Meteorological data set automatic construction method and system based on modal bridging

Abstract

The invention relates to a meteorological data set automatic construction method and system based on modal bridging, and aims to meet the multi-modal large model training requirement in the meteorological field, and the method and system realize automatic conversion from an original meteorological image to a structured expert reasoning text through deep fusion of image and text information. The method comprises five stages of data preprocessing, image-text semantic modeling, causal reasoning generation, consistency screening and parallel processing, key meteorological elements are extracted by using a multi-modal model, a chained thinking process is constructed through a language model with meteorological knowledge, chained reasoning annotation is realized, cross-modal semantic alignment and a multi-round reasoning mechanism are introduced, and a multi-modal semantic model is established. The method is advantaged in that high-quality samples are screened in combination with rules and models, automation, high consistency and good expansibility are realized, manual annotation cost is substantially reduced, and the method is suitable for large-scale meteorological reasoning multi-modal data set construction.

Description

Technical Field

[0001] The present invention relates to the field of meteorological data processing, and in particular to a method and system for automatically constructing a meteorological data set based on modal bridging. Background Art

[0002] With the rapid development of artificial intelligence (AI) technology, the application prospects of large multimodal language models (such as vision-language models) in intelligent reasoning, disaster identification, and automatic analysis in the meteorological field are becoming increasingly broad. Traditional methods for constructing datasets in the meteorological field mainly rely on meteorological experts to manually analyze and annotate meteorological images and observation data. This is inefficient and costly, and it is difficult to meet the demand for large-scale, high-quality data. In recent years, continuous advances in technologies such as cross-modal understanding, image-text alignment, and Chain of Thought (CoT) have provided new ideas for the automated construction of meteorological datasets. Modal bridging technology can automatically integrate meteorological images and text information, using large models to achieve efficient and accurate data annotation and knowledge reasoning, thereby reducing dependence on human experts and improving the speed, scale, and quality of data construction.

[0003] At present, the relevant technical solutions for automatic labeling and knowledge reasoning of multimodal data mainly include: Manual expert labeling: Under the traditional method, it mainly relies on meteorological experts to manually analyze and label each piece of data. It is suitable for small-scale, high-precision tasks, but has limited efficiency in large-scale and diverse scenarios. Weak supervision and transfer learning methods: Some studies have tried to use a small number of labeled samples to generate more data through transfer learning or semi-automatic methods to improve labeling efficiency, but the complexity and accuracy of the reasoning chain are still limited. Large image-text multimodal models: such as CLIP, BLIP, LLaVA, etc., can achieve joint understanding of meteorological images and text descriptions, and have certain automatic generation of labels and reasoning capabilities, but most existing models lack customized training for meteorological field knowledge. Automated data labeling platform: Combining manual review and machine generation to achieve preliminary labeling and subsequent correction of data sets, which can improve the efficiency of data set construction to a certain extent.

[0004] Deficiencies of existing technical solutions. High degree of manual dependence: Traditional solutions are highly dependent on meteorological experts, have a long labeling cycle, and are difficult to adapt to the needs of continuously expanding data scale. Limited automated reasoning capabilities: Existing multimodal models perform well in general scenarios, but still have problems with accuracy and depth in meteorological-specific knowledge reasoning and chain labeling, and lack effective integration of meteorological professional knowledge. Labeling consistency and standardization are difficult to guarantee: Semi-automatic or manual labeling is prone to style inconsistencies, broken reasoning chains, etc., which affect the overall quality of the data set. Insufficient scalability and portability: Some automatic labeling tools are developed for specific tasks and are difficult to flexibly adapt to diverse meteorological scenarios and new reasoning tasks. Summary of the Invention

[0005] To address the shortcomings of the existing technology, this paper proposes a method for automatically constructing a meteorological dataset based on modal bridging. The automatic construction method includes five stages: data preparation and preprocessing, modal bridging, detail reasoning and thought chain generation, consistency verification and sample screening, and parallel processing. Specifically,

[0006] Step 1: Data preparation and preprocessing stage: meteorological images and corresponding text description data are collected. Through preprocessing such as outlier removal, spatiotemporal matching, and format standardization, a unified original meteorological multimodal dataset is formed.

