Remote sensing image processing system and method based on intelligent agent

Abstract

The invention belongs to the technical field of remote sensing image intelligent processing, and discloses a remote sensing image processing system and method based on an intelligent agent, and the system architecture comprises an operator layer, an interaction layer, a data adaptation layer and a dynamic process execution engine. The operator layer comprises an operator registration module, a code generation module and a test sandbox module; the interaction layer comprises a natural language interaction module, a process automatic generation module and a visual process module; and the data adaptation layer comprises a data interface specification module and field adaptation middleware. Compared with a traditional remote sensing image processing system, the method is high in operator expansibility, simple in data adaptation and low in man-machine interaction threshold, and is particularly suitable for scenes such as a satellite ground system, an emergency monitoring platform and environment remote sensing monitoring which need to quickly respond to complex processing requirements; and a core technical support is provided for constructing a new-generation intelligent remote sensing processing cloud platform.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image intelligent processing, and in particular relates to a remote sensing image processing system and method based on intelligent agent. Background Art

[0002] Traditional remote sensing image processing systems have long faced three major technical bottlenecks, severely hampering industry efficiency and emergency response capabilities. First, they rely on fixed operator libraries for process development. Algorithm expansion requires manual compilation and deployment by professionals, resulting in long rollout times of several days for new operators. This static architecture cannot meet the demands for rapid algorithm iteration in scenarios like disaster monitoring. For example, statistics from the China Resources Satellite Application Center show that 30% of processing time in emergency missions is consumed by operator adaptation. Second, the input and output formats of heterogeneous operators vary significantly, and there is a lack of unified standards for parameters such as band configuration, coordinate system, and resolution for different sensor data (such as multispectral and synthetic aperture radar SAR). This leads to a processing interruption rate as high as 23%. Existing solutions rely on manual adaptation code, requiring an additional one to two hours of format conversion for each task. Finally, there is a significant lag between algorithm development cycles and mission requirements. Traditional development models, which require multiple stages including requirements analysis, code development, and testing and deployment, are unable to meet the real-time processing demands of emergencies such as forest fires.

[0003] The current wave of large language models and the rapid development of research related to intelligent agents offer new approaches to addressing these issues. Breakthroughs in natural language understanding enabled by large models have enabled remote sensing image processing systems to accurately parse ambiguous natural language instructions, such as "extract flooded areas," and map them to specific algorithm combinations. The dynamic decision-making capabilities of intelligent agents, combined with reinforcement learning path planning, can increase operator matching accuracy to 92%. More importantly, code generation technology based on large models, combined with a sandbox testing environment, can automatically deploy new operators within two hours, a 90% speedup compared to traditional methods. Summary of the Invention

[0004] The purpose of the present invention is to provide a remote sensing image processing system and method based on intelligent agents. Compared with traditional remote sensing image processing systems, the present invention has strong operator scalability, simple data adaptation, and low human-computer interaction threshold. It is particularly suitable for scenarios such as satellite ground systems, emergency monitoring platforms, and environmental remote sensing monitoring that require rapid response to complex processing requirements, providing core technical support for building a new generation of intelligent remote sensing processing cloud platforms.

[0005] The technical solution adopted by the present invention is a remote sensing image processing system based on intelligent agents. The system architecture includes an operator layer, an interaction layer, a data adaptation layer and a dynamic process execution engine. The operator layer includes an operator registration module, a code generation module and a test sandbox module. The interaction layer includes a natural language interaction module, a process automation generation module and a visualization process module. The data adaptation layer includes a data interface specification module and a field adaptation middleware. The operator registration module is used to implement dynamic registration and intelligent retrieval of operator metadata; the code generation module generates operator code by integrating the Qwen3 model and presetting specific templates in the remote sensing field; the generated operator code is unit tested through the test sandbox module, and is saved after passing the test. If it fails, it is manually corrected and then tested again; the test sandbox module performs code inspection and code operation inspection by building a basic verification system and a business verification system. After the inspection is completed, the operator registration is automatically completed through the gray release mechanism; The natural language interaction module uses a hybrid expert model to parse the natural language instructions input by the user, converts them into structured parameters, integrates a confidence assessment mechanism, and scores the confidence of the parsing results. When the confidence is below the threshold, multiple rounds of dialogue are triggered, and semantic completion is achieved by combining with the knowledge graph to achieve accurate understanding of the user's intention. The process automation generation module is used to achieve task decomposition and process generation. The visual process module is used to realize process DAG interaction and real-time process monitoring. The data interface specification module is used to unify data interface specifications; the field adaptation middleware is used to implement automatic detection and mapping of input and output fields, dynamic format conversion and adaptation, and metadata repair; The dynamic process execution engine assigns priorities based on the resource requirements and dependencies of operators, monitors the execution status of operators, and dynamically adjusts resource allocation. When an operator fails to execute in the process, the engine backtracks to the most recent checkpoint and calls the process automation generation module of the interaction layer to recommend an alternative operator to continue execution.

[0006] Furthermore, the operator registration module includes an operator meta-information dynamic registration unit and an intelligent retrieval unit; the operator meta-information dynamic registration unit supports dynamic operator loading and version management, structures the operator meta-information into JSON format, and vectorizes the structured operator meta-information through the Jina-Embeddings-v3 model and stores it in a vector database; the intelligent retrieval unit is based on RAG technology, inputs operator meta-information, operator historical execution information and resource consumption, and realizes priority sorting retrieval.

[0007] Furthermore, in the test sandbox module: The basic verification system ensures that the operator code is free of basic defects and meets the requirements of operational stability through static code analysis, unit test coverage, and dynamic memory detection; The business verification system conducts end-to-end testing by building a small remote sensing dataset to ensure that the operator's output results in real remote sensing tasks meet business logic and accuracy requirements; The specific implementation process of the grayscale release mechanism is as follows: The new operator first enters "shadow mode" and executes in parallel with the existing operator, comparing the results. When the difference exceeds the threshold, the manual correction process is triggered; After manually correcting the code, the test is completed through code inspection and code operation inspection in the test sandbox. After the test passes, it is added to the process.

