Low-altitude remote sensing image-oriented ground feature fine-grained attribute extraction method

Abstract

The invention discloses a low-altitude remote sensing image-oriented ground feature fine-grained attribute extraction method, which comprises the following steps of: constructing a multi-modal data cue word template guided by a fused thinking chain, constructing an attribute extraction network based on a multi-modal large model, executing thinking chain reasoning of the data cue word template, and evaluating a reasoning effect. And optimizing the attribute extraction network by adopting a progressive optimization strategy, deploying the optimized attribute extraction network, and carrying out low-altitude remote sensing ground feature fine-grained attribute extraction. According to the method, the dependence of an attribute extraction network on large-scale annotation data is remarkably reduced, and the general MLLM is enabled to quickly adapt to the field of low-altitude remote sensing with extremely low annotation cost and calculation overhead; generalization and expansion capabilities of diversified and long-tail ground feature attributes are improved; meanwhile, by designing an effective guiding mechanism, the MLLM guiding problem caused by large difference and high professionality in the field of low-altitude remote sensing images is solved, and the reliability and effect of Prompt are improved.

Description

Technical Field

[0001] This invention relates to the field of fine-grained attribute extraction technology for ground features in low-altitude remote sensing images, and particularly to a method for fine-grained attribute extraction of ground features from low-altitude remote sensing images. Background Technology

[0002] Low-altitude remote sensing technology (generally referring to the acquisition of remote sensing data using low-altitude platforms such as drones) has been widely used in urban planning, agricultural surveys, and disaster assessment due to its advantages of high resolution, high flexibility, and relatively low cost. Refined interpretation of ground features (such as buildings, roads, and vegetation) in low-altitude remote sensing images to obtain their fine-grained attributes (such as the precise shape and material of building roofs, road surface types, and tree species) is crucial for constructing high-precision digital twin cities and achieving accurate scene understanding and 3D reconstruction.

[0003] Existing technical solutions mainly face the following limitations and challenges: 1. Strong data dependence and high annotation costs: Existing methods mostly rely on deep convolutional neural networks for fully supervised training, requiring a large amount of pixel-level or instance-level finely annotated data. Fine-grained attribute annotation of low-altitude remote sensing data has a high professional threshold, long cycle, and extremely high cost, making it difficult to acquire on a large scale. 2. Poor model generalization ability and scalability: Traditional methods usually train a single dedicated model for a specific region, specific sensor, or specific task, resulting in high customization and development costs. They are often difficult to directly transfer to new and unseen scenarios, have weak processing ability for long-tail distributions (i.e., uncommon or diverse categories) in ground object attributes, and have poor task scalability. 3. Single modality and failure to fully utilize prior knowledge: Existing methods usually only utilize image information and fail to effectively combine multimodal prior knowledge such as text descriptions and Geographic Information System (GIS) data for joint reasoning, limiting the model's understanding and generalization ability. 4. Direct application of large models faces domain differences and efficiency issues: Although multimodal large models (MLLMs) demonstrate powerful zero-shot / few-shot learning capabilities in natural image understanding, their direct application in low-altitude remote sensing is ineffective. This is because low-altitude imagery differs significantly from natural images in terms of domain (e.g., shooting perspective, ground scale, texture features). Fine-tuning all parameters of the entire large model would incur enormous computational costs, storage requirements, and resource consumption, making it difficult to popularize in practical applications.

[0004] For example, invention application No. 202411326298.5 discloses a method for extracting fine-grained ground features from remote sensing images based on feature transformation and knowledge enhancement. Using this method, the network's ability to distinguish and extract fine-grained ground features can be improved, achieving automated and intelligent identification of fine-grained ground features, and exhibiting high accuracy in identifying ground features with high inter-class similarity. However, this method also suffers from several drawbacks: single modality in ground feature attribute extraction, high sample dependence, failure to fully utilize prior knowledge, and domain differences and inefficiencies when directly applying large models.

[0005] Based on the above challenges, this invention proposes a prompt word guidance and fine-tuning strategy specifically designed for low-altitude ground feature attribute extraction. It aims to efficiently adapt to general multimodal large models and achieve efficient, accurate, and flexible extraction of fine-grained attributes of low-altitude remote sensing ground features using very few labeled samples. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to provide a method for extracting fine-grained attributes of ground features from low-altitude remote sensing images, thereby solving the problems of single modality, failure to fully utilize prior knowledge, and domain differences and inefficiency in the direct application of large models in existing low-altitude remote sensing ground feature attribute extraction methods.

