A fog area identification method and system based on multispectral and spatiotemporal feature fusion

Abstract

The application discloses a kind of fog area identification method and system based on multispectral and space-time feature fusion, it is related to satellite remote sensing technical field, for the single spectrum of existing satellite fog identification method is identified, poor stability, easy to confuse target distinguish ability is insufficient, FY4B GHI fast scanning data advantage is not fully utilized, the present application is based on FY4B GHI minute level multispectral fast scanning data, fusion visible light spatial feature, near-infrared interference suppression feature, long-wave infrared thermal feature and multi-temporal space-time evolution feature, by pixel-by-pixel joint determination and morphological processing, time persistence verification is output after fog identification result.The present application has significantly improved the distinguish ability of fog area, thin cloud, snow ground and bright ground, the recognition accuracy is improved, the misidentification rate is reduced, suitable for high-frequency continuous monitoring of land fog, sea fog, can directly serve traffic meteorology, aviation operation guarantee, port shipping scheduling and other businesses.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing technology, and in particular to a fog area identification method and system based on multispectral fast scan data from the FY4B (Fengyun-4B) Geostationary High-speed Imager (GHI), which integrates visible light, near-infrared, long-wave infrared, and spatiotemporal feature fusion. Background Technology

[0002] Fog is a low-visibility weather phenomenon formed by a large number of tiny water droplets or ice crystals suspended near the ground, significantly impacting road traffic, aviation operations, port shipping, and refined urban meteorological services. Traditional ground-based observation stations can provide high-precision records of visibility and weather phenomena, but their spatial distribution is limited, making it difficult to meet the needs for large-scale, continuous, and refined fog monitoring. Geostationary meteorological satellites, with their advantages of wide observation range and strong temporal continuity, are an important means of conducting dynamic monitoring of fog areas.

[0003] Based on domestic and international research and operational applications, existing satellite fog identification methods can be broadly categorized into the following three types:

[0004] (1) Identification methods based on visible light reflectance. This type of method utilizes the characteristics of fog in the visible light band, such as high reflectance, grayish-white color, and smooth texture. It combines brightness threshold, color consistency, and texture features for identification. The advantage is that the fog area boundary is relatively clear and the spatial details are good. However, it is significantly affected by the solar altitude angle, observation geometry, and the reflectance characteristics of the underlying surface. It is not usable at night and is prone to misjudgment in scenarios such as snow, bright ground, low clouds, and high reflectance clouds.

[0005] (2) Identification methods based on infrared brightness temperature or brightness temperature difference. These methods typically use the brightness temperature or brightness temperature difference features of infrared channels such as 3.7μm, 10.8μm, and 12.0μm to identify fog and low-level clouds. They are suitable for use at night, at dawn and dusk, and when visible light conditions are insufficient. However, their identification effect is greatly affected by channel configuration, surface temperature, cloud top height, seasonal changes, and sea and land background. When relying solely on infrared thresholds, they are prone to insufficient regional adaptability.

[0006] (3) Recognition methods based on objects, time series or machine learning. These methods usually regard fog as a low-level fog target that is near the ground, water phase, layered, persistent and slowly evolving. They comprehensively utilize spatial morphology, texture structure, temporal continuity, motion changes and sample learning results for recognition, which can improve the recognition ability in complex scenes. However, they often require stable multi-temporal data, relatively complete ground real-time samples or multi-channel observation conditions. When directly transferred to different satellite sensors and different regional business scenarios, targeted adaptation is still required.

[0007] The FY4B GHI rapid imager possesses high temporal and spatial resolution, enabling rapid regional scanning and minute-level continuous observation of the study area. Its channel configuration simultaneously covers the visible, near-infrared, and long-wave infrared bands. The visible light channel can be used to acquire brightness, color, and texture details in fog areas; the near-infrared channel provides auxiliary information for suppressing interference from cirrus clouds, snowfields, ice clouds, and low clouds; and the 10.3–12.5 μm long-wave infrared channel provides information on cloud formations and surface thermal radiation. Therefore, when using the FY4B GHI for fog identification, it is not advisable to rely solely on a single visible light, single infrared, or single temporal feature. Instead, its high-frequency, multi-channel observation advantages should be fully utilized to construct an identification method that synergistically integrates visible light spatial details, near-infrared interference removal, long-wave infrared thermal constraints, and minute-level spatiotemporal evolution features.

[0008] Based on the above, it is necessary to propose a fog recognition method that integrates visible light, infrared, and spatiotemporal features for FY4B GHI fast scan data. While inheriting the ability to identify details in visible light, this method introduces long-wave infrared thermal feature constraints and combines near-infrared interference suppression, minute-level time stability, and motion feature analysis to improve the ability to distinguish fog areas from easily confused targets such as high clouds, thin clouds, low clouds, snowfields, and bright surfaces, thereby achieving high-frequency, continuous, and stable recognition of fog areas. Summary of the Invention

[0009] In view of this, the present invention aims to address the problems of insufficient stability of single-spectral band identification, inadequate fusion of multi-source features, and insufficient utilization of the advantages of FY4B GHI fast scan data in satellite fog identification services. It provides a fog area identification method and system based on the fusion of visible light, infrared, and spatiotemporal features of FY4B GHI. Based on FY4B GHI fast scan data, and focusing on the characteristics of fog in satellite imagery, such as high visible light reflectivity, relatively uniform color, smooth texture, stable infrared thermal features, and slow temporal changes, a fog area identification method is constructed that fuses visible light, near-infrared, long-wave infrared, and spatiotemporal features to improve the accuracy, continuity, and operational applicability of fog identification.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] In a first aspect, the present invention provides a fog area identification method based on the fusion of multispectral and spatiotemporal features, implemented using GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, comprising the following steps:

[0012] S1. Data reading and preprocessing: Read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid.

