A hierarchical collaborative multi-agent method and system for on-orbit processing of remote sensing data

Abstract

The present application belongs to the technical field of intelligent processing of remote sensing data, and particularly relates to a layered collaborative multi-agent method and system for on-orbit processing of remote sensing data; comprising: parsing a remote sensing task from a user instruction and decomposing the remote sensing task into multiple sub-targets; a management agent constructing a structured global state from the user instruction, multi-modal remote sensing data and the current sub-target; the management agent reasoning the structured global state to obtain routing decisions and control decisions and routing the sub-targets to corresponding expert agents; the expert agents executing the sub-targets and feeding back structured data to the management agent; the management agent verifying the structured data, and the management agent and the expert agent performing adaptive control according to the verification results, updating the structured global state and advancing to the next sub-target; the cycle is iterated until all sub-targets are processed, and the management agent generates a final answer; the present application effectively blocks the error propagation problem of long-range tasks and overcomes the workflow, knowledge and control mismatch problem.

Description

Technical Field

[0001] This invention belongs to the field of intelligent remote sensing data processing technology, specifically relating to a hierarchical collaborative multi-agent method and system for on-orbit processing of remote sensing data. Background Technology

[0002] Remote sensing (RS) has become the infrastructure for Earth observation, environmental monitoring, disaster assessment, and geospatial intelligence. With the rapid development of observation platforms and sensor modalities, modern remote sensing applications increasingly rely on large-scale, heterogeneous, and multimodal data, making analysis workflows increasingly complex. Many practical remote sensing applications involve far more than a single sensing action; they often require multi-step reasoning, tool calls, interpretation of intermediate results, and iterative decision-making across multiple processing stages, forming long processing chains with strong upstream and downstream coupling. For example, in flood mapping tasks, errors in the early spectral analysis stage (such as misclassifying cloud shadows as water bodies) directly affect downstream water index calculations, inundation area estimations, and the final report.

[0003] In recent years, multimodal large language models (MLLMs) have demonstrated powerful capabilities in remote sensing perception tasks (visual question answering, scene understanding, object detection, and semantic segmentation). Representative systems such as RSChatGPT explore remote sensing-specific multimodal alignment and instruction fine-tuning, LHRS-Bot implements a VGI-enhanced remote sensing multimodal language model, and SkyEyeGPT unifies instruction fine-tuning for remote sensing vision-language tasks. Furthermore, RS-Agent is evolving towards executable remote sensing agents, exploring multi-step tool use and geospatial reasoning; Earth-Agent and ThinkGeo provide benchmark evaluation frameworks for structured tool use and multi-step planning.

[0004] Furthermore, in terms of multi-agent collaboration, existing research has introduced multi-expert debate mechanisms to enhance reasoning diversity, and some works (such as GeoAgent and HTAM) utilize hierarchical coordinator-executor architectures for geospatial analysis. Remote sensing agents (such as RS-Agent) using a single large language model as the central controller, supplemented by retrieval augmentation generation (RAG) for tool selection, have improved tool integration and multi-round interaction capabilities. However, whether it's multi-expert debate, a general hierarchical coordinator-executor architecture, or a flat agent with a "single controller + tool library" structure, their role divisions mainly mimic general social divisions of labor or remain at a unified tool scheduling level. They do not organize expert divisions of labor based on the inherent stage dependencies of the remote sensing processing chain, nor do they provide progressive constraint verification of intermediate products. In terms of training optimization, recent studies such as ToolRL and AgentPO have shown that relying solely on sparse supervision at the final answer level is too coarse for optimizing the behavior of complex agents. Process-level fine-grained feedback is more conducive to improving tool invocation strategies, but currently there is no remote sensing agent scheme that unifies hierarchical coordination, process-level verification, and fine-grained reinforcement learning within the same framework.

[0005] While the above-mentioned solutions have made some progress in remote sensing capabilities and tool integration, they still face the following key challenges in ensuring reliable execution of long-link tasks:

[0006] (1) Workflow distribution: Remote sensing tasks generally follow a structured processing chain (spectral analysis → physical inversion → statistical analysis). Existing systems still rely on general intelligent agent design (such as ReAct, Plan-and-Solve, etc.), without explicit modeling of dependency relationships or long-link workflow organization, and the tool space lacks functional division according to processing stages.

[0007] (2) Knowledge mismatch: The large language model, which is the core of the intelligent agent, is usually not fully aligned with the professional knowledge in the field of remote sensing. It cannot truly understand the physical mechanisms (such as the physical constraints of surface temperature inversion), spectral constraints (such as the band selection rules of vegetation index) or professional data processing procedures unique to remote sensing.

[0008] (3) Mismatch control: Existing frameworks mainly emphasize local action selection or final answer evaluation, with limited support for intermediate state verification and execution failure recovery. The lack of a stepwise verification mechanism leads to the continuous propagation and accumulation of early errors in the workflow, which seriously affects the validity of downstream analysis results.

[0009] In summary, the remote sensing field urgently needs a remote sensing data processing method that can organize tasks according to the inherent structure of the remote sensing workflow, has the ability to align domain knowledge, and implement step-by-step verification and adaptive error correction during execution, in order to overcome the bottleneck of existing remote sensing agents in terms of reliability in long-link task execution. Summary of the Invention

[0010] To address the shortcomings of existing technologies, particularly the limitations of current remote sensing agents in adapting to the constraints of computing power, storage, and model deployment resources in the on-orbit environment of satellites, this invention proposes a hierarchical collaborative multi-agent method and system for on-orbit processing of remote sensing data. This method includes:

[0011] S1: Acquire user commands and multimodal remote sensing data, parse the remote sensing task from the user commands, and decompose the remote sensing task into multiple sub-targets;

[0012] S2: The management agent constructs a structured global state based on user commands, multimodal remote sensing data, and the current sub-target;

[0013] S3: The management agent reasones about the structured global state to obtain routing and control decisions; based on the routing and control decisions, it routes sub-goals to the corresponding expert agents.

[0014] S4: The expert agent executes sub-goals and feeds back structured data to the management agent;

[0015] S5: The management agent verifies the structured data and obtains the verification results;

[0016] S6: The management agent and expert agent execute adaptive control based on the verification results, complete the processing of the current sub-objective, update the structured global state, and advance to the next sub-objective;

[0017] S7: Repeat steps S3 to S6 until all sub-goals have been processed and the management agent generates the final answer.

[0018] Preferably, in step S2, the structured global state is represented as follows:

[0019]

[0020] in, This represents the structured global state of the current step. Indicates user commands, This represents multimodal remote sensing input data. Indicates the current sub-target. This represents the structured global memory of the current step. This represents the set of dynamic constraints for the current step.

