Multi-agent-driven remote sensing image intelligent processing algorithm autonomous training verification system

Abstract

The invention relates to a multi-agent-driven remote sensing image intelligent processing algorithm autonomous training verification system. The system comprises a task input and large model agent planning module, a data screening and preprocessing module, an algorithm model screening and super parameter configuration module, a model training and testing module and a comparative analysis and report generation module. According to the method, the whole process of data preparation, algorithm matching, parameter configuration, parallel training, test verification and report generation can be intelligently completed through multi-agent cooperative work, manual intervention is remarkably reduced, the model research and development efficiency and quality are improved, and it is ensured that the process is traceable and the result is reproducible.

Description

Technical Field

[0001] This invention relates to the field of intelligent remote sensing image technology, specifically to an autonomous training and verification system for a multi-agent driven intelligent remote sensing image processing algorithm. Background Technology

[0002] In fields such as military reconnaissance, emergency rescue, environmental monitoring, space debris monitoring, and astronomical observation, the demand for intelligent remote sensing image processing models is becoming increasingly diverse and specialized. Efficient model training and validation are crucial for the practical application of these technologies. However, existing training and validation processes for intelligent remote sensing image processing models suffer from significant bottlenecks, primarily in the following aspects:

[0003] The process is complex and lacks automation and intelligence: Traditional model training involves multiple stages such as data screening, preprocessing, algorithm selection, hyperparameter configuration, training, and testing. These stages are usually operated independently by humans, and there is a lack of unified and intelligent connection and scheduling between the stages, which can easily lead to problems such as configuration errors, version confusion, and process interruption. It relies heavily on the personal experience and repetitive work of algorithm engineers.

[0004] Inefficiency and low resource utilization: A large amount of manpower is spent on repetitive tasks such as data preprocessing, hyperparameter tuning, and result comparison, making it difficult to reflect the innovative value of high-level algorithm engineers. At the same time, the allocation of computing resources (such as GPUs) is often static or based on simple rules, and cannot be dynamically optimized according to task priority and real-time status, resulting in idle resources or resource contention.

[0005] The lack of integrated, low-barrier intelligent platforms: Existing tools or platforms (such as some AutoML tools) either require users to have high programming skills and algorithmic knowledge, or their functions are relatively limited and cannot cover the entire lifecycle of remote sensing image processing models, from task understanding to report generation. For domain experts (such as remote sensing interpretation experts), they are well-versed in business needs and data characteristics, but find it difficult to directly translate their professional knowledge into an executable model training process.

[0006] To address the above problems, existing technologies have proposed some improvement solutions.

[0007] For example, Chinese invention patent CN117876840A discloses a method and system for rapid training of remote sensing basic models based on template editing. This solution allows users to select templates according to task requirements by pre-defining templates that integrate multiple algorithm frameworks and mirrors, thereby quickly configuring and starting training tasks, and provides visual monitoring functions. This technology reduces the complexity of environment configuration to a certain extent.

[0008] For example, Chinese invention patent CN120803471A discloses a method for automated orchestration and deployment of artificial intelligence algorithms with visual interaction. This method introduces an agent based on a large language model to understand user tasks, automates the deployment process of existing algorithms, generates a directed acyclic graph (DAG), and uses containerization technology for resource scheduling and task execution.

[0009] However, the aforementioned existing technologies still have the following shortcomings:

[0010] (1) Lack of end-to-end intelligent driving: Chinese invention patent with publication number CN117876840A focuses on simplifying configuration through preset templates, but its process initiation still relies on manual selection and assembly, and does not have the fully autonomous and intelligent driving capability from natural language task description to the final model report generation. It does not solve the automation problem of upstream key links such as task understanding, intelligent data matching, and automatic algorithm selection and optimization.

[0011] (2) Failure to achieve specialized multi-agent collaboration for remote sensing model training: Although the Chinese invention patent with publication number CN120803471A introduces intelligent agents for process orchestration, its application scenario focuses on the deployment process orchestration of existing algorithms or models, rather than the full lifecycle management of model training from scratch for the specific field of remote sensing image processing. It does not design dedicated intelligent agents (such as data screening and analysis agents, preprocessing strategy configuration agents) for the characteristics of remote sensing data, nor does it construct a collaborative decision-making and execution system that runs through the data, algorithm, training, and testing stages.

[0012] (3) Insufficient depth of intelligent decision-making: Existing solutions rely heavily on rule bases or simple matching for core decision-making points such as algorithm selection and hyperparameter configuration. They lack the ability to use advanced technologies such as Large Language Models (LLM) to deeply analyze task requirements and data characteristics and generate diverse and interpretable candidate strategies. At the same time, they lack unified monitoring and intelligent optimization suggestions for the parallel training process.

[0013] (4) Limited traceability and reproducibility support: Existing systems often do not record the data version, parameter configuration and intermediate results in the training process, making it difficult to reproduce the model results, which is not conducive to algorithm iteration and experience accumulation.

[0014] Therefore, there is an urgent need for a solution that can deeply integrate large-scale model intelligence with professional domain knowledge to achieve a high degree of automation, intelligence and adaptability in the entire process of remote sensing image processing algorithms from task input to verification report. Summary of the Invention

[0015] In view of the above-mentioned technical problems, this invention proposes a multi-agent driven autonomous training and verification system for intelligent remote sensing image processing algorithms. Through the collaborative work of multiple agents, it can automatically analyze user task requirements and intelligently complete the entire process of data preparation, algorithm matching, parameter configuration, parallel training, testing and verification, and report generation. This significantly reduces manual intervention, improves the efficiency and quality of model development, and ensures that the process is traceable and the results are reproducible.

