On-orbit three-dimensional modeling method and system

Abstract

The invention relates to the technical field of three-dimensional modeling, in particular to an on-orbit three-dimensional modeling method and system. The method comprises the following steps: mapping a task demand input by a user into a task specification object; carrying out mixed retrieval on the task template library, carrying out atlas consistency verification on the retrieved task templates, and sorting candidate templates after the verification is passed; expanding the high-score candidate template into an executable high-score operator sequence to serve as a task chain to be output; mapping each operator in the task chain into an MCP calling node to obtain an MCP calling chain; and calling a data processing tool and a three-dimensional modeling tool to generate a preliminary three-dimensional model, then performing incremental correction and local optimization through a dynamic feedback updating mechanism, and downloading the updated incremental three-dimensional data to a ground data center. Automatic task analysis is achieved, and manual intervention is not needed. The method has the lightweight modeling capability, knowledge-driven complementation can be realized, and a collaborative mode of on-orbit quick response and ground fine reconstruction is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional modeling, in particular to an on-orbit three-dimensional modeling method and system. BACKGROUND

[0002] With the rapid development of remote sensing satellite technology, three-dimensional geographic information acquisition is increasingly widely used in disaster emergency, city monitoring, national defense security and other fields. At present, the mainstream three-dimensional modeling methods, such as point cloud reconstruction technology based on motion recovery structure and multi-view stereo vision, and commercial software platforms such as RealityCapture and Pix4D, generally rely on ground computing environment for processing. These methods usually require professional personnel to configure complex processes and adjust parameters when implemented, including camera parameter setting, matching point screening and error optimization, and multiple interactive links. This processing mode highly dependent on manual intervention makes it difficult to adapt to the autonomous operation requirements in the on-orbit environment.

[0003] At the same time, completely downloading massive raw image data to the ground for processing will occupy huge satellite-ground communication bandwidth and ground storage resources, and introduce significant processing time delay. In disaster monitoring, battlefield situation awareness and other application scenarios with extremely high time efficiency requirements, it cannot meet the real-time response requirements of minute-level acquisition of local three-dimensional information. SUMMARY

[0004] In order to solve the above problems, the present application provides an on-orbit three-dimensional modeling method and system, wherein the method comprises the following steps: S1, using controlled semantic analysis, mapping the user input task requirement into a task specification object; S2, based on the task specification object, performing hybrid retrieval on the task template library, and performing graph consistency verification on the retrieved task templates regarding template preconditions and data dependencies, and after the verification is passed, sorting the candidate templates; S3, expanding the sorted high-score candidate templates into executable high-score operator sequences through a task planning method based on hierarchical task network and / or planning domain description language, and outputting the high-score operator sequences as task chains; S4, mapping each operator in the task chain into an MCP (Model Context Protocol) call node, and generating a call order according to the directed acyclic graph topology order of the task chain and resource availability to obtain an MCP call chain; S5, calling data processing tools and three-dimensional modeling tools according to the MCP call chain to generate a preliminary three-dimensional model, and then performing incremental correction and local optimization on multi-temporal images through a dynamic feedback updating mechanism, and downloading the updated incremental three-dimensional data to a ground data center.

[0005] Specifically, the structure of the task specification object includes at least the fields of task intention, imaging modality, preconditions and postconditions; When the AI Agent performs the mixed retrieval of the task template library in S2, it includes: Based on the task specification object, the task template library is retrieved in parallel or in series using a triple retrieval mechanism, and the retrieval results of the triple retrieval mechanism are integrated to obtain the task template, wherein the triple retrieval mechanism includes a semantic vector index, a structured inverted index and a task graph index, the semantic vector index vectorizes the description and label of the task template through an embedding model to realize semantic similarity matching, the structured inverted index quickly filters the task intention, the imaging modality, the preconditions and the postconditions field, and the task graph index constructs a searchable graph according to the preconditions and the postconditions to realize subgraph matching and dependency verification.