[0013] S2. Visible light spatial feature extraction: Based on the visible light channel data of FY4B GHI, the highlight features, color consistency features and texture smoothness features of the fog area are extracted respectively. The three types of features are jointly judged pixel by pixel to generate visible light candidate fog area masks.

[0014] S3. Near-infrared cloud phase and interference suppression feature extraction: Using the cirrus cloud channel and near-infrared channel data of FY4B GHI, cirrus cloud interference targets are removed, and low clouds, snow, ice clouds and highly reflective surfaces are distinguished to obtain candidate fog areas after interference suppression.

[0015] S4. Long-wave infrared thermal feature extraction: Based on the long-wave infrared channel data of FY4B GHI, extract the brightness temperature threshold feature and infrared local smoothing feature, and remove low-temperature high cloud and thick cloud interference targets.

[0016] S5. Multi-temporal stability and motion characteristic analysis: Using FY4B GHI minute-level continuous fast scan images, the visible light temporal stability characteristics and optical flow motion characteristics of the target area are calculated to screen out the target areas that are persistent and move slowly.

[0017] S6. Comprehensive Fog Area Judgment: Perform pixel-by-pixel joint logical judgment on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features, and multi-temporal spatiotemporal features to generate initial candidate fog areas;

[0018] S7. Post-processing: The initial candidate fog regions are sequentially processed by spatial morphology, connected component filtering, and temporal persistence verification, and the final fog recognition result is output.

[0019] As a further aspect of the present invention, in step S1, the number of consecutive time intervals read is 5-10, and the time interval between adjacent time intervals is 1 minute; the channels read include B01 panchromatic channel, B02 0.445-0.495μm blue light channel, B03 0.52-0.57μm green light channel, B04 0.62-0.67μm red light channel, B05 1.371-1.386μm cirrus channel, B06 1.58-1.64μm near-infrared channel, and B07 10.3-12.5μm long-wave infrared channel.

[0020] As a further aspect of the present invention, the specific steps of data preprocessing in step S1 are as follows:

[0021] S11. Read the original observation data and corresponding calibration table of each channel, convert the original digital values ​​of the B01-B06 reflectivity channels into reflectivity data, and convert the original digital values ​​of the B07 long-wave infrared channel into brightness temperature data.

[0022] S12. Based on the data quality identifier and physical value range, fill values, invalid values, saturated values, and abnormal pixels that exceed the reasonable range of atmospheric and surface temperatures are set to null values;

[0023] S13. Using the identification grid of the visible light channel as a unified processing grid, the long-wave infrared brightness temperature data is resampled to the unified processing grid through nearest neighbor interpolation or bilinear interpolation.

[0024] S14. Arrange the multi-channel data from multiple consecutive time periods in chronological order of observation time to form a time series dataset with dimensions of [number of time periods, number of channels, image height, and image width].

[0025] As a further aspect of the present invention, the specific steps for extracting visible light spatial features in step S2 are as follows:

[0026] S21. Highlight Feature Extraction: Construct an RGB true-color image by using channels B02, B03, and B04 as R, G, and B components, respectively. After normalization to the range of 0-255, convert it into a grayscale image. Use the contrast-limited adaptive histogram equalization (CLAHE) method to enhance the local contrast of the grayscale image. Calculate the 75th percentile of the brightness of all effective pixels in the enhanced image as the adaptive brightness threshold. Pixels with brightness values ​​greater than this threshold are marked as high-brightness pixels.

[0027] S22. Color uniformity feature extraction: Calculate the standard deviation of the three components R, G, and B for each pixel. Pixels with a standard deviation less than 10 are marked as color uniform pixels.

[0028] S23. Texture smoothing feature extraction: The grayscale brightness image is convolved using a 3×3 Laplacian operator. The absolute value of the convolution response is taken as the texture intensity value. Pixels with a texture intensity value less than 1 are marked as texture smoothing pixels.

[0029] S24. Perform a pixel-by-pixel logical AND operation on the masks of bright pixels, uniform color pixels, and smooth texture pixels to obtain visible light candidate haze masks.

[0030] As a further aspect of the present invention, in step S21, the parameters of the contrast-limited adaptive histogram equalization method are set as follows: the grayscale image is divided into an 8×8 local grid, and the contrast limitation parameter is 2.0.

[0031] As a further aspect of the present invention, the specific steps for extracting near-infrared interference suppression features in step S3 are as follows:

[0032] S31. Cirrus cloud removal: Read the reflectance data of channel B05. Considering the influence of observation angle and regional atmospheric background, the cirrus cloud removal threshold T_cirrus is set to 0.02~0.04. When the reflectance of channel B05 is greater than or equal to T_cirrus, the pixel is marked as a cirrus cloud interference pixel and removed from the candidate fog area.