[0016] The technical solution to the technical problem of this invention is: a multi-agent driven remote sensing image intelligent processing algorithm autonomous training and verification system, comprising:

[0017] The task input and large model agent planning module is used to receive training task instructions input by the user, and to parse task requirements, generate structured task chains and arrange and schedule task processes through the large model overall planning agent.

[0018] The data filtering and preprocessing module is used to automatically filter remote sensing image training samples, statistically analyze data, generate and execute data preprocessing strategies according to task requirements, so as to output a standardized training dataset.

[0019] The algorithm model selection and hyperparameter configuration module is used to automatically select matching baseline algorithms and pre-trained models, and automatically generate hyperparameter configuration strategies and candidate configuration lists based on task requirements and data characteristics.

[0020] The model training and testing module is used to automatically allocate resources, perform parallel training, monitor the training process, and automatically test and record the trained model based on the candidate configuration list and available computing power.

[0021] The comparative analysis and report generation module is used to comprehensively analyze the test results of all training tasks and automatically generate model performance evaluation reports.

[0022] According to one technical solution of the present invention, the task input and large model agent planning module further includes:

[0023] The Executive Director intelligent agent is used to monitor task execution progress, pass parameters and resources between modules, and handle task execution exceptions.

[0024] According to one technical solution of the present invention, the data filtering and preprocessing module includes:

[0025] The training data filtering and analysis agent is used to generate structured filtering conditions and acquire data according to task requirements, as well as to perform multi-dimensional statistical analysis on the data.

[0026] A preprocessing strategy configuration agent is used to automatically generate data preprocessing strategy configuration files based on data analysis results.

[0027] A preprocessing execution agent is used to perform preprocessing operations according to the policy configuration file to generate a standardized dataset.

[0028] According to one technical solution of the present invention, the algorithm model selection and hyperparameter configuration module includes:

[0029] Algorithm model selection agent is used to parse task requirements using a large language model and automatically match and recommend algorithms and models;

[0030] Train a hyperparameter policy configuration agent to automatically generate a list of candidate hyperparameter configurations and configuration files based on task, data, and algorithm information.

[0031] According to one technical solution of the present invention, the model training and testing module includes:

[0032] A training computing power planning and allocation agent is used to allocate training resources based on the candidate configuration list and the system computing power.

[0033] An automated training execution agent is used to receive datasets, algorithm files, configuration files, and computing resources, and to perform model training.

[0034] A training process supervisory agent is used to monitor the training status in real time, analyze training process information, and provide optimization suggestions.

[0035] The model testing and recording agent is used to automatically configure the test environment, call the test dataset to perform model verification, and record the test results in a structured manner.

[0036] According to one technical solution of the present invention, the comparative analysis and report generation module includes:

[0037] The training task result analysis agent is used to summarize and compare the training and testing data of all sub-tasks horizontally, and output a ranking list of models.

[0038] The model description report generation agent is used to automatically retrieve traceability information and generate a model description report according to the template based on the model selected by the user.

[0039] According to one aspect of the present invention, a method for autonomous training and verification of a multi-agent driven intelligent remote sensing image processing algorithm is proposed, comprising the following steps:

[0040] Step S1: Receive training task instructions input by the user in natural language or template format, automatically parse the task requirements through the large model overall planning intelligent agent, generate a structured task chain, and perform task flow orchestration and scheduling.

[0041] Step S2: Based on the parsed task requirements, the data filtering agent automatically filters remote sensing image training samples that meet the conditions from the database and performs statistical analysis; based on the analysis results, the preprocessing configuration agent automatically generates a data preprocessing strategy, and the preprocessing execution agent executes the strategy to generate a standardized training dataset.

[0042] Step S3: The algorithm model screening agent automatically selects matching baseline algorithms and pre-trained models from the algorithm library and model library according to the task requirements; the hyperparameter configuration agent automatically generates diverse hyperparameter configuration strategies based on the task requirements, dataset features and selected algorithms, forming a candidate configuration list.

[0043] Step S4: The training computing power planning agent automatically allocates resources and schedules the parallel training of multiple models based on the candidate configuration list and available computing power resources; during the training process, the training monitoring agent monitors the training status and performance in real time; after the training is completed, the model testing agent automatically calls the test dataset to verify the model and records the test results.

[0044] Step S5: Conduct a comprehensive comparative analysis of the test results of all training sub-tasks and generate a model description report that includes model performance evaluation and optimization suggestions.

[0045] According to one technical solution of the present invention, step S1 specifically includes:

[0046] Step S11: The overall planning agent of the large model performs semantic parsing on the task text input by the user and extracts key information such as task type, data requirements and resource constraints.

[0047] Step S12: Based on the parsed key information, generate a structured task chain containing multiple ordered subtasks;

[0048] Step S13: Assign each subtask in the structured task chain to the corresponding execution agent module and generate a visual task flow script;

[0049] Step S1 further includes:

[0050] An executive director agent is set up to monitor the execution progress of the structured task chain, pass task parameters and resources between the execution agents, and trigger a retry mechanism for failed task stages.

[0051] According to one technical solution of the present invention, in step S2, the data filtering agent extracts keywords through natural language processing, generates structured filtering conditions including file format, resolution and data type, and retrieves data from the database based on these conditions;

[0052] The data filtering agent performs multi-dimensional statistical analysis on the acquired dataset to automatically identify problems such as missing samples or class imbalance.

[0053] The preprocessing strategy includes at least one of image cropping and sample augmentation.

[0054] According to a technical solution of the present invention, in step S3, the algorithm model screening agent uses a large language model to parse the task requirements, generate detailed screening conditions, and automatically match candidate solutions in the algorithm library and model library.

[0055] The hyperparameter configuration agent automatically generates hyperparameter configuration strategies, including model structure parameters and training strategy parameters, based on task requirements, data characteristics, and algorithm architecture, and forms a batch training candidate configuration list sorted by priority.

