Remote sensing image interpretation methods based on large model inference and tool enhancement

Abstract

This invention proposes a remote sensing image interpretation method based on large-model reasoning and tool enhancement, comprising the following steps: Based on system prompts and input remote sensing image features, a large language model is used as an active decision-maker to generate initial reasoning nodes for the reasoning tree; based on the current state of the reasoning tree, the active decision-maker performs reasoning and generates new reasoning nodes, thereby expanding the reasoning tree; based on node scores, a pruning strategy is used to eliminate reasoning paths with low scores; the above steps are repeated until the reasoning termination condition is met, and the path with the highest cumulative score is selected as the optimal reasoning path from the reasoning tree; key information is extracted and organized into a structured intelligence report for output. This invention explicitly models the complete interpretation chain of "thinking-tool-observation" through a tree-like reasoning mechanism, making the generation logic of each intelligence conclusion clearly visible, thus solving the trust crisis and insufficient interpretability problems caused by the "black box" decision-making of traditional deep learning models.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image interpretation, and in particular to a remote sensing image interpretation method based on large model reasoning and tool enhancement. BACKGROUND

[0002] Remote sensing image interpretation is a comprehensive technology that combines physical property analysis, interpretation sign system, human-computer interaction, and case practice. Its goal is to obtain effective information of earth observation through the combination of multi-source data and various methods.

[0003] Since the beginning of the 21st century, with the widespread application of remote sensing satellites, aerial sensors, and unmanned aerial vehicle platforms, human ability to obtain information about the earth's surface has significantly increased. Remote sensing images have become an important data source in the fields of land resource monitoring, urban planning, environmental protection, disaster assessment, and military reconnaissance. At the same time, the volume of remote sensing data continues to grow, the resolution continues to improve, and the time and spatial scales covered continue to expand, making remote sensing interpretation technology crucial in both scientific research and practical applications.

[0004] However, remote sensing image interpretation still faces a series of technical challenges. On the one hand, due to limitations in imaging conditions and hardware capabilities, remote sensing images often have problems such as low resolution, noise interference, and cloud and fog obstruction, which lead to decreased accuracy in ground object recognition and scene understanding. On the other hand, multi-source heterogeneous remote sensing data differ in waveband characteristics, imaging mechanisms, and time spans, significantly increasing the difficulty of cross-source information fusion and unified modeling. These problems make it difficult for traditional interpretation methods relying on visual interpretation or single deep learning models to balance accuracy, efficiency, and scalability.

[0005] In recent years, with the rapid development of deep learning and computer vision, methods based on convolutional neural networks (CNN) and Transformers have made breakthroughs in tasks such as remote sensing image classification, target detection, and change detection. However, these methods generally rely on large-scale labeled data and are prone to insufficient generalization performance in complex and varied remote sensing scenarios. At the same time, existing models lack explicit modeling of the reasoning chain in the interpretation process, resulting in insufficient explainability.

[0006] Therefore, there is an urgent need for a remote sensing image interpretation method that does not rely on large-scale labeled data, has strong generalization performance in complex scenarios, and has strong explainability. SUMMARY

[0007] The present application proposes a remote sensing image interpretation method based on large model reasoning and tool enhancement, solving the problems of existing interpretation methods relying on large-scale labeled data, poor generalization performance in complex scenarios, and insufficient explainability.

[0008] The technical solution of the present application is implemented as follows:

[0009] The first aspect of the present application provides a remote sensing image interpretation method based on large model inference and tool enhancement, comprising the following steps:

[0010] S1, input and initialization: input remote sensing image, and initialize large language model as active decision maker, generate initial inference node of inference tree based on preset system prompt word and input remote sensing image feature;

[0011] S2, node inference and tree expansion: based on the current inference tree state, the active decision maker performs inference and generates new inference nodes, thereby expanding the inference tree; each node of the inference tree represents a structured inference unit containing thinking, tool call decision, tool observation result and node score;

[0012] S3, path evaluation and pruning: when the inference process of step S2 produces multiple candidate branches, evaluate the candidate paths based on node scores, and eliminate low-score paths using pruning strategy;

[0013] S4, termination judgment and output: repeat steps S2 to S3 until the inference termination condition is met, then select the path with the highest cumulative score from the inference tree as the optimal inference path, and generate a structured intelligence report based on the path.