[0007] Step 2: Modal bridging stage, using multimodal models to analyze meteorological image content and generate detailed description text of meteorological images containing key meteorological elements;

[0008] Step 21: Multimodal model inference: The original weather image and the corresponding time label are used as input to the multimodal model to build a preliminary cross-modal understanding of the current weather situation. The multimodal model identifies and annotates significant meteorological features in the weather image. Based on the built-in image attention mechanism of the multimodal model, it locates key areas in the weather image and outputs structured semantic text.

[0009] Step 22: Generate a detailed description text of the meteorological image. Based on the structured semantic text, the regional scope, time information and meteorological background of the original meteorological image are combined to construct a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image.

[0010] Step 3: Detailed reasoning and thought chain generation phase: The detailed description text of the meteorological image and the original structured labels are input into a pre-trained language model with meteorological domain knowledge. This model then sequentially completes professional reasoning of meteorological image elements and constructs a thought chain of causal logic, generating chained reasoning text that conforms to the logic of meteorological experts.

[0011] Step 31: Detail inference: receiving the meteorological image detail description text and the original structured label as input, a general pre-trained language model with meteorological domain knowledge parses the meteorological elements involved in the detail description text and outputs intermediate semantic information;

[0012] Step 32: Thought chain generation. After completing the detailed-level factor reasoning, the pre-trained language model logically organizes and linguistically expands the intermediate semantic information. Based on the chain reasoning structure of observation-interpretation-deduction-judgment, it gradually simulates the analytical process of meteorological experts.

[0013] Step 33: The pre-trained language model finally outputs a structured chain reasoning text, which is represented in a standard annotation format.

[0014] Step 4: Consistency verification and sample screening: The generated chain reasoning text is evaluated for consistency and rationality through the rule system and fine-tuned discriminant model. High-quality samples with standardized structure, reasonable logic, and accurate expression are selected to construct the chain reasoning dataset.

[0015] Step 41: The rule system quickly checks the format and content logic of the chain reasoning text, fine-tunes the discriminant model to further determine the semantic consistency, logical closure, and rationality of meteorological knowledge in the text, and obtains the judgment result and semantic score for each chain reasoning text;

[0016] Step 42: After obtaining the structural judgment results and semantic scores of each chain reasoning sample, high-quality samples are screened to form the target chain reasoning dataset;

[0017] Step 5: Parallel processing and scheduling phase. In this phase, multi-threading and multi-GPU parallel processing mechanisms are introduced to divide the processing steps from Step 1 to Step 4 into independent task units. These units are then scheduled in batches to run simultaneously on multiple threads or GPU devices, enabling efficient batch processing of large-scale raw meteorological datasets.

[0018] The automatic construction system of meteorological dataset based on modal bridging includes a multimodal model, a detail generation module, a pre-trained language model, a parallel processing module, a rule system and a fine-tuning discrimination module, wherein:

[0019] The multimodal model is used to analyze and semantically model the pre-processed meteorological images during the modal bridging phase, identify and annotate significant meteorological features in the meteorological images, and locate key areas in the meteorological images based on the built-in image attention mechanism module of the multimodal model, outputting them as structured semantic text.

[0020] The detail generation module is used to generate detailed description text of meteorological images. Based on the structured semantic text generated by the multimodal model, it introduces a weather term dictionary and professional sentence templates, combines the regional scope, time information and meteorological background of the original meteorological image, and constructs a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image;

[0021] The pre-trained language model completes multiple rounds of reasoning and causal chain construction on meteorological image information based on detailed description text and original structured labels, and outputs chain reasoning text that simulates the analysis process of meteorological experts;

[0022] The rule system and fine-tuning discriminant module are used to jointly verify the chain reasoning text, eliminating samples that do not conform to the label structure, have poor semantic consistency, or have incomplete reasoning chains. High-quality text that conforms to professional logic, expression standards, and closed reasoning loops is retained to form the target chain reasoning dataset.

[0023] The parallel processing module is used to divide the data processing of each module into independent units and perform multi-threaded and multi-GPU parallel processing in batches.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The modal bridging method proposed in this paper automatically integrates meteorological imagery and text information. Combined with a multimodal large language model, it enables chained reasoning and annotation, significantly reducing the need for manual intervention by meteorological experts. This solution supports multi-task parallel processing, improves data construction efficiency, significantly reduces labor and time costs, and effectively adapts to the needs of rapidly generating large-scale datasets.

[0026] 2. The method of the present invention introduces cross-modal semantic alignment and multi-round reasoning mechanisms in the automatic annotation process to improve the consistency and accuracy of the annotation results, ensure the integrity of the knowledge chain, and enhance the professionalism and generalization ability of model training.