[0008] Furthermore, the natural language interaction module uses a hybrid expert model to parse the natural language instructions input by the user and converts them into structured parameters. The specific steps include: Receive user natural language instructions; Extract semantic features through the BERT+BiLSTM model and perform text encoding; Combined with the current remote sensing image metadata, spatial context analysis is performed to complete abnormal context correction and coordinate system constraints, including: Abnormal context correction: If the user's natural language command conflicts with the current remote sensing image metadata context, multiple rounds of dialogue are triggered to confirm the intention; Coordinate system constraints: Convert ambiguous expressions in user natural language instructions into specific geographic coordinate ranges to avoid misunderstandings caused by coordinate system differences; Output structured parameters.

[0009] Furthermore, the confidence assessment mechanism specifically includes: (1) Semantic uncertainty quantification; (2) Spatial context conflict detection; (3) Verification of knowledge graph semantic coverage; (4) Dynamic threshold decision-making.

[0010] Furthermore, the process automation generation module includes a task decomposition unit and a process generation unit; the task decomposition unit decomposes the complex task into atomic subtask chains based on the thinking chain technology; after receiving the atomic subtask chain, the process generation unit automatically matches the operator through the hybrid expert model input system prompt words, operator meta-information and historical execution information of the operator, and then matches the optimal operator and optimizes the task execution path through the ReAct framework combined with the improved A* algorithm path cost function, sets the operator's configuration parameters and running resources, and generates the optimal process; when the registered operator is not found in the operator library, the code generation sandbox is triggered and verified. After the verification is completed, the newly generated operator is registered to the operator library and added to the process. Finally, the loop termination condition is designed to prevent infinite reasoning; the visual process module includes a process DAG interaction unit and a process real-time monitoring unit; the process DAG interaction unit allows the user to adjust the operator order and operator configuration parameters by dragging and dropping and using a form to modify the processing flow; the process real-time monitoring unit supports real-time performance monitoring and marks abnormal nodes.

[0011] Furthermore, the task decomposition unit decomposes the complex task into atomic subtask chains based on the thought chain technology, and the specific steps include: Receive the converted structured parameters; Retrieve historical similar process task cases through retrieval enhancement generation technology; The system constructs structural parameters, historical similar process task cases and current remote sensing image metadata as prompt words, builds task execution paths based on thought chain technology, and decomposes complex tasks into atomic subtask chains; Outputs a chain of atomic subtasks.

[0012] Furthermore, the field adaptation middleware includes an automatic detection and mapping unit for input and output fields, a dynamic format conversion and adaptation unit, and a metadata repair unit; the automatic detection and mapping unit for input and output fields detects the differences in the names, data types, dimensions, spatial references, and metadata labels of input and output fields between operators in real time; understands the field semantics through a rule engine or a large model, and automatically associates synonymous fields; the dynamic format conversion and adaptation unit automatically inserts forced type conversion logic, band order adjustment, and coordinate system conversion operators when it detects that there is a data type inconsistency between adjacent operator inputs and outputs; resolves multidimensional data conflicts through structural adaptation, or unifies resolution through interpolation / resampling; the metadata repair unit inherits and supplements from upstream data or global configuration when key metadata is lost in the operator output, and automatically adds semantic labels to unlabeled data.

[0013] A remote sensing image processing method based on an intelligent agent is implemented according to the above-mentioned remote sensing image processing system based on an intelligent agent, comprising the following steps: S1, at the operator layer, register the prepared operators through the operator registration module and vectorize and store the operator metadata; S2, the user inputs natural language instructions at the interaction layer; S3, the natural language interaction module of the interaction layer, uses a hybrid expert model to parse the natural language instructions input by the user and convert them into structured parameters. If the confidence level is lower than the threshold, it triggers multiple rounds of dialogue and semantic completion to achieve accurate understanding of the user's intention. S4, after receiving the structured parameters converted by the natural language interaction module, the process automation generation module of the interaction layer decomposes the complex task into atomic subtask chains based on the thought chain technology; S5: The process automation generation module in the interaction layer automatically matches operators using a hybrid expert model input, system prompts, operator metadata, and historical operator execution information. It then uses the ReAct framework combined with an improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The field configuration information of the process is then modified and confirmed using the field adaptation middleware in the data adaptation layer, and the process description file is finally output. S6: When no operator is matched in the operator library, the code is triggered to generate a sandbox and verify it. After the verification is completed, the newly generated operator is registered with the operator library and added to the process; S7, after receiving the generated process, the visualization process module of the interaction layer displays it on the user interface through a directed acyclic graph (DAG). Users can modify the operator order and operator configuration parameters in the process by dragging and dropping and using forms. S8, the dynamic process execution engine allocates operator scheduling priorities based on the operator's resource requirements, dependencies, and current system resources, monitors the operator's execution status, dynamically adjusts resource allocation, and implements directed acyclic graph (DAG) task scheduling. It automatically saves the intermediate state to the distributed cache Redis after completing every three atomic subtasks or taking 120 seconds. If an operator fails during process execution, the engine backtracks to the most recent checkpoint and calls the process automation generation module to recommend an alternative operator for continued execution. It also records the failure log and updates historical performance indicators.

[0014] Furthermore, in S5, after receiving the atomic subtask chain, the process automation generation module of the interaction layer automatically matches operators by inputting system prompt words, operator meta-information, and historical execution information of the operators through the hybrid expert model. Then, the specific steps of matching the optimal operator and optimizing the task completion path by combining the improved A* algorithm path cost function with the ReAct framework include: S51, builds the initial decision environment, receives the atomic subtask chain, and dynamically loads the environment context; The dynamic loading environment context includes: (1) Real-time status of the operator library; (2) System-level constraints; (3) Data characteristics; S52, the closed-loop decision-making process of the ReAct framework, cyclically executes the following four stages until the operator matching of all atomic subtasks is completed: The first stage is the reasoning stage: based on the current atomic subtask and environmental context, a decision-making chain is generated and structured decision options are output. If no matching operator is found in the operator library, the code is triggered to generate a sandbox and verify it. After verification is complete, the newly generated operator is registered with the operator library and added to the process. The second stage is the action stage: calling the improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The specific steps include: (1) Based on the dynamic topology reconstruction mechanism, the data dependency between atomic subtasks is analyzed, and the directed acyclic graph (DAG) topology is reconstructed by splitting and merging branches. The atomic subtasks without data coupling are split into independent execution flows to achieve multi-operator parallel computing. (2) Path cost optimization; The path cost function based on the improved A* algorithm is: Among them, Cost is the path cost of the operator; is the historical average time consumed by the operator; F rate is the historical failure rate of the operator; R comp is the real-time resource competition coefficient; is the time cost weight; is the failure penalty weight; k, m, n are coefficients, which can be set; gpu util is GPU utilization; cpu util is the CPU utilization; mem util is the memory utilization; (3) Path search process; Construct a directed acyclic graph (DAG) with atomic subtasks as nodes, initialize the open list, iteratively expand the path, and select the node with the smallest cost function f(n) each time; Where f(n) is the total cost of executing the current node and completing the task; g(n) is the actual cost from the starting point to the current node; h(n) is the estimated cost of the remaining subtasks; When resource conflicts are detected, an adaptive resource scheduling strategy is implemented to dynamically adjust the real-time resource competition coefficient R. comp and reorder them; (4) Operator matching; For each operator subtask, select the operator with the smallest path cost from the operator candidate pool; The third phase is the observation phase: verifying that operators match constraints and generating exception reports when conflicts are detected; The fourth stage is the feedback adjustment stage: dynamic adaptation is triggered according to the conflict type, and is carried out in the following four aspects: Insert degradation operators when resource conflicts occur; Insert format conversion operators when data is incompatible; Insert automatic transposition operator when data bit depth is incompatible; For other conflict types that are not preset, the conflict report is added to the environment context and returned to the first stage.