[0006] Specifically, when the AI Agent performs S2, it also performs S6 in parallel, which includes: According to the task specification object, knowledge is retrieved from an extended knowledge base and used as a constraint condition for constraining the task template, the data processing tool and the three-dimensional modeling tool, wherein the extended knowledge base includes at least one of city planning rules, historical three-dimensional models and geographic information system data.

[0007] Specifically, S3 further includes: When the high-resolution operator sequence is output as the task chain, a probe subtask is inserted in the high-resolution operator sequence for quickly evaluating on-orbit environmental state parameters, and the adaptive branch selection of the task chain is driven according to the evaluation results; Through constraint planning or heuristic scheduling methods, when the task chain is generated, multi-objective scheduling optimization is performed to generate a near-optimal task chain that meets multi-objective constraints, with the optimization of time delay, downlink bandwidth, energy consumption and quality indicators as the target.

[0008] Specifically, the data processing tool includes a plurality of special processing modules controlled by a modular protocol (MCP), and the special processing modules include: MCP-SAR, for processing SAR remote sensing data, performing radiation correction, geometric correction and interference analysis; MCP-VIS, for processing visible light images, completing cloud removal, orthorectification and registration; MCP-TIR, for processing thermal infrared images, for night modeling or ground surface thermal feature supplement; MCP-3D Matching, for multi-view geometry sparse point cloud matching; MCP-3D Gaussian Splatting, for efficient rendering and continuous 3D reconstruction; MCP-Data Overlay, for multi-modal overlay and data fusion; MCP-Data Compression, for data compression and optimization; The AI Agent invokes the specialized processing module according to the MCP call chain, and outputs the processed standardized data to the three-dimensional modeling tool for use in the modeling process.

[0009] Specifically, each specialized processing module declares its capabilities by providing a standardized tool capability declaration file, each tool capability declaration file including Schema definition of input and output, adjustable parameters and value range, resource profile, telemetry field, error model and fallback list, and declaring version and compatibility strategy; The specialized processing module supports idempotent operation, health check, checkpoint output and permission audit, and provides cost prompt capability through Dryrun mode The Schema definition of input and output in the tool capability declaration file is used for interface alignment and automatic verification, the adjustable parameters and value range are used for parameter binding in the planning stage, the resource profile is used for resource allocation in multi-target scheduling optimization, the telemetry field is used for real-time monitoring of module execution state and triggering of quality gating, the error model and fallback list are used to define failure fallback path and retry strategy, and the version and compatibility strategy are used to ensure compatibility of module invocation; The idempotent operation is implemented by generating an idempotent key through hashing of input and parameters, and repeated invocation of the same input returns consistent results; The health check regularly reports module state indicators; The checkpoint output supports breakpoint resume of task execution; The Dry run mode verifies the feasibility of input, parameters, resource estimation and data path without executing core calculation.

[0010] Specifically, the S1 includes: converting user natural language input into structured task specification object through constraint prompt and JSON Schema verification, the task specification object containing at least part of fields of task intent, area, time window, imaging modality, quality target, latency target, resource constraint, privacy constraint and expected output, and strictly mapping user input to unambiguous structured data through controlled semantic parsing.

[0011] Specifically, the AI Agent includes the following when performing the graph consistency check on the retrieved task template in the S2 regarding the template preconditions and data dependencies: checking the satisfiability of the task template in terms of input data availability, tool capability matching, and dependency closure, using the task graph index to construct the template preconditions and postconditions into a searchable graph, and verifying the dependency consistency of data flow and task chain through subgraph matching.

