Fog recognition method for automatically generating a hint based on ground observation data and satellite images

Abstract

This invention discloses a fog identification method that automatically generates prompts based on ground observation data and satellite images. The method includes: data acquisition, acquiring satellite remote sensing images and ground meteorological observation station data; prompt generation, generating initial prompts based on ground observation data through logical judgment, and filtering and optimizing the prompts using satellite image features; feature extraction, extracting high-frequency features from satellite images through frequency domain transformation, and fusing them with deep semantic features to obtain enhanced features; and fog area identification, inputting the optimized guidance prompts and enhanced features into a segmentation model and outputting fog area identification results. This invention solves the problems of relying on manual annotation and the difficulty of multi-source data fusion in fog identification by automatically generating guidance prompts from ground observation data and combining them with frequency domain enhanced features, achieving high-precision automatic monitoring of sea fog without human intervention.

Description

Technical Field

[0001] This invention relates to the field of data image processing technology, and in particular to a fog recognition method (LF-SAM) that automatically generates prompts based on ground observation data and satellite images. Background Technology

[0002] Fog is an important atmospheric phenomenon highly relevant to ecosystems and the environment. Fog occurring near the ground reduces visibility, worsens air quality, and significantly impacts transportation, power facilities, human health, and agricultural production. With advancements in satellite remote sensing technology, fog detection based on remote sensing imagery has been extensively studied in recent years. Researching fog detection methods based on satellite remote sensing technology is of great significance for disaster prevention and mitigation, and for the sustainable development of the national economy and society.

[0003] Traditional fog detection methods typically employ thresholding techniques, determining the presence of fog by setting thresholds based on different pixel intensities for mid-to-high clouds, low clouds, and fog. However, these methods not only rely on empirical threshold selection but also suffer from relatively low computational efficiency. In recent years, with the rapid development of artificial intelligence technology, deep learning-based semantic segmentation methods have gradually been applied to the field of fog detection.

[0004] Convolutional Neural Networks (CNNs) and Visual Transformers (ViTs), as two mainstream methods of deep learning, have performed well in many fields. However, these methods are highly dependent on large-scale, precisely labeled pixel-level data. Due to the complexity of fog formation and its susceptibility to environmental changes, accurate labeling of fog areas is extremely costly and requires a large amount of pixel-level labeled data, which is a significant limitation in practical applications. Limited by the time, conditions, and labeling costs of fog formation, obtaining a large number of pixel-level labels for training is difficult in practice, affecting the widespread application and accuracy of deep learning methods. Therefore, some weakly supervised and unsupervised learning methods have gradually gained attention. These methods generate pseudo-labels through point marking, bounding box marking, or coarse masks, reducing labeling costs. However, these methods still have shortcomings in utilizing auxiliary information from ground observation stations, failing to fully integrate multi-source data and limiting their detection performance under complex meteorological conditions. Furthermore, the distribution of ground observation stations leads to incomplete and uneven fog monitoring coverage in reality. Although using multi-channel observation images from geostationary meteorological satellites and semantic segmentation can identify fog areas and overcome the limitations of ground observation stations, the existing technology still needs improvement.

[0005] Therefore, there is an urgent need for a fog identification method that can automatically generate multimodal guidance prompts, which can eliminate the dependence on large-scale manual pixel-level labeling, and fully combine the frequency domain characteristics of ground observation stations and satellites to achieve high-precision and automated monitoring of sea fog areas under complex meteorological conditions. Summary of the Invention