[0015] Furthermore, the specific steps of S6 include: S61, construct code generation prompt words; S62, inputting the code generation prompt words into the Qwen3 large model to complete the code writing, and performing code checking and code running checking after the code writing is completed; S63: After the inspection is completed, the operator is automatically registered through the grayscale release mechanism and added to the process. If the inspection fails, the code is manually corrected and re-inspected after the modification is confirmed.

[0016] Furthermore, in S8, the priority calculation of operator scheduling adopts a three-level priority calculation model, and the priority weight formula is: Among them, Priority is the scheduling priority of the operator in this task process; DepPriority is the dependency priority; ResPriority is the resource priority; DynamicAdj is the dynamic adjustment factor.

[0017] The beneficial effects of the present invention are: Compared with traditional remote sensing image processing systems, the operator scalability of this invention is strong, data adaptation is simple, and the threshold for human-computer interaction is low. It is particularly suitable for scenarios such as satellite ground systems, emergency monitoring platforms, and environmental remote sensing monitoring that require rapid response to complex processing requirements, providing core technical support for building a new generation of intelligent remote sensing processing cloud platforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 Schematic diagram of data interaction of the system of the present invention.

[0019] Figure 2 Schematic diagram of the process of the present invention. DETAILED DESCRIPTION

[0020] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] Intelligent agent is an important concept in the field of computer science and artificial intelligence in recent years. It refers to a computing entity that resides in a specific environment, can perceive the environment, and can run autonomously to achieve a series of goals on behalf of its designer or user.

[0022] A remote sensing image processing system based on intelligent agent, the system architecture includes interaction layer, dynamic process execution engine, operator layer and data adaptation layer, the following combination Figure 1 The remote sensing image processing system based on intelligent agent of the present invention is described in detail: 1. Operator Layer The operator layer includes the operator registration module, the code generation module, and the test sandbox module.

[0023] 1. Operator Registration Module The operator registration module is used to implement dynamic registration and intelligent retrieval of operator metadata.

[0024] The operator registration module includes an operator meta-information dynamic registration unit and an intelligent retrieval unit.

[0025] (1) Operator meta-information dynamic registration unit It supports dynamic operator loading and version management, structures operator metadata into JSON format, and vectorizes the structured operator metadata through the Jina-Embeddings-v3 model and stores it in a vector database.

[0026] Operator metadata includes input and output data types, configuration parameter templates, resource requirements (such as CPU / GPU quotas and memory limits), and dependencies (third-party library versions).

[0027] (2) Intelligent retrieval unit Based on RAG technology, operator meta-information, operator historical execution information, resource consumption and other data are input to achieve priority sorting retrieval.

[0028] 2. Code generation module Operator code is generated by integrating the Qwen3 model and presetting specific templates in the remote sensing field (such as GDAL data reading and writing interface, SNAP operator adaptation, coordinate conversion, bit depth conversion and band reorganization, etc.); the generated operator code is unit tested through the test sandbox module. If the test passes, it is saved. If it fails, it is manually corrected and then tested again.

[0029] 3. Test sandbox module By building a basic verification system and a business verification system, code inspection and code operation inspection are carried out. After the inspection is completed, the operator registration is automatically completed through the grayscale release mechanism.

[0030] The basic verification system is mainly to ensure that the operator code has no basic defects and meets the requirements of operational stability.

[0031] The basic verification system includes static code analysis, unit test coverage and dynamic memory detection.

[0032] Static code analysis uses the SonarQube+Pylint code checking tools to detect syntax errors, security vulnerabilities, code standards, and third-party dependency conflicts.

[0033] Unit test coverage tests the code by generating different test cases at the method level (such as boundary tests, normal tests, exception tests, etc.).

[0034] Dynamic memory detection monitors the memory leak threshold (for example, memory growth ≤ 2% after 10 consecutive runs) by injecting abnormal data (such as null pointers and very large matrices).

[0035] The specific steps to build a basic verification system include: The MCP Server program integrates the SonarQube+Pylint code checking tools to perform static analysis on the automatically generated operator code.

[0036] Generate different test cases at the function level and class level based on the operator code to test the entire operator code.

[0037] Build a Docker container for the operator to run independently, inject abnormal data into the operator, and monitor the memory leakage threshold of the container.

[0038] The business verification system conducts end-to-end testing by constructing small remote sensing datasets (such as Landsat-8 sub-areas) to ensure that the output results of operators in real remote sensing tasks meet business logic and accuracy requirements.

[0039] The specific steps to build a business verification system include: Construct small test datasets for different types of operators. The test datasets include data, accuracy indicators, and timeliness indicators.

[0040] Build a Docker container for the operator to run independently, inject the test dataset, run the operator, and record the operator's running results and timeliness results.

[0041] Result verification: The operator's running results are compared with the true values ​​in the test dataset. If the comparison results meet the threshold, timeliness and resource constraints are checked based on the operator's running timeliness results.

[0042] The specific implementation process of the grayscale release mechanism is as follows: The new operator first enters "shadow mode" and executes in parallel with the existing operators to compare the differences in the results.