[0014] Specifically, in step S2, the construction process of the inference tree is dynamic and iterative;

[0015] In the first t round of inference, the inference tree is represented as:

[0016] ;

[0017] Wherein, represents t the set of inference nodes generated at time t; represents the set of directed edges between inference nodes, describing the parent-child relationship between inference nodes;

[0018] The i th inference node is represented as:

[0019] ;

[0020] Wherein, represents the structured decision of the large language model at the i th inference node, represents the tool call decision of the large language model at the i th inference node, represents the tool call decision of the large language model at the iThe structured observations returned after each inference node invokes the tool. Indicates the first i The score or confidence level of each inference node.

[0021] Furthermore, in step S2, the iterative expansion process of the reasoning tree is as follows:

[0022] Structured decision-making of inference nodes The generation method is as follows:

[0023] ;

[0024] in, Representing a large language model, x The input consists of remote sensing image features, and prompt is the system prompt word. This represents the state of the reasoning tree at the previous moment;

[0025] Tool call decision Compared with the observation results The generation method is as follows:

[0026] , ;

[0027] in, Call the policy function for the tool, based on the current decision. and toolset Output tool call decision ; For toolset The tools called in These are the parameters used when calling the tool;

[0028] During the inference process, when multiple available tools or strategies exist, parallel generation is performed based on the current node state. k Use multiple child nodes to explore different reasoning paths:

[0029] ;

[0030] New node It is inserted into the tree and connected to its parent node. Update the set of nodes and edges:

[0031] ;

[0032] .

[0033] Specifically, the tool calls are implemented through a standardized model context protocol interface to decouple the tool from the main inference process and enable plug-and-play functionality;

[0034] The tools include at least one or more of the following: open object detection tools, image super-resolution tools, remote sensing image declouding / denoising tools, image cropping tools, external web search engines, or retrieval enhancement generation modules.

[0035] Specifically, in step S3, the node score is calculated using a comprehensive scoring function;

[0036] For the t Round of reasoning j candidate nodes Its rating The calculation formula is as follows:

[0037] ;

[0038] in, Indicates observation results based on tools The calculation tool uses validity scores. Represents node-based decision making The calculated context semantic consistency score, Indicates node state Align with mission objectives The correlation score; , , These are adjustable weighting coefficients used to control the degree of influence of each score item on the overall score;

[0039] The pruning strategy adopts the Top-K strategy, Beam Search strategy, or UCB strategy. After each round of inference expansion, the system calculates the comprehensive score of all candidate nodes and retains only the top-scoring paths for further expansion.

[0040] Specifically, in step S4, the inference termination condition includes at least one of the following:

[0041] a) The mission objective has been determined to be completed;

[0042] b) The depth of the reasoning tree reaches the preset maximum value;

[0043] c) The reasoning tree can no longer be effectively expanded;

[0044] Select the path with the highest cumulative score from the reasoning tree as the optimal reasoning path. The chosen strategy is:

[0045] ;

[0046] in, P Representation of reasoning tree Any complete path from the root node to a leaf node in the dataset. For path P Middle node v The score, the optimal path It is the path with the highest total node score among all these paths.

[0047] Specifically, in step S4, the structured intelligence report is organized in SOAP format, including:

[0048] The theme specifies the core target or area of ​​intelligence focus;

[0049] Observe and list the objective facts and data obtained through the toolchain;

[0050] Evaluation, an analysis and judgment based on observations;

[0051] A plan is a set of recommendations or decisions for follow-up actions based on the current assessment.

[0052] Optionally, after step S4, a model transfer and adaptation step may also be included:

[0053] S5, Knowledge Tree Transfer Fine-tuning: Using the optimal reasoning path as a supervision signal, knowledge distillation and supervision fine-tuning are performed on the lightweight student model to transfer the reasoning ability of the large model.

[0054] S6, Accuracy judgment: Evaluate the performance indicators of the fine-tuned student model on the validation task. If the performance meets the standard, proceed to the next step; otherwise, return to step S7 to adjust the training strategy.

[0055] S7, Lightweight Model Building: Performs quantization and pruning operations on student models that meet performance standards to generate lightweight models suitable for resource-constrained platforms.

[0056] A second aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the remote sensing image interpretation method.