[0007] This invention provides a method for extracting fine-grained attributes of ground features from low-altitude remote sensing images.

[0008] First aspect: A method for extracting fine-grained attributes of ground features from low-altitude remote sensing images, including: S1. Construct a multimodal data prompt word template that integrates thought chain guidance; S2. Construct an attribute extraction network based on a multimodal large model and perform thought chain reasoning based on data prompt word templates; S3. Evaluate the reasoning performance and use a progressive optimization strategy to optimize the attribute extraction network; S4. Deploy the optimized attribute extraction network to extract fine-grained attributes of ground features from low-altitude remote sensing.

[0009] In one embodiment of the present invention, step S1 includes the following steps: S11. Acquire low-altitude remote sensing images; S12. Construct a set of image-text pairs describing the fine-grained attributes of ground features; S13. Based on the image-text pair description set, design a prompting project, integrate the thinking chain of low-altitude remote sensing image analysis, and form a data prompt word template.

[0010] In one embodiment of the present invention, the data prompt word template includes: The system role definition section clarifies the professional role played by the model; The task definition section clearly explains the specific tasks involved in this analysis. The reasoning steps guidance section of the thought chain uses step-by-step, structured natural language to clearly guide the model to simulate the analytical logic and process of experts; The structured output constraint part outputs results in a specific structured format through the constraint attribute extraction network.

[0011] In one embodiment of the present invention, the reasoning step guidance portion of the thought chain includes the following reasoning steps: S21. Confirm that the two images are basically consistent in terms of viewpoint and range, and perform preliminary registration; S22. Visual feature comparison of the building's outline shape, roof features, facade materials, and shadow geometry; S23. Based on the visual feature comparison results of S22, conduct comprehensive logical reasoning by combining architectural common sense.

[0012] In one embodiment of the present invention, the structured output constraint part, the constraint attribute extraction network outputs the results in JSON format, including change type, region, attributes of each period and natural language description.

[0013] In one embodiment of the present invention, the multimodal large model in S2 adopts Qwen2.5-VL or InternVL-2.0.

[0014] In one embodiment of the present invention, the incremental optimization strategy in S3 includes: zero-sample testing, few-sample optimization, and efficient parameter fine-tuning, wherein: Zero-sample testing involves directly using thought chain clues for reasoning. If the results are satisfactory, this approach will be adopted directly. The few-shot optimization is as follows: if the zero-shot test is not effective, add image-text example pairs to the prompt words to provide the model with more explicit contextual learning examples, and test the effect again.

[0015] Efficient parameter fine-tuning involves introducing a lightweight adapter or low-rank adaptation technique to update only a very small number of parameters if optimization with a few samples still does not meet the requirements, thereby enabling the model to adapt more deeply to the low-altitude remote sensing field.

[0016] In one embodiment of the present invention, the fine-grained attributes in S4 include the roof shape and material of the building, the surface type and number of lanes of the road, the tree species identification of the vegetation, and the variation type of the building.

[0017] Second aspect: An electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, performs the steps of the method provided in the first aspect.

[0018] Third aspect: A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method provided in the first aspect.

[0019] The beneficial effects of this invention are: 1. The method of this invention significantly reduces the model's dependence on large-scale labeled data, enabling general-purpose MLLMs to be quickly adapted to the low-altitude remote sensing field with extremely low labeling costs and computational overhead. This invention introduces efficient parameter fine-tuning techniques, requiring only a very small number of parameters (less than 1% of the total model parameters) to train, combined with a small number of samples (typically hundreds of images) to achieve excellent performance. This reduces the computational overhead and storage requirements for model fine-tuning by 1-2 orders of magnitude, making high-precision ground feature interpretation possible with limited budgets and computational resources, greatly promoting the practical application of the technology. The method provided by this invention brings breakthroughs in reducing dependence, improving accuracy, enhancing generalization, ensuring reliability, and optimizing deployment, providing an efficient, economical, and flexible solution for fine-grained attribute extraction of low-altitude remote sensing ground features, with extremely high practical value and broad industrialization prospects.

[0020] 2. The method of this invention demonstrates superior generalization and extension capabilities for diverse and long-tailed land cover attributes, enabling a unified model to flexibly handle multiple attribute extraction tasks for various land cover types such as buildings, roads, and vegetation, avoiding the complex deployment of a separate model for each task. This method exhibits excellent task scalability and scene generalization capabilities, effectively addressing long-tail demands. This invention achieves a paradigm shift from a single model for a single task to a single model for multiple tasks. Only the input natural language prompts need to be modified; the model can be flexibly extended to new land cover types or attribute tasks without retraining. Based on the powerful prior knowledge and semantic understanding capabilities of the pre-trained large model, this method exhibits excellent zero-shot and few-shot generalization capabilities for unknown scenes, novel sensors, and rare land cover categories, solving the fundamental pain point of poor generalization performance in traditional dedicated models.