[0043] When the difference exceeds a threshold (for example, the intersection-over-union ratio (IOU) < 0.9), the manual correction process is triggered.

[0044] After manually correcting the code, the test is completed through code inspection and code operation inspection in the test sandbox. After the test passes, it is added to the process.

[0045] 2. Interaction Layer The interaction layer includes a natural language interaction module, a process automation generation module, and a visual process module.

[0046] 1. Natural language interaction module A hybrid expert model is used to parse natural language instructions input by users (for example, "extract flooded areas") and convert them into structured parameters ({algorithm type: water body detection, band requirement: SWIR}). A confidence assessment mechanism is integrated to assign a confidence score (0-1) to the parsing results. When the score is below the threshold, multiple rounds of dialogue are triggered (for example, do you mean vegetation cover or water body extraction?). The knowledge graph is then combined to achieve semantic completion [for example, mapping ambiguous natural language instructions ("extract water bodies") into structured parameters (band = SWIR / NIR, algorithm = NDWI / MNDWI)] to accurately understand the user's intention.

[0047] The natural language interaction module uses a hybrid expert model to parse the natural language instructions input by the user and convert them into structured parameters. The specific steps include: Receive user natural language instructions.

[0048] The BERT+BiLSTM model is used to extract semantic features and perform text encoding.

[0049] Combined with the current remote sensing image metadata, spatial context analysis is performed to mainly complete abnormal context correction and coordinate system constraints, including: Abnormal context correction: If the user's natural language command conflicts with the current remote sensing image metadata context (for example, requesting "extract snow" but the image was taken in the summer and in a tropical area), multiple rounds of dialogue will be triggered to confirm the intent; Coordinate system constraints: Convert ambiguous expressions in user natural language instructions (such as "southwest region") into specific geographic coordinate ranges to avoid misunderstandings caused by coordinate system differences.

[0050] Output structured parameters.

[0051] The confidence assessment mechanism specifically includes: (1) Semantic uncertainty quantification The BERT-BiLSTM dual-channel semantic encoding architecture is used to implement initial confidence calculation, where the BERT layer is responsible for extracting the global semantic features of the user instruction (for example, mapping "extract floods" to "extract water bodies"), and the BiLSTM layer captures local context dependencies (for example, identifying the spatial orientation attributes of "southwest"). Calculated by inverse ratio of information entropy: Formula 1 Among them, Entropy is the information entropy function, which is used to quantify the uncertainty of probability distribution; P(y) is the intent classification probability distribution output by dual-channel semantic encoding.

[0052] A higher entropy value indicates stronger semantic uncertainty and lower basic confidence.

[0053] Formula 2 Where n is the total number of intent categories; represents the predicted probability of the i-th type of intent.

[0054] (2) Spatial context conflict detection Logical conflict determination is performed based on a pre-built remote sensing rule base containing over 300 specialized constraints (e.g., "summer + snow extraction = high conflict probability"). Conflict probabilities are categorized as extremely high (0.8), high (0.7), medium (0.5), low (0.2), extremely low (0.1), and no conflict (0). The system loads the current remote sensing image metadata (including geographic region, season, sensor type, etc.) in real time and performs conflict scanning using the rule engine: The initial confidence correction is implemented using the product decay model: Formula 3 Among them, Conf adj is the revised initial confidence value; conflict_score is the conflict probability. For example, when a conflict value of 0.2 is detected, the original confidence score of 0.86 will be revised to 0.688.

[0055] (3) Verification of knowledge graph semantic coverage Build a domain knowledge graph to achieve standardized terminology mapping. The graph nodes include remote sensing entities (such as sensors, bands, algorithms), data products, and processing flows.

[0056] Coverage calculation is implemented using the entity matching rate model: Formula 4 For example, the instruction "Use SWIR and NIR for water extraction" contains the entities [SWIR, NIR, water extraction]. The knowledge graph matching results are SWIR → shortwave infrared (match), NIR → near infrared (match), water extraction → NDWI (match), so Coverage = 1.0.

[0057] Confidence after coverage compensation final The correction formula is: Formula 5 The confidence level of the previous correction of 0.688 is increased to 0.888 after coverage compensation.

[0058] (4) Dynamic threshold decision Set up a scenario-adaptive confidence threshold system: the threshold for directly outputting structured parameters is 0.7 (which can be set manually). If it is lower than 0.7, multiple rounds of dialogue will be directly triggered.

[0059] 2. Process automation generation module Used to achieve task decomposition and process generation.

[0060] The process automation generation module includes a task decomposition unit and a process generation unit.

[0061] (1) Task decomposition unit Based on the chain thinking technique, complex tasks are decomposed into atomic subtask chains (for example, "flood assessment" is decomposed into: sub_tasks = ["radiation correction", "water index calculation", "change detection", "flood range statistics"]). The specific steps include: Receives the converted structured parameters.

[0062] Retrieve historical similar process task cases through retrieval enhancement generation technology.

[0063] Structural parameters, historical similar process task cases and current remote sensing image metadata are constructed as system prompts, and task execution paths are constructed based on thinking chain technology, which decomposes complex tasks into atomic subtask chains.

[0064] Outputs a chain of atomic subtasks.

[0065] (2) Process generation unit After receiving the atomic subtask chain, the system prompt words, operator meta-information and historical execution information of the operator are input through the hybrid expert model to automatically match the operator. Then, the ReAct framework is combined with the improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The operator's configuration parameters and running resources are set to generate the optimal process. When the registered operator is not found in the operator library, the code generation sandbox is triggered and verified. After the verification is completed, the newly generated operator is registered with the operator library and added to the process. Finally, the loop termination condition is designed (maximum number of iterations = 5) to prevent infinite reasoning.

[0066] 3. Visual process module Used to implement process DAG (directed acyclic graph) interaction and real-time process monitoring.

[0067] The visual process module includes a process DAG interaction unit and a process real-time monitoring unit.

[0068] (1) Process DAG interaction unit Allows users to adjust the operator order and operator configuration parameters by dragging and dropping and using forms to modify the processing flow.

[0069] (2) Process real-time monitoring unit Supports real-time performance monitoring (such as monitoring node time consumption and resource utilization) and marking abnormal nodes (such as operator failure).

[0070] 3. Data Adaptation Layer The data adaptation layer includes the data interface specification module and the field adaptation middleware.