[0012] In addition, the system of the present application operates by the above-mentioned method, including: a user demand layer for mapping the user input task demand into a task specification object using controlled semantic parsing; based on the task specification object, performing hybrid retrieval on the task template library, and performing graph consistency check on the retrieved task template regarding the template preconditions and data dependencies, and after the check passes, sorting the candidate templates; expanding the sorted high-score candidate templates into executable high-score operator sequences through a task planning method based on a hierarchical task network and / or a planning domain description language, and outputting the high-score operator sequences as task chains; an agent scheduling layer for mapping each operator in the task chain to an MCP call node, and generating a call order according to the directed acyclic graph topology order of the task chain and resource availability to obtain an MCP call chain; calling the data processing tools and three-dimensional modeling tools according to the MCP call chain to generate a preliminary three-dimensional model, and then performing incremental correction and local optimization on the multi-temporal image through a dynamic feedback updating mechanism, and downloading the updated incremental three-dimensional data to the ground data center; a tool layer including an extensible modular tool set composed of a plurality of special-purpose processing modules.

[0013] The present application has the following technical effects: The present application introduces an AI Agent-driven automated task analysis and scheduling mechanism, which realizes autonomous processing from user demand to three-dimensional model generation in an on-orbit environment, effectively overcoming the strong dependence of existing ground modeling methods on human intervention; Through the cooperation of task planning and modular tool chains, data filtering and preliminary modeling are completed on the satellite, significantly reducing the bandwidth pressure and processing delay of raw data download; At the same time, with the help of dynamic feedback and incremental updating capability, real-time task demand can be responded to within minutes, significantly improving the timeliness and practicality of on-orbit three-dimensional modeling. BRIEF DESCRIPTION OF DRAWINGS

[0014] The above and other objects, features and advantages of the present application will become readily apparent from the following detailed description read in conjunction with the accompanying drawings. In the drawings, several embodiments of the application are illustrated by way of example and not limitation. Like or corresponding reference numerals primarily designate the same or corresponding parts throughout the several views.

[0015] Figure 1 is a step flow chart of a three-dimensional modeling system in orbit in an embodiment of the application; Figure 2 is a framework schematic diagram of a three-dimensional modeling method in orbit in an embodiment of the application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0017] The present application provides a systematic solution to the key problems existing in the prior art three-dimensional modeling technology in orbit. The traditional three-dimensional modeling method relies heavily on ground computing platform and manual intervention, and cannot meet the requirements of autonomy, real-time and resource constraints in orbit. Specifically, the prior art faces three fundamental problems: first, the modeling process needs the continuous participation of professionals in parameter adjustment and process control, which is essentially in conflict with the unattended operation mode in orbit; second, the amount of raw remote sensing data is extremely large, and complete download will cause extreme tension of communication bandwidth and significant increase of processing time delay; third, the response period from data acquisition to model generation is too long, which is difficult to meet the urgent demand of minutes-level response in emergency monitoring scenarios. The root cause of these problems lies in the fact that the existing technical architecture fails to effectively integrate intelligent task planning, modular tool scheduling and dynamic optimization capabilities into the in-orbit processing link.

[0018] The core of the present application is to build a complete three-dimensional modeling system in orbit, as shown in Figure 1 , which realizes autonomous processing of the whole process from task reception to model generation with AIAgent as the intelligent scheduling center. In the specific implementation process, the steps shown in Figure 2 are performed: S1, controlled semantic analysis is adopted to map the user input task demand into a task specification object MissionSpec; S2, based on the task specification object MissionSpec, mixed retrieval is performed on the task template library, and the retrieved task templates are subjected to graph consistency verification on the template preconditions and data dependencies, and after the verification is passed, the candidate templates are sorted; S3, expanding the sorted high-score candidate templates into executable high-score operator sequences by a task planning method based on hierarchical task network (HTN) and / or planning domain description language (PDDL), and outputting the high-score operator sequences as task chains; S4, mapping each operator in the task chains into a MCP invocation node, and generating an invocation order according to the directed acyclic graph (DAG) topology order of the task chains and resource availability to obtain a MCP invocation chain; S5, invoking the data processing tool and the three-dimensional modeling tool according to the MCP invocation chain to generate a preliminary three-dimensional model, and then performing incremental correction and local optimization on the multi-temporal image through a dynamic feedback updating mechanism, and downloading the updated incremental three-dimensional data to a ground data center.