[0021] Preferably, in step S5, the process of the management agent validating the structured data includes:

[0022] Three complementary checks are performed: model validity check, sub-objective consistency check, and constraint satisfaction check. The constraint satisfaction check specifically includes: verifying the physical consistency of the inversion relationship, the spectral correctness of band usage and index calculation, and the statistical rationality of the aggregation model and output distribution.

[0023] Output "pass" only if all three conditions are met: pattern validity, sub-objective consistency, and constraint satisfaction. Output "uncertain" if none of the three conditions are violated but the confidence level is insufficient. Output "fail" if any constraint is violated.

[0024] Preferably, step S6 specifically includes:

[0025] If the verification result is "fail", attribution-based replanning is triggered. Attribution-based replanning includes: the management agent locates the failure level; when the failure is attributed to the parameter level, the expert agent performs tool parameter repair and retryes in place; when the failure is attributed to the tool or routing level, an alternative route is executed; when the failure is attributed to the sub-goal level, the sub-goal is further decomposed or the sub-goal is adjusted, and a new route is re-executed; all intermediate results of the execution are written into the structured global memory in the structured global state.

[0026] If the verification result is uncertain, cross-checking is triggered. Cross-checking includes: the management agent reassigning the current sub-objective along with its output structured data to another expert with overlapping verification capabilities, or having the same expert independently recalculate the key quantities under alternative inference paths and alternative tools, and comparing the results of the two paths. If they are consistent within the tolerance, the result is upgraded to pass; otherwise, it is downgraded to fail. All intermediate results of the execution are written into the structured global memory in the structured global state.

[0027] If the verification result is pass, the structured data of the current sub-target is written into the structured global memory in the structured global state and the process is advanced to the next sub-target.

[0028] Preferably, the expert agents include surface analysis agents, physical inversion agents, and statistical analysis agents. The surface analysis agent is used to process perceived and analyzed sub-targets, including target detection, semantic segmentation, region of interest and mask generation, and spectral index analysis. The physical inversion agent is used to process physical inversion and quantitative remote sensing product generation sub-targets, including surface temperature inversion, atmospheric correction, and radiometric calibration. The statistical analysis agent is used to process data aggregation and cleaning, statistical testing and correlation analysis, report-level index calculation, and comparative sub-targets. The three types of expert agents integrate 120 remote sensing professional tools, with backend support provided by three MCP tool servers.

[0029] Preferably, the management agent and the expert agent share the same basic large language model and parameter adapter, and distinguish between management layer roles and expert layer roles through role prompts and decoding configurations; its training process includes hierarchical reinforcement learning with expert-to-workflow alignment fine-tuning and validation guidance.

[0030] The expert-to-workflow alignment fine-tuning comprises two sequentially executed supervised fine-tuning phases: the first phase is the remote sensing expert knowledge alignment phase, which constructs a remote sensing professional knowledge system and generates knowledge supervision samples based on this system. This is used to supervise and fine-tune a shared basic large language model, resulting in an expert model with remote sensing domain knowledge. The second phase is the remote sensing workflow alignment phase, which uses the expert model as an initialization model. Workflow supervision samples are constructed based on real remote sensing task pipelines or multi-agent demonstration logs. These samples include workflow selection samples and workflow sorting samples. A structured mapping is established between the workflow supervision samples and the corresponding knowledge branches in the remote sensing professional knowledge system. The initialization model is then further supervised and fine-tuned, enabling the model to learn the mapping relationship from remote sensing task semantics to executable tool processes, resulting in a shared strategy model.

[0031] The validation-guided hierarchical reinforcement learning involves using the GRPO algorithm to perform validation-guided hierarchical reinforcement learning on the shared policy model, thereby training the management agent and the expert agent.

[0032] Furthermore, the process of using the GRPO algorithm to validate and guide the hierarchical reinforcement learning of the shared policy model includes:

[0033] For the same remote sensing task, multiple complete execution trajectories are sampled based on the shared policy model. Each execution trajectory includes sub-objective decomposition, expert routing, and termination control decisions generated by the management agent, as well as tool selection, parameter construction, and tool execution results generated by the expert agent. Based on the step-level verification results in each execution trajectory, the management layer trajectory-level reward and the expert layer step-level reward are calculated respectively. The management layer trajectory-level reward is used to evaluate the global coordination quality of the entire execution trajectory, and the expert layer step-level reward is used to evaluate the quality of single-step tool invocation. Based on the management layer trajectory-level reward and the expert layer step-level reward, the intra-group relative advantage of the management layer and the expert layer is estimated to obtain the intra-group relative advantage of the management layer and the intra-group relative advantage of the expert layer.

[0034] The management strategy loss is calculated based on the relative advantage within the management group, and the expert strategy loss is calculated based on the relative advantage within the expert group. The management strategy loss and the expert strategy loss are then weighted and fused at the loss level to obtain a mixed loss. The shared strategy model is updated according to the mixed loss, enabling the management agent to obtain trajectory-level global coordination capabilities and enabling the expert agent to obtain step-level tool execution capabilities.

[0035] Furthermore, the formulas for calculating the trajectory-level reward for management and the step-level reward for expert levels are expressed as follows:

[0036]

[0037]

[0038] in, This indicates management-level trajectory rewards. Indicates routing reward, Indicates a binary reward. Indicates boundary reward, , and These represent route weights, binary weights, and boundary weights, respectively. , and These represent the expert-level step-level rewards for tool invocation correctness at levels one, two, and three, respectively. This indicates the original execution feedback. This represents the normalized parameter quality score;

[0039] The formulas for obtaining the relative advantage within the management group and the relative advantage within the expert group are expressed as follows:

[0040]

[0041]

[0042] in, Indicates the number of rollouts within the same group; Indicates the first The relative advantage of the management team within the rollout; Indicates the first The management trajectory-level rewards for the rollout; Indicates the first The rollout in the 1st The relative advantages within a group in the steps of calling up expert-level tools; Indicates the first The rollout in the 1st Expert-level step-by-step rewards for each tool invocation step; and These represent the mean and standard deviation of the trajectory-level rewards for the current management group, respectively. This represents a small constant to prevent the denominator from being zero; and These represent the current expert groupings. The mean and standard deviation of the rewards for each tool call step. This represents the stability threshold for low-variance groups.

[0043] Furthermore, the formula for calculating the mixing loss is expressed as:

[0044]

[0045]

[0046]

[0047] in, This indicates a strategic loss by management; This indicates a slight loss to the expert group; This represents the weighted average of management strategy losses and expert strategy losses; This indicates the parameters of the shared strategy model that are currently pending update; Indicates the rollout number for management. and generation location Seeking expectations; Indicates the rollout number for the expert level. Tool Call Steps Index and generation location Seeking expectations; Indicates the management's first The rollout in the 1st The probability ratio of the strategies at each generation location; Indicates the expert level The rollout in the 1st The first tool call step, the first The probability ratio of the strategies at each generation location; This represents the truncation function; Indicates the cutoff coefficient; This indicates the weight of management loss in the mixed loss.