[0006] This invention addresses the problem of traditional deep learning models' over-reliance on pixel-level manual annotation and difficulty in integrating multi-source meteorological data in fog detection tasks. It proposes a fog recognition method based on ground observation data and satellite imagery that automatically generates prompts. An automatic prompt generation mechanism is employed, automatically converting meteorological elements such as visibility and humidity from ground observation stations into spatial guidance prompts. False prompts are eliminated through satellite channel filtering masks, achieving automatic prompting without manual intervention. Simultaneously, this invention introduces a high-frequency feature extraction module, utilizing Discrete Cosine Transform (DCT) to capture the fine texture features of cloud and fog boundaries, and deeply fusing them with semantic features extracted by Transformer, enhancing the model's ability to identify sea fog edges. Furthermore, for large-scale meteorological scenarios, this invention employs a multimodal feature alignment strategy to correlate point-like ground observation knowledge with area-like satellite image features, achieving accurate segmentation of fog areas on complex underlying surfaces. Through this method, this invention can eliminate reliance on expensive manual annotation while fully leveraging the complementary advantages of ground station and satellite data, achieving efficient and high-precision automatic monitoring of sea fog areas.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a fog identification method that automatically generates prompts based on ground observation data and satellite images, comprising the following steps:

[0009] Data acquisition steps: Acquire satellite remote sensing images and ground meteorological observation station data for the target area;

[0010] Prompt generation steps: Based on ground meteorological observation station data, initial prompt information is generated through logical judgment, and the initial prompt information is filtered and optimized using the features of satellite remote sensing imagery to obtain optimized guidance prompts;

[0011] Feature extraction steps: Perform frequency domain transformation on satellite remote sensing images to extract high-frequency features, and fuse them with the deep semantic features of the images to obtain enhanced features;

[0012] Fog area recognition steps: Input the optimized guidance prompts and enhanced features into the segmentation model, and the segmentation model outputs the fog area recognition results.

[0013] Furthermore, in the prompt generation step, the logical determination is based on one or more combinations of visibility, humidity, dew point temperature and weather phenomenon codes from ground meteorological observation station data to generate positive and / or negative prompt points.

[0014] Furthermore, in the prompt generation step, the filtering optimization includes: extracting channel features corresponding to high cloud coverage areas in satellite remote sensing images to generate a filter mask; based on the filter mask, removing positive prompt points covered by high clouds, and / or retaining prompt points consistent with image features.

[0015] Furthermore, in the feature extraction step, the frequency domain transformation adopts discrete cosine transform; the high-frequency features are obtained by inverse transform after filtering out low-frequency components through a high-pass filter.

[0016] Furthermore, in the feature extraction step, the fusion method of the high-frequency features and deep semantic features is as follows: the high-frequency features and the embedded features are fused, and after nonlinear transformation, additional features are output, and the additional features are injected into each transformation layer of the segmentation model.

[0017] Furthermore, the segmentation model is built on the architecture of segmenting everything, the optimized guidance prompts are used as input to the prompt encoder, the enhanced features are used as input to the image encoder, and the fog area segmentation result is output through the mask decoder.

[0018] Secondly, this invention proposes a fog recognition system that automatically generates prompts based on ground observation data and satellite images, comprising:

[0019] Data acquisition module: used to acquire satellite remote sensing images and ground meteorological observation station data of the target area;

[0020] Prompt generation module: used to generate initial prompt information based on the data from the ground meteorological observation station through logical judgment, and to filter and optimize the initial prompt information using the features of satellite remote sensing images to obtain optimized guidance prompts;

[0021] Feature enhancement module: used to perform frequency domain transformation on the satellite remote sensing image to extract high-frequency features, and fuse them with the deep semantic features of the image to obtain enhanced features;

[0022] Fog Area Recognition Module: This module takes the optimized guidance prompts and enhanced features as input to the segmentation model and outputs the fog area recognition result.

[0023] Furthermore, the prompt generation module includes:

[0024] Logical decision unit: used to generate positive and / or negative alert points based on one or more combinations of visibility, humidity, dew point temperature and weather phenomenon codes from ground meteorological observation station data;

[0025] Mask generation unit: used to extract high cloud coverage areas based on channel features of satellite remote sensing imagery and generate a filter mask;

[0026] Filtering optimization unit: used to filter the initial prompt information based on the filter mask, and remove prompt points that conflict with image features.

