proposed for automatically executing intelligent agent processes based on large models in geospatial analysis. The system aims to address the problems of high technical barriers, low automation, poor flexibility, and difficulty in guaranteeing the reliability of results in traditional geospatial analysis techniques by constructing a hierarchical, multi-agent collaborative architecture and introducing a self-verifying quality control mechanism.

[0026] In one or more embodiments of this application, the concept of this system includes deconstructing the complex spatial analysis task process into three independent yet collaborative layers: a Planning Layer, an Execution Layer, and a Review Layer. This forms a complete and automated workflow from user requirement analysis, task decomposition and planning, multi-role collaborative execution, to result review and feedback. It should be noted that each layer has a corresponding intelligent agent; for example, the Planning Layer has a Planning Intelligent Agent, the Execution Layer has an Execution Intelligent Agent, and the Review Layer has a Review Intelligent Agent. The use of functional layer descriptions is for the purpose of clearly illustrating the workflow; where there is no contradiction, the descriptions of functional layers can be replaced with descriptions of intelligent agents (e.g., replacing the Planning Layer with a Planning Intelligent Agent, the Execution Layer with an Execution Intelligent Agent, and the Review Layer with a Review Intelligent Agent).

[0027] In one or more embodiments of this application, the action plan formulated by the planning layer is passed to the execution layer. This layer mimics the "division of labor and collaboration" model of a human expert team, with a manager agent at its core, and multiple execution agents with different professional skills under its command. The manager agent is responsible for receiving and managing the entire action plan, and according to the plan steps, it accurately schedules and assigns specific sub-tasks to the most suitable execution agents.

[0028] In one or more embodiments of this application, the high-order, natural language-based analysis requirements input by the user are first deeply parsed and structurally decomposed through a planning layer. This layer uses a task parsing module to evaluate the complexity of the task, identify the task category, and determine the required data type. Subsequently, based on the results of the task parsing module, the task planning module formulates detailed, orderly, and executable action plans for the decomposed subtasks. This design, which separates planning from execution, effectively solves the problems of rigid planning and poor flexibility in traditional methods, enabling the system to intelligently generate matching execution plans for tasks of varying difficulty.

[0029] In one or more embodiments of this application, the collaborative workflow logic of the intelligent agents is decoupled from the underlying large language model implementation, thereby supporting the following two core deployment modes: 1. Cloud API mode: In this deployment mode, the core reasoning and content generation functions of each intelligent agent in the system are implemented by calling the application programming interface (API) of an external commercial large language model provider. The system encapsulates the role prompts, current task context, and data to be processed of each intelligent agent into standard API requests and sends them to the cloud model service. The advantage of this mode is that it can instantly utilize the most advanced model capabilities without investing a large amount of local computing resources for model training and maintenance, making it suitable for application scenarios that require rapid prototyping, pursue ultimate performance, and have low data privacy requirements. 2. Localized deployment mode: In this deployment mode, the system supports replacing the large language model relied upon by the intelligent agents with an open-source or self-developed model deployed on a local server or in a private cloud environment. By modifying the system configuration, API calls are directed to the endpoint of the local model service. The advantage of this mode is that it can ensure data privacy and security, as all data processing is completed in the local network environment and is not transmitted to external third parties. This mode is suitable for scenarios with strict data security requirements or specific scenarios that need to run offline in an environment without network connection.

[0030] Please refer to Figure 1 , Figure 1A schematic diagram of the structure of a geospatial analysis intelligent agent system based on a large model, as provided in this application, is shown. The system includes: a planning agent, an execution agent, and a review agent.

[0031] In this embodiment, the planning agent is used to decompose and generate a structured multi-step action plan from geospatial data and natural language spatial analysis task instructions through semantic understanding and complexity assessment.

[0032] In this embodiment, the execution agent includes multiple professional execution agents of various specialties, which are used to execute multiple tasks and obtain multiple task results based on a multi-step action plan; the review agent includes the multiple professional review agents, and the review agents correspond one-to-one with the professional execution agents.