[0056] In step S4, the training computing power planning agent automatically selects executable training subtasks from the candidate configuration list based on the algorithm type and available computing power, and allocates computing resources to each subtask.

[0057] During model training, the training monitoring agent collects and visualizes model performance metrics, resource usage, and training process curves in real time.

[0058] The model testing agent automatically calls the standardized test dataset corresponding to the task to test the trained model and records the test results in a structured manner, supporting result tracking and reproduction.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] The present invention discloses a multi-agent driven autonomous training and verification system for remote sensing image intelligent processing algorithms. By introducing a large model chief planning agent and an executive director agent, the system can understand user natural language commands and intelligently decompose, arrange and monitor the entire training and verification process, freeing algorithm engineers from tedious and repetitive process operations, and realizing end-to-end full-process automation and intelligence.

[0061] This invention constructs a specialized multi-agent collaborative system for the remote sensing field. For key aspects of remote sensing image processing model training, it designs multiple dedicated agents for data filtering and analysis, preprocessing strategy configuration, algorithm model selection, hyperparameter configuration, training monitoring, and model testing. These agents independently perform their own functions while also collaborating, deeply integrating domain knowledge with AI decision-making capabilities, significantly improving the specialization and intelligence of each stage.

[0062] The system of this invention supports parallel training based on candidate lists with diverse hyperparameter configurations, and dynamically allocates resources by the training computing power planning agent, realizing an efficient experimental mode of "one-time configuration, multiple parallel groups", which greatly shortens the model iteration cycle, improves the utilization rate of expensive computing resources such as GPUs, and thus improves the efficiency of model development and resource utilization.

[0063] This invention provides a structured record and management of all key data, configuration parameters, intermediate results, and logs throughout the entire process. The final model description report meticulously records the data source, hyperparameters, training process, testing performance, and resource consumption, ensuring complete traceability of the experiment and accurate model reproduction. This strongly supports continuous algorithm iteration and team collaboration, thereby enhancing process traceability and result reproducibility.

[0064] This invention enables domain experts to drive complex model training and verification processes without needing in-depth knowledge of programming and algorithm details through natural language interaction and a visual interface. This promotes the rapid application and popularization of artificial intelligence technology in vertical fields such as remote sensing, lowers the professional technical threshold, and enhances the user experience. Attached Figure Description

[0065] Figure 1 This is a schematic diagram of the composition of a multi-agent driven remote sensing image intelligent processing algorithm autonomous training and verification system in one embodiment of the present invention.

[0066] Figure 2 This is a schematic diagram illustrating the workflow of a multi-agent driven remote sensing image intelligent processing algorithm autonomous training and verification system in one embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0068] like Figure 1 and Figure 2 As shown, this invention provides a multi-agent (large model agent) driven autonomous training and verification system and method for remote sensing image intelligent processing algorithms. Its core lies in constructing an autonomous system where multiple agents work collaboratively. These agents are divided into planning, execution, and supervision layers according to their functions, jointly completing a complete closed loop from user task input to final report generation.

[0069] The present invention provides an autonomous training and verification system for a multi-agent driven remote sensing image intelligent processing algorithm, comprising:

[0070] The task input and large model agent planning module is used to receive training task instructions input by the user, and to parse task requirements, generate structured task chains and arrange and schedule task processes through the large model overall planning agent.

[0071] The data filtering and preprocessing module is used to automatically filter remote sensing image training samples, statistically analyze data, generate and execute data preprocessing strategies according to task requirements, so as to output a standardized training dataset.

[0072] The algorithm model selection and hyperparameter configuration module is used to automatically select matching baseline algorithms and pre-trained models, and automatically generate hyperparameter configuration strategies and candidate configuration lists based on task requirements and data characteristics.

[0073] The model training and testing module is used to automatically allocate resources, perform parallel training, monitor the training process, and automatically test and record the trained model based on the candidate configuration list and available computing power.

[0074] The comparative analysis and report generation module is used to comprehensively analyze the test results of all training tasks and automatically generate model performance evaluation reports.

[0075] In some embodiments of the present invention, the task input and large model agent planning module further includes: an executive director agent, used to monitor task execution progress, transfer parameters and resources between modules, and handle task execution anomalies.

[0076] The Executive Director Agent occupies a central coordinating position in the architecture, communicating with all execution-layer Agents via event-driven mechanisms or message queues. When a training execution Agent reports a task failure, the Executive Director Agent first attempts to retry according to a pre-defined strategy. If the retry fails several times, the task is marked as failed, and the lead planning Agent and the user are notified. Simultaneously, it ensures that the output of upstream agents is accurately transmitted to downstream agents.

[0077] In some embodiments of the present invention, the data filtering and preprocessing module includes:

[0078] The training data filtering and analysis agent is used to generate structured filtering conditions and acquire data according to task requirements, as well as to perform multi-dimensional statistical analysis on the data.

[0079] A preprocessing strategy configuration agent is used to automatically generate data preprocessing strategy configuration files based on data analysis results.

[0080] A preprocessing execution agent is used to perform preprocessing operations according to the policy configuration file to generate a standardized dataset.

[0081] In some embodiments of the present invention, the algorithm model selection and hyperparameter configuration module includes:

[0082] Algorithm model selection agent is used to parse task requirements using a large language model and automatically match and recommend algorithms and models;

[0083] Train a hyperparameter policy configuration agent to automatically generate a list of candidate hyperparameter configurations and configuration files based on task, data, and algorithm information.

[0084] In some embodiments of the present invention, the model training and testing module includes:

[0085] A training computing power planning and allocation agent is used to allocate training resources based on the candidate configuration list and the system computing power.

[0086] An automated training execution agent is used to receive datasets, algorithm files, configuration files, and computing resources, and to perform model training.

[0087] A training process supervisory agent is used to monitor the training status in real time, analyze training process information, and provide optimization suggestions.