[0048] A hierarchical collaborative multi-agent system for on-orbit processing of remote sensing data is disclosed. This system implements the aforementioned method and is deployed in a satellite on-orbit remote sensing data processing environment or an on-board edge computing environment. It includes a task input module, a management agent module, an expert agent module, a tool service module, a structured global state module, and a result output module. The task input module sends user commands and multimodal remote sensing data to the management agent module. The management agent module parses the remote sensing task and generates sub-targets, making routing and control decisions based on the structured global state. The expert agent module receives the sub-targets and invokes the tool service module to execute remote sensing processing tools. The tool service module returns structured results. The management agent module performs step-by-step verification and adaptive control on the structured results and writes intermediate results, verification results, and recovered trajectories into the structured global state module. When all sub-targets are completed, the result output module outputs the final remote sensing data processing result.

[0049] The beneficial effects of this invention are as follows:

[0050] 1. Architecture Level: This invention makes the hierarchical division of labor strictly isomorphic to the remote sensing processing chain and reuses a single backbone model. The system internalizes the stage dependency relationship of "spectral analysis → physical inversion → statistical analysis" into structural priors. While significantly reducing the dependence on manual orchestration and the overhead of multi-model collaboration, it realizes the automated decomposition and low-resource deployment of long-link remote sensing tasks in remote sensing data processing scenarios with limited computing and storage resources, such as satellite on-orbit processing and on-board edge computing.

[0051] 2. Knowledge Level: This invention utilizes a two-stage fine-tuning process that aligns the knowledge system with the structured sequence of executable tools. The system not only incorporates knowledge of remote sensing physical mechanisms, spectral rules, and professional processes, but also establishes a traceable mapping from domain knowledge to tool call sequences. This makes the translation of natural language instructions into standardized tool call sequences traceable, fundamentally overcoming the knowledge mismatch of general models in the field of remote sensing.

[0052] 3. Control level: The present invention adopts memory-driven and constraint-bound stepwise verification and three-state control logic, which enables the system to determine the rationality of the result with remote sensing semantics at the moment of tool invocation or intermediate product generation, and to perform parameter repair, alternative routing or sub-target re-decomposition according to failure attribution, thereby blocking the error in place before it spreads across stages, significantly improving the reliability of long-link task execution.

[0053] 4. Training Level: This invention uses validation-guided hierarchical reinforcement learning with a decoupled design of trajectory-level coordinated rewards and step-level hierarchical gating rewards. Combined with role-aware group advantage estimation and hybrid loss joint optimization, it resolves the credit allocation problem of mutual interference between global routing and local tool execution under sparse long-link supervision, realizes stable collaborative training of the two-layer strategy, and significantly improves the decision quality of complex remote sensing tasks. Attached Figure Description

[0054] Figure 1 This is a diagram illustrating the overall architecture of the multi-agent hierarchical collaborative remote sensing data processing method in this invention.

[0055] Figure 2 This is a flowchart of the expert-to-workflow alignment fine-tuning data construction and training process in this invention. Detailed Implementation

[0056] 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.

[0057] This invention proposes a hierarchical collaborative multi-agent method and system for on-orbit processing of remote sensing data. It aims to address the issues of workload mismatch, knowledge mismatch, and control mismatch in the execution of long-link tasks by existing remote sensing agents, and further adapt to the application requirements of limited computing, storage, and model deployment resources in the satellite on-orbit environment. This invention uses a large language model as its core to construct a multi-agent architecture with hierarchical planning and decision-making capabilities, as well as the execution and verification control capabilities of domain-specific tools. This enables intelligent processing of the entire process, from user natural language input and multimodal remote sensing data to structured tool call chains and final analysis results.

[0058] Within the task cycle of the intelligent agent, the user submits a request, i.e., inputs a user command, such as "Analyze the land use change trend of a certain region from 2020 to 2023 and assess the impact of urban expansion on vegetation cover," and provides multi-temporal remote sensing imagery. The remote sensing hierarchical collaborative intelligent agent first decomposes the task into multiple sub-objectives (multi-temporal land classification → vegetation index calculation → change detection and statistics → trend assessment), and then sequentially routes each sub-objective to the corresponding expert intelligent agent for execution. After each step is completed, the management layer verifies the intermediate results, confirming the consistency of the task execution trajectory and remote sensing-specific constraints such as physical consistency, spectral correctness, and statistical reasonableness before proceeding to the next step. If the verification fails, replanning or repair is triggered, ultimately outputting a complete analysis report. Figure 1 As shown, the method includes the following:

[0059] S1: Acquire user commands and multimodal remote sensing data, parse the remote sensing task from the user commands, and decompose the remote sensing task into multiple sub-targets.

[0060] A two-layer collaborative architecture is adopted, including a management layer and an expert layer. All roles share the same large language model backbone and LoRA adapter, and are distinguished only by role-specific system prompts and decoding settings. This allows a single model instance to support multiple expert roles, avoiding the resource overhead of deploying multiple models in parallel.

[0061] The management layer is instantiated by an orchestration agent, which serves as the core controller of the entire agent system. Its main function is to parse user instructions, extract task intent, obtain the user's remote sensing task, transform the remote sensing task requirements into an executable sequence of sub-objectives, and implement verification and adaptive control during subsequent execution.

[0062] S2: The management agent constructs a structured global state based on user instructions, multimodal remote sensing data, and the current sub-target.

[0063] The management agent maintains a structured global state:

[0064]

[0065] in, This represents the structured global state of the current step. Indicates user commands, This represents multimodal remote sensing input data. Indicates the current sub-target. This represents the structured global memory of the current step. This represents the set of dynamic constraints for the current step.

[0066] The dynamic constraint set is initialized from domain prior constraints parsed from user commands and multimodal input metadata (such as sensor type, band configuration, geographical range, unit system, etc.). During execution, it is incrementally updated and tightened with each step's intermediate evidence and verification conclusions (e.g., once a band is occupied, the reasonable range of a physical quantity is further narrowed by the previous step's result). Its constraint types cover four categories: physical consistency, spectral bands, statistical rationality, and interface patterns. Unlike traditional execution logs, Organized into a queryable, structured memory, it supports historically-aware decision-making in subsequent steps.

[0067] S3: The management agent reasones about the structured global state to obtain routing and control decisions; based on the routing and control decisions, it routes the sub-goals to the corresponding expert agents.

[0068] The management agent reasones about the structured global state and outputs a routing decision. (Selecting the expert agent to execute the current sub-objective) and control decision (Continue execution, trigger verification, reschedule, or terminate the task). The output decision is formalized as follows: ,in To arrange role prompts for intelligent agents, For sharing strategy.