[0019] In this embodiment, the user proposes a modeling requirement through natural language, for example, three-dimensional reconstruction of a certain urban area. The AIAgent first converts the user's natural language input into a structured task specification object MissionSpec through a controlled semantic parsing technology, constraint prompts and JSON Schema verification. The task specification object MissionSpec includes at least some of the fields of task intent, area, time window, imaging modality, quality target, time delay target, resource constraint, privacy constraint and expected output, and the user input is strictly mapped into unambiguous structured data through controlled semantic parsing. The accuracy of semantic mapping is ensured through constraint prompts and JSON Schema verification, avoiding ambiguity in natural language understanding.

[0020] Based on the task specification object MissionSpec, the system starts a hybrid retrieval mechanism to screen suitable processing flows from the task template library. This stage adopts a three-fold retrieval strategy in parallel. In the structure of the task specification object MissionSpec, at least the fields of task intent, imaging modality, preconditions and postconditions are included. When the AI Agent performs hybrid retrieval on the task template library in S2, it includes: The task template library is searched in parallel or in series based on a mission specification object MissionSpec using a triple search mechanism, and a task template is obtained by integrating the search results of the triple search mechanism, wherein the triple search mechanism includes a semantic vector index, a structured inverted index, and a task graph index; the semantic vector index performs vectorization representation on the description and label of the task template by using an embedding model to achieve semantic similarity matching; the structured inverted index performs rapid filtering on the task intent, imaging modality, precondition, and postcondition fields; and the task graph index constructs a searchable graph according to the preconditions and postconditions of the task template to achieve subgraph matching and dependency relationship verification. The semantic vector index is responsible for matching similar tasks from a semantic level, the structured inverted index performs rapid filtering on explicit task attributes, and the task graph index ensures the feasibility of the task chain by analyzing the dependency relationship between templates. This multi-dimensional search mechanism improves the accuracy and efficiency of template selection. The candidate templates obtained by searching also need to be subjected to graph consistency verification to verify their satisfiability in terms of input data availability, tool capability matching, and dependency relationship closure, and only the templates that pass the verification will enter the subsequent processing link.

[0021] In this embodiment, when performing the graph consistency verification of the searched task template in S2 regarding the template preconditions and data dependencies, the AI Agent includes: checking the satisfiability of the task template in terms of input data availability, tool capability matching, and dependency relationship closure, constructing the template preconditions and postconditions into a searchable graph using the task graph index, and verifying the dependency consistency of the data flow and the task chain through subgraph matching.

[0022] The knowledge enhancement layer RAG of the system provides domain knowledge support for the modeling process through retrieval enhancement generation technology, and the AI Agent also performs S6 in parallel when performing S2, S6 includes: retrieving knowledge from an extended knowledge base according to the mission specification object MissionSpec, and using the knowledge as a constraint condition for constraining the task template, data processing tool, and three-dimensional modeling tool, wherein the extended knowledge base includes at least one of city planning rules, historical three-dimensional models, and geographic information system data. When processing a three-dimensional modeling task of a city area, the system can retrieve relevant city planning rules, building specifications, historical three-dimensional models, and other information from the extended knowledge base, and integrate these knowledge as constraint conditions into the modeling process. For example, when encountering a region with missing image data, the system can make reasonable inferences based on prior knowledge such as building height restrictions in city planning, to improve the integrity and rationality of the generated model.