[0027] 3. Compared with traditional manual labeling or weakly supervised amplification methods, the present invention has good scalability and versatility, can flexibly adapt to diverse meteorological scenarios and various reasoning tasks, and provides a solid data foundation for intelligent reasoning and application innovation of multimodal large models in the weather field. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 is a process flow chart of the automatic construction method of the present invention;

[0029] Figure 2 It is a structural diagram of the automatic construction system of the present invention. DETAILED DESCRIPTION

[0030] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.

[0031] The following is a detailed description with reference to the accompanying drawings.

[0032] The present invention mainly solves the problems of strong dependence on meteorological experts and low efficiency of manual data annotation in the process of building multimodal datasets in the meteorological field. Traditional multimodal dataset generation methods often require the participation of a large number of meteorological experts, and have high requirements for manpower and professional knowledge, which makes it difficult to meet the needs of rapid construction of large-scale datasets. With the rapid development of multimodal large language model technology, it has brought about improvements in the ability to understand multi-source data and reason about knowledge, and also provided a new technical path for the automated construction of datasets. The modal bridging method can effectively integrate meteorological images and text information. Without relying on experts to manually annotate one by one, it automatically performs cross-modal understanding and chain reasoning annotation through the model, thereby greatly improving the efficiency of multimodal dataset generation. Faced with the actual situation of large data volume and strong heterogeneity in the meteorological field, traditional manual annotation methods are difficult to support the continuously growing data demand and can no longer meet the requirements of large model training for real-time construction of diversified and high-quality data. Therefore, it is necessary to make full use of the automatic annotation capability of modal bridging, combined with methods such as parallel processing, to maximize the construction speed and scale of multimodal datasets in the meteorological field, and promote the application and development of multimodal large language models in the field of meteorological intelligent reasoning.

[0033] In view of the shortcomings of the existing technology, the present invention proposes a method for automatically constructing a chain reasoning text dataset based on the modal bridging method. The processing flow is as follows: Figure 1 As shown below, combined Figure 1 The proposed automatic construction method is explained in detail.

[0034] Step 1: Data preparation and preprocessing: Meteorological images and corresponding text description data are collected. A unified original meteorological multimodal dataset is formed through preprocessing including outlier removal, spatiotemporal matching, and format standardization. Original meteorological datasets containing meteorological images (such as radar images and satellite cloud images) and corresponding text descriptions (such as weather conditions and disaster reports) are collected and processed for outlier processing, spatiotemporal matching, and format standardization. Outlier processing includes removing missing values, invalid images, and non-meteorological observation data; spatiotemporal matching synchronizes images and text descriptions through timestamp alignment and spatial grid reprojection; format standardization includes image size unification, text encoding consistency, and field structure standardization to ensure consistency in data input, thereby obtaining structured original meteorological data input.

[0035] Step 2: In the modal bridging phase, the meteorological image content is analyzed through a multimodal model to generate detailed description text of the meteorological image containing key meteorological elements. This specifically includes multimodal model reasoning and meteorological image detailed description text generation. The meteorological image is input into a multimodal large language model with a large number of parameters. The large model's cross-modal understanding and text generation capabilities are used to automatically output a description text containing rich meteorological image details. The detailed description text specifically includes the numerical characteristics and extreme values ​​of various meteorological parameters (such as temperature, wind speed, air pressure, wind shear, etc.); the spatial distribution and distribution pattern of each parameter; the manifestation of meteorological phenomena corresponding to each region and administrative unit; and indicators such as the morphology, location, and intensity of important weather systems.

[0036] The main task of the modal bridging stage is to input the preprocessed meteorological image data into a multimodal model with text and image understanding capabilities, such as Qwen2.5-VL-72B-Instruct. The multimodal model analyzes and semantically models the image content, and outputs detailed description text that can accurately describe key meteorological elements in the image (such as wind shear, precipitation structure, unstable areas, etc.) as an intermediate semantic representation.

[0037] Step 21: Multimodal model inference: The original meteorological image (such as radar reflectivity map, satellite cloud map, wind field map, etc.) and the corresponding time label are used as the input of the multimodal model to build a preliminary cross-modal understanding of the current weather situation. The multimodal model identifies and annotates the significant meteorological features in the meteorological image, locates the key areas in the meteorological image based on the built-in image attention mechanism module of the multimodal model, and outputs structured semantic text.