[0071] 1. Data interface specification module Used to unify data interface specifications, including coordinate system (EPSG code), resolution, band order (BGR / RGB) and data type (uint16 / float32 type).

[0072] 2. Field Adaptation Middleware Used to implement automatic detection and mapping of input and output fields, dynamic format conversion and adaptation, and metadata repair.

[0073] The field adaptation middleware includes an automatic detection and mapping unit for input and output fields, a dynamic format conversion and adaptation unit, and a metadata repair unit.

[0074] (1) Automatic detection and mapping unit of input and output fields Detect in real time the differences in the names, data types, dimensions, spatial references, and metadata tags of input and output fields between operators (for example, operator A outputs a field named "NDVI," while operator B requires an input field named "Vegetation Index"). Understand field semantics through a rule engine or large model, and automatically associate synonymous fields (for example, the "SWIR" band with the "Shortwave Infrared" band).

[0075] Large-scale model understanding of field semantics: LLMs are used to parse the deeper meaning of remote sensing data fields, addressing the semantic gap between heterogeneous operators. The core goal is to map fields from different sources and with different names (e.g., "NDVI" and "Vegetation Index") into a unified semantic space, enabling intelligent association and adaptation of fields, thus overcoming the limitations of traditional rule engines.

[0076] (2) Dynamic format conversion and adaptation unit When data type inconsistencies are detected between adjacent operator inputs and outputs, operators are automatically inserted for forced type conversion logic (e.g., float32 to uint16 type), band order adjustment (e.g., BGR to RGB), and coordinate system conversion (e.g., WGS84 to UTM coordinate system). Multidimensional data conflicts are resolved through structural adaptation (e.g., splitting a three-band RGB color image into single-band inputs), or unifying the resolution (e.g., 10m→5m) through interpolation / resampling.

[0077] (3) Metadata repair unit In the case that the operator output is missing key metadata (such as coordinate system), it is inherited and supplemented from upstream data or global configuration (such as the EPSG code associated with the original image), and semantic labels are automatically added to unlabeled data (such as renaming "Band_1" to "NIR" band).

[0078] 4. Dynamic Process Execution Engine Priorities are assigned based on the resource requirements and dependencies of operators, the execution status of operators is monitored, and resource allocation is dynamically adjusted to implement directed acyclic graph (DAG) task scheduling. When an operator fails to execute in the process (for example, due to insufficient memory), the engine backtracks to the most recent checkpoint and calls the process automation generation module to recommend alternative operators (for example, replacing "DeepLabV3+" with "Lightweight Mobile-UNet") to continue execution.

[0079] A remote sensing image processing method based on intelligent agent is implemented by the remote sensing image processing system based on intelligent agent, such as Figure 2 As shown, the specific steps include: S1, at the operator layer, register the prepared operators through the operator registration module and vectorize and store the operator metadata.

[0080] Operator files include Python executable code files and operator configuration files.

[0081] The operator configuration file defines the operator type, input, output, required resources, configuration parameters, and description information.

[0082] S2, the user inputs natural language instructions in the interaction layer.

[0083] S3, the natural language interaction module of the interaction layer uses a hybrid expert model to parse the natural language instructions input by the user and convert them into structured parameters. If the confidence level is lower than the threshold, multiple rounds of dialogue and semantic completion are triggered to achieve accurate understanding of the user's intentions.

[0084] S4, after receiving the structured parameters converted by the natural language interaction module, the process automation generation module of the interaction layer decomposes the complex task into atomic subtask chains based on the thinking chain technology.

[0085] S5, the process automation generation module of the interaction layer automatically matches operators through the hybrid expert model input system prompt words, operator metadata and historical execution information of operators (such as the number of successful executions, the number of failed executions, the reason for failure, the execution time, etc.). Then, the ReAct framework is combined with the improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. Then, the field configuration information of the process is modified and confirmed through the field adaptation middleware of the data adaptation layer, and finally the process description file is output [the process description file is in JSON format, including the operator binding sequence, operator parameter configuration, operator resource allocation plan (such as GPU, memory, CPU quota and checkpoint interval) and operator exception handling record (such as the backup operator list)].

[0086] After receiving the atomic subtask chain, the process automation generation module in the interaction layer automatically matches operators using a hybrid expert model input, system prompts, operator metadata, and historical operator execution information. The ReAct framework, combined with the improved A* algorithm path cost function, matches the optimal operator and optimizes the task completion path. The specific steps include: S51, builds the initial decision environment, receives the atomic subtask chain (e.g. ["radiation correction", "water index calculation"]), and dynamically loads the environment context; The dynamic loading environment context includes: (1) Real-time status of the operator library [including operator availability, resource usage, operator historical execution logs, and operator meta-information; the operator meta-information structure is JSON data, including input / output specifications (such as data type, band order, spatial reference), resource requirements (such as GPU video memory > 4GB), algorithm dependencies (such as TensorFlow 2.15 version), and historical performance indicators (such as a success rate of 92.3% and an average execution time of 23s); the operator historical execution log is dynamically updated runtime data, using a sliding window to record the last 100 execution records, including actual resource usage (such as a CPU peak of 85%), abnormal patterns (such as a memory leak probability of 0.7%), and context relevance (such as 89% compatibility with the NDWI algorithm)].

[0087] (2) System-level constraints (maximum delay ≤ 300 seconds, accuracy threshold ≥ 0.9).

[0088] (3) Data characteristics (upstream input data type, auxiliary data information, spatial reference system, output data type).

[0089] S52, the closed-loop decision-making process of the ReAct framework, cyclically executes the following four stages until the operator matching of all atomic subtasks is completed: The first stage is the reasoning stage: based on the current atomic subtask and environmental context, a decision-making thinking chain is generated and a structured decision option ({action type: operator selection, candidate operator: ["NDWI_V2.1", "NDWI_Lite"]}) is output. When no operator is matched in the operator library, the code generation sandbox is triggered and verified. After the verification is completed, the newly generated operator is registered with the operator library and added to the process.

[0090] The second stage is the action stage: calling the improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The specific steps include: (1) Based on the dynamic topology reconstruction mechanism, the data dependency relationship between atomic subtasks is analyzed, and the directed acyclic graph (DAG) topology structure is reconstructed by splitting and merging branches. The atomic subtasks without data coupling are split into independent execution flows to realize multi-operator parallel computing.