[0057] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the remote sensing image interpretation method.

[0058] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0059] (1) This invention explicitly models the complete interpretation chain of "thinking-tool-observation" through a tree-like reasoning mechanism, making the generation logic of each intelligence conclusion clearly visible, fundamentally solving the trust crisis and insufficient interpretability problem caused by the "black box" decision-making of traditional deep learning models;

[0060] (2) The present invention adopts a dynamic iteration-based reasoning tree construction and expansion mechanism. The system can autonomously plan and explore a variety of interpretation strategies, effectively cope with the challenges of target scale, pose variation and task steps in remote sensing images, and significantly enhance the logical reasoning and generalization performance of the model in complex and varied remote sensing scenarios.

[0061] (3) By introducing a pruning mechanism based on multi-factor fusion scoring, this invention can quickly select the optimal interpretation link from a large number of candidate paths, effectively avoid the unnecessary consumption of computing resources, ensure efficient convergence of the reasoning process, and maintain the continuity and consistency of decision-making through path memory.

[0062] (4) This invention distills the complex reasoning ability of a large model into a lightweight student model through a knowledge tree transfer mechanism, and after quantization and compression, enables the high-performance interpretation technology to run stably on platforms with strictly limited resources such as edge and spaceborne, which greatly expands the practical application boundaries of this method. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of a remote sensing image interpretation method based on large model inference and tool enhancement according to the present invention.

[0065] Figure 2 This is a schematic diagram of the workflow architecture of a remote sensing image interpretation method based on large model inference and tool enhancement according to the present invention;

[0066] Figure 3 This is a schematic diagram of the migration and quantization guided by the inference path in an embodiment of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0068] Reference Figure 1 , 2 The first aspect of this invention provides a remote sensing image interpretation method based on large model inference and tool enhancement, comprising the following steps:

[0069] S1, Input and Initialization: Input the remote sensing image and initialize the large language model as an active decision-maker. Based on the preset system prompt words and the features of the input remote sensing image, generate the initial inference nodes of the inference tree.

[0070] S2, Node Reasoning and Tree Expansion: Based on the current state of the reasoning tree, the active decision-maker performs reasoning and generates new reasoning nodes, thereby expanding the reasoning tree; each node of the reasoning tree represents a structured reasoning unit that includes thinking, tool call decisions, tool observation results, and node scoring;

[0071] The construction process of the reasoning tree is dynamic and iterative;

[0072] In the t In round-robin reasoning, the reasoning tree is represented as:

[0073] ;

[0074] in, express t The set of inference nodes that have been generated at any given time; This represents the set of directed edges between inference nodes, used to describe the parent-child relationships between inference nodes;

[0075] No. i Each inference node Represented as:

[0076] ;

[0077] in, The large language model is represented in the first... i Structured decision-making for each inference node The large language model is represented in the first... i Tool call decision for each inference node The large language model is represented in the first... i The structured observations returned after each inference node invokes the tool. Indicates the first iThe score or confidence level of each inference node.

[0078] The iterative expansion process of the reasoning tree is as follows:

[0079] Structured decision-making of inference nodes The generation method is as follows:

[0080] ;

[0081] in, Representing a large language model, x The input consists of remote sensing image features, and prompt is the system prompt word. This represents the state of the reasoning tree at the previous moment;

[0082] Based on the tool call decision of the current node in the reasoning tree, the corresponding remote sensing image processing tool or external knowledge retrieval tool is called, and the returned structured observation results are filled into the node;

[0083] Tool call decision Compared with the observation results The generation method is as follows:

[0084] , ;

[0085] in, Call the policy function for the tool, based on the current decision. and toolset Output tool call decision ; For toolset The tools called in These are the parameters used when calling the tool;

[0086] The tool calls are implemented through a standardized model context protocol interface to decouple the tool from the main inference process and enable plug-and-play functionality;

[0087] The tools include at least one or more of the following: open object detection tools, image super-resolution tools, remote sensing image declouding / denoising tools, image cropping tools, external web search engines, or retrieval enhancement generation modules.

[0088] During the inference process, when multiple available tools or strategies exist, parallel generation is performed based on the current node state. k Use multiple child nodes to explore different reasoning paths:

[0089] ;

[0090] New node It is inserted into the tree and connected to its parent node. Update the set of nodes and edges:

[0091] ;

[0092] .