[0033] S32. Screening of low cloud and strong reflective targets: Read the reflectivity data of channel B06. Considering the regional and seasonal differences in the response of low clouds, fog, snow, ice clouds and bright surfaces in the near-infrared channel, the near-infrared candidate threshold T_nir is set to 0.2 to 0.4. When the reflectivity of channel B06 is greater than T_nir, the pixel is regarded as a candidate pixel for low cloud and fog.

[0034] S33. Visible light reflectance auxiliary constraint: Read the reflectance data of the B02 channel. Considering that the visible light reflectance is affected by the solar altitude angle, surface type, sea and land background and fog thickness, the visible light high reflectance threshold T_vis is set to 0.20~0.45. When the reflectance of the B02 channel is greater than T_vis, the pixel is regarded as a candidate pixel for high reflectance.

[0035] As a further aspect of the present invention, the specific steps for extracting long-wave infrared thermal features in step S4 are as follows:

[0036] S41. Long-wave infrared brightness temperature preprocessing: The brightness temperature data of channel B07 is resampled to a unified visible light grid, and a 3×3 window median filter is used to smooth local isolated abnormal brightness temperatures, and invalid and abnormal brightness temperature values ​​are removed.

[0037] S42. Low-temperature target removal: Read the brightness temperature data of channel B07. Considering that the brightness temperature of fog areas is closely related to season, land and sea background, terrain height, near-surface air temperature, diurnal variation, and fog thickness, the brightness temperature threshold T_bt is set to 270-275K. When the brightness temperature of channel B07 is less than or equal to T_bt, the pixel is marked as a low-temperature high-cloud or thick-cloud interference pixel and removed from the candidate fog area. When the brightness temperature of channel B07 is greater than T_bt, it is identified as meeting the thermal characteristics of near-surface fog or low-cloud fog targets and continues to participate in subsequent comprehensive judgment.

[0038] S43. Infrared Local Smoothing Feature Extraction: Using each pixel as the center, calculate the local standard deviation of brightness temperature using a 5×5 window, and retain only pixels with a local standard deviation less than a set threshold as candidate pixels for thermal smoothing.

[0039] As a further aspect of the present invention, the specific steps of the multi-temporal stability and motion characteristic analysis in step S5 are as follows:

[0040] S51. Visible light temporal stability analysis: Establish a temporal buffer queue with a length of 5 frames to store grayscale brightness images for 5 consecutive time intervals; calculate the temporal standard deviation of grayscale brightness values ​​for each pixel in 5 consecutive frames, and mark pixels with a temporal standard deviation greater than 0 and less than 8 as temporally stable pixels.

[0041] S52. Calculation of Optical Flow Motion Features: Using the current image and images from the previous 4 minutes as input, the Farneback dense optical flow algorithm is employed to calculate the pixel-by-pixel horizontal displacement component u and the vertical displacement component v, according to the formula... Calculate the synthetic motion velocity V, and mark pixels with a motion velocity less than 2.0 pixels as slowly moving pixels.

[0042] As a further aspect of the present invention, in step S6, the pixels determined to be initial candidate fog regions must simultaneously meet all of the following conditions:

[0043] (1) The luminance value enhanced by CLAHE is greater than the 75th percentile of the effective pixel luminance;

[0044] (2) The standard deviation of the RGB three-channel components is less than 10;

[0045] (3) The Laplacian texture intensity value is less than 1;

[0046] (4) The time standard deviation of visible light grayscale brightness is greater than 0 and less than 8;

[0047] (5) The motion velocity calculated based on the Farneback optical flow method is less than 2.0 pixels;

[0048] (6) The reflectivity of channel B05 is less than the cirrus removal threshold T_cirrus, where T_cirrus is set to 0.02 to 0.04;

[0049] (7) The reflectivity of the B06 channel is greater than the near-infrared candidate threshold T_nir, where T_nir is set to 0.2 to 0.4;

[0050] (8) The reflectivity of the B02 channel is greater than the visible light high reflectivity threshold T_vis, where T_vis is set to 0.20 to 0.45;

[0051] (9) The brightness temperature of channel B07 is greater than the brightness temperature threshold T_bt, where T_bt is set to 270~275K;

[0052] (10) The reflectivity or brightness temperature change within a local 5×5 window meets the preset smoothness condition.

[0053] As a further aspect of the present invention, the specific steps of post-processing in step S7 are as follows:

[0054] S71. Spatial morphological processing: A 5×5 square structuring element is used to perform morphological closing operations on the initial candidate fog area to fill the small holes inside the fog area and connect the local broken areas.

[0055] S72. Connected Component Filtering: Mark connected components in the morphologically processed binary image, count the number of pixels in each connected component, and remove small patches with an area of ​​less than or equal to 50 pixels.

[0056] S73. Time persistence verification: Count the number of times the same pixel is identified as fog in the most recent 5 consecutive time intervals, and retain only the pixels that are identified as fog at least 3 times as the final fog area;

[0057] S74. Output a fog recognition binary raster product with a uniform spatial resolution, where 1 represents a fog area and 0 represents a non-fog area.

[0058] As a further aspect of the present invention, the method is applicable to remote sensing monitoring of land fog and sea fog, and the output fog identification results can be used for dynamic tracking of heavy fog processes, highway traffic meteorological services, aviation operation support, port shipping scheduling, and identification of the impact of low cloud and fog in ecological environment remote sensing services.