[0038] First, the original meteorological image and time information are packaged and sent to a multimodal model with visual-language understanding capabilities for analysis. The multimodal model uses the existing image-text joint pre-training capabilities to identify and annotate significant meteorological features in meteorological images (such as wind shear areas, precipitation belts, convective cores, etc.). During the recognition and annotation process, the image attention mechanism module built into the multimodal model is used as an intermediate tool to locate key areas in the meteorological image and improve the multimodal model's ability to pay attention to meteorological image details. The output of the multimodal model is a structured intermediate semantic representation text that summarizes the main weather phenomena, spatial distribution characteristics, and potential risk warnings reflected in the current meteorological image. The structured semantic text will serve as the input basis for the downstream detail generation module.

[0039] Step 22: Generate detailed description text of the meteorological image. Based on the structured semantic text and combined with the regional scope, time information and meteorological background of the original meteorological image, construct a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image.

[0040] First, structured semantic text and optional regional information (such as province, city, and weather station number) are fed into the detail generation module as prompt input. The module generates the details using a large language model fine-tuned by instructions, enabling step-by-step development and causal expression. To improve generation quality, a weather terminology dictionary and professional sentence templates are introduced as auxiliary components during the detail generation process to standardize terminology and enhance language consistency. These weather terminology dictionaries and professional sentence templates are based on existing text data.

[0041] The final detailed description text output will cover all types of meteorological parameters identified in the meteorological image (such as extreme wind speed, distribution of cold and warm fronts, precipitation intensity, uplift triggering conditions, etc.), and will be presented in a logically clear and terminologically accurate manner. It is the core text basis for subsequent meteorological reasoning.

[0042] Steps 21 and 22 embody the cross-modal semantic alignment mechanism. Specifically, the multimodal model realizes semantic modeling of meteorological image content by constructing an embedding space between images and text. With the help of the image attention mechanism, the model can accurately locate key meteorological elements in the image (such as convective core, front, echo structure), and perform semantic matching in combination with the weather terminology dictionary to automatically complete the alignment from image content to professional terminology. Subsequently, combined with the regional scope and time information of the image, a detailed description text in the style of meteorological experts is generated to achieve deep fusion and unified expression of multimodal information.

[0043] Step 3: In the detailed reasoning and thought chain generation stage, the detailed description text of the meteorological image and the original structured labels are input into a pre-trained language model with meteorological domain knowledge, and the professional reasoning of meteorological image elements and the construction of the causal logic thought chain are completed in sequence to generate a chain reasoning text that conforms to the logic of meteorological experts.

[0044] In the detailed reasoning and thought chain generation stage, the image detail description text generated in the modal bridging stage and the original structured text information (such as labels such as the affected area and severity) are input into a pre-trained language model with meteorological domain knowledge. Without relying on explicit rule templates, the model completes multiple rounds of reasoning and causal chain construction on the image information, and outputs chain reasoning text that simulates the analysis process of meteorological experts. The chain reasoning text includes the logical reasoning process and conclusive judgment results, and is output in a structured format.

[0045] The detailed description of the meteorological image output in step 2, along with the original text labels (such as "affected area" and "severity"), is fed into a large-scale pre-trained language model capable of meteorological reasoning. This model, requiring no additional construction, directly leverages existing large-scale models trained or fine-tuned with high-quality meteorological reasoning data, enabling it to logically understand and express weather evolution. Upon receiving the input data, the model automatically identifies key meteorological information and causal relationships within the detailed description. It then performs multiple rounds of reasoning around factors such as regional assessment and disaster-causing conditions to generate a complete, expert-level analysis.

[0046] This process doesn't rely on explicit rule templates, but rather is accomplished by the model's inherent meteorological knowledge and language organization capabilities. The output text includes a professional interpretation of key features in the image, regional impact analysis, and logic for determining severe weather levels. The results are presented in a structured format and can be directly used for subsequent evaluation and training.

[0047] Step 31: Detail inference, receiving the meteorological image detail description text and the original structured labels (such as area range, degree of attention) as input, and using a general pre-trained language model with meteorological domain knowledge to parse the meteorological elements involved in the detail description text (such as wind field structure, humidity gradient, echo morphology, etc.) and output intermediate semantic information.