[0021] 3. The method of this invention extracts fine-grained attributes, surpassing coarse-grained classification, and accurately identifies specific attributes of land features such as material, shape, and type. This invention overcomes the limitations of traditional methods in terms of coarse-grained attribute extraction, achieving highly flexible and adaptable fine-grained understanding. This invention flexibly defines query tasks through natural language prompts, enabling the model to simultaneously output multiple refined attributes of land features. Using a more unified model provides a data foundation with far greater information than traditional methods for applications such as building digital twin cities and high-precision scene reconstruction, achieving a crucial leap from perception to cognition.

[0022] 4. The method of this invention designs an effective guidance mechanism to overcome the challenges of MLLM guidance caused by the large differences and high professionalism in the field of low-altitude remote sensing images, and improves the reliability and effectiveness of prompts.

[0023] 5. The method of this invention improves the controllability and reliability of model reasoning through the Chain of Thought (CoT) framework, resulting in outputs that are more consistent with professional logic. One of the core innovations of this invention lies in embedding the analytical thinking of low-altitude remote sensing experts into the model reasoning process through structured prompts. This guidance effectively constrains the model's thinking path, making its analysis process more interpretable and its outputs closer to professional judgment, significantly reducing the randomness and error rate of generated content. Simultaneously, the mandatory structured output greatly facilitates the automatic parsing and integration of the backend system, meeting the high requirements of stability and reliability in production environments. Attached Figure Description

[0024] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the principle framework of the method of the present invention; Figure 3 This is a schematic diagram of the structure of the electronic device of the present invention. Detailed Implementation

[0025] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0026] Existing methods for extracting fine-grained attributes of ground features from low-altitude remote sensing images suffer from several problems: reliance on large-scale labeled data, complex deployment of one model per task, poor fine-grained attribute extraction performance, and a lack of effective guidance mechanisms.

[0027] To address the aforementioned problems, this invention provides a method for extracting fine-grained attributes of ground features from low-altitude remote sensing images. This method utilizes expert-guided thought process prompts to guide a large multimodal model through progressive and reliable reasoning, thereby achieving efficient and high-precision extraction of refined attributes of ground features.

[0028] Example 1: A method for extracting fine-grained attributes of ground features from low-altitude remote sensing images, the principle and technical architecture of which are as follows: Figure 2 As shown, the flowchart is as follows Figure 1 As shown, the method includes the following steps: S1. Construct a multimodal data prompt template guided by integrated thinking chain.

[0029] First, low-altitude remote sensing images are acquired and preprocessed, including noise removal, geometric correction, and radiometric correction, to improve image quality and usability.

[0030] Then, based on innovative Prompt Engineering technology, a Prompt Engineering design is used to construct a set of image-text pairs describing the fine-grained attributes of the land features. The image-text pairs in this set accurately reflect various fine-grained attributes of the land features, including but not limited to the roof shape and material of buildings, the surface type and number of lanes of roads, the tree species identification of vegetation, and the variation type of buildings.

[0031] Then, based on the image-text pair description set, combined with cueing engineering, and integrating the chain-of-thought (CoT) of low-altitude remote sensing image analysis, a data cue word template is formed. This template incorporates the logical reasoning process of the chain of thought, guiding the multimodal large model to reason according to specific analysis steps and ideas when processing low-altitude remote sensing images.

[0032] Furthermore, the data prompt word template includes the following parts: In the system role definition section, the professional role played by the model is clearly defined, such as "You are a professional low-altitude remote sensing image analysis expert".

[0033] The task definition section clearly explains the specific task of this analysis, such as "Compare the following two sets of time series images and identify the changes in buildings within a specified area".

[0034] The reasoning process guidance section of the thought chain uses step-by-step, structured natural language to clearly guide the analytical logic and process of the attribute extraction network simulation expert.

[0035] The structured output constraint section constrains the attribute extraction network to output results in a specific structured format. Strictly constrained attribute extraction networks must output results in a specific structured format (such as JSON), including one or more of the following: change type, region, period attributes, natural language description, and confidence level, ensuring the output is machine-readable, standardized, and reliable.