[0088] The model testing and recording agent is used to automatically configure the test environment, call the test dataset to perform model verification, and record the test results in a structured manner.

[0089] In some embodiments of the present invention, the comparative analysis and report generation module includes:

[0090] The training task result analysis agent is used to summarize and compare the training and testing data of all sub-tasks horizontally, and output a ranking list of models.

[0091] The model description report generation agent is used to automatically retrieve traceability information and generate a model description report according to the template based on the model selected by the user.

[0092] According to one aspect of the present invention, a method for autonomous training and verification of a multi-agent driven intelligent remote sensing image processing algorithm is proposed, comprising the following steps:

[0093] Step S1: User command input and intelligent parsing for remote sensing image intelligent processing task.

[0094] It receives training task instructions from users in natural language or template format, automatically parses task requirements through the large model overall planning intelligent agent, generates structured task chains, and performs task flow orchestration and scheduling.

[0095] This stage is mainly completed by the task input and large model intelligent agent planning module, whose goal is to understand the user's intent and plan the execution scheme.

[0096] Users can input training task instructions through the system's web interface or API, either in natural language (e.g., "Train a model for detecting changes in urban buildings, requiring the use of satellite imagery with a resolution better than 1 meter from the past three years") or using structured templates.

[0097] The large-scale model master planner agent (referred to as the "master planner agent") receives user instructions, automatically parses task requirements, breaks down and orchestrates the intelligent image processing algorithm training task flow, and generates a structured task chain. If the user input is irrelevant to model training, no plan is generated. This agent is built on a large language model and possesses powerful semantic understanding capabilities. Its workflow is as follows:

[0098] (1) Semantic parsing and key information extraction: The chief planning agent performs deep semantic parsing on the input task text to identify and extract key information such as task type (e.g., change detection, target recognition, land cover classification, etc.), data requirements (e.g., image type, resolution, time range, region), performance expectations, and resource constraints (e.g., maximum training time, available GPU memory). For example, from the example instruction above, the task type can be extracted as "change detection", the data requirements are "last three years", "resolution better than 1 meter", "satellite image", and the object is "urban buildings".

[0099] (2) Structured Task Chain Generation: Based on the extracted key information, the chief planning agent generates a structured task chain. This task chain defines the ordered sequence of subtasks required to complete this training task, for example: [Subtask 1: Data Filtering] → [Subtask 2: Data Preprocessing] → [Subtask 3: Algorithm Selection] → [Subtask 4: Hyperparameter Configuration] → [Subtask 5: Model Training] → [Subtask 6: Model Testing] → [Subtask 7: Report Generation]. Each subtask is associated with specific inputs, outputs, and constraints.

[0100] (3) Task Distribution and Visualization: The chief planning agent distributes the structured task chain to the executive director agent, and presents the task chain in a user-readable visual flowchart on the human-computer interaction interface, supporting manual review and adjustment by the user. If the user input is irrelevant to model training, the process terminates and prompts the user.

[0101] The Executive Director Agent, acting as the central controller for process execution, is responsible for receiving the structured task chain generated by the Chief Planner Agent. Its main uses include:

[0102] Task parameter transmission: Accurately transmit the parameters required by each link in the task chain (such as data filtering conditions and algorithm identifiers) to the corresponding execution layer Agent.

[0103] Resource matching and coordination: Based on task requirements, coordinate and match the computing and storage resources required by each executing agent.

[0104] Process progress monitoring: Real-time monitoring of the execution status of the entire task chain.

[0105] Anomaly Handling and Retry: When an Agent fails to execute a certain step (such as data acquisition timeout or training process crash), the Executive Director Agent automatically triggers the retry mechanism (such as rescheduling tasks or using backup data) and rolls back to the previous stable state to ensure the continuity and reliability of the task process, and reports the anomaly information to the Chief Planner Agent or the user.

[0106] The overall planning agent enables precise understanding and structured decomposition of fuzzy and complex natural language task instructions, transforming user requirements into a sequence of instructions that machines can execute precisely. The introduction of the executive director agent ensures the reliable progress and anomaly recovery of multi-stage, long-process tasks, forming the cornerstone of stable system operation and achieving intelligent task planning.

[0107] Step S2: Data screening and preprocessing for remote sensing image intelligent processing tasks.

[0108] Based on the parsed task requirements, the data filtering agent automatically selects training samples of remote sensing images that meet the conditions from the database and performs statistical analysis. Based on the analysis results, the preprocessing configuration agent automatically generates a data preprocessing strategy, and the preprocessing execution agent executes the strategy to generate a standardized training dataset.

[0109] Step S2 is performed by the data filtering and preprocessing module, which aims to automatically acquire and prepare a high-quality training dataset that meets the task requirements. The data filtering and preprocessing module mainly includes three agents: a training data filtering and analysis agent, a preprocessing strategy configuration agent, and a preprocessing execution agent.

[0110] The training data filtering and analysis agent receives task requirements from the executive director agent, which are then parsed by the chief planner agent. Its workflow is as follows:

[0111] (1) Generation of structured filtering conditions: The agent uses natural language processing technology to transform the unstructured data requirements in the task into structured filtering conditions that can be queried in the database. The filtering conditions also include file format, storage path mode, etc.

[0112] (2) Data retrieval and acquisition: Based on the generated structured filtering conditions, the Agent initiates a query request to the connected remote sensing image database or data management platform to retrieve and obtain a list of datasets that meet the conditions and related metadata.

[0113] (3) Multidimensional statistical analysis of data: Automated multidimensional statistical analysis of the acquired dataset. Statistical content includes, but is not limited to: total number of images, sample quantity and distribution of each target category (e.g., buildings, roads, water bodies), proportion of images from different data sources or resolutions, and image size distribution. This analysis aims to automatically identify potential problems in the dataset, such as severe imbalance of sample categories, missing samples in certain key categories, and overly singular data sources.