[0023] After the task chain is generated, the system maps each operator in the task chain to a corresponding MCP call node, and generates a specific call order according to the directed acyclic graph (DAG) topology of the task chain and the current resource availability. The construction of the MCP call chain adopts a standardized interface definition, and each specialized processing module explicitly specifies its input and output formats, adjustable parameter ranges, resource requirements, and other key information through a tool capability declaration file Manifest. This standardized design ensures seamless collaboration between different modules and provides good scalability for the system. During module calling, the system uses an idempotent key mechanism to ensure that repeated calls with the same input return consistent results, effectively improving the fault tolerance and execution stability of the system. Specifically, when outputting the high-score operator sequence as the task chain, a probe sub-task is inserted into the high-score operator sequence to quickly evaluate the on-orbit environmental state parameters, and the adaptive branch selection of the task chain is driven according to the evaluation results. Through constraint programming or heuristic scheduling methods, multi-objective scheduling optimization is performed to generate near-optimal task chains that meet multi-objective constraints by optimizing time delay, downlink bandwidth, energy consumption, and quality indicators during task chain generation.

[0024] Specifically, each specialized processing module declares its capabilities by providing a standardized tool capability declaration file Manifest, which includes Schema definitions for input and output, adjustable parameters and value ranges, resource profiles, telemetry fields, error models, and fallback lists, as well as version and compatibility policies. Specialized processing modules support idempotent operations, health checks, checkpoint outputs, and permission audits, and provide cost suggestion capabilities through a Dry run mode Among them, the Schema definition of input and output in the tool capability declaration file Manifest is used for interface alignment and automatic verification, the adjustable parameters and value ranges are used for parameter binding in the constraint programming stage, the resource profile is used for resource allocation in multi-objective scheduling optimization, the telemetry field is used for real-time monitoring of module execution status and triggering of quality gating, the error model and fallback list are used to define failure fallback paths and retry strategies, and the version and compatibility policies are used to ensure the compatibility of module calling; Idempotent operations are implemented by generating an idempotent key through hashing of input and parameters, and repeated calls with the same input return consistent results; Health checks periodically report module state indicators; Checkpoint output supports breakpoint resume of task execution; Dry run mode verifies the feasibility of input, parameters, resource estimates, and data paths without performing core calculations.

[0025] In the data processing and 3D modeling stages of this embodiment, the system is equipped with multiple dedicated processing modules to handle different types of remote sensing data. The data processing tools include multiple dedicated processing modules controlled by the MCP via a modular protocol. These dedicated processing modules include: MCP-SAR is used to process SAR remote sensing data and perform radiometric correction, geometric correction, and interferometric analysis. MCP-VIS is used to process visible light images, performing cloud removal, orthorectification, and registration. MCP-TIR is used to process thermal infrared images for nighttime modeling or to supplement surface thermal features; MCP-3D Matching is used for multi-view geometric sparse point cloud matching. MCP-3D Gaussian Splatting is used for efficient rendering and continuous 3D reconstruction. MCP-Data Overlay is used for multimodal overlay and data fusion. MCP - Data Compression, used for data compression and optimization; The AI ​​Agent invokes a dedicated processing module based on the MCP call chain and outputs the processed standardized data to the 3D modeling tool for use in the modeling process.

[0026] The MCP-SAR module specifically processes synthetic aperture radar data, performing operations such as radiometric correction, geometric correction, and interferometric analysis. The MCP-VIS module is responsible for cloud removal, orthorectification, and registration of visible light imagery. The MCP-TIR module focuses on the analysis of thermal infrared imagery, providing support for nighttime modeling or supplementing surface thermal features. These modules adopt a unified interface specification, ensuring both flexibility in invocation and consistency in data processing quality. The MCP-3D Matching module achieves sparse point cloud matching through multi-view geometry methods to construct a preliminary 3D structure; the MCP-3D Gaussian Splatting module is responsible for efficient rendering and continuous 3D reconstruction, enhancing the realism and visualization of the model. These modules were designed with full consideration of the computational resource limitations of the on-orbit environment, achieving a good balance between processing efficiency and model quality through algorithm optimization. In addition, the system also includes the MCP-Data Overlay module for multimodal data fusion and the MCP-DataCompression module for data compression and optimization, ensuring that the generated 3D model can be efficiently downloaded to the ground.