[0033] In this embodiment, the reviewing agent is used to: review the results of multiple tasks; if the review fails, the review result is returned to the executing agent to continue updating the task results.

[0034] In this embodiment, the planning agent, the execution agent, and the review agent are all agents. The agents are based on a language model and have pre-set role information, workflows, and tool sets for the language model to understand and invoke.

[0035] In this embodiment, the working process of the intelligent agent includes: selecting a target tool from the tool set based on the natural language function description of the natural language instruction and the tool set; the intelligent agent converts the natural language instruction into a structured computing task according to the parameter specification of the selected target tool, and obtains the execution result of the computing task, wherein the execution result includes the execution status and returned data.

[0036] As an example, the agent structure comprises three basic components: 1. Role Prompt: A set of preset text instructions used to configure the agent's behavioral patterns, defining the agent's functional positioning, work objectives, capability boundaries, and input / output specifications; 2. Agent Workflow: A directed flowchart defining the agent's internal processing logic, consisting of one or more processing nodes, specifying the internal work sequence of receiving input, performing inference, calling tools, and generating output; 3. Tool: A set of functional interfaces for the agent to interact with the external computing environment, through which the agent performs specific actions such as file operations, data processing, or knowledge retrieval.

[0037] In other words, all intelligent agents in this system, regardless of their level or specific role, follow a standardized component structure design. This achieves a high degree of modularity and scalability for the system.

[0038] The tool invocation mechanism provides the foundation for automation and intelligence, ensuring that the reasoning ability of the language model can be effectively translated into accurate and reliable physical or logical operations. The entire tool invocation process is designed as a closed-loop process that includes tool definition, selection, execution, and feedback.

[0039] As an example, the available tools are standardized and encapsulated. Each tool provides three core pieces of information: a unique tool identifier, natural language text describing its function and applicable scenarios, and a parameter specification defining the input parameter structure and output data format. When an agent needs to perform an operation within its internal workflow, it selects the most suitable tool from its list of accessible tools based on the goal and context of the current subtask, using the natural language function descriptions. The agent formats the required input data into a structured invocation request according to the parameter specification of the selected tool. This request is sent to the system backend, which actually executes the functional logic encapsulated by the tool in a secure, sandboxed computing environment. After execution, the backend encapsulates the execution result, including the execution status (success or failure), return data, and any possible exception information, into a structured response object and returns it to the invoking agent. The agent parses this response object to obtain the execution result of its action and adds this result as new information to its context, providing a basis for decision-making for its next reasoning and actions.

[0040] In this embodiment, the toolset includes a data understanding toolset; the data understanding toolset integrates a geospatial data processing library, which is used to parse and extract metadata information of geospatial data. The metadata information includes one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial range, number of bands and resolution of raster data.

[0041] Data understanding tools are used to empower agents to autonomously understand the objects they are operating on before performing tasks. When processing a spatial data file, the agent can invoke such tools. These tools integrate specialized geospatial data processing libraries that can automatically parse and extract metadata information from the data, such as coordinate reference system, geometric type, attribute field structure, spatial extent, and the number and resolution of raster data bands. The tools return this information in structured text format, enabling the agent to formulate more accurate and efficient processing strategies based on the inherent characteristics of the data.

[0042] For example, please refer to Figure 2 , Figure 2 This diagram illustrates the workflow of the system, which includes the following steps:

[0043] Step S11: Task Submission. The user inputs a geospatial analysis task to be performed in natural language and simultaneously uploads one or more raw spatial data files required to perform the task.

[0044] Step S12: Planning Layer Processing. The system first activates the planning layer. The task parsing module within the planning layer performs semantic analysis and intent recognition on the user-input natural language task instructions, assesses the overall complexity of the task, identifies its analysis category, and verifies the type and completeness of the uploaded data. Based on the parsing results, the task planning module decomposes the complex overall task objective into a series of independent sub-tasks with a logical sequence, and generates a structured overall action plan containing specific execution steps. This plan is then transmitted to the execution layer.