[0114] (4) Output of analysis results: The original data storage location information, dataset metadata and multi-dimensional statistical analysis report are encapsulated into structured formats such as JSON and sent to the preprocessing strategy configuration agent and the executive director agent.

[0115] The preprocessing strategy configuration agent receives data analysis results from the training data filtering and analysis agent. Its workflow is as follows:

[0116] (1) Policy knowledge retrieval: The Agent has a built-in or connected preprocessing knowledge base containing experience in the field of remote sensing image processing. Based on the data analysis results (such as identifying problems such as sample imbalance and inconsistent image sizes), it retrieves or combines the corresponding preprocessing strategies from the knowledge base to generate them.

[0117] (2) Preprocessing strategy generation: Automatically generate one or more data preprocessing strategies. The strategy content is specific. For example, for the problem of imbalanced samples, the strategy may include data augmentation operations such as "image rotation, mirror flipping, and color dithering" for a minority of categories; for the problem of image size, the strategy may include "uniformly cropping to 512x512 pixels". The strategy also specifies the calling order of each preprocessing tool (such as OpenCV functions and custom scripts), the parameter configuration of each step (such as cropping size and rotation angle range), and the estimated resource requirements.

[0118] (3) Configuration file generation: The generated preprocessing strategy is output as a standardized configuration file, supporting JSON or YAML format, which facilitates subsequent Agent parsing and execution.

[0119] The preprocessing execution agent receives scheduling instructions from the execution director agent, the raw data path provided by the training data filtering and analysis agent, and the policy configuration file generated by the preprocessing policy configuration agent. Its workflow is as follows:

[0120] (1) Task execution and resource allocation: Based on the priority and the limit on the number of parallel tasks issued by the Executive Director Agent, schedule computing resources.

[0121] (2) Standardized dataset generation: Following the steps and parameters in the strategy configuration file, the corresponding preprocessing tools are called sequentially to process the raw data. The processing is logged in real time.

[0122] (3) Exception handling and feedback: Automatically alert to exceptions such as resource conflicts and tool call failures that occur during the processing, and trigger retry mechanism or rollback operation.

[0123] (4) Output of results: Finally, a standardized dataset that can be directly used for model training is generated, and the dataset storage path and other information are returned to the Executive Director Agent for use in subsequent stages.

[0124] Through the collaborative action of three intelligent agents, the data preparation process is fully automated and intelligent. It not only accurately locates the required data but also intelligently diagnoses data quality issues and automatically generates corresponding strategies, ensuring the quality of the data input to the model. This provides a fundamental guarantee for the subsequent model training effect and significantly reduces the time users spend on data cleaning and preparation.

[0125] Step S3: Selection of algorithm models and configuration of hyperparameters for intelligent remote sensing image processing tasks.

[0126] The algorithm model screening agent automatically selects matching baseline algorithms and pre-trained models from the algorithm and model libraries according to task requirements; the hyperparameter configuration agent automatically generates diverse hyperparameter configuration strategies based on task requirements, dataset features, and the selected algorithm, forming a candidate configuration list.

[0127] Step S3 is executed by the algorithm model selection and hyperparameter configuration module, whose purpose is to automatically match the most suitable algorithm architecture and configure diverse training parameters for the current task. The algorithm model selection and hyperparameter configuration module mainly includes an algorithm model selection agent and a training hyperparameter strategy configuration agent.

[0128] The algorithm model filters the Agent, which receives task requests from the master planning Agent. Its workflow is as follows:

[0129] (1) Detailed filtering conditions generation: The large language model is used to perform a deeper analysis of the task requirements and generate detailed conditions for querying the algorithm library and model library.

[0130] (2) Intelligent matching and recommendation: Based on refined conditions, matching queries are performed from algorithm and model libraries that integrate various deep learning framework models and traditional machine learning algorithms. Multiple baseline algorithms and pre-trained models with high matching degrees are retrieved.

[0131] (3) Visualized Result Output: The retrieved candidate algorithms and models, along with their matching scores, algorithm descriptions, advantages and disadvantages, and applicable scenarios, are displayed in a visual format on the human-computer interaction interface. The system supports user intervention, allowing users to change the final selected algorithm and model by adjusting query conditions or making direct selections.

[0132] The training hyperparameter strategy configuration agent receives task information from the executive director agent, dataset information from the training data filtering and analysis agent, and information on the algorithm model selected by the algorithm model filtering agent. Its workflow is as follows:

[0133] (1) Multi-factor comprehensive analysis: The Agent comprehensively analyzes the characteristics of the task, the dataset, and the selected algorithm model architecture.

[0134] (2) Hyperparameter strategy generation: Based on the analysis results, a variety of hyperparameter configuration strategies are automatically inferred and generated. These strategies cover model structure parameters and training strategy parameters. For example, for small datasets, a conservative strategy of "using a small learning rate, strong data augmentation, and early stopping strategy" may be generated; for large datasets, an aggressive strategy of "using a large batch size and learning rate warm-up" may be generated.

[0135] (3) Candidate List and File Generation: The generated hyperparameter strategies are combined into a candidate configuration list and prioritized according to the estimated effect or risk. At the same time, a corresponding hyperparameter configuration file (TXT, JSON or YAML format) that can be directly read by the training script is generated for each strategy in the list. The candidate list is also presented to the user in a visual way, allowing the user to view, adjust and confirm it.

[0136] Intelligent matching enhanced by large models can quickly locate potential options from massive algorithms; and the automatic generation of strategies based on multi-dimensional information can systematically and exploratoryly construct the hyperparameter space, avoiding the blindness and limitations of manual debugging.

[0137] Step S4: Training and testing of the remote sensing image intelligent processing task model.