[0093] S3, Path Evaluation and Pruning: When multiple candidate branches are generated during the reasoning process in step S2, the candidate paths are evaluated based on node scores, and a pruning strategy is used to eliminate paths with low scores and retain paths with high scores to continue expansion.

[0094] The node score is calculated using a comprehensive scoring function;

[0095] For the t Round of reasoning j candidate nodes Its rating The calculation formula is as follows:

[0096] ;

[0097] in, Indicates observation results based on tools The calculation tool uses validity scores. Represents node-based decision making The calculated context semantic consistency score, Indicates node state Align with mission objectives The correlation score; , , These are adjustable weighting coefficients used to control the degree of influence of each score item on the overall score;

[0098] The pruning strategy adopts the Top-K strategy, Beam Search strategy, or UCB strategy. After each round of inference expansion, the system calculates the comprehensive score of all candidate nodes and retains only the top-scoring paths for further expansion.

[0099] S4, Termination Judgment and Output: Repeat steps S2 to S3 until the reasoning termination condition is met. Then, select the path with the highest cumulative score from the reasoning tree as the optimal reasoning path and generate a structured intelligence report based on this path.

[0100] The reasoning termination condition includes at least one of the following:

[0101] a) The mission objective has been determined to be completed;

[0102] b) The depth of the reasoning tree reaches the preset maximum value;

[0103] c) The reasoning tree can no longer be effectively expanded;

[0104] Select the path with the highest cumulative score from the reasoning tree as the optimal reasoning path. The chosen strategy is:

[0105] ;

[0106] in, P Representation of reasoning tree Any complete path from the root node to a leaf node in the dataset. For path P Middle node v The score, the optimal path It is the path with the highest total node score among all these paths.

[0107] Then, based on the optimal reasoning path, key information is extracted and organized into a structured intelligence report for output.

[0108] The structured intelligence report is organized in SOAP format and includes:

[0109] S (Subject) specifies the core target or area of ​​intelligence focus;

[0110] O (Observation): Listing objective facts and data obtained through the toolchain;

[0111] A (Assessment) refers to the analysis and judgment based on observations.

[0112] P (Plan) refers to a plan, which is a recommendation or decision for follow-up action based on the current assessment.

[0113] Optionally, such as Figure 3 As shown, after step S4, a model transfer and adaptation step may also be included:

[0114] S5, Knowledge Tree Transfer Fine-tuning: Using the optimal reasoning path as a supervision signal, knowledge distillation and supervision fine-tuning are performed on the lightweight student model to transfer the reasoning ability of the large model.

[0115] The loss function used in the supervised fine-tuning It consists of the following parts:

[0116] ;

[0117] in, The standard autoregressive language modeling loss is used to guide the model to generate reasonable text; Tree backtracking consistency loss is used to ensure that the reasoning steps generated by the student model are consistent with the logical structure of the reasoning tree constructed during training. To improve the stability of model language generation, the loss for labeled autoregressive language is increased by using a standard target output path. and For the corresponding loss weighting coefficient.

[0118] S6, Accuracy judgment: Evaluate the performance indicators of the fine-tuned student model on the validation task. If the performance meets the standard, proceed to the next step; otherwise, return to step S7 to adjust the training strategy.

[0119] S7, Lightweight Model Building: Performs quantization and pruning operations on student models that meet performance standards to generate lightweight models suitable for resource-constrained platforms.

[0120] This embodiment uses "port vessel identification" as a specific application scenario to fully and in detail reproduce the execution process of the technical solution of the present invention in practical applications, demonstrating the complete reasoning chain from initial image input to final intelligence generation (the path scoring and pruning processes are omitted in this case, and one optimal reasoning path is selected for illustration):

[0121] First round of reasoning: Initial target detection and analysis

[0122] The system received a satellite image description from the user: "This is a high-resolution satellite image showing a large ship docked in a port, with a large number of aircraft parked on the deck in a regular arrangement. There are white numbers at the stern, but the image is not clear enough to identify the specific numbers. There are roads, buildings and vehicles in the surrounding area, and the environment is well-organized."