[0069] Unlike general-purpose intelligent agents that progressively select the tool space, the routing in this invention is based on the stage affiliation of the sub-target in the "spectral analysis → physical inversion → statistical analysis" processing chain, combined with the products of completed steps in the global memory and the constraint state for joint determination. This ensures that the routing structure remains isomorphic to the remote sensing processing chain, thereby guaranteeing that upstream products (such as masks and inversion products) are correctly delivered to the corresponding experts in the downstream stages. Specifically: the functional attributes of the current sub-target in the remote sensing processing chain are determined based on the structured global state. If the sub-target involves surface analysis operations such as target detection, semantic segmentation, or spectral index calculation, it is routed to the surface analysis intelligent agent; if it involves physical inversion operations such as surface temperature inversion, atmospheric correction, or radiometric calibration, it is routed to the physical inversion intelligent agent; if it involves statistical analysis operations such as document retrieval, statistical testing, spatial aggregation, trend calculation, or report generation, it is routed to the statistical analysis intelligent agent.

[0070] Based on routing and control decisions, sub-goals are routed to the corresponding expert agents.

[0071] S4: The expert agent executes the sub-goals and feeds back structured data to the management agent.

[0072] The expert layer is organized into three function-oriented expert families, or three types of expert intelligent agents, and is divided according to the typical progress of remote sensing workflow (spectral analysis → physical inversion → statistical analysis):

[0073] 1. Surface Resolution Agent: Corresponding to the "Spectral Resolution" stage of the remote sensing processing chain, it handles perceived and resolved sub-targets, including target detection, semantic segmentation, region of interest (ROI) and mask generation, and spectral index resolution (such as NDVI and NDWI calculation). Its tool domains correspond one-to-one with this stage, covering visual resolution, spatial region processing, and index calculation primitives, containing a total of 42 tools.

[0074] 2. Physical Inversion Agent: Corresponding to the "physical inversion" stage of the remote sensing processing chain, it processes physical inversion and generates sub-targets from quantitative remote sensing products, such as land surface temperature (LST) inversion, atmospheric correction, and radiometric calibration. Its tool domain covers physical inversion operators, calibration and normalization routines, and quality control checks, comprising a total of 18 tools.

[0075] 3. Statistical Analysis Agent: Corresponding to the "Statistical Analysis" stage of the remote sensing processing chain, it handles sub-objectives such as data aggregation and cleaning, statistical testing and correlation analysis, and report-level indicator calculation and comparison. Its tool domain covers tabular and time-series processing, statistical inference, and controlled file I / O, containing a total of 60 tools.

[0076] The three types of expert agents integrate 120 remote sensing professional tools, with backend support (standardized interface) provided through three MCP (Model Context Protocol) tool servers. Each tool defines a unified input / output pattern (JSON Schema), including tool name, parameter type, required / optional tags, and return value structure. Tools are categorized by expert family: the surface analysis agent manages 42 tools such as detection, segmentation, and index calculation; the physical inversion agent manages 18 tools such as physical inversion, calibration, and quality control; and the statistical analysis agent manages 60 tools such as statistical testing, data aggregation, and report generation.

[0077] Each expert agent receives sub-goals from the orchestrating agent. Based on the current global state (in For user commands, Input data for multimodal remote sensing, For the current sub-goal, For structured global memory, (As a dynamic constraint set), it selects appropriate tools from the MCP tool servers under its jurisdiction and constructs the invocation parameters to generate tool calls. The MCP server parses the call request, executes the corresponding tool, and returns a structured result containing execution status, output data, and diagnostic information. The system then verifies the parameters using the orchestration agent. If an invalid parameter is encountered, the system allows up to two automatic retries.

[0078] S5: The management agent verifies the structured data and obtains the verification results.

[0079] After each tool execution step is completed, the management agent receives the execution result, i.e., the returned structured data (including tool call records, output results, and execution evidence), and retrieves relevant global memory. Execute verification and judgment ,in Output for the current step. .

[0080] The validator performs three complementary checks:

[0081] The schema validity check verifies whether the output format and type conform to the JSON Schema defined by the tool, ensuring that the output meets the requirements for required fields and data types.

[0082] Sub-goal consistency check checks whether the output meets the semantic expectations of the current sub-goal, ensuring that the output advances or completes the current sub-goal;

[0083] The constraint satisfaction check examines whether constraints in areas such as numerical range, physical units, and spectral band selection are met, ensuring that the results remain physically, spectrally, and statistically reasonable.

[0084] In remote sensing workflows, constraint satisfaction checks specifically include: verifying the physical consistency of inversion relationships (e.g., whether surface temperature values ​​are within a reasonable range), the spectral correctness of band usage and index calculations (e.g., whether the correct near-infrared and red bands were used in NDVI calculations), and the statistical rationality of aggregation patterns and output distributions.

[0085] Formalistically, the verification result can be viewed as the conjunction of three checks: pass is output only if model validity, sub-objective consistency, and constraint satisfaction are all true; uncertain is output if none of the three are violated but confidence is insufficient (e.g., a key physical quantity approaches the reasonable range boundary); and fail is output if any hard constraint is violated (e.g., the inverted physical quantity goes out of bounds, or the exponent calculation uses the wrong band). This three-state judgment allows the verification to distinguish between "clear errors" and "questionable errors," and provides a basis for subsequent differentiated recovery strategies.

[0086] S6: The management agent and expert agent execute adaptive control based on the verification results, complete the processing of the current sub-objective, update the structured global state, and advance to the next sub-objective.

[0087] Based on the verification results The orchestration agent performs adaptive control, specifically:

[0088] If the verification result is "fail", attribution-based replanning is triggered. Attribution-based replanning includes: the management agent locates the failure level; when the failure is attributed to the parameter level (such as band path error or parameter out of bounds), the expert agent performs tool parameter repair and retryes in place; when the failure is attributed to the tool or routing level, an alternative route is executed.

[0089] When the issue is attributed to the sub-goal level (unreachable or overly coarse-defined sub-goal), the sub-goal is re-decomposed or adjusted, and a new route is re-executed. Specifically, when the orchestration agent verifies that the current sub-goal itself cannot be directly executed by the expert agent, or that the current sub-goal is too coarse-grained, lacks necessary preconditions, or has unclear constraints, the orchestration agent, based on structured global memory state and failure diagnostic information, re-plans the sub-goal. This includes: breaking down the overly coarse sub-goal into multiple finer-grained executable sub-goals, supplementing missing preconditions, adjusting the execution order of sub-goals, and tightening or clarifying the constraints of sub-goals. For example, if the atomic goal is to analyze vegetation change, but verification shows that the goal is too coarse, it can be decomposed into multiple executable sub-goals such as "acquiring multi-temporal images, calculating NDVI, statistically analyzing regional averages, calculating change trends, and generating analysis conclusions."

[0090] If the verification result is uncertain, cross-checking is triggered. The orchestration agent schedules another expert agent to cross-check the result of this step to confirm its reliability. Specifically, the management agent reassigns the current sub-objective along with its output structured data to another expert with overlapping verification capabilities, or has the same expert independently recalculate the key quantities under alternative inference paths and alternative tools, and compares the results of the two paths. If they are consistent within the tolerance, the result is upgraded to pass; otherwise, it is downgraded to fail.