[0138] The training computing power planning agent automatically allocates resources and schedules the parallel training of multiple models based on the candidate configuration list and available computing power resources; during the training process, the training monitoring agent monitors the training status and performance in real time; after the training is completed, the model testing agent automatically calls the test dataset to verify the model and records the test results.

[0139] Step S4 is executed by the model training and testing module, with the aim of efficiently and reliably executing multiple training tasks and objectively evaluating model performance. The model training and testing module comprises four agents: a training computing power planning and allocation agent, an automated training execution agent, a training process supervision agent, and a model testing and recording agent.

[0140] The training computing power planning and allocation agent receives information from the executive director agent, the data filtering and analysis agent, and the algorithm model filtering agent. Its workflow is as follows:

[0141] (1) Resource awareness and subtask selection: The agent is aware of the status of the available computing resources pool in the system in real time (such as the number of idle GPUs, memory size, etc.). Combining the priority of the hyperparameter candidate configuration list generated in step S3 and the estimated resource consumption of each configuration, the agent automatically selects a set of training subtasks that can be executed in parallel under the current resource conditions.

[0142] (2) Dynamic resource allocation: Allocate computing resources reasonably for each executable training subtask and generate resource allocation vouchers.

[0143] (3) Task scheduling: Send the resource allocation scheme and task start instructions to the automated training execution agent through the message queue or callback interface.

[0144] The automated training execution agent receives dataset files from the data filtering and analysis agent, algorithm files from the algorithm model filtering agent, hyperparameter configuration files from the hyperparameter configuration agent, and computing resources allocated to the training computing power planning and allocation agent. Its workflow is as follows:

[0145] (1) Environment preparation and training start-up: Based on the allocated resources, in an independent container or computing environment, automatically configure Python dependency libraries, deep learning framework version, load data, algorithm code and hyperparameter configuration, and start the model training process.

[0146] (2) Multi-group parallel training: Supports the simultaneous initiation of parallel training of multiple models based on various hyperparameter configurations selected by the user, which greatly improves experimental efficiency.

[0147] (3) Process monitoring and log push: During the training process, training logs and resource usage are collected in real time and pushed to the training process supervision agent in real time. At the same time, the progress bar and curve on the task monitoring page are dynamically updated for users to view in real time.

[0148] (4) Results and Abnormal Output: After training is completed, the final model weight file, training log, and final evaluation metrics are output. If a failure occurs during training, a detailed error log and the reason for the interruption are output and reported to the Executive Director Agent for abnormal handling.

[0149] The training process supervision agent monitors all running training tasks uniformly based on the list of training subtasks. This includes the following processes:

[0150] (1) Real-time data acquisition: Real-time collection of performance metrics, resource usage data and training configuration information for each training task from the automated training execution agent.

[0151] (2) Intelligent Analysis and Suggestions: Automatically analyze the collected data, generate training trend analysis, and identify potential anomalies (such as loss explosion, no longer improving accuracy) or optimization points (such as the learning rate may be too high). Based on the analysis results, provide training optimization suggestions to users through the interface or reports.

[0152] (3) Visualization: The monitoring data is displayed in the form of charts on the monitoring panel.

[0153] Model testing and recording: The agent starts automatically after model training is complete. The process includes:

[0154] (1) Automated testing environment configuration: Based on the parameters of the training subtasks, automatically configure the hardware and software environment required for model testing, including the versions of algorithm dependency libraries, the testing toolchain, and bind computing resources. Strictly verify the executability of test scripts and the accessibility of test data paths.

[0155] (2) Standardized test execution: Automatically associate and load the standardized test dataset that has been prepared in advance for the task.

[0156] (3) Testing and Structured Records: Execute the model's inference on the test set and calculate various performance metrics. During the testing process, capture raw logs and performance data, and generate structured test records according to information such as model version and task ID. Store these records in a database or file system for easy subsequent querying, comparison, and reproduction.

[0157] Through the above process, large-scale, automated parallel execution of training tasks and refined process management are achieved. Dynamic resource planning improves cluster utilization; multi-task parallelism accelerates experimental iteration; real-time monitoring and intelligent analysis help users grasp the training status in a timely manner and make intervention or optimization decisions; automated testing and recording ensure the objectivity of model evaluation and the traceability of results.

[0158] Step S5: Comparative analysis and report generation of intelligent remote sensing image processing tasks.

[0159] By comprehensively comparing and analyzing the test results of all training sub-tasks, a model description report is generated, which includes model performance evaluation and optimization suggestions.

[0160] Step S5 is performed by the comparative analysis and report generation module, which comprehensively evaluates all experiments and generates final results that can be delivered or archived. This module consists of two agents: a training task result analysis agent and a model description report generation agent.

[0161] The training task result analysis agent summarizes the full-link information of all training sub-tasks.

[0162] (1) Data summary: Gather the data filtering results, preprocessing strategies, algorithm architectures used, hyperparameter configurations, training process logs, resource consumption curves and final model test results for each subtask.

[0163] (2) Horizontal comparative analysis: The performance of all models on the test set is compared horizontally. The analysis dimensions include main performance indicators, resource efficiency, model complexity, etc.

[0164] (3) Sorting and display: Output a sorted list of models according to their performance, and clearly display the relative advantages and disadvantages of each model in different dimensions through visual charts (such as radar charts and bar charts) to assist users in making decisions.

[0165] The model description report generation agent generates a report based on the final model (or multiple models) selected by the user from the sorted list.

[0166] (1) Source information aggregation: Automatically retrieve complete source information related to the selected model, including: the version and path of the training data used, details of the data preprocessing strategy, source code identification of the algorithm model, the final hyperparameter configuration, complete training process logs, detailed test index results, and hardware resource consumption statistics.