[0123] The active decision-maker generates an initial reasoning node, with the following thought process: "I need to extract port vessel intelligence based on the image description. This is a large vessel docked in the port, with aircraft densely stored on its deck. There is ambiguous numbering information at the stern, and the current image content is insufficient to confirm the vessel's identity or specific affiliation. I will call the target detection tool to identify key targets such as the hull structure, numbered areas, and port buildings for subsequent numbering extraction and intelligence judgment."

[0124] The system invokes the target detection tool, inputting the original satellite image and detection prompts: "ships, cargo ships, aircraft, ship stern number, aircraft stern number, ship stern mark, aircraft mark, port structures".

[0125] The object detection tool returned the following observation: This is a high-resolution satellite image showing a large ship docked in the harbor. The image contains multiple colored boxes that mark key identification areas and their category information.

[0126] The main vessel area is a large area, categorized as "cargo ship", with a confidence level of 0.379 and coordinates (129, 67, 1042, 409), covering the entire large vessel;

[0127] A building in the middle of the port is labeled as "port building" with a confidence level of 0.315 and coordinates (357,512,527,649).

[0128] The stern area is marked as "stern mark" with a confidence level of 0.276 and coordinates (972,342,1030,388), which is speculated to be the possible location of the hull number;

[0129] The area of ​​the front hull clamps partially frames the tail of the aircraft, categorized as "aircraft tail number", with a confidence level of 0.274 and coordinates (270, 369, 315, 410), which is suspected to contain the tail number information of the aircraft.

[0130] The small buildings in the nearshore area in the upper left corner are selected, categorized as "port buildings", with a confidence level of 0.263 and coordinates (73,0,151,46).

[0131] Second round of reasoning: Hull structure reinforcement analysis

[0132] Based on the observations from the first round of reasoning, the active decision-maker generates a second reasoning node. The thought process is as follows: Upon receiving information from the object detection tool, the following key regions were identified in the image:

[0133] The “Cargo Ship” area (129,67,1042,409) covers the entire large vessel and is the main basis for identifying the hull structure and platform level.

[0134] The "Flight Tail Number" area (270, 369, 315, 410) is located at the tail of the deck aircraft and is suspected to contain a number.

[0135] "Stern Markings" area (972,342,1030,388): Stern number area;

[0136] Several areas of "port buildings": to help determine the nature and layout of the port.

[0137] The current priority is to identify the overall configuration features of large vessels. Therefore, I will crop the "cargo ship area" and perform binarization enhancement processing for subsequent vessel type identification and structural confirmation.

[0138] The system invokes the cropping and enhancement tool, inputs the original satellite image and the coordinates of the ship's main body area (129,67,1042,409), and enables the binarization enhancement function.

[0139] The tool returned the following observations: The binarized and enhanced image shows that the hull structure lines are clear, the deck outline is regular, the boundary of the stern marking area is significant, the outlines of the densely distributed aircraft on the deck are complete, and the upper platform and antenna equipment are clearly visible. The overall deck layout, hull outline and aircraft form are consistent with the typical appearance characteristics of a certain type of large ship.

[0140] Third round of reasoning: Confirmation of aircraft model

[0141] Based on the results of the second round of reasoning, the active decision-maker generates a third reasoning node. The thought process is as follows: "I have received a binarized enhanced image of the hull area, showing a large ship docked in the port. The hull structure is clear, the deck outline is regular, and the boundary of the marked area at the stern is significant. The outlines of the densely distributed aircraft on the deck are complete, and the upper platform and antenna equipment are clearly visible. The overall deck layout, hull outline, and aircraft form conform to the typical appearance characteristics of a certain type of large ship. The next step is to extract the aircraft area for model confirmation to verify the consistency between the aircraft and the platform."

[0142] The system calls the image cropping tool. Input the original satellite image and the coordinates of the spacecraft region (270,369,315,410).

[0143] The tool returned the observation results: The enlarged view of the aircraft shows a jet aircraft painted in light gray. The fuselage structure is compact, and the overall outline and body structure height are consistent with the typical appearance characteristics of a certain type of aircraft.

[0144] Fourth round of reasoning: Extraction of numbered regions

[0145] Based on the results of the third round, the active decision-maker generates the fourth reasoning node. The thought process is as follows: "The enlarged view of the aircraft shows a light gray jet aircraft with a compact fuselage. The overall structure conforms to the typical characteristics of a certain type of aircraft. The next step is to perform a cropping operation on the numbered area (972, 342, 1030, 388) at the stern to obtain a clear numbered image for subsequent identification."