[0027] Furthermore, the feature enhancement module includes:

[0028] Frequency domain transformation unit: used to perform discrete cosine transform on satellite remote sensing images to obtain a frequency domain representation;

[0029] High-frequency extraction unit: used to filter out low-frequency components in the frequency domain representation by a high-pass filter, while retaining high-frequency components;

[0030] Inverse Transform Unit: Used to perform inverse discrete cosine transform on the filtered frequency domain representation to obtain high-frequency feature maps;

[0031] Feature fusion unit: used to fuse high-frequency features with deep semantic features of the image to generate enhanced features.

[0032] Furthermore, the fog area recognition module includes a segmentation model, which is based on a segmentation-all model architecture and includes a prompt encoder, an image encoder, and a mask decoder; the optimized guidance prompt is input to the prompt encoder, the enhanced feature is input to the image encoder, and the mask decoder outputs the fog area segmentation result based on the encoded features.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] (1) This invention proposes a fog identification framework based on automatic prompts from multi-source data. By fully combining the high-precision location information of ground observation stations with the wide coverage advantage of geostationary meteorological satellites, an automatic prompt generation and filtering mechanism is designed to effectively solve the problem of deep learning models' dependence on large-scale pixel-level manual annotation and improve the model's automated monitoring capability under complex meteorological backgrounds.

[0035] (2) The present invention realizes a refined modeling method that integrates spatial and frequency domain features. A spatial semantic extraction module based on Transformer and a frequency domain feature extraction module based on Discrete Cosine Transform (DCT) are designed to independently capture the global context information and local high-frequency texture details of the image, avoiding interference caused by similar spectral features of clouds and fog, thereby improving the segmentation accuracy of sea fog edges and fine structures.

[0036] (3) This invention introduces a transfer learning architecture of interactive large model (SAM). By transforming meteorological business logic into understandable guiding signals, the general segmentation large model is accurately transferred to specific meteorological tasks. Its powerful zero-shot segmentation capability is used to deal with fog area identification under different underlying surface and illumination conditions, thereby improving the robustness and generalization performance of the model in business applications. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0038] Figure 1 This is an overall architecture diagram of the fog recognition method that automatically generates prompts based on ground observation data and satellite images, provided in an embodiment of the present invention.

[0039] Figure 2 The channel filtering process provided in the embodiments of the present invention.

[0040] Figure 3 The process of the high-frequency extractor provided in the embodiments of the present invention.

[0041] Figure 4 The high-frequency feature extraction process provided in the embodiments of the present invention. Detailed Implementation

[0042] To better understand this technical solution, the method of the present invention will be described in detail below with reference to the accompanying drawings.

[0043] This invention proposes a fog identification method (LF-SAM) that automatically generates prompts based on ground observation data and satellite imagery. The specific technical solution is as follows:

[0044] (1) Problem definition and data preparation

[0045] This invention addresses the problem that deep learning fog recognition models rely excessively on expensive pixel-level manual annotation and have difficulty effectively integrating multi-source meteorological data. It proposes a fog recognition method (LF-SAM) that automatically generates prompts based on ground observation station data and satellite images. By converting ground station observation data into spatial guidance prompts and fusing high-frequency texture features from satellite images, it achieves high-precision automatic monitoring of complex sea fog areas without human intervention.

[0046] Data Preparation: Within a single day, we selected data within the range of 0:00 to 8:00 (UTC time), with hourly intervals. Since fog within the same day exhibits temporal continuity, its distribution is similar. To avoid similarity between the training and validation sets, we avoided selecting data from close dates and times within the same year.

[0047] The data used in this invention mainly comes from real-time observation data from ground meteorological observation stations and observation data from the FY-4A geostationary meteorological satellite. Details are as follows:

[0048] Ground-based observation data: Records include visibility, humidity, dew point temperature, weather codes, and other meteorological elements closely related to fog formation. Ground-based observation data has the advantages of high accuracy and real-time performance, providing a direct characterization of atmospheric conditions.