[0045] Step S13: Execution and Review Loop. The manager agent within the execution layer receives and acts as the overall scheduler of the overall action plan. The manager agent reads the first subtask to be executed from the plan and, based on the nature of the subtask, selects an agent with the corresponding professional capabilities from its pool of managing execution agents (e.g., selecting an "analyst" to perform a data analysis task, or selecting a "cartographer" to perform a visualization task). The manager agent dispatches the detailed instructions and required data paths for the subtask to the selected execution agent. Upon receiving the instructions, the execution agent autonomously completes the task operation and generates the corresponding output (e.g., an intermediate data file, a map image, or a piece of analytical text). This output is returned by the execution agent to the manager agent, which then forwards it to the review layer. The review agent in the review layer, corresponding to the role of the execution agent (e.g., a "code reviewer" or "image reviewer"), receives the output, evaluates its professionalism, accuracy, and standardization, and generates a review report containing a "pass" or "revision" conclusion.

[0046] Step S14: Iterative Judgment. If the review conclusion is "Revision," the task and review comments are reassigned to the original executing agent, requiring it to make modifications. After the executing agent completes the modifications, it re-enters the execution and review process of Step S13 until the output of the subtask receives a "Pass" rating from the reviewer. If the review conclusion is "Pass," this conclusion is sent to the manager. Subsequently, the manager agent checks whether there are any unexecuted subsequent subtasks in the overall action plan. If so, the next subtask is selected, and the processes of Steps S13 and S14 are repeated; if all subtasks have been marked as completed, it means that the overall task has been successfully executed, and the process proceeds to the final step.

[0047] Step S15: Result Delivery. Once all subtasks have been reviewed and completed, the managing agent is responsible for collecting and integrating all validated final outputs. The system presents these outputs to the user in a complete and well-organized final result format (e.g., a comprehensive report including analysis descriptions, key charts, and data attachments), completing the entire process of this automated geospatial analysis.

[0048] It should be noted that the approach provided in this embodiment offers a hierarchical, multi-agent collaborative architecture by planning agents, executing agents, reviewing agents, and a set of tools to support agents. It also introduces a self-verifying quality control mechanism to address the problems of high technical barriers, low automation, poor flexibility, and difficulty in guaranteeing the reliability of results in traditional geospatial analysis techniques.

[0049] In some embodiments, the toolset further includes a system interaction toolset and a retrieval toolset; wherein the system interaction tools are used to provide controlled interfaces for the agent to interact with the underlying operating system and file system, and the types of controlled interfaces include, but are not limited to: reading, writing and deleting files, creating and querying directories, and executing external code scripts; the retrieval toolset is used to provide the agent with an interface to access external knowledge bases.

[0050] The system interaction tools provide intelligent agents with controlled interfaces to interact with the underlying operating system and file system. Specific functions include, but are not limited to: reading, writing, and deleting files; creating and querying directories; and executing external code scripts. All interactions occur in an isolated environment to ensure system security and stability.

[0051] Specialized knowledge retrieval tools provide agents with the ability to dynamically access external knowledge bases, compensating for the static and limited nature of knowledge within large language models. When an agent needs to apply a specialized function or comply with specific technical standards, it generates a query request to access the internal documentation interfaces provided by that function library, thereby providing real-time, highly reliable technical support for generating accurate and compliant code or analysis steps.

[0052] In some embodiments, the executing agent includes a manager agent, which is used to: receive and manage the multi-step action plan, generate a task sequence according to the multi-step action plan, and schedule and assign tasks in the task sequence to corresponding professional executing agents; the professional executing agents include an analyst agent, which is used to process and analyze geospatial data, generate and execute geospatial analysis code by calling tools, and generate analysis results; the professional executing agents include a cartographer agent, which is used to visualize the analysis results and generate visualization results including maps and charts; the professional executing agents include a reporter agent, which is used to integrate the analysis process, analysis results, and visualization results to generate a structured analysis report.