[0059] Secondly, this invention also provides a fog area recognition system based on the fusion of multispectral and spatiotemporal features, implemented using GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, comprising:

[0060] The data reading and preprocessing module is used to read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid.

[0061] The visible light spatial feature extraction module is used to extract the highlight features, color consistency features and texture smoothness features of the fog area based on the visible light channel data of FY4B GHI. The module performs pixel-by-pixel joint judgment on the three types of features to generate visible light candidate fog area masks.

[0062] The near-infrared cloud phase and interference suppression module is used to eliminate cirrus cloud interference targets using the cirrus cloud channel and near-infrared channel data of FY4B GHI, and to help distinguish low clouds, snow, ice clouds and highly reflective ground surfaces to obtain candidate fog areas after interference suppression.

[0063] The long-wave infrared thermal feature extraction module is used to extract brightness temperature threshold features and infrared local smoothing features from long-wave infrared channel data based on FY4B GHI, and to remove low-temperature high-cloud and thick-cloud interference targets.

[0064] The multi-temporal stability and motion feature analysis module is used to calculate the visible light temporal stability features and optical flow motion features of the target area using FY4B GHI minute-level continuous fast scan images, and to screen out the target areas that are persistent and move slowly.

[0065] The fog area comprehensive judgment module is used to perform pixel-by-pixel joint logical judgment on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features and multi-temporal spatiotemporal features to generate initial candidate fog areas.

[0066] The post-processing module is used to perform spatial morphological processing, connected component filtering, and temporal persistence verification on the initial candidate fog areas in sequence, and output the final fog recognition result.

[0067] As a further embodiment of the present invention, the number of consecutive time intervals read by the data reading and preprocessing module is 5-10, and the time interval between adjacent time intervals is 1 minute; the channels read include B01 panchromatic channel, B02 0.445-0.495μm blue light channel, B03 0.52-0.57μm green light channel, B04 0.62-0.67μm red light channel, B05 1.371-1.386μm cirrus channel, B06 1.58-1.64μm near-infrared channel and B07 10.3-12.5μm long-wave infrared channel.

[0068] As a further aspect of the present invention, the data reading and preprocessing module is specifically used to perform the following operations:

[0069] Read the original observation data and corresponding calibration table of each channel, convert the original digital values ​​of the reflectivity channels B01-B06 into reflectivity data, and convert the original digital values ​​of the long-wave infrared channel B07 into brightness temperature data.

[0070] Based on the data quality identifier and physical value range, fill values, invalid values, saturated values, and abnormal pixels that exceed the reasonable range of atmospheric and surface temperatures are set to null values;

[0071] Using the visible light channel identification grid as a unified processing grid, long-wave infrared brightness temperature data is resampled to the unified processing grid through nearest neighbor interpolation or bilinear interpolation.

[0072] The multi-channel data from multiple consecutive time periods are arranged in chronological order of observation time to form a time series dataset with dimensions of [number of time periods, number of channels, image height, and image width].

[0073] As a further aspect of the present invention, the visible light spatial feature extraction module is specifically used to perform the following operations:

[0074] RGB true-color images were constructed by using channels B02, B03, and B04 as R, G, and B components, respectively. After normalization to the range of 0-255, the images were converted into grayscale images. The contrast-limited adaptive histogram equalization (CLAHE) method was used to enhance the local contrast of the grayscale images. The 75th percentile of the brightness of all effective pixels in the enhanced image was calculated as the adaptive brightness threshold. Pixels with brightness values ​​greater than this threshold were marked as bright pixels.

[0075] Calculate the standard deviation of the three components R, G, and B for each pixel. Pixels with a standard deviation less than 10 are marked as color-uniform pixels.

[0076] A 3×3 Laplacian operator is used to perform convolution operation on the grayscale brightness image. The absolute value of the convolution response value is taken as the texture intensity value. Pixels with a texture intensity value less than 1 are marked as texture smooth pixels.

[0077] A pixel-by-pixel logical AND operation is performed on the masks of bright pixels, uniform color pixels, and smooth texture pixels to obtain visible light candidate haze masks.

[0078] As a further embodiment of the present invention, in the visible light spatial feature extraction module, the parameters of the contrast-limited adaptive histogram equalization method are set as follows: the grayscale image is divided into an 8×8 local grid, and the contrast limit parameter is 2.0.

[0079] As a further aspect of the present invention, the near-infrared cloud phase and interference suppression module is specifically used to perform the following operations:

[0080] Read the reflectance data of channel B05. Considering the influence of observation angle and regional atmospheric background, the cirrus cloud removal threshold T_cirrus is set to 0.02~0.04. When the reflectance of channel B05 is greater than or equal to T_cirrus, the pixel is marked as a cirrus cloud interference pixel and removed from the candidate fog area.

[0081] The reflectance data of channel B06 is read. Considering the regional and seasonal differences in the response of low clouds, fog, snow, ice clouds and bright surfaces in the near-infrared channel, the near-infrared candidate threshold T_nir is set to 0.2 to 0.4. When the reflectance of channel B06 is greater than T_nir, the pixel is regarded as a candidate pixel for low cloud and fog.

[0082] The reflectance data of the B02 channel is read. Considering that the visible light reflectance is affected by the solar altitude angle, surface type, sea and land background and fog thickness, the visible light high reflectance threshold T_vis is set to 0.20~0.45. When the reflectance of the B02 channel is greater than T_vis, the pixel is regarded as a candidate pixel for high reflectance.