[0049] FY-4A satellite observation data: Satellite imagery within the latitude range of 30°N to 40°N and longitude range of 110°E to 120°E was selected. To utilize the high reflectivity of the visible light channel and the difference in channel brightness temperature (the brightness temperature of fog is significantly higher than that of mid-to-high clouds), this invention specifically selects data from the daytime period (UTC 0 to 8) to improve the ease of identification in fog monitoring.

[0050] Spatiotemporal sample selection: Sample dates were selected based on the fog chronology from 2020 to 2022. Time sampling was done at hourly intervals, and a strategy was adopted to avoid similar dates within the same year and adjacent times within the same day to eliminate sample similarity caused by the temporal continuity of fog and ensure the independence of the training and validation sets.

[0051] Multi-dimensional feature fusion basis: The point-like high-precision observations from ground stations and the large-area coverage features of satellite images together constitute the multi-source data basis of this invention. The pseudo-color RGB image synthesized by the satellite's 2nd, 3rd, and 13th channels can intuitively distinguish between low clouds, fog areas (white), and high clouds (blue).

[0052] (2) Overall framework design

[0053] This invention proposes a three-stage recognition framework of "multimodal cueing - frequency domain enhancement - interactive segmentation". This framework decomposes the fog recognition task into three core stages: data processing, network training, and network testing. By combining ground station meteorological elements with satellite imagery features, it achieves high-precision automated segmentation. First, in the data processing module, ground observation station data and FY4A satellite observation data are input into the cue generator. Through logical judgment and channel filtering, a pseudo-color image with cueing is automatically generated, while the original satellite data is converted into a pseudo-color image. Next, in the network training module, the generated cue image is fed into the cue encoder for encoding. In parallel, a Transformer block combined with a high-frequency extractor is used to extract spatial and frequency domain features from the pseudo-color image. These features are converged into a mask decoder to predict the sea fog region, and the network parameters are continuously optimized by calculating a loss function. Finally, in the network testing module, real-time acquired satellite images and ground observation station data are input into the trained model. The model directly outputs the recognition result of the sea fog region, achieving end-to-end automated monitoring. The overall architecture process is as follows: Figure 1 As shown.

[0054] (3) Data processing module

[0055] For ease of annotation, we selected channels 2, 3, and 13 of the fourteen channels of FY-4A and synthesized them into a pseudo-color RGB image. In the pseudo-color image, low clouds and fog areas are displayed in white, high clouds are displayed in a striking blue, and the ground surface without cloud or fog cover is displayed in orange or red.

[0056] The cue generator produces a pseudo-color image with cues: First, it generates a sequence of candidate cues. If interpolation is used directly to infer fog regions, the sparse point distribution leads to inaccurate inferences. However, when used as cues, these observations provide sufficient location information. We define two types of input formats:

[0057]

[0058] Where V represents visibility, H represents humidity, D represents dew point temperature, and W represents weather phenomenon code.

[0059] Next, channel filtering is performed. In real-world images, the results from ground observation stations often conflict with image features. The most common issue is high cloud cover obscuring fog. To ensure consistency in the model's recognition of image features, we assume that even if the ground observation station shows severe fog, but it is covered by high clouds, it can still be classified as fog-free. We extract grayscale images from channels 2 and 13, perform simple binarization and synthesis, and generate a filter mask. Based on the filter mask, incorrect prompts are filtered out. In the ground observation station prompts, green represents positive points, white represents negative points, and blue represents filtered points. The filter mask filters out positive points that do not belong to the white area and negative points that do not belong to the black area. Finally, the final prompt from the ground observation station is obtained. The filtering process is as follows: Figure 2 As shown.