[0053] In some embodiments, the reviewing agent includes a code reviewer agent for reviewing geospatial analysis code generated by the analyst agent; an image reviewer agent for reviewing visualization results; and a report reviewer agent for reviewing analysis reports. If the review result indicates failure, the manager agent re-issues the task to the appropriate professional execution agent for modification based on the feedback generated by the reviewing agent.

[0054] Optionally, three types of intelligent agents can be preset: 1. Analyst: Responsible for spatial data processing and core analysis, autonomously generating and executing code to complete tasks such as data transformation and spatial calculations by calling tools. 2. Cartographer: Responsible for visualizing the analysis results, autonomously generating professional maps and charts that conform to cartographic standards. 3. Reporter: Responsible for integrating the analysis process, data, and visualization results, automatically generating structured and complete analysis reports. In this way, the system decomposes the large and complex analysis task into multiple intelligent agents, solving the problem of low efficiency caused by relying on a single intelligent agent to perform all operations in traditional methods.

[0055] To address the inherent reliability issues in automated processes, this system incorporates a Review Layer. This layer serves as a quality assurance mechanism for the system's output. Any intermediate or final results generated by the execution layer (such as code, maps, and reports) are submitted to the Review Layer. The Review Layer includes reviewer agents, such as Code Reviewers, Image Reviewers, and Report Reviewers, each corresponding to an execution agent. These reviewers conduct multi-dimensional evaluations of the results and generate "Pass" or "Revise" review decisions along with specific modification suggestions. If revisions are required, the administrator will reassign the task to the appropriate execution agent based on feedback for modification and optimization. This closed-loop mechanism of "execution-review-feedback-re-execution" constitutes the core of the system's self-verification and self-optimization, thereby ensuring the professionalism and reliability of the final output.

[0056] In some embodiments, the planning agent is used to: invoke the semantic understanding capabilities of the language model to analyze the natural language spatial analysis task instructions, invoke data understanding tools to extract metadata from geospatial data, and generate the multi-step action plan.

[0057] For example, please refer to Figure 3 A schematic diagram of the planning layer's processing flow. Figure 3 The process shown includes:

[0058] Step 31: The system starts up, and the planning layer receives initial input containing the user's natural language task description and related spatial data.

[0059] Step 32: The planning agent in the planning layer performs task parsing on the received input. This process calls on the semantic understanding capabilities of a large language model to analyze the text, and at the same time calls on data understanding tools to extract metadata from the data file, thereby forming an understanding of the task complexity and analysis category.

[0060] Step 33: Based on the analysis results, plan the intelligent agent generation task scheme. This scheme describes in detail the execution steps of each sub-task and defines the preconditions and input / output specifications required for its execution.

[0061] Step 34: The planner agent outputs this complete action plan to the execution layer.

[0062] In some embodiments, the professional execution agent is used to: invoke a data understanding tool to obtain data attributes of the input data; invoke a retrieval tool to obtain technical specifications or function usage; the professional execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result; in response to the execution result indicating successful task execution, submits the generated execution result to the corresponding review agent for review; wherein, the computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment, and after execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result; wherein, the execution result includes the execution status and returned data.

[0063] For example, please refer to Figure 4 A schematic diagram of the processing flow of the intelligent agent. Figure 4 The process shown includes:

[0064] Step 41: The management agent within the execution layer acts as a scheduling unit, reads the action plan, and, based on the task dependencies defined in the plan and the current execution status of the system, decides and assigns the first or next subtask to be executed to the designated professional execution agent, such as the analyst agent.

[0065] Step 42: After receiving the subtask instruction, the designated execution agent first calls the data understanding tool to accurately grasp the specific attributes of the input data, and, if necessary, calls the professional knowledge retrieval tool to obtain the technical specifications or function usage required to complete the task.