[0146] The system calls the image cropping tool. Input the original satellite image and the coordinates of the stern numbered area (972,342,1030,388).

[0147] The tool returned the observation results: A magnified view of the stern area shows clearly visible white lines forming the number "" on the deck. The number is displayed in a large font with an outline and is located in the center of the end of the deck.

[0148] Fifth round of reasoning: External intelligence retrieval

[0149] Based on the results of the fourth round, the active decision-maker generates the fifth reasoning node, with the following thought process: "I have confirmed the stern number through the image." This is crucial identification information. Next, I need to use this number to query relevant vessel background information, including class, name, and home port. This data will help me comprehensively analyze the current vessel's identity and status, ensuring the accuracy and completeness of the intelligence. Therefore, I will use a search tool to query the "number". "Large ship home port" to obtain detailed public information.

[0150] The system invokes the search tool; enter the keyword "number" to search. "Home port for large ships".

[0151] The search tool returned the following observations: Based on publicly available information from multiple channels and a professional data transmission system, it was confirmed that the ship is a large vessel of a certain type, with the stern number " "Extensively documented in numerous documents and publicly available data. The vessel's home port is a certain port. Historical trajectory data shows that the vessel has recently been anchored in port and has not undertaken any long-duration maritime missions. Based on current satellite imagery and historical data, it is inferred that the vessel's number..." The large vessel is currently docked at a port and is expected to remain in port status in the short term.

[0152] Sixth round of reasoning: Comprehensive intelligence generation

[0153] Based on the results of the first five rounds, the active decision-maker generates the sixth inference node, with the following thought process: "I have received complete intelligence information: the satellite image was taken in mid-May 2025, showing the number as..." A large ship is docked in the port, with approximately 30 aircraft neatly arranged on its deck. The deck structure is clearly visible. Based on the serial numbers and aircraft types, the ship is confirmed to be a certain type of large vessel, with its home port being a certain port. Currently, the ship is at anchor and shows no obvious signs of activity. Information is sufficient; inference is now complete, and a detailed SOAP intelligence report is generated.

[0154] The system automatically integrates all observations and generates an intelligence report in SOAP format.

[0155] Subject (S): Source: High-resolution port satellite imagery from May 2025;

[0156] Observation (O): Observed the number " A large ship was docked at a port, with about 30 aircraft parked on its deck in an orderly manner, and the deck structure was complete and clear.

[0157] Assessment (A): Analysis confirms that the ship is a certain type of large vessel and its home port is located in a certain port;

[0158] Plan (P): It is predicted that the vessel is currently anchored in port with no signs of activity and no plans for long-distance voyages are expected in the near future.

[0159] This implementation plan demonstrates the entire process from coarse-grained identification to refined judgment, combining perception, enhancement, reasoning, retrieval, and structured output, and possesses good versatility, accuracy, and practicality. It can also be further expanded to various application scenarios such as ship classification, port monitoring, and automatic generation of target intelligence.

[0160] This invention achieves stepwise inference, dynamic tool invocation, and lightweight optimization of remote sensing images through a layered architecture, ensuring that the system can still generate high-precision and interpretable interpretation results in complex environments and has cross-platform deployment capabilities.

[0161] The system framework of this invention adopts a five-layer layered architecture, including:

[0162] Layer 1, Reasoning and Control Layer: This layer explicitly models the interpretation process using a tree-like reasoning mechanism, responsible for scheduling and path generation for multiple rounds of thinking. It decomposes the reasoning process into a cycle of "thinking—tools—observation," generating multiple candidate reasoning nodes at each step and expanding them layer by layer along the path. By modeling tool calls and environmental interactions as tree nodes, this layer ensures the structured and transparent nature of the interpretation process, allowing the final intelligence results to be traced back to their generation process.

[0163] Layer 2, the perception and representation layer, is responsible for preprocessing remote sensing images and extracting multimodal features, covering spectral features, spatial geometric features, and time-series features. During inference, this layer also handles candidate path pruning and memory management. A scoring mechanism evaluates the credibility of different paths, and a Top-K pruning strategy is used to select the optimal path, ensuring efficient convergence of the inference process. Simultaneously, a memory module is used to store intermediate states and historical inference information, reducing redundant computation and improving the consistency and stability of the interpretation results.