[0091] (2) Path cost optimization; The path cost function based on the improved A* algorithm is: Formula 6 Formula 7 Among them, Cost is the path cost of the operator; F is the historical average time consumed by the operator (seconds); rate is the historical failure rate of the operator (0~1); Rcomp is the real-time resource competition coefficient; is the time cost weight (initial value is 1.0, and the value is 1.5 when GPU / CPU utilization exceeds 80%); is the failure penalty weight (the default value is 2.0, which is configurable); k, m, and n are coefficients, which can be set. The initial values ​​are k=0.8, m=0.5, and n=0.7. For CPU operators, k=0, m=0.7, and n=0.8; for GPU operators, util is the utilization of GPU; cpu util is the CPU utilization; mem util The memory utilization.

[0092] (3) Path search process; Construct a directed acyclic graph (DAG) with atomic subtasks as nodes, initialize the open list, iteratively expand the path, and select the node with the smallest cost function f(n) each time.

[0093] Formula 8 Among them, f(n) is the total cost of executing the current node and completing the task; g(n) is the actual cost from the starting point to the current node; h(n) is the estimated cost of the remaining subtasks (that is, the cost from the current node to the end node).

[0094] When resource conflicts are detected (e.g. GPU, CPU, memory utilization exceeds 80%), an adaptive resource scheduling strategy is implemented to dynamically adjust the real-time resource contention coefficient R. comp and reorder them.

[0095] (4) Operator matching; For each operator subtask, select the operator with the smallest path cost from the operator candidate pool.

[0096] Dynamic topology reconstruction mechanism: When the cumulative path cost function value exceeds the preset threshold, the data dependencies between atomic subtasks are analyzed, and the directed acyclic graph (DAG) topology structure is reconstructed by splitting and merging branches. The atomic subtasks without data coupling are split into independent execution flows to achieve multi-operator parallel computing.

[0097] Adaptive resource scheduling strategy: real-time resource competition coefficient R comp Dynamically refresh based on the cluster's real-time load (refresh cycle ≤ 5 seconds). When the GPU / CPU / memory utilization is monitored to be greater than or equal to 80% for 30 seconds, the real-time resource competition coefficient R is automatically increased. comp The value increases to 1.5 times the baseline value, triggering the operator degradation strategy (replacing GPU operators with CPU operators).

[0098] The third phase is the observation phase: verifying operator matching constraints [including data compatibility (upstream input data type, auxiliary data information, spatial reference system, output data type), resource feasibility (GPU, memory, number of CPU cores), and algorithm dependency satisfaction (third-party library version)], and generating an exception report when a conflict is detected (for example: {conflict type: insufficient resources, operator: NDWI_V2.1, required GPU: 4GB, available GPU: 3.2GB}).

[0099] The fourth stage is the feedback adjustment stage: dynamic adaptation is triggered according to the conflict type, mainly in the following four aspects: Insert a downgrade operator when resource conflicts occur (when the GPU does not meet the requirements, use the CPU for calculation).

[0100] Insert format conversion operators (TIFF to JPEG / TIFF to PNG formats) when data is incompatible.

[0101] Insert an automatic conversion operator (float32 to UINT16 type) when the data bit depth is incompatible.

[0102] For other conflict types that are not preset, the conflict report is added to the environment context and returned to the first stage.

[0103] In step S6, if no operator is matched in the operator library, the code is triggered to generate a sandbox and verify it. After the verification is complete, the newly generated operator is registered with the operator library and added to the process. The specific steps include: S61, build code to generate prompt words (for example: you are an expert in radiation correction in the field of remote sensing. The input data is: *.tif, the data metadata is: {number of bands: 3, bit depth: 8, band order: BGR}, resource requirements are: {CPU: 4, memory: 500MB}, language requirement: Python, please write an operator to complete the radiation correction of the data).

[0104] S62, inputting the code generation prompt words into the Qwen3 large model to complete the code writing, and performing code checking and code running checking after the code writing is completed.

[0105] S63: After the check is completed, the operator registration is automatically completed through the grayscale release mechanism. If the check fails, the code is manually corrected and re-checked after the modification is confirmed.

[0106] S7, after the visualization process module of the interaction layer receives the generated process, it is visualized on the user interface through a directed acyclic graph (DAG), supporting users to modify the operator order and operator configuration parameters in the process by dragging and dropping and using forms.

[0107] S8, the dynamic process execution engine allocates operator scheduling priorities based on the operator's resource requirements, dependencies, and current system resources, monitors the operator's execution status, dynamically adjusts resource allocation, and implements directed acyclic graph (DAG) task scheduling. After completing every three atomic subtasks or taking 120 seconds, the engine automatically saves the intermediate state [the state between the start and end of the entire process execution, the execution status of a certain operator or certain operators (execution success, execution failure, executing, etc.)] to the distributed cache Redis. If an operator fails during process execution (for example, due to insufficient memory), the engine backtracks to the most recent checkpoint and calls the process automation generation module to recommend an alternative operator for continued execution. At the same time, it records the failure log and updates historical performance indicators.

[0108] Since the process only specifies the execution order of operators in the current process, and when multiple processes are issued at the same time, it is necessary to calculate the scheduling priority of the operator based on the current resource status of the system. Once the scheduling priority of the operator is determined, the operator is directly scheduled to the corresponding node for execution.

[0109] The priority calculation of operator scheduling adopts a three-level priority calculation model, and the priority weight formula is: Formula 9 Among them, Priority is the scheduling priority of the operator in this task process; DepPriority is the dependency priority. According to the node types in the directed acyclic graph (DAG), they are divided into head nodes, independent nodes (predecessor nodes have been completed), and ordinary nodes (predecessor nodes have not been completed). The DepPriority value of the head node is 1.0, the DepPriority value of the independent node is 0.8, and the DepPriority value of the ordinary node is 0.5. ResPriority is the resource priority, and the calculation formula is: Formula 10 ReqResource is the normalized resource value requested by the operator (weighted sum of GPU memory + number of CPU cores + Mem); ClusterMax is the weighted sum of the maximum resource capacity of a single node in the system; DynamicAdj is a dynamic adjustment factor with an initial value of 0. When cluster idle resources (memory / CPU / GPU utilization) are greater than 50%, the value is 0.1. When the task is urgent, the value is 0.3.