[0060] (4) Network training module

[0061] The input consists of a processed pseudo-color image and a pseudo-color image with added prompts. Feature extraction is performed using a transformer block. Since clouds and fog are difficult to distinguish in pseudo-color images, and texture details are not readily apparent in RGB space, a high-frequency extractor is designed to differentiate fog areas. This extracts features from the frequency domain as a supplement to the RGB space features. Deep and high-frequency features of the image are then input into the transformer block.

[0062]

[0063] in It is a linear layer used to generate features for specific tasks. It is an embedded feature. It is a high-frequency feature. This refers to the output features added to each transformer layer of the model. The high-frequency extractor is mainly responsible for deeply integrating the deep features of the image with the frequency domain features to compensate for the insufficient discernibility of cloud and fog textures in the conventional spatial domain. This process uses embedded features and high-frequency features as initial inputs. First, the two features are fused through an addition operation, and then fed into a linear layer (MLP) for task-specific feature mapping. To enhance the model's non-linear expressive power, the fused features are processed by the GELU activation function, and finally, an MLP layer outputs the final additional features. These output features are added to each transformer layer of the model, providing fine-grained texture detail support and helping the model accurately distinguish fog boundaries in complex meteorological backgrounds. The high-frequency extractor process is as follows: Figure 3 As shown.

[0064] Taking image X as input, the high-frequency extractor first uses Discrete Cosine Transform (DCT) to transform it from the RGB domain to the frequency domain:

[0065]

[0066] The low-frequency features of the image after DCT transformation are concentrated in the upper left corner. Therefore, we use a high-pass filter to eliminate the low-frequency components, and then use inverse DCT transformation to restore the filtered image to the RGB domain.

[0067]

[0068] in Indicates a high-pass filter. It is a manually designed threshold used to control the low-frequency components to be filtered out.

[0069] The high-frequency feature extraction process is achieved through frequency domain transformation and filtering. First, the pseudo-color image X is taken as input and transformed from the RGB domain to the frequency domain using Discrete Cosine Transform (DCT), generating a frequency domain image. In the matrix after DCT transformation, low-frequency energy is usually concentrated in the upper left corner. Therefore, a pre-designed high-pass filter (with a manually set threshold) is used to filter out these low-frequency components, retaining only the high-frequency information representing edge and texture details. Finally, the filtered frequency domain signal is restored to the RGB space using inverse discrete cosine transform, obtaining the high-frequency feature image. This process can extract the fine edge features of clouds and fog from the smooth background, significantly improving the model's accuracy in recognizing the edges and fine structures of sea fog. The high-frequency feature extraction process is as follows: Figure 4 As shown.

[0070] (5) Network testing module

[0071] The satellite images and corresponding ground observation station information are input into the trained model, and the model's output is the prediction result for the sea fog area.

[0072] This invention presents the first automated framework for applying the Segmentation-All-Aspects Model (SAM) to sea fog identification tasks. It proposes an interactive segmentation method based on multimodal information assistance, achieving automatic generation and optimization of prompts by establishing a collaborative mechanism between ground station meteorological element logical judgment and satellite channel masking filtering. Compared to traditional supervised learning methods that rely solely on pixel-level labels, this framework effectively solves the problem of high annotation costs and demonstrates significant performance advantages in sea fog monitoring tasks.

[0073] Meanwhile, this invention designs a high-frequency feature extractor based on Discrete Cosine Transform (DCT). Addressing the challenges of highly similar spectral features and blurred spatial textures in cloud and fog images, this invention innovatively introduces frequency domain processing technology. Through DCT transformation, it accurately extracts high-frequency features from satellite imagery and uses them as an important supplement to deep semantic features. This design significantly improves the model's accuracy in recognizing sea fog edges and their fine textures, greatly enhancing the segmentation performance of the SAM model in specific meteorological domains.