[0036] Furthermore, the reasoning steps in the thought chain guidance section include the following steps: First, confirm that the two images are basically consistent in terms of viewpoint and range, and perform preliminary registration; Secondly, visual feature comparisons were made of the building's outline shape, roof features, facade materials, and shadow geometry; Then, based on the above visual feature comparison results, and combined with architectural common sense, comprehensive logical reasoning is performed (e.g., a change in roof material but an unchanged outline may indicate roof renovation; an enlarged overall outline may indicate expansion). An example of a prompt template design specifically applied to building roof types: <|im_start|>system You are a professional low-altitude remote sensing image analysis expert, specializing in the detailed interpretation of building changes. Please strictly follow the following steps for analysis and output the results in the required format.

[0037] **Task Definition:** By comparing the two sets of time-series low-altitude remote sensing images (T1 and T2), identify and describe the changes in buildings within the specified area.

[0038] **Output format requirements (must be strictly followed):** { "change_type": "[Add building | Demolish building | Renovate building | Roof renovation | No change]", "change_region": Briefly describe the geographical location or building area that has changed, such as 'a detached building slightly east of the center of the image', "t1_attributes": { "existence": "exists / does not exist", "roof_shape": "[flat roof | sloping roof | dome | cannot be determined]", "roof_material": "[Concrete | Roofing tile | Metal | Cannot determine]" }, "t2_attributes": { "existence": "exists / does not exist", "roof_shape": "[flat roof | sloping roof | dome | cannot be determined]", "roof_material": "[Concrete | Roofing tile | Metal | Cannot determine]" }, "description": "A detailed natural language description of the change process, including significant changes in size, outline, and other salient features." } **Analysis Steps:** 1. **Image Registration and Preliminary Judgment:** Confirm that the perspectives of the two images are basically the same, and make a preliminary judgment on whether there are any macroscopic changes.

[0039] 2. **Detailed Feature Comparison:** Analyzes the building's outline, roof shape, material texture, shadows, and other features one by one.

[0040] 3. **Logical Reasoning:** Based on the feature comparison results, reasoning is performed in conjunction with common architectural knowledge (e.g., if the roof material and shape change at the same time, it may be a complete renovation).

[0041] 4. **Structured Output:** Fill the inference results into the JSON format described above.

[0042] <|im_end|> <|im_start|>user <|vision_start|><|image_pad|><|vision_end|> S2. Construct an attribute extraction network based on a multimodal large model and perform thought chain reasoning based on data prompt word templates.

[0043] A pre-trained multimodal large model (such as Qwen2.5-VL, InternVL-2.0, etc.) is used as the core backbone network. The image constructed in step S1 and the data prompt word template fused with the thought chain are input into this network. During the inference process, the model's built-in reasoning ability is activated and guided by the provided thought chain steps, thereby performing a more reliable and interpretable step-by-step analysis and ultimately generating a structured output result.

[0044] S3. Evaluate the reasoning performance and use a progressive optimization strategy to optimize the attribute extraction network.

[0045] To evaluate the accuracy of the output of step S2, this invention employs a progressive optimization strategy with increasing cost, specifically including... Zero-Shot testing: First, use thought chain prompts directly for reasoning. If the results are satisfactory, then adopt this approach directly, as it has the lowest cost.

[0046] Few-Shot Optimization: If zero-shot results are unsatisfactory, add 1-2 image-text pairs to the prompt words to provide the model with more explicit contextual learning examples, and test the effect again.

[0047] Efficient Parameter Fine-Tuning (PEFT): If a small number of samples still do not meet the requirements, efficient fine-tuning is initiated. This is achieved using a small amount of labeled data by introducing a lightweight adapter or low-rank adaptation (LoRA) technique. The lightweight adapter is a small neural network module inserted into the Transformer layer of the large multimodal model.

[0048] Only a very small number of parameters are updated to allow the model to be more deeply adapted to the low-altitude remote sensing field. After fine-tuning, the original thought chain prompt template will perform even better. The low-altitude remote sensing image to be processed and the designed thought chain prompts are input into the model after the above process to obtain the model extraction results.

[0049] S4. Deploy the optimized attribute extraction network to extract fine-grained attributes of ground features from low-altitude remote sensing.