[0027] In addition, the system features a dynamic feedback update mechanism. After generating the initial 3D model, the system can perform incremental correction and local optimization based on multi-temporal imagery. This mechanism continuously monitors modeling quality indicators and automatically triggers corresponding compensation measures when model deviations or data gaps are detected. For example, if cloud cover is detected causing a decrease in local model accuracy, the system can autonomously schedule data from other temporal phases or modalities for supplementary processing, ensuring the stability of the output model's quality. Simultaneously, the incremental update mechanism ensures that only the changed parts need to be reprocessed and re-uploaded, significantly reducing data transmission volume.

[0028] Obviously, the embodiments described above are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0029] It should be understood that when the terms "first," "second," etc., are used in the claims, description, and drawings of this application, they are only used to distinguish different objects and not to describe a specific order. The terms "comprising" and "including" used in the description and claims of this application indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

Claims

1. An on-orbit 3D modeling method, executed by an AI Agent, characterized in that, When the AI ​​Agent executes the method, it includes the following steps: S1. Use controlled semantic parsing to map user-inputted task requirements into task specification objects; S2. Based on the task specification object, perform a mixed search on the task template library, and perform a graph consistency check on the retrieved task templates regarding template preconditions and data dependencies. After the check passes, sort the candidate templates. S3. Expand the sorted high-scoring candidate templates into an executable high-scoring operator sequence using a task planning method based on a hierarchical task network and / or a planning domain description language, and output the high-scoring operator sequence as a task chain. S4. Map each operator in the task chain to an MCP call node, and generate a call order based on the directed acyclic graph topology order and resource availability of the task chain to obtain the MCP call chain. S5. According to the MCP call chain, call the data processing tool and the 3D modeling tool to generate a preliminary 3D model. Then, perform incremental correction and local optimization on the multi-temporal images through a dynamic feedback update mechanism, and download the updated incremental 3D data to the ground data center.

2. The method according to claim 1, characterized in that, The structure of the task specification object includes at least the following fields: task intent, imaging modality, preconditions, and postconditions. When the AI ​​Agent performs a mixed search of the task template library in step S2, it includes: Based on the task specification object, a triple retrieval mechanism is used to perform parallel or serial retrieval on the task template library. The retrieval results of the triple retrieval mechanism are combined to obtain the task template. The triple retrieval mechanism includes a semantic vector index, a structured inverted index, and a task graph index. The semantic vector index uses an embedding model to vectorize the description and tags of the task template to achieve semantic similarity matching. The structured inverted index filters the fields of task intent, imaging modality, preconditions, and postconditions. The task graph index constructs a searchable graph of the task template according to the preconditions and postconditions to achieve subgraph matching and dependency verification.

3. The method according to claim 1, characterized in that, While executing S2, the AI ​​Agent also executes S6 in parallel, wherein S6 includes: Knowledge is retrieved from the extended knowledge base based on the task specification object, and the knowledge is used as a constraint to constrain the task template, the data processing tool, and the 3D modeling tool. The extended knowledge base includes at least one of urban planning rules, historical 3D models, and geographic information system data.

4. The method according to claim 1, characterized in that, S3 further includes: When the high-resolution operator sequence is output as the task chain, a probe subtask is inserted into the high-resolution operator sequence to evaluate the on-orbit environmental state parameters and drive the adaptive branch selection of the task chain based on the evaluation results. By using constraint programming or heuristic scheduling methods, when generating the task chain, multi-objective scheduling optimization is performed with the goal of optimizing latency, downlink bandwidth, energy consumption and quality indicators to generate a near-optimal task chain that satisfies multi-objective constraints.