[0074] Furthermore, this invention proposes a fog recognition system that automatically generates prompts based on ground observation data and satellite imagery, comprising:

[0075] Data acquisition module: used to acquire satellite remote sensing images and ground meteorological observation station data of the target area;

[0076] Prompt generation module: used to generate initial prompt information based on the data from the ground meteorological observation station through logical judgment, and to filter and optimize the initial prompt information using the features of satellite remote sensing images to obtain optimized guidance prompts;

[0077] Feature enhancement module: used to perform frequency domain transformation on the satellite remote sensing image to extract high-frequency features, and fuse them with the deep semantic features of the image to obtain enhanced features;

[0078] Fog Area Recognition Module: This module takes the optimized guidance prompts and enhanced features as input to the segmentation model and outputs the fog area recognition result.

[0079] The prompt generation module includes:

[0080] Logical decision unit: used to generate positive and / or negative alert points based on one or more combinations of visibility, humidity, dew point temperature and weather phenomenon codes from ground meteorological observation station data;

[0081] Mask generation unit: used to extract high cloud coverage areas based on channel features of satellite remote sensing imagery and generate a filter mask;

[0082] Filtering optimization unit: used to filter the initial prompt information based on the filter mask, and remove prompt points that conflict with image features.

[0083] The feature enhancement module includes:

[0084] Frequency domain transformation unit: used to perform discrete cosine transform on satellite remote sensing images to obtain a frequency domain representation;

[0085] High-frequency extraction unit: used to filter out low-frequency components in the frequency domain representation by a high-pass filter, while retaining high-frequency components;

[0086] Inverse Transform Unit: Used to perform inverse discrete cosine transform on the filtered frequency domain representation to obtain high-frequency feature maps;

[0087] Feature fusion unit: used to fuse high-frequency features with deep semantic features of the image to generate enhanced features.

[0088] The fog area recognition module includes a segmentation model, which is based on a segmentation-all model architecture and includes a prompt encoder, an image encoder, and a mask decoder. The optimized guidance prompt is input to the prompt encoder, the enhanced feature is input to the image encoder, and the mask decoder outputs the fog area segmentation result based on the encoded features.

[0089] Compared with the prior art, the present invention has the following advantages:

[0090] (1) Highly efficient automated identification and monitoring capabilities: This invention establishes an automatic prompt generation mechanism to directly convert meteorological elements such as visibility and humidity from ground stations into spatial guidance signals, realizing full-process automation from data input to sea fog segmentation result output. Compared with the traditional method that relies on manual input of prompts, this significantly improves the real-time performance and efficiency of sea fog monitoring in meteorological operations.

[0091] (2) Significantly reduced data annotation costs: By automatically generating supervision signals using existing ground observation station data, this invention eliminates the dependence of deep learning models on large-scale pixel-level manual annotation. This not only solves the pain point of scarce high-quality labeled samples in the meteorological field, but also significantly reduces the resource investment in model training and iteration.

[0092] (3) Fine texture feature capture and edge segmentation accuracy: A high-frequency feature extraction module based on discrete cosine transform (DCT) is introduced, which can effectively compensate for the problem of similar cloud and fog features and blurred boundaries in satellite images under conventional airspace. Through deep fusion of frequency domain and spatial domain features, the model has higher recognition accuracy when processing sea fog edges and fine structures, and the average intersection-union ratio (mIoU) is excellent.

[0093] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended 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. However, 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 fog identification method that automatically generates prompts based on ground observation data and satellite imagery, characterized in that, Includes the following steps: Data acquisition steps: Acquire satellite remote sensing images and ground meteorological observation station data for the target area; Prompt generation steps: Based on ground meteorological observation station data, initial prompt information is generated through logical judgment, and the initial prompt information is filtered and optimized using the features of satellite remote sensing imagery to obtain optimized guidance prompts; Feature extraction steps: Perform frequency domain transformation on satellite remote sensing images to extract high-frequency features, and fuse them with the deep semantic features of the images to obtain enhanced features; Fog area recognition steps: Input the optimized guidance prompts and enhanced features into the segmentation model, and the segmentation model outputs the fog area recognition results.