[0083] As a further aspect of the present invention, the long-wave infrared thermal feature extraction module is specifically used to perform the following operations:

[0084] The brightness temperature data of channel B07 was resampled to a unified visible light grid, and a 3×3 window median filter was used to smooth local isolated abnormal brightness temperatures, eliminating invalid and outlier brightness temperature values.

[0085] Brightness temperature data from channel B07 is read. Considering that brightness temperature in fog areas is closely related to season, land-sea background, terrain elevation, near-surface air temperature, diurnal variation, and fog thickness, the brightness temperature threshold T_bt is set to 270–275 K. When the brightness temperature of channel B07 is less than or equal to T_bt, the pixel is marked as a low-temperature high-cloud or thick-cloud interference pixel and removed from the candidate fog area. When the brightness temperature of channel B07 is greater than T_bt, it is identified as meeting the thermal characteristics of near-surface fog or low-cloud fog targets and continues to participate in subsequent comprehensive judgment.

[0086] Centered on each pixel, a 5×5 window is used to calculate the local standard deviation of brightness temperature, and only pixels with a local standard deviation less than a set threshold are retained as candidate pixels for thermal smoothing.

[0087] As a further aspect of the present invention, the multi-temporal stability and motion characteristic analysis module is specifically used to perform the following operations:

[0088] Establish a time buffer queue with a length of 5 frames to store grayscale brightness images for 5 consecutive time intervals; calculate the time standard deviation of grayscale brightness values ​​of each pixel in 5 consecutive frames, and mark pixels with a time standard deviation greater than 0 and less than 8 as time-stable pixels.

[0089] Using the current image and images from the previous 4 minutes as input, the Farneback dense optical flow algorithm is employed to calculate the pixel-by-pixel horizontal displacement component u and the vertical displacement component v, according to the formula... Calculate the synthetic motion velocity V, and mark pixels with a motion velocity less than 2.0 pixels as slowly moving pixels.

[0090] As a further aspect of the present invention, in the fog area comprehensive determination module, the pixels determined as initial candidate fog areas must simultaneously meet all of the following conditions:

[0091] (1) The luminance value enhanced by CLAHE is greater than the 75th percentile of the effective pixel luminance;

[0092] (2) The standard deviation of the RGB three-channel components is less than 10;

[0093] (3) The Laplacian texture intensity value is less than 1;

[0094] (4) The time standard deviation of visible light grayscale brightness is greater than 0 and less than 8;

[0095] (5) The motion velocity calculated based on the Farneback optical flow method is less than 2.0 pixels;

[0096] (6) The reflectivity of channel B05 is less than the cirrus removal threshold T_cirrus, where T_cirrus is set to 0.02 to 0.04;

[0097] (7) The reflectivity of the B06 channel is greater than the near-infrared candidate threshold T_nir, where T_nir is set to 0.2 to 0.4;

[0098] (8) The reflectivity of the B02 channel is greater than the visible light high reflectivity threshold T_vis, where T_vis is set to 0.20 to 0.45;

[0099] (9) The brightness temperature of channel B07 is greater than the brightness temperature threshold T_bt, where T_bt is set to 270~275K;

[0100] (10) The reflectivity or brightness temperature change within a local 5×5 window meets the preset smoothness condition.

[0101] As a further aspect of the present invention, the post-processing module is specifically used to perform the following operations:

[0102] A morphological closing operation was performed on the initial candidate fog region using a 5×5 square structuring element to fill the small holes inside the fog region and connect the local broken areas.

[0103] Connected component labeling is performed on the morphologically processed binary image, the number of pixels in each connected component is counted, and small patches with an area of ​​less than or equal to 50 pixels are removed.

[0104] The number of times the same pixel is identified as fog in the most recent 5 consecutive time intervals is counted, and only pixels that are identified as fog at least 3 times are retained as the final fog area;

[0105] Output a fog recognition binary raster product with a uniform spatial resolution, where 1 represents a fog area and 0 represents a non-fog area.

[0106] As a further embodiment of the present invention, the system also includes a result output and application module, which is used to push the final fog identification result to the fog process dynamic tracking system, the highway traffic meteorological service platform, the aviation operation support system, the port shipping scheduling system and the ecological environment remote sensing business platform to realize real-time identification and early warning of the impact of low cloud fog.

[0107] Compared with existing technologies, the fog area identification method and system based on the fusion of FY4B GHI visible light, infrared and spatiotemporal features provided by the present invention have the following beneficial effects:

[0108] 1. This invention uses the long-wave infrared brightness temperature of the 7th channel of FY4B GHI as the core thermal feature constraint, and constructs a multi-dimensional fusion criterion system of visible light spatial details, near-infrared interference elimination, long-wave infrared thermal constraint and minute-level spatiotemporal evolution. This significantly improves the ability to distinguish fog areas from easily confused targets such as high clouds, thin clouds, low clouds, snow and bright ground surfaces, and reduces the false recognition rate.

[0109] 2. Targeting the minute-level continuous observation characteristics of FY4B GHI, this invention designs a stability analysis based on 5-frame time series and an optical flow motion feature extraction method with a 4-minute interval. This effectively utilizes the advantage of the temporal continuity of the data, solves the problem of insufficient stability in single-frame identification, and reduces the missed identification rate. An adaptive brightness threshold based on image statistics is used instead of a fixed threshold, which reduces the impact of solar altitude angle, observation geometry, and regional background differences on the identification results. The adaptability of the algorithm in different seasons and regions is significantly improved.