[0091] Whether the verification result is "fail" or "uncertain", all intermediate results of the execution (including failure records and recovery traces) are written to the structured global memory in the structured global state. This is achieved by writing all intermediate results to... This enables subsequent decisions to have historical awareness rather than relying solely on partial information from the current step.

[0092] If the verification result is pass, the structured data of the current sub-target is written into the structured global memory in the structured global state and the process is advanced to the next sub-target.

[0093] S7: Repeat steps S3 to S6 until all sub-goals have been processed and the management agent generates the final answer.

[0094] Steps S3-S6 are performed iteratively until all sub-goals are completed; the management agent generates the final answer based on the execution trajectory and intermediate results in the structured global memory, and the system returns the final answer to the user through the terminal interface.

[0095] The following example illustrates the implementation process of the method of the present invention:

[0096] The user inputs the command "Using Landsat 8 thermal infrared and reflectance data of New York City in 2022, estimate the land surface temperature (LST) using a single-channel method based on NDVI and thermal band 10, then calculate the average LST for summer and autumn respectively, and calculate the difference between their means to assess seasonal temperature changes," while specifying the path of the corresponding image data. After receiving the command, the orchestration agent performs semantic parsing and decomposes the task into four phased sub-objectives: acquiring data files, land surface temperature inversion, seasonal mean statistics, and difference analysis.

[0097] The first step involves the orchestration agent routing the "acquire data files" sub-objective to the statistical analysis agent. The statistical analysis agent autonomously selects the `get_filelist` tool based on the sub-objective's semantics, inputs the data directory path, and returns a list of all Landsat 8 image files in that directory (including multiple .tif files such as thermal and near-infrared bands). The verification module confirms that the output file format and quantity meet expectations and then determines the pass.

[0098] The second step involves the orchestration agent routing the "land surface temperature inversion" sub-objective to the physical inversion agent. Based on the sub-objective requirements and the file list obtained in the previous step, the physical inversion agent autonomously selects the `lst_single_channel` tool, inputting parameters such as the thermal band path, near-infrared band path, red band path, and output path. It then performs LST inversion based on a single-channel algorithm, generating a pixel-by-pixel land surface temperature raster file. The verification module checks whether the LST output value range is within a reasonable physical range (e.g., 200K–350K), and if confirmed, determines it to pass.

[0099] The third step involves the orchestration agent routing the "seasonal average statistics" sub-objective to the statistical analysis agent. The statistical analysis agent autonomously selects the `calc_batch_image_mean_mean` tool, passing the pixel-by-pixel land surface temperature raster files to the summer and autumn LST raster file lists respectively. It calculates the pixel-by-pixel mean of the LST images for each time phase and then takes the global mean, returning a summer average LST of 298.85K and an autumn average LST of 286.43K. The verification module confirms that the tool call is correct and that both means are within a reasonable temperature range before determining a pass.

[0100] In the fourth step, the orchestration agent routes the "difference analysis" sub-objective to the statistical analysis agent. The statistical analysis agent autonomously selects the difference tool, inputs the two means obtained in the previous step, and calculates the seasonal mean difference as 12.42K. The validation module confirms the difference result is reasonable and then determines it to pass.

[0101] After all sub-objectives have been executed, the orchestration agent summarizes all execution evidence and generates the final answer: "The mean difference between the average LST in summer and autumn is 12.42K.", along with the complete execution trajectory, intermediate output data, and execution log, which is then returned to the user via the terminal.

[0102] In some preferred embodiments of the present invention, the intelligent agent system is based on a hierarchical multi-agent architecture, combined with two training phases: expert-to-workflow alignment fine-tuning and verification-guided hierarchical reinforcement learning. These phases jointly drive the operation of six functional stages: task understanding, sub-objective decomposition, expert routing, tool execution, intermediate verification, and result output, achieving continuous mapping from user semantics to remote sensing processing instructions. The management agent and expert agent of the present invention share the same basic large language model and parameter adapter, and distinguish between management layer roles and expert layer roles through role prompts and decoding configurations. The two-layer collaborative architecture of the present invention completes joint training through two training phases: expert-to-workflow alignment fine-tuning and verification-guided hierarchical reinforcement learning. The specific process includes:

[0103] 1. Experts fine-tune workflow alignment.

[0104] like Figure 2 As shown, to align the general large language model with the workflow-driven characteristics of remote sensing tasks, this invention proposes a two-stage supervised fine-tuning strategy that executes sequentially. This strategy does not isolate domain knowledge injection and workflow training; instead, it first obtains an expert model with remote sensing expertise through basic knowledge alignment. Then, using this expert model as initialization, workflow alignment continues under the same knowledge classification system constraints, thereby establishing an explicit, structured mapping between "domain knowledge" and "executable tool sequences."

[0105] Phase 1 (Remote Sensing Expert Knowledge Alignment Phase): Constructing a structured remote sensing knowledge system and injecting expert knowledge.

[0106] A remote sensing professional knowledge system was constructed, and knowledge supervision samples were generated based on this system. Supervised fine-tuning of a shared basic large language model was then performed to obtain an expert model possessing knowledge in the remote sensing domain. Specifically, the knowledge space was organized into a three-level classification system, including 3 L1 topics (remote sensing fundamentals, digital image processing, and intelligent interpretation and applications), 11 L2 modules, and 28 L3 domains. Knowledge content was systematically collected from four authoritative remote sensing textbooks and processed through a five-stage pipeline: ① Content extraction: extracting original text and diagram descriptions from the textbooks; ② Normalization: standardizing terminology, symbols, and formats; ③ Semantic segmentation: dividing knowledge blocks into coherent paragraphs according to knowledge units; ④ Style rewriting: transforming textbook descriptions into expressions oriented towards intelligent agent reasoning; ⑤ Metadata annotation: labeling each knowledge block with its L1-L3 classification and source. Finally, a remote sensing professional knowledge corpus was constructed, containing 1,583 semantically coherent knowledge blocks, and further generating 5,687 knowledge center training samples. By using the training samples from the aforementioned knowledge center, the shared base model is fine-tuned under supervision to obtain an expert model with knowledge of the remote sensing domain.

[0107] Phase Two (Remote Sensing Workflow Alignment Phase): Based on the expert model obtained in Phase One, the workflow is further fine-tuned and supervised to align remote sensing knowledge with executable remote sensing processing procedures.

[0108] Specifically, the expert model obtained in Phase 1 was used as the initial model, and the L1-L3 remote sensing knowledge classification system established in Phase 1 was used as the basis for organizing the workflow samples. The workflow supervision and fine-tuning data were sourced from the multi-agent execution logs of real remote sensing application pipelines and were reviewed and verified by two remote sensing experts. The data was organized into two complementary formats: workflow selection questions (684 samples, requiring the model to select the correct processing flow from candidate solutions) and workflow ranking questions (309 samples, requiring the model to recover the correct tool usage sequence), totaling 993 samples, covering 14 remote sensing task families.