2. The fog identification method based on ground observation data and satellite imagery for automatically generating prompts according to claim 1, characterized in that, In the prompt generation step, the logical determination is based on one or more combinations of visibility, humidity, dew point temperature and weather phenomenon codes from ground meteorological observation station data to generate positive and / or negative prompt points.

3. The fog identification method based on ground observation data and satellite imagery for automatically generating prompts according to claim 1, characterized in that, In the prompt generation step, the filtering optimization includes: extracting channel features corresponding to high cloud coverage areas in satellite remote sensing images and generating a filter mask; based on the filter mask, removing positive prompt points covered by high clouds and / or retaining prompt points consistent with image features.

4. The fog identification method based on ground observation data and satellite imagery for automatically generating prompts according to claim 1, characterized in that, In the feature extraction step, the frequency domain transformation adopts discrete cosine transform; the high-frequency features are obtained by inverse transformation after filtering out low-frequency components through a high-pass filter.

5. The fog identification method based on ground observation data and satellite imagery for automatically generating prompts according to claim 1, characterized in that, In the feature extraction step, the fusion method of the high-frequency features and deep semantic features is as follows: the high-frequency features and the embedded features are fused, and after nonlinear transformation, additional features are output and injected into each transformation layer of the segmentation model.

6. The fog identification method based on ground observation data and satellite imagery for automatically generating prompts according to claim 1, characterized in that, The segmentation model is built on the architecture of the segmentation model. The optimized guidance prompts are used as inputs to the prompt encoder, and the enhanced features are used as inputs to the image encoder. The fog area segmentation results are output through the mask decoder.

7. A fog recognition system that automatically generates prompts based on ground observation data and satellite imagery, characterized in that, include: Data acquisition module: used to acquire satellite remote sensing images and ground meteorological observation station data of the target area; Prompt generation module: used to generate initial prompt information based on the data from the ground meteorological observation station through logical judgment, and to filter and optimize the initial prompt information using the features of satellite remote sensing images to obtain optimized guidance prompts; Feature enhancement module: used to perform frequency domain transformation on the satellite remote sensing image to extract high-frequency features, and fuse them with the deep semantic features of the image to obtain enhanced features; Fog Area Recognition Module: This module takes the optimized guidance prompts and enhanced features as input to the segmentation model and outputs the fog area recognition result.

8. The fog recognition system based on ground observation data and satellite imagery for automatically generating prompts according to claim 7, characterized in that, The prompt generation module includes: Logical decision unit: used to generate positive and / or negative alert points based on one or more combinations of visibility, humidity, dew point temperature and weather phenomenon codes from ground meteorological observation station data; Mask generation unit: used to extract high cloud coverage areas based on channel features of satellite remote sensing imagery and generate a filter mask; Filtering optimization unit: used to filter the initial prompt information based on the filter mask, and remove prompt points that conflict with image features.

9. The fog recognition system based on ground observation data and satellite imagery for automatic generation of prompts according to claim 7, characterized in that, The feature enhancement module includes: Frequency domain transformation unit: used to perform discrete cosine transform on satellite remote sensing images to obtain a frequency domain representation; High-frequency extraction unit: used to filter out low-frequency components in the frequency domain representation by a high-pass filter, while retaining high-frequency components; Inverse Transform Unit: Used to perform inverse discrete cosine transform on the filtered frequency domain representation to obtain high-frequency feature maps; Feature fusion unit: used to fuse high-frequency features with deep semantic features of the image to generate enhanced features.

10. The fog recognition system based on ground observation data and satellite imagery for automatically generating prompts according to claim 7, characterized in that, The fog area recognition module includes a segmentation model, which is based on a segmentation-all model architecture and includes a prompt encoder, an image encoder, and a mask decoder. The optimized guidance prompt is input to the prompt encoder, the enhanced feature is input to the image encoder, and the mask decoder outputs the fog area segmentation result based on the encoded features.

Images