[0050] After deploying the optimized attribute extraction network, image data acquired through low-altitude remote sensing is received in real time. Based on the features and patterns learned during the previous optimization process, the attribute extraction network performs detailed attribute analysis on ground features in the images. For different types of ground features, such as buildings, vegetation, and roads, the network can accurately identify their boundaries and features, accurately presenting the fine-grained attribute changes of ground features in low-altitude remote sensing images. For example, the change type may cover new construction, demolition, and renovation, and the attributes for each period include detailed information such as the height, area, and material of buildings.

[0051] Furthermore, based on the multimodal large model, natural language descriptions are used to further explain these changes. For example, it can be explained that the construction of new buildings in a certain area is due to urban planning expansion, or that renovations are to improve the functionality of buildings. In this way, strong data support and clear information references can be provided for research and decision-making in related fields, making ground feature analysis based on low-altitude remote sensing imagery more scientific, accurate, and efficient.

[0052] This invention, through the embedding of expert thought processes from low-altitude remote sensing images into prompt word templates, fundamentally guides the model to generate more reliable outputs that align with domain knowledge by embedding professional analysis processes (registration, feature comparison, and logical reasoning) into the prompt words, thus solving the problem of insufficient professional reasoning capabilities in general-purpose models. Simultaneously, this invention achieves a unification of standardization and intelligence; the prompt word templates not only transform diverse remote sensing tasks into a unified vLLM-processable format but also endow the model with expert-level thinking modes through natural language commands, combining the advantages of standardization and intelligence. Furthermore, for model optimization, an economical and efficient incremental optimization path is adopted: clearly defining the optimization order of prompt engineering first, followed by fine-tuning, prioritizing the use of zero-cost and low-cost prompt word design to solve problems, and systematically reducing dependence on labeled data and computing power.

[0053] The method of this invention significantly improves the professional analysis capabilities and output reliability of large models in the field of low-altitude remote sensing, and provides a brand-new technical path for efficient and accurate interpretation of ground features.

[0054] Application Experiment 1: Extraction of building roof attributes based on adapter.

[0055] Data Preparation: Collect 100 low-altitude remote sensing images of urban areas, and have experts annotate the shape (flat roof, pitched roof, dome) and material (metal, asphalt, tile, concrete) of building roofs. Design a special extraction prompt template; the content of the building roof attribute extraction prompt template is shown below: <|im_start|>system You are a professional low-altitude remote sensing image analysis expert. Please strictly follow the following steps to analyze the roof material of buildings in the image and output the results in JSON format.

[0056] **Analysis Steps:** 1. Location: First, identify the main buildings in the image.

[0057] 2. Focus: Concentrate your attention on the roof area of ​​the building.

[0058] 3. Analysis: Carefully observe the visual characteristics of the roof, including its texture, reflectivity, color, and pattern.

[0059] 4. Reasoning: Infer the material based on visual features (options: metal, asphalt, tile, concrete).

[0060] 5. Output: Strictly adhere to the JSON format below for output results.

[0061] **Output format (must be strictly followed):** {"roof_material": "[metal / asphalt / tiles / concrete]"} <|im_end|> <|im_start|>user Here are two analysis examples; please study the analysis patterns within them: Example 1: <Image 1: Silver High-Reflectivity Flat Roof> Analysis: The roof surface is smooth and highly reflective, which is a typical characteristic of metallic materials.

[0062] Output: {"roof_material": "metal"} Example 2: <Image 2: Dark gray grainy textured roof> Analysis: The roof is dark gray in color and has a matte, granular texture, which is consistent with the characteristics of asphalt material.

[0063] Output: {"roof_material": "asphalt"} Now, please analyze the following new image: <Image_3: Remote sensing image to be analyzed> Please determine the material of the roof of the building in the picture.

[0064] <|im_end|> Attribute extraction and effect evaluation: Extraction process: The constructed prompts (including system instructions, few sample examples, and new images to be analyzed) are directly input into the pre-trained Qwen2.5-VL-3B model without any fine-tuning.

[0065] Model output: The model will follow the thought chain in the prompt words to reason and output results that strictly conform to the JSON format, for example: {"roof_material": "tile"}.

[0066] Performance evaluation: Run the above process on 50 test images and compare the model's prediction results with the expert-annotated ground truth labels. Calculate the accuracy metric.

[0067] Results: Experiments show that, using only the carefully designed prompts mentioned above, the model can achieve an accuracy of over 90% on this test set, fully meeting the expected evaluation criteria of the project (e.g., >85% is considered satisfactory).

[0068] Application Experiment 2: Road Change Detection Based on LoRA.