[0110] In summary, the present invention provides a fog area identification method and system based on the fusion of visible light, infrared and spatiotemporal features of FY4B GHI. By fusing multi-dimensional features, the method improves the distinguishability, has high computational efficiency, and can be directly integrated into the FY4B satellite service processing system. This enables high-frequency, continuous and automatic monitoring of fog areas, thereby improving the accuracy, continuity and service applicability of fog identification.

[0111] These or other aspects of the invention will become more apparent from the following description of embodiments. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0112] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. In the drawings:

[0113] Figure 1 This is a flowchart of a fog area identification method based on the fusion of multispectral and spatiotemporal features according to the present invention.

[0114] Figure 2 This is a sub-flowchart of visible light spatial feature extraction in a fog area identification method based on multispectral and spatiotemporal feature fusion according to the present invention.

[0115] Figure 3 This is a sub-flowchart of long-wave infrared thermal feature extraction in a fog area identification method based on multispectral and spatiotemporal feature fusion according to the present invention.

[0116] Figure 4 This is a flowchart of the post-processing in a fog area identification method based on multispectral and spatiotemporal feature fusion according to the present invention. Detailed Implementation

[0117] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0118] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0119] This invention aims to address the problems in satellite fog identification services, such as insufficient stability of single-spectral band identification, inadequate fusion of multi-source features, and insufficient utilization of the advantages of FY4B GHI fast scan data, and solves the following technical problems:

[0120] Recognition methods based on visible light rely primarily on features such as brightness, uniform color, and smooth texture, and are insufficient in distinguishing targets such as high clouds, bright ground surfaces, snow, low clouds, and fog.

[0121] Single threshold or single spectral band methods are sensitive to changes in solar altitude angle, twilight, shadowed areas, surface temperature, and observational geometry, and suffer from insufficient threshold stability.

[0122] Many methods rely on single-frame identification or simple multi-frame rule fusion, which do not fully utilize the temporal continuity, stability, and slow-moving characteristics of FY4B GHI minute-level fast scan data.

[0123] Existing business methods still have room for improvement in areas such as continuous verification in foggy areas, infrared thermal uniformity constraints, low-temperature high-cloud removal, and snow / bright surface suppression, and are prone to false positives and false negatives.

[0124] To this end, this invention proposes a fog recognition method that integrates FY4B GHI visible light, near-infrared, long-wave infrared and multi-temporal spatiotemporal features to improve the accuracy, continuity and business applicability of fog recognition.

[0125] Based on FY4B GHI fast scan data, this invention constructs a fog recognition method that integrates visible light, near-infrared, long-wave infrared and spatiotemporal features, taking into account the characteristics of fog in satellite imagery, such as high visible light reflectivity, relatively uniform color, smooth texture, stable infrared thermal features and slow temporal changes.

[0126] The core idea of ​​this invention is to fully utilize the advantages of FY4B GHI's minute-level fast scanning and multispectral coverage to construct a multi-dimensional feature fusion criterion system that integrates visible light spatial details, near-infrared interference removal, long-wave infrared thermal constraint, and minute-level spatiotemporal evolution. The overall process is as follows: First, FY4B GHI data from multiple consecutive time intervals are read, calibrated, invalid values ​​are removed, and spatiotemporal matching is performed to form a unified multi-channel, multi-temporal data foundation. Second, the visible light channel is used to extract brightness, color, and texture features of fog areas; the near-infrared channel is used to assist in removing interfering targets such as cirrus clouds, snow, ice clouds, and highly reflective surfaces; and the long-wave infrared channel is used to extract brightness and thermal uniformity features to enhance the ability to distinguish between high clouds, thick clouds, and ordinary low clouds. Third, temporal stability and motion feature analysis are conducted using continuous fast-scan images to identify target areas with persistent, slow-changing, and low-moving characteristics. Finally, a comprehensive judgment is made on multiple features, and the final fog identification result is generated through morphological processing, connected component filtering, and persistence verification.

[0127] Compared with existing algorithms based solely on visible light and spatiotemporal features, this invention uses the long-wave infrared brightness temperature of the 7th channel of the GHI as a thermal feature constraint for fog areas, and introduces a multi-dimensional fusion criterion of visible light, near-infrared, long-wave infrared, temporal stability, and optical flow motion to improve the ability to distinguish fog areas from high clouds, bright surfaces, snowfields, and ordinary low clouds.

[0128] See Figures 1 to 4 As shown, this invention provides a fog area identification method based on the fusion of multispectral and spatiotemporal features, implemented using GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, including the following steps:

[0129] S1. Data Reading and Preprocessing: Read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid.

[0130] In this step, the input data consists of multiple consecutive FY4B GHI area fast scan L1 HDF data. The number of consecutive time intervals read is 5-10, and the time interval between adjacent time intervals is 1 minute or the nearest neighbor fast scan interval available for the service. The channels read include B01 full-color channel, B02 0.445~0.495μm blue light channel, B03 0.52~0.57μm green light channel, B04 0.62~0.67μm red light channel, B05 1.371~1.386μm cirrus channel, B06 1.58~1.64μm near-infrared channel, and B07 10.3~12.5μm long-wave infrared channel.