[0167] (2) Structured report generation: According to the predefined or user-selected report template, the above traceability information is filled into the corresponding chapters of the report, and a structured model description report is automatically generated. The core content of the report includes model overview, data usage instructions, model architecture and parameters, training process analysis, test performance evaluation, and resource consumption summary.

[0168] (3) Recording of relationships: The relationship between model version, training task ID and dataset version used is clearly recorded in the report and system metadata to ensure rapid backtracking and accurate reproduction of historical experiments.

[0169] (4) Results archiving: The generated model files and test reports are automatically pushed to the model repository for version management and a download link is provided. The report supports export in multiple formats (such as PDF and HTML).

[0170] Step S5 automates the summarization and knowledge accumulation of experimental results. Through systematic comparative analysis, the optimal model can be quickly identified from a large number of experiments, aiding decision-making. The automatically generated detailed report not only serves as a project deliverable but, more importantly, establishes a complete traceability chain of "data-algorithm-parameter-results," greatly enhancing the repeatability of scientific research and engineering practice and promoting team knowledge sharing and the accumulation of algorithmic assets.

[0171] Steps S1 to S5 constitute a complete large-scale agent-driven intelligent remote sensing image processing algorithm.

[0172] In some embodiments of the present invention, step S1 specifically includes:

[0173] Step S11: The overall planning agent of the large model performs semantic parsing on the task text input by the user and extracts key information such as task type, data requirements and resource constraints.

[0174] Step S12: Based on the parsed key information, generate a structured task chain containing multiple ordered subtasks;

[0175] Step S13: Assign each subtask in the structured task chain to the corresponding execution agent module and generate a visual task flow script.

[0176] In some embodiments of the present invention, step S1 further includes:

[0177] An executive director agent is set up to monitor the execution progress of the structured task chain, pass task parameters and resources between the execution agents, and trigger a retry mechanism for failed task stages.

[0178] In some embodiments of the present invention, in step S2, the data filtering agent extracts keywords through natural language processing, generates structured filtering conditions including file format, resolution and data type, and retrieves data from the database based on these conditions.

[0179] In some embodiments of the present invention, the data filtering agent performs multi-dimensional statistical analysis on the acquired dataset to automatically identify problems such as missing samples or class imbalance.

[0180] The preprocessing strategy includes at least one of image cropping and sample augmentation.

[0181] In some embodiments of the present invention, in step S3, the algorithm model filters the intelligent agent by using a large language model to parse the task requirements, generate detailed filtering conditions, and automatically match candidate solutions in the algorithm library and model library.

[0182] The hyperparameter configuration agent automatically generates hyperparameter configuration strategies, including model structure parameters and training strategy parameters, based on task requirements, data characteristics, and algorithm architecture, and forms a batch training candidate configuration list sorted by priority.

[0183] In some embodiments of the present invention, in step S4, the training computing power planning agent automatically selects executable training subtasks from the candidate configuration list according to the algorithm type and available computing power, and allocates computing resources to each subtask.

[0184] In some embodiments of the present invention, during the model training process, the training monitoring agent collects and visualizes the model performance indicators, resource usage, and training process curves in real time.

[0185] In some embodiments of the present invention, the model testing agent automatically calls the standardized test dataset corresponding to the task to test the trained model and records the test results in a structured manner, supporting result tracking and reproduction.

[0186] In some embodiments of the present invention, in step S5, the model description report automatically records training data source information, hyperparameter configuration, training process log, test indicators and hardware resource consumption, and records the association between the model version and the dataset.

[0187] According to one aspect of the present invention, an electronic device is provided, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in the memory; when the electronic device is running, the processor executes the one or more computer programs stored in the memory to enable the electronic device to execute a multi-agent driven remote sensing image intelligent processing algorithm autonomous training and verification method as described in any of the above technical solutions.

[0188] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0189] The memory can be an internal storage unit of the terminal device, such as a hard drive or RAM. Alternatively, it can be an external storage device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory can include both internal and external storage units. The memory is used to store the computer program and other programs and data required by the terminal device. It can also be used to temporarily store data that has been output or will be output.

[0190] According to one aspect of the present invention, a computer-readable storage medium is provided for storing computer instructions, which, when executed by a processor, implement an autonomous training and verification method for a multi-agent driven remote sensing image intelligent processing algorithm as described in any of the above technical solutions.

[0191] For example, computer-readable storage media can be read-only memory (ROM), random access memory (RAM), read-only optical disc (CD-ROM), magnetic tape, floppy disk, and optical data storage devices. They can be implemented using computer-executable program code, thus allowing them to be stored in a storage device for execution by a computing device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Therefore, this invention is not limited to any particular hardware and software combination.

[0192] In summary, this invention proposes a multi-agent-driven autonomous training and verification system for intelligent remote sensing image processing algorithms. It introduces an innovative agent-driven mechanism to achieve automatic task parsing, process orchestration, and intelligent decision-making, significantly reducing human intervention. The system possesses multi-dimensional data and model adaptation capabilities, supporting parallel training of multiple types of remote sensing images, multiple algorithm structures, and multiple tasks. The system features fully automated management, from data screening to report generation, enabling an efficient and traceable model development process. Furthermore, the system exhibits high scalability and reproducibility, supporting rapid migration to different task scenarios and the reproduction of model results.

[0193] Furthermore, it should be noted that the embodiments of the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code.

[0194] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0195] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing terminal equipment to cause a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0196] It should also be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0197] Finally, it should be noted that the above description represents a preferred embodiment of the present invention. It should be pointed out that although preferred embodiments have been described, those skilled in the art, once they understand the basic inventive concept of the present invention, can make various improvements and modifications without departing from the principles described herein. These improvements and modifications should also be considered within the scope of protection of the present invention. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the embodiments of the present invention.