[0109] A structured mapping is established between the supervised workflow samples and the corresponding knowledge branches in the remote sensing professional knowledge system. The initial model is then further fine-tuned under supervision, enabling the model to learn the mapping relationship from remote sensing task semantics to executable tool processes, resulting in a shared policy model. The key to Phase Two is that each workflow sample is explicitly anchored to the corresponding branch of the knowledge classification system from Phase One, further binding the remote sensing domain knowledge injected in Phase One to specific tool processes in Phase Two. Thus, the model learns not isolated process templates, but rather establishes a correspondence between the remote sensing knowledge basis of "why it is handled this way" and the executable tool sequence of "how it is handled," thereby achieving a mapping from natural language remote sensing tasks to professional processing flows and tool invocation sequences.

[0110] 2. Validation-guided hierarchical reinforcement learning

[0111] To further improve workflow-level decision-making quality beyond supervised fine-tuning, this invention proposes a verification-guided hierarchical reinforcement learning strategy. Its core idea is to separate and optimize trajectory-level global coordination from step-level tool execution, and to design a role-aware grouping optimization strategy. This ensures precise alignment between credit allocation and the two-layer collaborative structure of the management and expert layers, preventing the confusion between global routing and local tool execution signals under sparse long-link supervision. In implementation, the GRPO (Group Relative Policy Optimization) algorithm is used to perform verification-guided hierarchical reinforcement learning on the shared policy model. Specifically:

[0112] (1) Layered Rewards: To avoid signal confusion and credit allocation mismatch, independent reward signals are defined for two layers. For the same remote sensing task, multiple complete execution trajectories are sampled based on the shared strategy model. Each execution trajectory includes sub-target decomposition, expert routing and termination control decisions generated by the management agent, as well as tool selection, parameter construction and tool execution results generated by the expert agent. The management layer trajectory-level reward and the expert layer step-level reward are calculated based on the step-level verification results (verification results of each step) in each execution trajectory. The management layer trajectory-level reward is used to evaluate the global coordination quality of the entire execution trajectory, and the expert layer step-level reward is used to evaluate the quality of single-step tool invocation.

[0113] Calculate the trajectory-level reward for management:

[0114]

[0115] in, This represents the routing reward, which measures the average stepwise correctness of the predicted route sequence relative to the reference route sequence at the alignment position. The reward is binary (+1 for a correct answer / -1 for an incorrect answer). This indicates a boundary reward, with an unnecessaryly long trajectory for which the penalty exceeds the tolerance boundary. , and These represent route weight, binary weight, and boundary weight, respectively. The default weight is... .

[0116] Calculate expert-level step-by-step rewards:

[0117] A domain-aware hierarchical gating mechanism is adopted to classify each tool call into three levels based on correctness. , and , respectively, represent the expert-level rewards when the correctness of the tool call is Tier 1 (illusionary out-of-domain tool, i.e., the prediction tool does not belong to the current expert's registered tool domain), Tier 2 (incorrect but in-domain tool, i.e., the prediction tool belongs to the current expert's domain but does not match the reference tool), and Tier 3 (correct tool).

[0118]

[0119]

[0120]

[0121] in, , and These represent the expert-level step-level rewards for tool invocation correctness at levels one, two, and three, respectively. This indicates the original execution feedback (failure / skip / success). ; This represents the quality score of the normalized parameter. This represents the quality score of the normalized parameters and is calculated only when the tool selection is correct, i.e., Tier 3. It is used to measure whether the tool call parameters generated by the expert agent are complete, correctly formatted, and whether the values ​​or paths are consistent with the reference parameters.

[0122] In some preferred embodiments of the present invention:

[0123]

[0124] in, This represents parameter key coverage, used to measure whether the predicted parameters include the key parameters required for reference tool invocation.

[0125]

[0126] This represents the parameter value matching degree, used to measure whether the predicted parameter value is consistent with or approximately consistent with the reference parameter value:

[0127]

[0128] in, This represents the set of parameter keys in the parameters called by the prediction tool. This represents the set of parameter keys in the reference tool call parameters; and Representing parameters respectively Predicted and reference values; This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise. For numerical parameters, approximate consistency can be determined by a preset tolerance; for discrete parameters such as file paths, band names, and tool options, consistency can be determined by whether the strings or normalized identifiers are consistent.

[0129] (2) Grouping advantage estimation of role perception

[0130] The management team groups all rollouts for the same task together. A rollout refers to a complete execution trajectory generated by the shared policy model for the same remote sensing task. This includes the sub-objective decomposition, expert routing and termination control decisions of the management agent, as well as the tool selection, parameter construction, tool execution results and intermediate verification records of the expert agent.

[0131] The relative advantage within the management group is calculated based on management rewards. This relative advantage is obtained by comparing different execution paths within the same task, using multiple complete execution paths as the grouping unit. Similarly, the relative advantage within the expert group is obtained by comparing tool calls at the same execution location within the same task and using the same tool call step index as the grouping unit.

[0132]

[0133] in, This represents the first sample obtained from the same task. A rollout, Indicates the number of rollouts within the same group; Indicates the first in the expert layer The steps for calling a tool. Indicates the first The relative advantage of the management team within the rollout; Indicates the first The management level of the rollout receives trajectory-level rewards; the expert layer is indexed and grouped according to the same task and the same step, allowing tool calls at the same execution location to be compared and calculating the relative advantage within the expert layer group:

[0134]

[0135] in, Indicates the first The rollout in the 1st The relative advantages within a group in the steps of calling up expert-level tools; Indicates the first The rollout in the 1st Expert-level step-by-step rewards for each tool invocation step; This represents a small constant to prevent the denominator from being zero; This represents the stability threshold for low-variance groups, used for stability advantage normalization calculations when the reward standard deviation within the expert layer group is too small. and These represent the mean and standard deviation of the trajectory-level rewards for the current management group, respectively, and are calculated as follows:

[0136]

[0137]

[0138] and These represent the current expert groupings. The mean and standard deviation of the rewards for each tool call step are calculated as follows:

[0139]

[0140]

[0141] The management and expert levels jointly update the shared strategy through a hybrid loss mechanism, ensuring independent normalization and stable optimization dynamics for both levels. The management strategy loss is calculated based on the relative advantage within the management group, and the expert strategy loss is calculated based on the relative advantage within the expert group. The management and expert strategy losses are then weighted and fused at the loss level to obtain the hybrid loss, expressed as:

[0142]

[0143]

[0144]