[0069] Data Preparation: Collect 200 pairs of remote sensing images (T1 and T2) of the same area at different time periods, and label the types of road changes (e.g., new construction, widening, closure, no change). For each image pair, construct a prompt text: "Compare these two time-series images and describe the changes in the road network. Options: A. New construction, B. Widening, C. Closure, D. No change."

[0070] Model Construction: The Qwen2-VL multimodal large model was selected. Without changing its network structure, LoRA technology was applied to the linear layers in the model, adding a set of low-rank matrices (rank r=8) in parallel.

[0071] Model Training and Inference: The model is trained using 200 sets of samples. During training, only the low-rank matrix parameters introduced by LoRA are updated. After training, the model can output structured descriptions of changes when given new image pairs and query text.

[0072] Experiments show that the above two implementation methods can achieve performance close to or reach that of the traditional method using 100% data for full parameter fine-tuning when only 1%-10% of the labeled data is used, which fully demonstrates the effectiveness and efficiency of the present invention.

[0073] The present invention also provides an electronic device, Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 3 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method: S1. Construct a multimodal data prompt word template that integrates thought chain guidance; S2. Construct an attribute extraction network based on a multimodal large model and perform thought chain reasoning based on data prompt word templates; S3. Evaluate the reasoning performance and use a progressive optimization strategy to optimize the attribute extraction network; S4. Deploy the optimized attribute extraction network to extract fine-grained attributes of ground features from low-altitude remote sensing.

[0074] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example: S1. Construct a multimodal data prompt word template that integrates thought chain guidance; S2. Construct an attribute extraction network based on a multimodal large model and perform thought chain reasoning based on data prompt word templates; S3. Evaluate the reasoning performance and use a progressive optimization strategy to optimize the attribute extraction network; S4. Deploy the optimized attribute extraction network to extract fine-grained attributes of ground features from low-altitude remote sensing.

[0076] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0077] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting fine-grained attributes of ground features from low-altitude remote sensing images, characterized in that, include: S1. Construct a multimodal data prompt word template that integrates thought chain guidance; S2. Construct an attribute extraction network based on a multimodal large model and perform thought chain reasoning based on data prompt word templates; S3. Evaluate the reasoning performance and use a progressive optimization strategy to optimize the attribute extraction network; S4. Deploy the optimized attribute extraction network to extract fine-grained attributes of ground features from low-altitude remote sensing.

2. The method according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire low-altitude remote sensing images; S12. Construct a set of image-text pairs describing the fine-grained attributes of ground features; S13. Based on the image-text pair description set, design a prompting project, integrate the thinking chain of low-altitude remote sensing image analysis, and form a data prompt word template.

3. The method according to claim 1, characterized in that, The data prompt word template includes: The system role definition section clarifies the professional role played by the model; The task definition section clearly explains the specific tasks involved in this analysis. The reasoning steps guidance section of the thought chain uses step-by-step, structured natural language to clearly guide the model to simulate the analytical logic and process of experts; The structured output constraint part outputs results in a specific structured format through the constraint attribute extraction network.

4. The method according to claim 3, characterized in that, The reasoning steps guided by the thought chain include the following steps: S21. Confirm that the two images are basically consistent in terms of viewpoint and range, and perform preliminary registration; S22. Visual feature comparison of the building's outline shape, roof features, facade materials, and shadow geometry; S23. Based on the visual feature comparison results of S22, conduct comprehensive logical reasoning by combining architectural common sense.

5. The method according to claim 3, characterized in that, The structured output constraint part, the constraint attribute extraction network outputs the results in JSON format, including change type, region, attributes of each period and natural language description.

6. The method according to claim 1, characterized in that, The multimodal large model in S2 adopts Qwen2.5-VL or InternVL-2.

0.

7. The method according to claim 1, characterized in that, The incremental optimization strategy in S3 includes: zero-sample testing, few-sample optimization, and efficient parameter fine-tuning, wherein: Zero-sample testing involves directly using thought chain clues for reasoning. If the results are satisfactory, this approach will be adopted directly. The few-shot optimization is as follows: if the zero-shot test results are not good, add image-text example pairs to the prompt words to provide the model with more explicit context learning examples, and test the results again; Efficient parameter fine-tuning involves introducing a lightweight adapter or low-rank adaptation technique to update only a very small number of parameters if optimization with a few samples still does not meet the requirements, thereby enabling the model to adapt more deeply to the low-altitude remote sensing field.

8. The method according to claim 1, characterized in that, The fine-grained attributes in S4 include the roof shape and material of buildings, the surface type and number of lanes of roads, the tree species identification of vegetation, and the type of building variation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 8.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images