[0131] In this embodiment, the specific steps of data preprocessing in step S1 are as follows:

[0132] S11. Read the original observation data and corresponding calibration table of each channel from the FY4B GHIHDF file, convert the original digital values ​​of the B01-B06 reflectivity channels into reflectivity data, and convert the original digital values ​​of the B07 long-wave infrared channel into brightness temperature data.

[0133] For the reflectivity channel, the raw digital values ​​are converted into reflectivity data according to the calibration table; for the long-wave infrared channel, the raw digital values ​​are converted into brightness temperature data according to the calibration table.

[0134] S12. Based on the data quality identifier and physical value range, fill values, invalid values, saturated values, and abnormal pixels that exceed the reasonable range of atmospheric and surface temperatures are set to null values.

[0135] This step is used to remove invalid and outlier values. Based on the data quality identifier or the original observation range, filled values, invalid values, saturated values, and obviously abnormal pixels are set to null values ​​and are not included in subsequent calculations. For reflectance data, outlier pixels less than 0 or significantly exceeding the physical range are removed; for brightness temperature data, outlier pixels that do not conform to the actual atmospheric and surface temperature ranges are removed.

[0136] S13. Using the identification grid of the visible light channel as a unified processing grid, the long-wave infrared brightness temperature data is resampled to the unified processing grid through nearest neighbor interpolation or bilinear interpolation.

[0137] In this step, during spatial positioning and matching, the spatial range and pixel location of the image are determined based on the latitude and longitude, row and column number information, or GEO-assisted positioning information in the FY4B GHI satellite channel data. Since different channels may have different spatial resolutions, this invention uses a visible light identification grid as a unified processing grid. Long-wave infrared brightness temperature data is resampled to the unified grid using nearest neighbor or bilinear interpolation, enabling data from each channel to participate in the comprehensive judgment pixel-by-pixel.

[0138] S14. Arrange the multi-channel data from multiple consecutive time periods in chronological order of observation time to form a time series dataset with dimensions of [number of time periods, number of channels, image height, and image width].

[0139] S2. Visible light spatial feature extraction: Based on the visible light channel data of FY4B GHI, the highlight features, color consistency features and texture smoothness features of the fog area are extracted respectively. The three types of features are jointly judged pixel by pixel to generate visible light candidate fog area masks.

[0140] In this step, RGB true-color images are constructed from B02, B03, and B04, and combined with the B01 full-color channel to form an enhanced brightness image. Fog areas in visible light images typically exhibit high brightness, a grayish-white color, and relatively smooth internal texture. Therefore, this invention extracts the spatial features of fog areas from three aspects: brightness, color, and texture. The specific steps for extracting visible light spatial features are as follows:

[0141] S21. Highlight Feature Extraction: Construct an RGB true-color image by using channels B02, B03, and B04 as R, G, and B components, respectively. After normalization to the range of 0-255, convert it into a grayscale image. Use the contrast-limited adaptive histogram equalization (CLAHE) method to locally enhance the contrast of the grayscale image. Calculate the 75th percentile of the brightness of all effective pixels in the enhanced image as the adaptive brightness threshold. Pixels with brightness values ​​greater than this threshold are marked as high-brightness pixels.

[0142] During highlight feature extraction, an RGB image is first constructed using the red, green, and blue light channels. Specifically, the red channel is used as the R component, the green channel as the G component, and the blue channel as the B component. These three channels are then uniformly normalized to convert them into an 8-bit image within the range of 0-255. The RGB image is then converted into a grayscale image, where the grayscale value represents the overall brightness of a pixel in the visible light image.

[0143] To enhance the local brightness difference between the foggy area and the surrounding background, this invention performs local contrast enhancement processing on the grayscale image. Specifically, it employs the Limiting Contrast Adaptive Histogram Equalization (CLAHE) method, which first divides the grayscale image into several small local regions, then enhances the brightness distribution of each small region while limiting the enhancement amplitude to avoid excessive amplification of noise. In this embodiment, the image can be divided into an 8×8 local grid, and the contrast limit parameter can be set to 2.0.

[0144] After local contrast enhancement, the brightness values ​​of all effective pixels in the enhanced image are statistically analyzed, and their 75th percentile is calculated as the adaptive brightness threshold. If the enhanced brightness value of a pixel is greater than this threshold, the pixel is considered to have fog highlight features and is recorded as a highlight pixel; otherwise, it is considered not to meet the highlight condition. This step yields the brightness feature mask.

[0145] In this embodiment, in step S21, the parameters of the contrast-limited adaptive histogram equalization method are set as follows: the grayscale image is divided into an 8×8 local grid, and the contrast limit parameter is 2.0.

[0146] S22. Color uniformity feature extraction: Calculate the standard deviation of the three components R, G, and B for each pixel. Pixels with a standard deviation less than 10 are marked as color uniform pixels.

[0147] When extracting color consistency features, fog areas in true-color images typically appear as grayish-white or milky-white, with minimal differences between the red, green, and blue channels. To quantitatively describe this feature, this invention calculates the standard deviation of each pixel across the R, G, and B components. Specifically, the calculation method involves first calculating the average of the R, G, and B values ​​for that pixel, then calculating the difference between each channel value and the average, and finally obtaining the RGB standard deviation for that pixel.