[0145] in, This indicates a strategic loss by management; This indicates the strategic loss of the expert layer; This represents the weighted average of management losses and expert losses; This indicates the parameters of the shared strategy model that are currently pending updating. Indicates the rollout number for management. and generation location Seeking expectations; Indicates the rollout number for the expert level. Tool Call Steps Index and generation location Seeking expectations. This represents the cutoff function in GRPO, used to limit the policy probability ratio to a certain value. Within the range; Indicates the cutoff coefficient; This represents the weight of management loss in the mixed loss; preferably, it takes... . Indicates the management's first The rollout in the 1st The policy probability ratio at each generation location is defined as:

[0146]

[0147] Indicates the expert level The rollout in the 1st The first tool call step, the first The policy probability ratio at each generation location is defined as:

[0148]

[0149] in, This represents the current sharing strategy model that needs to be optimized. This indicates the old strategy model used when sampling to generate the current execution trajectory; Indicates the first The remote sensing task input corresponding to the sampling execution trajectory; Indicates that management is in The token output by the generated location; This indicates the sequence of tokens that has been generated before this position; Indicates the expert level in the first The first tool call step The token output by the generated location; This indicates a sequence of tokens that has been generated prior to this position in the tool's call step.

[0150] Among them, the management layer strategy loss is used to optimize sub-objective decomposition, expert routing, and termination control strategies; the expert layer strategy loss is used to optimize tool selection, parameter construction, and tool execution strategies; and the weighted fusion of loss levels is used to maintain the independent normalization of the management layer trajectory-level rewards and the expert layer step-level rewards, so as to avoid the confusion between trajectory-level global coordination signals and step-level tool execution signals.

[0151] Based on the hybrid loss update sharing strategy model, the management agent acquires trajectory-level global coordination capabilities, and the expert agent acquires step-level tool execution capabilities; the trained management agent and expert agent are used to execute the above-mentioned multi-agent hierarchical collaborative remote sensing data processing method.

[0152] This invention also proposes a hierarchical collaborative multi-agent system for on-orbit processing of remote sensing data. This system is used to execute the above-mentioned hierarchical collaborative remote sensing data processing method of multi-agent systems. The system is deployed in a satellite on-orbit remote sensing data processing environment or an on-board edge computing environment. It includes: a task input module, a management agent module, an expert agent module, a tool service module, a structured global state module, and a result output module.

[0153] The task input module is used to acquire user instructions and multimodal remote sensing data, and send the user instructions and multimodal remote sensing data to the management intelligent agent module;

[0154] The management agent module is used to parse the remote sensing task from the user instructions, decompose the remote sensing task into multiple sub-targets, generate routing and control decisions based on the structured global state, route the current sub-target to the corresponding expert agent module, and perform step-level verification and adaptive control on the structured data fed back by the expert agent module.

[0155] The expert agent module is used to receive sub-targets assigned by the management agent module, call the corresponding remote sensing processing tools to execute the sub-targets, and feed back structured data to the management agent module;

[0156] The tool service module is used to provide a remote sensing professional tool execution interface, receive tool call requests from the expert intelligent agent module, execute the corresponding remote sensing processing tool, and return structured results including execution status, output data and diagnostic information;

[0157] The structured global state module is used to store user commands, multimodal remote sensing data, current sub-targets, structured global memory, dynamic constraint sets, intermediate results, verification results, and recovery trajectories, and provides historical perception decision-making basis to the management intelligent agent module;

[0158] The result output module is used to output the final remote sensing data processing result generated by the management agent module after all sub-targets have been completed.

[0159] In the scenario of satellite on-orbit remote sensing data processing, the management agent module and the expert agent module share the same basic large language model and parameter adapter. The system distinguishes the identities and functions of different agents through role prompts, decoding configurations, and hierarchical reinforcement learning loss guided by verification, thereby reducing the overhead of model storage, inference computation, and multi-model collaboration in the on-orbit environment.

[0160] In summary, this invention achieves reliable long-link solution from multimodal remote sensing input to structured final answer by combining a closed-loop control mechanism of "intent understanding - stage-aware routing - tool execution - step-level verification" with hierarchical reinforcement learning that aligns and fine-tunes the two-stage expert-to-workflow and verification guidance. This effectively blocks error propagation and overcomes the mismatch between workflow, knowledge and control, and has good application prospects.

[0161] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data, characterized in that, The steps include: S1: Acquire user commands and multimodal remote sensing data, parse the remote sensing task from the user commands, and decompose the remote sensing task into multiple sub-targets; S2: The management agent constructs a structured global state based on user commands, multimodal remote sensing data, and the current sub-target; S3: The management agent reasones about the structured global state to obtain routing and control decisions; based on the routing and control decisions, it routes sub-goals to the corresponding expert agents. S4: The expert agent executes sub-goals and feeds back structured data to the management agent; S5: The management agent verifies the structured data and obtains the verification results; S6: The management agent and expert agent execute adaptive control based on the verification results, complete the processing of the current sub-objective, update the structured global state, and advance to the next sub-objective; S7: Repeat steps S3 to S6 until all sub-goals have been processed and the management agent generates the final answer.

2. The hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 1, characterized in that, In step S2, the structured global state is represented as follows: ； in, This represents the structured global state of the current step. Indicates user commands, This represents multimodal remote sensing input data. Indicates the current sub-target. This represents the structured global memory of the current step. This represents the set of dynamic constraints for the current step.

3. The hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 1, characterized in that, In step S5, the process of the management agent verifying the structured data includes: Three complementary checks are performed: model validity check, sub-objective consistency check, and constraint satisfaction check. The constraint satisfaction check specifically includes: verifying the physical consistency of the inversion relationship, the spectral correctness of band usage and index calculation, and the statistical rationality of the aggregation model and output distribution. Output "pass" only if all three conditions are met: pattern validity, sub-objective consistency, and constraint satisfaction. Output "uncertain" if none of the three conditions are violated but the confidence level is insufficient. Output "fail" if any constraint is violated.

4. The hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 1, characterized in that, Step S6 specifically includes: If the verification result is "fail", attribution-based replanning is triggered. Attribution-based replanning includes: the management agent locates the failure level; when the failure is attributed to the parameter level, the expert agent performs tool parameter repair and retryes in place; when the failure is attributed to the tool or routing level, an alternative route is executed; when the failure is attributed to the sub-goal level, the sub-goal is further decomposed or the sub-goal is adjusted, and a new route is re-executed; all intermediate results of the execution are written into the structured global memory in the structured global state. If the verification result is uncertain, cross-checking is triggered. Cross-checking includes: the management agent reassigning the current sub-objective along with its output structured data to another expert with overlapping verification capabilities, or having the same expert independently recalculate the key quantities under alternative inference paths and alternative tools, and comparing the results of the two paths. If they are consistent within the tolerance, the result is upgraded to pass; otherwise, it is downgraded to fail. All intermediate results of the execution are written into the structured global memory in the structured global state. If the verification result is pass, the structured data of the current sub-target is written into the structured global memory in the structured global state and the process is advanced to the next sub-target.