[0066] Step 43: After obtaining sufficient contextual information, the executing agent uses its internal large language model to generate the core output for completing the subtask. For example, the analyst agent generates an executable geospatial analysis code script.

[0067] Step 44: The agent invokes the system interaction tool to execute the generated core artifacts in a controlled sandbox environment and captures the execution results, which may be new data files, images, or text.

[0068] Step 45: After confirming that the task has been successfully executed, the executing agent submits the generated final output to the review layer for quality verification.

[0069] Upon receiving the deliverables from the execution layer, the review layer initiates its closed-loop quality verification and feedback process. For an example, please refer to [link / reference needed]. Figure 5 A schematic diagram of the processing flow of the review agent (i.e., the auditing agent). Figure 5 The process shown includes:

[0070] Step 51: The professional review agent (such as a code reviewer or an image reviewer) corresponding to the type of output receives the output and its original subtask instructions.

[0071] Step 52: The review agent calls the corresponding evaluation model to conduct multi-dimensional inspection of the output. For example, the image reviewer uses a multimodal large language model to comprehensively evaluate the map image from the dimensions of symbol system, color scheme, element integrity, etc., in accordance with the text requirements of the sub-task instructions.

[0072] Step 53: After the evaluation is completed, the review agent generates a structured review report containing a clear conclusion (pass or revise) and specific feedback.

[0073] Step 54: The review agent performs the distribution operation based on the review conclusion: If the conclusion is "passed", a task completion signal is sent to the management agent of the execution layer to trigger the scheduling of subsequent tasks; if the conclusion is "revise", a review report containing detailed modification suggestions is sent back to the corresponding execution agent, which will start a new round of correction and execution cycle based on the feedback (return to step 42) until the output passes the review.

[0074] In some embodiments, the system includes an agent registration portal and a tool registration portal; the agent registration portal is used to: input the role information, workflow, and list of callable tools for a new role; the tool registration portal is used to: encapsulate custom functions or APIs of external services into new tools; the new tools have tool identifiers, natural language text describing the tool's functions, and parameter specifications defining the input and output.

[0075] The system can be extended to include intelligent agents or tools.

[0076] As an example, this system supports adding new specialized agent roles to the framework through a standardized process. Specifically, the implementation is as follows: First, define a new, goal-oriented role prompt for the new role. Second, configure a dedicated toolset for the new role, which can include general tools from the system as well as new tools specifically developed for that role. Finally, if the new role's output requires professional verification, a new review agent and its review criteria can be defined in the review layer. Through this process, the system can easily extend its capabilities to new professional fields without altering the core collaboration framework.

[0077] As an example, this system provides an open tool management mechanism that allows developers to continuously expand the system's functionality. Specifically, developers can encapsulate a custom function (such as a piece of Python code) or an external service API into a new tool through a standardized registration process. During registration, the tool needs to provide its unique tool identifier, a natural language text accurately describing its functionality, and a parameter specification defining its input and output. This mechanism ensures that the system's analytical and operational capabilities can be continuously and dynamically enhanced.

[0078] Therefore, the modular design of this system provides a clear and efficient path for the expansion of system functions, mainly reflected in the expansion capabilities of the intelligent agent roles and tool libraries.

[0079] Please refer to Figure 6 , Figure 6 An exemplary workflow of the geospatial analysis method based on a large model provided in this application is illustrated. Figure 6 In the process shown, the method includes:

[0080] Step 61: Receive the natural language spatial analysis task instruction input by the user in natural language form.

[0081] Step 62: Obtain one or more geospatial data.

[0082] Step 63: Using a planning agent, geospatial data and natural language spatial analysis task instructions are decomposed and generated into a structured multi-step action plan through semantic understanding and complexity assessment.

[0083] Step 64: Utilize an execution agent that includes multiple professional execution agents to execute multiple tasks and obtain multiple task results based on a multi-step action plan.

[0084] Step 65: Use multiple professional review agents to review the results of multiple tasks; if the review fails, return the review result to the executing agent to continue updating the task results.