[0048] The pre-trained language model automatically determines the mechanisms of action of various factors in the current context and internally models the relationships between these factors (e.g., frontal triggering and convective lift, wind shear and storm intensification), generating intermediate semantic information. This reasoning process demonstrates the model's ability to understand the disaster-causing mechanisms underlying meteorological imagery and provides semantic support for subsequent reasoning chain construction.

[0049] Step 32: Thought chain generation. After completing the reasoning of elements at the detail level, the pre-trained language model logically organizes and linguistically expands the intermediate semantic information. Based on the chain reasoning structure of observation-interpretation-deduction-judgment, it gradually simulates the analysis process of meteorological experts.

[0050] The analysis process includes inferring weather system evolution trends from image features, analyzing potential disaster paths, predicting potentially affected areas and their impact, and forming a quantitative assessment of the disaster's severity. The entire process embodies a chain of reasoning: observation—interpretation—deduction—judgment. The pre-trained language model sequentially unfolds the analysis process, generating natural, clearly structured, and progressively reasoned text, showcasing the entire process from data interpretation to conclusion.

[0051] Step 33: The pre-trained language model finally outputs a structured chain reasoning text, which is represented in a standard annotation format.

[0052] Specific and complete reasoning process <think> ...< / think> The label package covers the gradual interpretation and judgment logic of key meteorological image elements; the final conclusion is <answer> ...< / answer> The label output includes standardized fields called areas_affected and concerning, which represent the affected area and the level of concern, respectively, facilitating subsequent automated evaluation and training. This output format is versatile and parseable, and can be used in a variety of multimodal task scenarios.

[0053] The multi-round reasoning mechanism is embodied in the detailed reasoning (step 31) and thought chain generation (steps 32 and 33) of step 3. The language model first performs a first round of element parsing and intermediate semantic information extraction (e.g., wind shear leading to enhanced convection) based on the input meteorological imagery's detailed description text and original label information. Subsequently, in the second round, the causal reasoning chain continues, encompassing regional impact analysis, disaster path determination, and disaster severity assessment. This step-by-step simulation of the meteorological expert's "from surface to core" and "from cause to effect" thought process ultimately generates a structured chain of reasoning text.

[0054] Step 4: In the consistency verification and sample screening stage, the generated chain reasoning text is evaluated for consistency and rationality through the rule system and fine-tuned discriminant model, and high-quality samples with standardized structure, reasonable logic, and accurate expression are selected to construct the chain reasoning dataset.

[0055] In the consistency verification and sample screening phase, the chain reasoning texts generated in the detail reasoning and thought chain generation phases are jointly verified using a structural rule system and a discriminant model. Samples that do not conform to the label structure, have poor semantic consistency, or have incomplete reasoning chains are eliminated, and high-quality texts that conform to professional logic, expression standards, and closed reasoning loops are retained as training samples.

[0056] The input is the output in step 3 <think> ...< / think> and <answer> ...< / answer> The evaluation process for structured text is completed by a "rule system" and a "fine-tuned discriminant model." The rule system is used to quickly identify content with formatting errors, missing structure, or obvious illogicality.

[0057] Step 41: The rule system quickly checks the format and content logic in the chain reasoning text, fine-tunes the discrimination model to further judge the semantic consistency, logical closure and rationality of meteorological knowledge in the text, and obtains the judgment result and semantic score of each chain reasoning text.

[0058] Step 41 systematically verifies the format, content logic and language quality of the chain reasoning text.

[0059] First, a set of manually designed rules are used to check the structure of the chain reasoning text, including but not limited to: (1) whether it contains the complete<think>and <answer>label; (2) whether the "affected area" or "severity" field is missing; (3) whether there is abnormal length, empty output, or obvious template language (such as repeated paragraphs, abnormal placeholders), etc. The rule system is implemented as a configurable module that can be adapted to different scenarios and standards.

[0060] The fine-tuned discriminative model is fine-tuned based on a general pre-trained language model and is capable of judging the semantic structure and reasoning logic in chain-based reasoning text. The final output is a set of high-quality, screened chain-based reasoning text samples with accurate content, fluent language, and complete reasoning, meeting the standards for subsequent training or evaluation.

[0061] Step 42: After obtaining the structural judgment result and semantic score of each chain reasoning sample, high-quality samples are screened to form the target chain reasoning dataset.