Claims

1. A self-training and verification system for a multi-agent-driven intelligent remote sensing image processing algorithm, characterized in that, include: The task input and large model agent planning module is used to receive training task instructions input by the user, and to parse task requirements, generate structured task chains and arrange and schedule task processes through the large model overall planning agent. The data filtering and preprocessing module is used to automatically filter remote sensing image training samples, statistically analyze data, generate and execute data preprocessing strategies according to task requirements, so as to output a standardized training dataset. The algorithm model selection and hyperparameter configuration module is used to automatically select matching baseline algorithms and pre-trained models, and automatically generate hyperparameter configuration strategies and candidate configuration lists based on task requirements and data characteristics. The model training and testing module is used to automatically allocate resources, perform parallel training, monitor the training process, and automatically test and record the trained model based on the candidate configuration list and available computing power. The comparative analysis and report generation module is used to comprehensively analyze the test results of all training tasks and automatically generate model performance evaluation reports.

2. The system according to claim 1, characterized in that, The task input and large model agent planning module also includes: The Executive Director intelligent agent is used to monitor task execution progress, pass parameters and resources between modules, and handle task execution exceptions.

3. The system according to claim 1, characterized in that, The data filtering and preprocessing module includes: The training data filtering and analysis agent is used to generate structured filtering conditions and acquire data according to task requirements, as well as to perform multi-dimensional statistical analysis on the data. A preprocessing strategy configuration agent is used to automatically generate data preprocessing strategy configuration files based on data analysis results. A preprocessing execution agent is used to perform preprocessing operations according to the policy configuration file to generate a standardized dataset.

4. The system according to claim 1, characterized in that, The algorithm model selection and hyperparameter configuration module includes: Algorithm model selection agent is used to parse task requirements using a large language model and automatically match and recommend algorithms and models; Train a hyperparameter policy configuration agent to automatically generate a list of candidate hyperparameter configurations and configuration files based on task, data, and algorithm information.

5. The system according to claim 1, characterized in that, The model training and testing module includes: A training computing power planning and allocation agent is used to allocate training resources based on the candidate configuration list and the system computing power. An automated training execution agent is used to receive datasets, algorithm files, configuration files, and computing resources, and to perform model training. A training process supervisory agent is used to monitor the training status in real time, analyze training process information, and provide optimization suggestions. The model testing and recording agent is used to automatically configure the test environment, call the test dataset to perform model verification, and record the test results in a structured manner.

6. The system according to claim 1, characterized in that, The comparative analysis and report generation module includes: The training task result analysis agent is used to summarize and compare the training and testing data of all sub-tasks horizontally, and output a ranking list of models. The model description report generation agent is used to automatically retrieve traceability information and generate a model description report according to the template based on the model selected by the user.

7. A method for autonomous training and verification of a multi-agent driven intelligent remote sensing image processing algorithm, characterized in that, Includes the following steps: Step S1: Receive training task instructions input by the user in natural language or template format, automatically parse the task requirements through the large model overall planning intelligent agent, generate a structured task chain, and perform task flow orchestration and scheduling. Step S2: Based on the parsed task requirements, the data filtering agent automatically filters remote sensing image training samples that meet the conditions from the database and performs statistical analysis; based on the analysis results, the preprocessing configuration agent automatically generates a data preprocessing strategy, and the preprocessing execution agent executes the strategy to generate a standardized training dataset. Step S3: The algorithm model screening agent automatically selects matching baseline algorithms and pre-trained models from the algorithm library and model library according to the task requirements; the hyperparameter configuration agent automatically generates diverse hyperparameter configuration strategies based on the task requirements, dataset features and selected algorithms, forming a candidate configuration list. Step S4: The training computing power planning agent automatically allocates resources and schedules the parallel training of multiple sets of models based on the candidate configuration list and available computing power resources; During training, the training monitoring agent monitors the training status and performance in real time; after training is completed, the model testing agent automatically calls the test dataset to verify the model and records the test results. Step S5: Conduct a comprehensive comparative analysis of the test results of all training sub-tasks and generate a model description report that includes model performance evaluation and optimization suggestions.

8. The method according to claim 7, characterized in that, Step S1 specifically includes: Step S11: The overall planning agent of the large model performs semantic parsing on the task text input by the user and extracts key information such as task type, data requirements and resource constraints. Step S12: Based on the parsed key information, generate a structured task chain containing multiple ordered subtasks; Step S13: Assign each subtask in the structured task chain to the corresponding execution agent module and generate a visual task flow script; Step S1 further includes: An executive director agent is set up to monitor the execution progress of the structured task chain, pass task parameters and resources between the execution agents, and trigger a retry mechanism for failed task stages.

9. The method according to claim 7, characterized in that, In step S2, the data filtering agent extracts keywords through natural language processing, generates structured filtering conditions including file format, resolution and data type, and retrieves data from the database based on these conditions. The data filtering agent performs multi-dimensional statistical analysis on the acquired dataset to automatically identify problems such as missing samples or class imbalance. The preprocessing strategy includes at least one of image cropping and sample augmentation.

10. The method according to claim 7, characterized in that, In step S3, the algorithm model filters the intelligent agent by using a large language model to parse the task requirements, generate detailed filtering conditions, and automatically match candidate solutions in the algorithm library and model library. The hyperparameter configuration agent automatically generates hyperparameter configuration strategies, including model structure parameters and training strategy parameters, based on task requirements, data characteristics, and algorithm architecture, and forms a batch training candidate configuration list sorted by priority. In step S4, the training computing power planning agent automatically selects executable training subtasks from the candidate configuration list based on the algorithm type and available computing power, and allocates computing resources to each subtask. During model training, the training monitoring agent collects and visualizes model performance metrics, resource usage, and training process curves in real time. The model testing agent automatically calls the standardized test dataset corresponding to the task to test the trained model and records the test results in a structured manner, supporting result tracking and reproduction.

Images