5. The hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 1, characterized in that, The expert agents include surface analysis agents, physical inversion agents, and statistical analysis agents. The surface analysis agent is used to process perceived and analyzed sub-targets, including target detection, semantic segmentation, region of interest and mask generation, and spectral index analysis. The physical inversion agent is used to process physical inversion and quantitative remote sensing product generation sub-targets, including surface temperature inversion, atmospheric correction, and radiometric calibration. The statistical analysis agent is used to process data aggregation and cleaning, statistical testing and correlation analysis, report-level index calculation, and comparative sub-targets. The three types of expert agents integrate 120 remote sensing professional tools, with backend support provided by three MCP tool servers.

6. The hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 1, characterized in that, The management agent and the expert agent share the same basic large language model and parameter adapter, and distinguish between management layer roles and expert layer roles through role prompt words and decoding configuration; Its training process includes hierarchical reinforcement learning with expert-to-workflow alignment fine-tuning and validation guidance: The expert-to-workflow alignment fine-tuning includes two sequentially executed supervised fine-tuning stages: the first stage is the remote sensing expert knowledge alignment stage, which constructs a remote sensing professional knowledge system, generates knowledge supervision samples based on the remote sensing professional knowledge system, and performs supervised fine-tuning on the shared basic large language model to obtain an expert model with remote sensing domain knowledge; The second stage is the remote sensing workflow alignment stage. Using the expert model as the initialization model, workflow supervision samples are constructed based on real remote sensing task pipelines or multi-agent demonstration logs. The workflow supervision samples include workflow selection samples and workflow sorting samples. A structured mapping is established between the workflow supervision samples and the corresponding knowledge branches in the remote sensing professional knowledge system. The initialization model is further supervised and fine-tuned to enable the model to learn the mapping relationship from remote sensing task semantics to executable tool processes, thus obtaining a shared policy model. The validation-guided hierarchical reinforcement learning involves using the GRPO algorithm to perform validation-guided hierarchical reinforcement learning on the shared policy model, thereby training the management agent and the expert agent.

7. A hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 6, characterized in that, The process of hierarchical reinforcement learning using the GRPO algorithm to validate a shared policy model includes: For the same remote sensing task, multiple complete execution trajectories are sampled based on the shared policy model. Each execution trajectory includes sub-objective decomposition, expert routing, and termination control decisions generated by the management agent, as well as tool selection, parameter construction, and tool execution results generated by the expert agent. Based on the step-level verification results in each execution trajectory, the management layer trajectory-level reward and the expert layer step-level reward are calculated respectively. The management layer trajectory-level reward is used to evaluate the global coordination quality of the entire execution trajectory, and the expert layer step-level reward is used to evaluate the quality of single-step tool invocation. Based on the management layer trajectory-level reward and the expert layer step-level reward, the intra-group relative advantage of the management layer and the expert layer is estimated to obtain the intra-group relative advantage of the management layer and the intra-group relative advantage of the expert layer. The management strategy loss is calculated based on the relative advantage within the management group, and the expert strategy loss is calculated based on the relative advantage within the expert group. The management strategy loss and the expert strategy loss are then weighted and fused at the loss level to obtain a mixed loss. The shared strategy model is updated according to the mixed loss, enabling the management agent to obtain trajectory-level global coordination capabilities and enabling the expert agent to obtain step-level tool execution capabilities.

8. A hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 7, characterized in that, The formulas for calculating the trajectory-level reward for management and the step-level reward for expert are expressed as follows: ； ； in, This indicates management-level trajectory rewards. Indicates routing reward, Indicates a binary reward. Indicates boundary reward, , and These represent route weights, binary weights, and boundary weights, respectively. , and These represent the expert-level step-level rewards for tool invocation correctness at levels one, two, and three, respectively. This indicates the original execution feedback. This represents the quality fraction of the normalized parameter; The formulas for obtaining the relative advantage within the management group and the relative advantage within the expert group are expressed as follows: ； ； in, Indicates the number of rollouts within the same group; Indicates the first The relative advantage of the management team within the rollout; Indicates the first The management trajectory-level rewards for the rollout; Indicates the first The rollout in the 1st The relative advantages within a group in the steps of calling up expert-level tools; Indicates the first The rollout in the 1st Expert-level step-by-step rewards for each tool invocation step; and These represent the mean and standard deviation of the trajectory-level rewards for management within the current management group, respectively. This represents a small constant to prevent the denominator from being zero; and These represent the current expert groupings. The mean and standard deviation of the rewards for each tool call step. This represents the stability threshold for low-variance groups.

9. A hierarchical collaborative multi-agent method for on-orbit processing of remote sensing data according to claim 7, characterized in that, The formula for calculating the mixing loss is expressed as: ； ； ； in, This indicates a strategic loss by management; This indicates the strategic loss of the expert layer; This represents the weighted average of management strategy losses and expert strategy losses; This indicates the parameters of the shared strategy model that are currently pending update; Indicates the rollout number for management. and generation location Seeking expectations; Indicates the rollout number for the expert level. Tool Call Steps Index and generation location Seeking expectations; Indicates the management's first The rollout in the 1st The probability ratio of the strategies at each generation location; Indicates the expert level The rollout in the 1st The first tool call step, the first The probability ratio of the strategies at each generation location; This represents the truncation function; Indicates the cutoff coefficient; This indicates the weight of management loss in the mixed loss.

10. A hierarchical collaborative multi-agent system for on-orbit processing of remote sensing data, characterized in that, The system is deployed in a satellite on-orbit remote sensing data processing environment or an on-board edge computing environment, and includes: a task input module, a management agent module, an expert agent module, a tool service module, a structured global state module, and a result output module; The task input module is used to acquire user instructions and multimodal remote sensing data, and send the user instructions and multimodal remote sensing data to the management intelligent agent module; The management agent module is used to parse the remote sensing task from the user instructions, decompose the remote sensing task into multiple sub-targets, generate routing and control decisions based on the structured global state, route the current sub-target to the corresponding expert agent module, and perform step-level verification and adaptive control on the structured data fed back by the expert agent module. The expert agent module is used to receive sub-targets assigned by the management agent module, call the corresponding remote sensing processing tools to execute the sub-targets, and feed back structured data to the management agent module; The tool service module is used to provide a remote sensing professional tool execution interface, receive tool call requests from the expert intelligent agent module, execute the corresponding remote sensing processing tool, and return structured results including execution status, output data and diagnostic information; The structured global state module is used to store user commands, multimodal remote sensing data, current sub-targets, structured global memory, dynamic constraint sets, intermediate results, verification results, and recovery trajectories, and provides historical perception decision-making basis to the management intelligent agent module; The result output module is used to output the final remote sensing data processing result generated by the management agent module after all sub-targets have been completed.

Images