[0110] Dynamic resource allocation adjustment means that after an operator completes execution, the system reclaims the allocated resources, recalculates the scheduling priorities of operators to be executed after the operator in the directed acyclic graph (DAG), and allocates and executes resources based on their scheduling priorities.

[0111] This invention incorporates the concepts and technologies of intelligent agents (AI Agents) into the intelligent processing of remote sensing data, evolving from the traditional "data input"-"fixed algorithm processing"-"result output" model to a higher-level "goal-driven, autonomous perception, intelligent decision-making, collaborative execution, continuous learning, and closed-loop feedback" intelligent system, providing a powerful technical framework and implementation path for building an automated, intelligent, adaptive, and collaborative next-generation remote sensing information processing platform.

[0112] The contents not described in detail in the specification of the present invention belong to the existing common technologies in this technical field.

Claims

1. A remote sensing image processing system based on intelligent agent, characterized in that: The system architecture includes an operator layer, an interaction layer, a data adaptation layer, and a dynamic process execution engine; the operator layer includes an operator registration module, a code generation module, and a test sandbox module; the interaction layer includes a natural language interaction module, a process automation generation module, and a visual process module; the data adaptation layer includes a data interface specification module and a field adaptation middleware; The operator registration module is used to implement dynamic registration and intelligent retrieval of operator metadata; the code generation module generates operator code by integrating the Qwen3 model and presetting specific templates in the remote sensing field; the generated operator code is unit tested through the test sandbox module, and is saved after passing the test. If it fails, it is manually corrected and then tested again; the test sandbox module performs code inspection and code operation inspection by building a basic verification system and a business verification system. After the inspection is completed, the operator registration is automatically completed through the gray release mechanism; The natural language interaction module uses a hybrid expert model to parse the natural language instructions input by the user, converts them into structured parameters, integrates a confidence assessment mechanism, and scores the confidence of the parsing results. When the confidence is below the threshold, multiple rounds of dialogue are triggered, and semantic completion is achieved by combining with the knowledge graph to achieve accurate understanding of the user's intention. The process automation generation module is used to achieve task decomposition and process generation. The visual process module is used to realize process DAG interaction and real-time process monitoring. The data interface specification module is used to unify data interface specifications; The field adaptation middleware is used to implement automatic detection and mapping of input and output fields, dynamic format conversion and adaptation, and metadata repair; The dynamic process execution engine assigns priorities based on the resource requirements and dependencies of operators, monitors the execution status of operators, and dynamically adjusts resource allocation; When an operator fails to execute in the process, the engine backtracks to the most recent checkpoint and calls the process automation generation module of the interaction layer to recommend an alternative operator to continue execution.

2. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: The operator registration module includes an operator meta-information dynamic registration unit and an intelligent retrieval unit; the operator meta-information dynamic registration unit supports dynamic operator loading and version management, structures the operator meta-information into JSON format, and vectorizes the structured operator meta-information through the Jina-Embeddings-v3 model and stores it in a vector database; The intelligent retrieval unit is based on RAG technology, inputs operator meta information, operator historical execution information and resource consumption, and implements priority sorting retrieval.

3. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: In the test sandbox module: The basic verification system ensures that the operator code is free of basic defects and meets the requirements of operational stability through static code analysis, unit test coverage, and dynamic memory detection; The business verification system conducts end-to-end testing by building a small remote sensing dataset to ensure that the operator's output results in real remote sensing tasks meet business logic and accuracy requirements; The specific implementation process of the grayscale release mechanism is as follows: The new operator first enters "shadow mode" and executes in parallel with the existing operator, comparing the results. When the difference exceeds the threshold, the manual correction process is triggered; After manually correcting the code, the test is completed through code inspection and code operation inspection in the test sandbox. After the test passes, it is added to the process.

4. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: The natural language interaction module uses a hybrid expert model to parse the natural language instructions input by the user and converts them into structured parameters. The specific steps include: Receive user natural language instructions; Extract semantic features through the BERT+BiLSTM model and perform text encoding; Combined with the current remote sensing image metadata, spatial context analysis is performed to complete abnormal context correction and coordinate system constraints, including: Abnormal context correction: If the user's natural language command conflicts with the current remote sensing image metadata context, multiple rounds of dialogue are triggered to confirm the intention; Coordinate system constraints: Convert ambiguous expressions in user natural language instructions into specific geographic coordinate ranges to avoid misunderstandings caused by coordinate system differences; Output structured parameters.

5. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: The confidence assessment mechanism specifically includes: (1) Semantic uncertainty quantification; (2) Spatial context conflict detection; (3) Verification of knowledge graph semantic coverage; (4) Dynamic threshold decision-making.

6. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: The process automation generation module includes a task decomposition unit and a process generation unit; the task decomposition unit decomposes a complex task into an atomic subtask chain based on the thought chain technology; after receiving the atomic subtask chain, the process generation unit automatically matches the operator by inputting the system prompt word, operator meta-information and the operator's historical execution information through the hybrid expert model, and then matches the optimal operator and optimizes the task execution path through the ReAct framework combined with the improved A* algorithm path cost function, sets the operator's configuration parameters and running resources, and generates the optimal process; when the registered operator is not found in the operator library, the code generation sandbox is triggered and verified. After the verification is completed, the newly generated operator is registered with the operator library and added to the process. Finally, the loop termination condition is designed to prevent infinite reasoning; The visual process module includes a process DAG interaction unit and a process real-time monitoring unit; the process DAG interaction unit allows users to adjust the operator order and operator configuration parameters by dragging and dropping and using a form to modify the processing flow; The process real-time monitoring unit supports real-time performance monitoring and marks abnormal nodes.

7. The remote sensing image processing system based on intelligent agent according to claim 6, characterized in that: The task decomposition unit is based on the thought chain technology, and the specific steps of decomposing a complex task into an atomic subtask chain include: Receive the converted structured parameters; Retrieve historical similar process task cases through retrieval enhancement generation technology; The system constructs structural parameters, historical similar process task cases and current remote sensing image metadata as prompt words, builds task execution paths based on thought chain technology, and decomposes complex tasks into atomic subtask chains; Outputs a chain of atomic subtasks.