[0085] Among them, the planning agent, the execution agent, and the review agent are all agents. The agents are based on the language model and have pre-set role information, workflows, and tool sets for the language model to understand and call.

[0086] The working process of the intelligent agent includes: selecting a target tool from the tool set based on natural language instructions and natural language function descriptions of the tool set; converting the natural language instructions into structured computational tasks according to the parameter specifications of the selected target tool, and obtaining the execution results of the computational tasks, wherein the execution results include execution status and returned data.

[0087] The toolset includes a data understanding toolset; the data understanding toolset integrates a geospatial data processing library, which is used to parse and extract metadata information from geospatial data. The metadata includes one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial extent, number of bands and resolution of raster data.

[0088] In some embodiments, the executing agent includes a manager agent, which receives and manages the multi-step action plan, generates a task sequence based on the multi-step action plan, and schedules and assigns tasks in the task sequence to corresponding professional executing agents; the professional executing agents include an analyst agent, which processes and analyzes spatial data, generates and executes geospatial analysis code by calling tools, and generates analysis results; the professional executing agents include a cartographer agent, which visualizes the analysis results and generates visualization results including maps and charts; and the professional executing agents include a reporter agent, which integrates the analysis process, analysis results, and visualization results to generate a structured analysis report.

[0089] In some embodiments, the professional execution agent is used to: invoke a data understanding tool to obtain data attributes of the input data; invoke a retrieval tool to obtain technical specifications or function usage; the professional execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result; in response to the execution result indicating successful task execution, submits the generated execution result to the corresponding review agent for review; wherein, the computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment, and after execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result; wherein, the execution result includes the execution status and returned data.

[0090] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A geospatial analysis intelligent agent system based on a large model, characterized in that, The system includes: Planning agents are used to decompose and generate structured, multi-step action plans from geospatial data and natural language spatial analysis task instructions through semantic understanding and complexity assessment. Execution agents, including specialized execution agents of various professional categories, are used to execute multiple tasks and obtain multiple task results based on a multi-step action plan; The review agents include review agents of various specialties, with a one-to-one correspondence between review agents and professional execution agents. The review agents are used to: review the results of multiple tasks; if the review fails, the review result is returned to the execution agent to continue updating the task results. The execution agent includes a manager agent, which is used to: receive and manage the multi-step action plan, generate a task sequence according to the multi-step action plan, and schedule and allocate the tasks in the task sequence to the corresponding professional execution agents. The planning agent, execution agent, and review agent are all agents. The agents are based on language models and have pre-set role information, workflows, and tool sets that the language models can understand and call. The working process of the intelligent agent includes: selecting a target tool from the tool set based on natural language instructions and a natural language function description of the tool set; the intelligent agent converts the natural language instructions into structured computational tasks according to the parameter specifications of the selected target tool, and obtains the execution results of the computational tasks, wherein the execution results include execution status and returned data; The toolset includes a data understanding toolset; the data understanding toolset integrates a geospatial data processing library, which is used to parse and extract metadata information from geospatial data. The metadata information includes one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial extent, number of bands and resolution of raster data; The toolset also includes a system interaction toolset and a retrieval toolset; wherein, the system interaction tools are used to provide controlled interfaces for the agent to interact with the underlying operating system and file system, and the types of controlled interfaces include, but are not limited to: reading, writing and deleting files, creating and querying directories, and executing external code scripts; The retrieval toolset is used to provide agents with interfaces to access external knowledge bases; The specialized execution agent includes an analyst role agent, which processes and analyzes geospatial data, generates and executes geospatial analysis code by calling tools, and generates analysis results. The specialized execution agent includes a cartographer role agent, which is used to visualize the analysis results and generate visualization results including maps and charts; The specialized execution agent includes a rapporteur agent, which integrates the analysis process, analysis results, and visualization results to generate a structured analysis report.