[0062] The screening process is as follows. First, all texts that fail the structural rule check (such as missing format, misplaced labels, empty output, etc.) are eliminated; second, a semantic scoring threshold is set, and only texts with scores higher than the set standard are retained. The screening process supports the introduction of multi-dimensional reference indicators, such as language fluency (which can be estimated by the log-likelihood of GPT-type models), information density (term coverage, causal chain length), consistency with the original image description (the image-text comparison module can be introduced for alignment evaluation), etc. The samples finally screened out constitute the target chain reasoning dataset, which has the characteristics of strong readability, clear logic, and high professionalism, and can be directly used to train or evaluate the meteorological reasoning ability of multimodal models.

[0063] Step 5: In the parallel processing pre-scheduling stage, a multi-threaded and multi-GPU parallel processing mechanism is introduced to divide the processing steps of steps 1 to 4 into independent task units. These units are scheduled in batches to run simultaneously on multiple threads or GPU devices, thus achieving efficient batch processing of large-scale raw meteorological data sets.

[0064] The parallel processing and scheduling stage supports multi-threaded and multi-GPU parallel processing in the data processing process by building a parallel processing and scheduling mechanism. It can significantly improve processing efficiency and system throughput when building large-scale chain reasoning data sets, and meet the requirements of high-quality corpus scale and diversity for multimodal large model training.

[0065] Multi-threaded and multi-GPU parallel processing enable efficient batch processing of large-scale raw meteorological datasets, significantly reducing overall build time, improving data construction efficiency, and adapting to large-scale multimodal training needs. This parallel strategy supports image sample-level granularity, offering excellent scalability and resource utilization, making it suitable for the continuous processing of weather data across multiple regions and time periods in real-world applications.

[0066] In terms of data expansion, the system supports task-level pipeline reuse and distributed execution, enabling automated processing of data across different time periods, weather types, or observation areas. Users can flexibly specify time ranges, parameter sets, or image types, and the system automatically batch-calls existing processes to build corresponding samples, ensuring the resulting dataset has sufficient coverage and representativeness to meet the data scale and diversity requirements for training large multimodal models.

[0067] Based on the semantic linkage of multimodal graphic and text information and the reasoning ability of pre-trained language models, the present invention can automatically complete the conversion process from meteorological images to structured chain reasoning text. The generated text has expert analysis logic and causal reasoning structure, which is suitable for training large-scale models with multimodal reasoning capabilities, and supports efficient expansion and batch generation on large-scale meteorological data sets.

[0068] The present invention also proposes a chain reasoning text dataset automatic construction system for implementing the above method, which includes: a multimodal model, a detail generation module, a pre-trained language model, a parallel processing module, a rule system and a fine-tuning discrimination module, wherein:

[0069] The multimodal model is used to analyze and semantically model the preprocessed meteorological images in the modal bridging stage, identify and annotate significant meteorological features in the meteorological images, and locate key areas in the meteorological images based on the built-in image attention mechanism module of the multimodal model, and output them as structured semantic text.

[0070] The detail generation module is used to generate detailed description text of meteorological images. Based on the structured semantic text generated by the multimodal model, it introduces a weather term dictionary and professional sentence templates, combines the regional scope, time information and meteorological background of the original meteorological image, and constructs a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image.

[0071] The pre-trained language model completes multiple rounds of reasoning and causal chain construction of meteorological image information based on detailed description text and original structured text information, and outputs chain reasoning text that simulates the analysis process of meteorological experts.

[0072] The rule system and fine-tuning discrimination module are used to jointly verify the chain reasoning text, eliminate samples that do not conform to the label structure, have poor semantic consistency or incomplete reasoning chain, and retain high-quality text that conforms to professional logic, expression specifications, and reasoning closed loop to form the target chain reasoning dataset.

[0073] The parallel processing module is used to divide the data processing of each module into independent units, and perform multi-threaded and multi-GPU parallel processing in batches to improve the system's processing efficiency and throughput.