8. The remote sensing image processing system based on intelligent agent according to claim 1, characterized in that: The field adaptation middleware includes an automatic detection and mapping unit for input and output fields, a dynamic format conversion and adaptation unit, and a metadata repair unit; the automatic detection and mapping unit for input and output fields detects the differences in the names, data types, dimensions, spatial references, and metadata labels of input and output fields between operators in real time; it understands the field semantics through a rule engine or a large model and automatically associates synonymous fields; the dynamic format conversion and adaptation unit automatically inserts forced type conversion logic, band order adjustment, and coordinate system conversion operators when it detects that there is a data type inconsistency between the input and output of adjacent operators; it resolves multidimensional data conflicts through structural adaptation, or unifies resolution through interpolation / resampling; the metadata repair unit inherits and supplements from upstream data or global configuration when key metadata is missing from the operator output, and automatically adds semantic labels to unlabeled data.

9. A remote sensing image processing method based on an intelligent agent, implemented by the remote sensing image processing system based on an intelligent agent according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1, at the operator layer, register the prepared operators through the operator registration module and vectorize and store the operator metadata; S2, the user inputs natural language instructions at the interaction layer; S3, the natural language interaction module of the interaction layer, uses a hybrid expert model to parse the natural language instructions input by the user and convert them into structured parameters. If the confidence level is lower than the threshold, it triggers multiple rounds of dialogue and semantic completion to achieve accurate understanding of the user's intention. S4, after receiving the structured parameters converted by the natural language interaction module, the process automation generation module of the interaction layer decomposes the complex task into atomic subtask chains based on the thought chain technology; S5: The process automation generation module in the interaction layer automatically matches operators using a hybrid expert model input, system prompts, operator metadata, and historical operator execution information. It then uses the ReAct framework combined with an improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The field configuration information of the process is then modified and confirmed using the field adaptation middleware in the data adaptation layer, and the process description file is finally output. S6: When no operator is matched in the operator library, the code is triggered to generate a sandbox and verify it. After the verification is completed, the newly generated operator is registered with the operator library and added to the process; S7, after receiving the generated process, the visualization process module of the interaction layer displays it on the user interface through a directed acyclic graph (DAG). Users can modify the operator order and operator configuration parameters in the process by dragging and dropping and using forms. S8, the dynamic process execution engine allocates operator scheduling priorities based on the operator's resource requirements, dependencies, and current system resources, monitors the operator's execution status, dynamically adjusts resource allocation, and implements directed acyclic graph (DAG) task scheduling. It automatically saves the intermediate state to the distributed cache Redis after completing every three atomic subtasks or taking 120 seconds. If an operator fails during process execution, the engine backtracks to the most recent checkpoint and calls the process automation generation module to recommend an alternative operator for continued execution. It also records the failure log and updates historical performance indicators.

10. The remote sensing image processing method based on intelligent agent according to claim 9, characterized in that: In S5, after receiving the atomic subtask chain, the process automation generation module of the interaction layer automatically matches operators by inputting system prompt words, operator meta-information, and historical execution information of the operators through the hybrid expert model. Then, the ReAct framework is combined with the improved A* algorithm path cost function to match the optimal operator and optimize the task completion path. The specific steps include: S51, builds the initial decision environment, receives the atomic subtask chain, and dynamically loads the environment context; The dynamic loading environment context includes: (1) Real-time status of the operator library; (2) System-level constraints; (3) Data characteristics; S52, the closed-loop decision-making process of the ReAct framework, cyclically executes the following four stages until the operator matching of all atomic subtasks is completed: The first stage is the reasoning stage: based on the current atomic subtask and environmental context, a decision-making chain is generated and structured decision options are output. If no matching operator is found in the operator library, the code is triggered to generate a sandbox and verify it. After verification is complete, the newly generated operator is registered with the operator library and added to the process. The second stage is the action stage: calling the improved A* algorithm path cost function to match the optimal operator and optimize the task execution path. The specific steps include: (1) Based on the dynamic topology reconstruction mechanism, the data dependency between atomic subtasks is analyzed, and the directed acyclic graph (DAG) topology is reconstructed by splitting and merging branches. The atomic subtasks without data coupling are split into independent execution flows to achieve multi-operator parallel computing. (2) Path cost optimization; The path cost function based on the improved A* algorithm is: Among them, Cost is the path cost of the operator; is the historical average time consumed by the operator; F rate is the historical failure rate of the operator; R comp is the real-time resource competition coefficient; is the time cost weight; is the failure penalty weight; k, m, n are coefficients, which can be set; gpu util is GPU utilization; cpu util is the CPU utilization; mem util is the memory utilization; (3) Path search process; Construct a directed acyclic graph (DAG) with atomic subtasks as nodes, initialize the open list, iteratively expand the path, and select the node with the smallest cost function f(n) each time; Where f(n) is the total cost of executing the current node and completing the task; g(n) is the actual cost from the starting point to the current node; h(n) is the estimated cost of the remaining subtasks; When resource conflicts are detected, an adaptive resource scheduling strategy is implemented to dynamically adjust the real-time resource competition coefficient R. comp and reorder them; (4) Operator matching; For each operator subtask, select the operator with the smallest path cost from the operator candidate pool; The third phase is the observation phase: verifying that operators match constraints and generating exception reports when conflicts are detected; The fourth stage is the feedback adjustment stage: dynamic adaptation is triggered according to the conflict type, and is carried out in the following four aspects: Insert degradation operators when resource conflicts occur; Insert format conversion operators when data is incompatible; Insert automatic transposition operator when data bit depth is incompatible; For other conflict types that are not preset, the conflict report is added to the environment context and returned to the first stage.

11. The remote sensing image processing method based on intelligent agent according to claim 9, characterized in that: The specific steps of S6 include: S61, construct code generation prompt words; S62, inputting the code generation prompt words into the Qwen3 large model to complete the code writing, and performing code checking and code running checking after the code writing is completed; S63: After the inspection is completed, the operator is automatically registered through the grayscale release mechanism and added to the process. If the inspection fails, the code is manually corrected and re-inspected after the modification is confirmed.

12. The remote sensing image processing method based on intelligent agent according to claim 9, characterized in that: In S8, the priority calculation of operator scheduling adopts a three-level priority calculation model, and the priority weight formula is: Among them, Priority is the scheduling priority of the operator in this task process; DepPriority is the dependency priority; ResPriority is the resource priority; DynamicAdj is the dynamic adjustment factor.

Images