[0164] Layer 3, Tools and Knowledge Layer: This layer standardizes and manages remote sensing tools through the MCP protocol, covering various operations such as detection, cropping, super-resolution, cloud removal / denoising, and retrieval enhancement. The design of this layer decouples the tools from the main inference process, making them plug-and-play. Simultaneously, this layer provides knowledge enhancement capabilities, enabling interaction with external knowledge bases and retrieval engines to supplement the inference process with prior knowledge and domain rules, thus forming an integrated "tool-knowledge" enhancement model that significantly improves the flexibility and robustness of remote sensing interpretation.

[0165] Layer 4, Quantization Transfer Layer: This layer utilizes the complete inference tree as a supervisory signal to perform distillation training and quantization on the lightweight model. Through distillation, the student model learns the inference path and intermediate results of the larger model, thereby reducing computational complexity while maintaining accuracy. Further quantization significantly compresses model parameters, making it adaptable to different computing power environments. Furthermore, this layer supports model transfer, enabling rapid deployment between edge devices and spaceborne platforms, ensuring the method's broad applicability and engineering value.

[0166] Layer 5, Intelligence Output Layer: Responsible for organizing the final results of multi-round reasoning into standardized, structured intelligence output. This layer not only outputs interpretation conclusions but also provides process explanations in a tree-like path format, enabling users to clearly understand the logical basis for intelligence generation. The output results conform to the standardized requirements of remote sensing intelligence expression, and can be used for scientific research and engineering tasks as well as directly support decision analysis, meeting the high requirements for reliability and traceability in practical scenarios.

[0167] To further enhance the versatility and adaptability of the multi-turn tree-structured reasoning agent method of this invention, the system has the following extension capabilities:

[0168] Extension 1: Dynamic parameter adjustment capability: The number of Top-K candidate selections and the scoring weights of each node in this system can be dynamically adjusted according to task type, complexity, or user preferences. For example, in scenarios requiring more divergent exploration, the system can appropriately increase the Top-K values ​​to expand the search space; while in accuracy-priority tasks, the weights can be reduced to enhance convergence speed and judgment stability.

[0169] Extension 2: Scalability and compatibility of the toolset: The toolset supports multiple extension methods, including the introduction of custom-developed remote sensing analysis tools or NLP tools; it is also compatible with third-party APIs, such as Google Earth Engine interface, commercial image recognition API, Gaode Map API, etc.

[0170] The tool's registration mechanism follows a unified description protocol, allowing for dynamic addition and deletion, and supports parallel remote loading and local caching.

[0171] Extension 3: Diverse Adaptation of Intelligence Output Formats: To adapt to the integration needs of different platforms and systems, the system supports outputting intelligence results in multiple structured formats such as JSON, XML, and CSV, and supports automatic embedding of chain-based reasoning process annotations, facilitating source tracing analysis and automated access.

[0172] Extension 4: Flexibility of inference process pruning strategies: In addition to the default Top-K pruning strategy, the system supports replacement with Beam Search, UCB (Upper Confidence Bound), and other strategy modules. Developers can switch pruning logic according to specific task objectives (such as enhancing accurate localization or improving distribution coverage) to enhance the model's generalization ability and efficiency performance under different tasks.

[0173] A second aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program executable on the processor, and the processor executes the computer program to implement the steps of the remote sensing image interpretation method.

[0174] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the remote sensing image interpretation method.