2. The system according to claim 1, characterized in that, The review agents include code reviewer agents, which are used to review geospatial analysis code generated by analyst role agents; The review agents include image reviewer agents, which are used to review visualization results; The review agents include report reviewer agents, which are used to review and analyze reports; If the review result indicates that the task is not approved, the manager agent will resend the task to the corresponding professional execution agent for modification based on the feedback generated by the review agent.

3. The system according to claim 1, characterized in that, The planning agent is used to: invoke the semantic understanding capabilities of the language model to analyze the instructions for the natural language spatial analysis task, invoke data understanding tools to extract metadata from the geospatial data, and generate the multi-step action plan.

4. The system according to claim 1, characterized in that, Specialized execution agents are used for: Use data understanding tools to obtain the data attributes of the input data; Use search tools to obtain technical specifications or function usage; The specialized execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result. In response to the execution result indicating that the task has been successfully executed, the generated execution result is submitted to the corresponding review agent for review; The computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment. After execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result. The execution result includes the execution status and returned data.

5. The system according to claim 1, characterized in that, The system includes an agent registration portal and a tool registration portal; The agent registration portal is used to: input the new role's role information, workflow, and list of available tools; The tool registration entry is used to: encapsulate custom functions or external service APIs into new tools; the new tools have tool identifiers, natural language text describing the tool's functions, and parameter specifications defining input and output.

6. A geospatial analysis method based on a large model, characterized in that, The method includes: Receive natural language spatial analysis task instructions from users in the form of natural language; Acquire one or more geospatial data; By using a planning agent, geospatial data and natural language spatial analysis task instructions are decomposed and generated into structured multi-step action plans through semantic understanding and complexity assessment. The execution agent includes a manager agent, which is used to receive and manage the multi-step action plan, generate a task sequence according to the multi-step action plan, and schedule and assign the tasks in the task sequence to the corresponding professional execution agents. By utilizing specialized execution agents encompassing multiple professional categories, and based on a multi-step action plan, multiple tasks are executed to obtain multiple task results; The review agent, which utilizes multiple professional review agents, reviews the results of multiple tasks; if the review fails, the review result is returned to the execution agent to continue updating the task results. Among them, the planning agent, the execution agent and the review agent are all agents. The agents are based on the language model and have pre-set role information, workflow and tool set for the language model to understand and call. The working process of the intelligent agent includes: selecting a target tool from the tool set based on natural language instructions and a natural language function description of the tool set; the intelligent agent converts the natural language instructions into structured computational tasks according to the parameter specifications of the selected target tool, and obtains the execution results of the computational tasks, wherein the execution results include execution status and returned data; The toolset includes a data understanding toolset; the data understanding toolset integrates a geospatial data processing library, which is used to parse and extract metadata information from geospatial data. The metadata includes one or more of the following: coordinate reference system, geometric type, attribute field structure, spatial extent, number of bands and resolution of raster data; The specialized execution agent includes an analyst role agent, which is used to process and analyze spatial data, generate and execute geospatial analysis code by calling tools, and generate analysis results. The specialized execution agent includes a cartographer role agent, which is used to visualize the analysis results and generate visualization results including maps and charts; The specialized execution agent includes a rapporteur agent, which integrates the analysis process, analysis results, and visualization results to generate a structured analysis report.

7. The method according to claim 6, characterized in that, Specialized execution agents are used for: Use data understanding tools to obtain the data attributes of the input data; Use search tools to obtain technical specifications or function usage; The specialized execution agent uses a language model to generate outputs for completing the role's tasks, generates a call request based on the outputs, sends the call request to the computing end, and receives the execution result. In response to the execution result indicating that the task has been successfully executed, the generated execution result is submitted to the corresponding review agent for review; The computing end executes the functional logic encapsulated by the target tool in a sandboxed computing environment. After execution, the computing end encapsulates the execution result into a structured response object and returns it to the initiating execution agent as the execution result. The execution result includes the execution status and returned data.

Images