5. The method according to claim 1, characterized in that, The data processing tool includes multiple dedicated processing modules controlled by a modular protocol for MCP calls. These dedicated processing modules include: MCP-SAR is used to process SAR remote sensing data and perform radiometric correction, geometric correction, and interferometric analysis. MCP-VIS is used to process visible light images, performing cloud removal, orthorectification, and registration. MCP-TIR is used to process thermal infrared images for nighttime modeling or to supplement surface thermal features; MCP-3D Matching is used for multi-view geometric sparse point cloud matching. MCP-3D Gaussian Splatting is used for efficient rendering and continuous 3D reconstruction. MCP-Data Overlay is used for multimodal overlay and data fusion. MCP - Data Compression, used for data compression and optimization; The AI ​​Agent invokes the dedicated processing module according to the MCP call chain and outputs the processed standardized data to the 3D modeling tool for use in the modeling process.

6. The method according to claim 5, characterized in that, Each of the dedicated processing modules declares its capabilities by providing a standardized tool capability declaration file. Each tool capability declaration file includes the schema definition of input and output, adjustable parameters and their value ranges, resource profiles, telemetry fields, error models and rollback lists, and declares the version and compatibility strategy. The dedicated processing module supports idempotent operations, health checks, checkpoint output, and permission auditing, and provides cost indication capabilities through Dry run mode. The input and output schema definitions in the tool capability declaration file are used for interface alignment and automatic verification; the adjustable parameters and value ranges are used for parameter binding in the constraint planning stage; the resource profile is used for resource allocation in multi-objective scheduling optimization; the telemetry field is used for real-time monitoring of module execution status and triggering quality gating; the error model and fallback list are used to define failure fallback paths and retry strategies; and the version and compatibility strategy are used to ensure the compatibility of module calls. The idempotent operation is achieved by hashing the input and parameters to generate an idempotent key, and repeated calls with the same input return consistent results; The health check periodically reports the module's status indicators; The checkpoint output supports resuming task execution from breakpoints; The Dry run mode verifies the feasibility of inputs, parameters, resource estimates, and data paths without performing core computations.

7. The method according to claim 1, characterized in that, S1 includes: converting user natural language input into a structured task specification object through constraint hints and JSONSchema validation, the task specification object containing at least some of the fields of task intent, region, time window, imaging modality, quality target, latency target, resource constraints, privacy constraints and expected output, and strictly mapping the user input into unambiguous structured data through controlled semantic parsing.

8. The method according to claim 1, characterized in that, When the AI ​​Agent performs graph consistency verification on the retrieved task template regarding template preconditions and data dependencies in step S2, it includes: Check the satisfiability of the task template in terms of input data availability, tool capability matching, and dependency closure. Use the task graph index to construct a searchable graph of the template preconditions and postconditions. Verify the dependency consistency of data flow and task chain through subgraph matching.

9. An on-orbit 3D modeling system, operating using the method described in any one of claims 1-8, characterized in that, include: The user requirements layer is used to map user-inputted task requirements into task specification objects using controlled semantic parsing. Based on the task specification object, a hybrid search is performed on the task template library, and the retrieved task templates are checked for consistency in terms of template preconditions and data dependencies. After the check passes, the candidate templates are sorted. The sorted high-scoring candidate templates are expanded into an executable sequence of high-scoring operators using a task planning method based on a hierarchical task network and / or a planning domain description language, and the high-scoring operator sequence is output as a task chain. The intelligent agent scheduling layer is used to map each operator in the task chain to an MCP calling node, and generate a calling order based on the topological order of the directed acyclic graph and resource availability of the task chain to obtain the MCP calling chain. The data processing tool and 3D modeling tool are invoked according to the MCP call chain to generate a preliminary 3D model. Then, the multi-temporal images are incrementally corrected and locally optimized through a dynamic feedback update mechanism, and the updated incremental 3D data is downloaded to the ground data center. The toolset layer includes an extensible modular toolset, which consists of multiple dedicated processing modules.

Images