[0175] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A remote sensing image interpretation method based on large model inference and tool enhancement, characterized in that, The method is performed by iteratively expanding the reasoning tree, comprising the following steps: S1, input and initialization: input the remote sensing image, and initialize the large language model as the active decision maker, generate the initial reasoning node of the reasoning tree based on the preset system prompt word and the input remote sensing image feature; S2, node reasoning and tree expansion: based on the current reasoning tree state, the active decision maker performs reasoning and generates new reasoning nodes, thereby expanding the reasoning tree; each node of the reasoning tree represents a structured reasoning unit containing thinking, tool call decision, tool observation result and node score; S3, path evaluation and pruning: when multiple candidate branches are generated in the reasoning process of step S2, the candidate paths are evaluated based on the node score, and the paths with low scores are eliminated by using the pruning strategy; The node score is calculated by a comprehensive score function; For the jth candidate node of the tth round of reasoning its score The calculation formula is as follows: ； wherein, represents the tool observation based outcome the computed tool usage effectiveness score, represents the node decision based outcome the computed contextual semantic consistency score, represents the node state the relevance score to the task goal ; , , are adjustable weight coefficients for controlling the influence degree of each score item on the comprehensive score; The pruning strategy adopts Top-K strategy, Beam Search strategy or UCB strategy, and the system calculates the comprehensive score of all candidate nodes after each reasoning expansion, and only keeps a few paths with high scores to continue expansion; S4, termination judgment and output: repeat steps S2 to S3 until the reasoning termination condition is met, then select the path with the highest cumulative score from the reasoning tree as the optimal reasoning path, and generate a structured intelligence report based on the path; S5, knowledge tree migration fine-tuning: use the optimal reasoning path as a supervision signal to perform knowledge distillation and supervised fine-tuning on the lightweight student model to migrate the reasoning ability of the large model; S6, accuracy judgment: evaluate the performance indicators of the fine-tuned student model on the validation task, if the performance meets the standard, go to the next step, otherwise return to step S7 to adjust the training strategy; S7, lightweight model construction: quantize and prune the student model with performance meeting the standard to generate a lightweight model suitable for resource-constrained platforms.

2. The method of claim 1, wherein, In step S2, the construction process of the reasoning tree is dynamic and iterative; In the tth reasoning round, the reasoning tree is represented as: ； wherein, represents a set of inference nodes generated at time t; represents a set of directed edges between inference nodes, used to describe the parent-child relationship between inference nodes; i-th inference node is represented as: ； wherein, represents a structured decision of the large language model at the i-th inference node, represents a tool invocation decision of the large language model at the i-th inference node, represents a structured observation returned by the large language model after invoking a tool at the i-th inference node, represents a score or confidence of the i-th inference node.

3. The method of claim 2, wherein the method further comprises: In step S2, the iterative expansion process of the reasoning tree is as follows: Structured decision making for inferencing nodes Generated in a manner that: ； wherein, represents a large language model, x is an input remote sensing image feature, prompt is a system prompt word, is the reasoning tree state of the previous moment; Tool invocation decision With the observation result The generation mode is: ， ； wherein, is a tool invocation policy function, based on the current decision and tool set outputs a tool invocation decision ; is a tool invoked from the tool set , and is a parameter used when invoking the tool. In the reasoning process, when there are multiple available tools or strategies, k child nodes are generated in parallel based on the current node state to explore different reasoning paths: ； new node is inserted into the tree and connected to its parent , the node and edge set are updated: ； 。 4. The method of claim 3, wherein the method further comprises: The tool call is realized through a standardized model context protocol interface to realize the decoupling and plug-and-play of the tool and the main reasoning process; The tool includes at least one or more of the following tools: open target detection tool, image super-resolution tool, remote sensing image cloud removal / noise removal tool, image cropping tool, external network search engine or retrieval enhancement generation module.

5. The method of claim 1, wherein, In step S4, the reasoning termination condition includes at least one of the following: a) the task target is judged to be completed; b) the depth of the reasoning tree reaches the preset maximum value; c) the reasoning tree cannot be expanded effectively any more; selecting the path with the highest accumulated score from the inference tree as the optimal inference path the selection strategy is ； where P denotes a reasoning tree any complete path from the root node to a leaf node in P, the score of a node v in a path P, the optimal path is the path in which the sum of the node scores is highest.

6. The method of claim 1, wherein, In step S4, the structured intelligence report adopts SOAP format organization, including: Subject, indicating the core object or area of interest of the intelligence; Observation, listing the objective facts and data obtained through the tool chain; Evaluation, analysis and judgment based on the observation results; planning, proposing subsequent action recommendations or decisions based on the current assessment.

7. An electronic device comprising a memory and a processor, said memory having stored therein a computer program executable on the processor, characterized in that, The processor implements the steps of the remote sensing image interpretation method as claimed in any one of claims 1 to 6 when executing the computer program.

8. A computer readable storage medium, the storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the remote sensing image interpretation method as claimed in any one of claims 1 to 6.

Images