[0074] It should be noted that the above-described specific embodiments are illustrative only. Those skilled in the art may devise various solutions based on the disclosure of the present invention, and such solutions fall within the scope of the present invention and are intended to be protected by the present invention. Those skilled in the art should understand that the present description and its accompanying drawings are intended to be illustrative only and are not intended to limit the scope of the claims. The scope of protection of the present invention is defined by the claims and their equivalents.< / answer> < / think>

Claims

1. A method for automatically constructing meteorological datasets based on modal bridging, characterized in that: The automatic construction method includes five stages: data preparation and preprocessing, modal bridging, detail reasoning and thought chain generation, consistency verification and sample screening, and parallel processing and scheduling. Specifically: Step 1: Data preparation and preprocessing stage: meteorological images and corresponding text description data are collected. Through preprocessing such as outlier removal, spatiotemporal matching, and format standardization, a unified original meteorological multimodal dataset is formed. Step 2: Modal bridging stage, using multimodal models to analyze meteorological image content and generate detailed description text of meteorological images containing key meteorological elements; Step 21: Multimodal model inference: The original weather image and the corresponding time label are used as input to the multimodal model to build a preliminary cross-modal understanding of the current weather situation. The multimodal model identifies and annotates significant meteorological features in the weather image. Based on the built-in image attention mechanism of the multimodal model, it locates key areas in the weather image and outputs structured semantic text. Step 22: Generate a detailed description text of the meteorological image. Based on the structured semantic text, the regional scope, time information and meteorological background of the original meteorological image are combined to construct a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image. Step 3: Detailed reasoning and thought chain generation phase: The detailed description text of the meteorological image and the original structured labels are input into a pre-trained language model with meteorological domain knowledge. This model then sequentially completes professional reasoning of meteorological image elements and constructs a thought chain of causal logic, generating chained reasoning text that conforms to the logic of meteorological experts. Step 31: Detail inference: receiving the meteorological image detail description text and the original structured label as input, a general pre-trained language model with meteorological domain knowledge parses the meteorological elements involved in the detail description text and outputs intermediate semantic information; Step 32: Thought chain generation. After completing the detailed-level factor reasoning, the pre-trained language model logically organizes and linguistically expands the intermediate semantic information. Based on the chain reasoning structure of observation-interpretation-deduction-judgment, it gradually simulates the analytical process of meteorological experts. Step 33: The pre-trained language model finally outputs a structured chain reasoning text, which is represented in a standard annotation format. Step 4: Consistency verification and sample screening: The generated chain reasoning text is evaluated for consistency and rationality through the rule system and fine-tuned discriminant model. High-quality samples with standardized structure, reasonable logic, and accurate expression are selected to construct the chain reasoning dataset. Step 41: The rule system quickly checks the format and content logic of the chain reasoning text, fine-tunes the discriminant model to further determine the semantic consistency, logical closure, and rationality of meteorological knowledge in the text, and obtains the judgment result and semantic score for each chain reasoning text; Step 42: After obtaining the structural judgment results and semantic scores of each chain reasoning sample, high-quality samples are screened to form the target chain reasoning dataset; Step 5: Parallel processing and scheduling phase. In this phase, multi-threading and multi-GPU parallel processing mechanisms are introduced to divide the processing steps from Step 1 to Step 4 into independent task units. These units are then scheduled in batches to run simultaneously on multiple threads or GPU devices, enabling efficient batch processing of large-scale raw meteorological datasets.

2. Automatic construction system of meteorological dataset based on modal bridging, characterized by: The automatic construction system performs the method according to claim 1, comprising a multimodal model, a detail generation module, a pre-trained language model, a parallel processing module, a rule system and a fine-tuning discrimination module, wherein: The multimodal model is used to analyze and semantically model the pre-processed meteorological images during the modal bridging phase, identify and annotate significant meteorological features in the meteorological images, and locate key areas in the meteorological images based on the built-in image attention mechanism module of the multimodal model to output structured semantic text. The detail generation module is used to generate detailed description text of meteorological images. Based on the structured semantic text generated by the multimodal model, it introduces a weather term dictionary and professional sentence templates, combines the regional scope, time information and meteorological background of the original meteorological image, and constructs a natural language text that conforms to the style of meteorological experts, namely the detailed description text of the meteorological image; The pre-trained language model completes multiple rounds of reasoning and causal chain construction on meteorological image information based on detailed description text and original structured labels, and outputs chain reasoning text that simulates the analysis process of meteorological experts; The rule system and fine-tuning discriminant module are used to jointly verify the chain reasoning text, eliminating samples that do not conform to the label structure, have poor semantic consistency, or have incomplete reasoning chains. High-quality text that conforms to professional logic, expression standards, and closed reasoning loops is retained to form the target chain reasoning dataset. The parallel processing module is used to divide the data processing of each module into independent units and perform multi-threaded and multi-GPU parallel processing in batches.

Images