[0148] When the RGB standard deviation of a pixel is less than the set threshold of 10, it indicates that the pixel has a relatively uniform color, which is consistent with the gray-white homogeneous color characteristics of a foggy area; when the RGB standard deviation is large, it indicates that the pixel has obvious color differences, which may be ground features, colored backgrounds, cloud boundaries or other non-fog targets.

[0149] S23. Texture smoothing feature extraction: The grayscale brightness image is convolved using a 3×3 Laplacian operator. The absolute value of the convolution response is taken as the texture intensity value. Pixels with a texture intensity value less than 1 are marked as texture smoothing pixels.

[0150] When extracting texture smoothing features, fog areas typically lack obvious edges of ground features and texture details, resulting in relatively gentle overall spatial variations. In contrast, areas such as urban buildings, mountains, bare land, vegetation boundaries, and cloud boundaries often exhibit strong local grayscale variations. To quantitatively describe the smoothing features of fog areas, this invention employs the Laplacian operator to calculate the local texture intensity of grayscale images.

[0151] In practice, a grayscale image is taken as input, and a second-order Laplacian derivative operation is performed on it. The Laplacian operator is used to calculate the degree of grayscale change between a pixel and its surrounding pixels; the more drastic the grayscale change, the larger the operator response value; the more gradual the grayscale change, the smaller the operator response value. In this embodiment, the following 3×3 Laplacian convolution template can be used:

[0152] ;

[0153] in, It is the Laplace operator. It is the input image.

[0154] For each pixel, the Laplacian response value of the pixel is obtained by convolving the template with the gray values ​​of its neighbors. The absolute value of the response value is then taken as the texture intensity value. If the texture intensity value of a pixel is less than a set threshold of 1, it indicates that the gray value variation of the pixel and its neighbors is small and the texture is relatively smooth, which conforms to the texture characteristics of a fog area. If the texture intensity value is large, it indicates that there are obvious edges or complex textures in the area, and it is not a priority candidate for a fog area. The texture smoothness feature mask is obtained through the above calculation.

[0155] S24. Perform a pixel-by-pixel logical AND operation on the masks of bright pixels, uniform color pixels, and smooth texture pixels to obtain visible light candidate haze masks.

[0156] For visible light candidate fog region generation, after extracting brightness features, color uniformity features, and texture smoothness features, this invention performs pixel-by-pixel joint judgment on the three feature masks. Only pixels that simultaneously meet the three conditions of high brightness, color uniformity, and texture smoothness are determined as visible light candidate fog regions. This result is used to characterize areas that may contain fog during the day or under good visible light observation conditions, and serves as the basis for subsequent near-infrared interference removal, long-wave infrared thermal constraint, and temporal feature analysis.

[0157] S3. Near-infrared cloud phase and interference suppression feature extraction: Using the cirrus cloud channel and near-infrared channel data of FY4B GHI, cirrus cloud interference targets are removed, and low clouds, snow, ice clouds and highly reflective surfaces are distinguished to obtain candidate fog areas after interference suppression.

[0158] In this step, when extracting near-infrared cloud phases and interference suppression features, cirrus clouds, snow, ice clouds, bright ground surfaces, and some highly reflective clouds in visible light images may have similar high brightness and low texture to fog areas, easily leading to misjudgment. To reduce such interference, this embodiment introduces 1.38μm and 1.6μm near-infrared channels for auxiliary identification.

[0159] In this embodiment, the specific steps for extracting near-infrared interference suppression features are as follows:

[0160] S31. Cirrus cloud removal: Read the reflectance data of channel B05. Considering the influence of observation angle and regional atmospheric background, the cirrus cloud removal threshold T_cirrus is set to 0.02~0.04. When the reflectance of channel B05 is greater than or equal to T_cirrus, the pixel is marked as a cirrus cloud interference pixel and removed from the candidate fog area.

[0161] When removing cirrus clouds, the 1.38μm cirrus channel is sensitive to thin cirrus clouds at higher altitudes, while the response of near-surface fog to this channel is usually weak. Therefore, this invention utilizes the 1.38μm channel to construct a cirrus cloud interference criterion.

[0162] In practice, the reflectance data of the 1.38μm channel is read, and each pixel is judged to determine whether its reflectance exceeds the cirrus removal threshold T_cirrus. That is, when the 1.38μm reflectance is less than the cirrus removal threshold T_cirrus, the pixel is considered to pass the cirrus removal condition; when it is greater than or equal to the cirrus removal threshold T_cirrus, it is marked as a cirrus interference pixel.

[0163] S32. Screening of low cloud and strong reflective targets: Read the reflectivity data of channel B06. Considering the regional and seasonal differences in the response of low clouds, fog, snow, ice clouds and bright surfaces in the near-infrared channel, the near-infrared candidate threshold T_nir is set to 0.2 to 0.4. When the reflectivity of channel B06 is greater than T_nir, the pixel is regarded as a candidate pixel for low cloud and fog.

[0164] When screening low-cloud and highly reflective targets, the 1.6μm channel can be used to assist in identifying low clouds, water clouds, ice clouds, snow cover, and some highly reflective ground surfaces. In the embodiment, when the reflectivity of the 1.6μm channel is greater than the near-infrared candidate threshold T_nir, the pixel is considered to have near-infrared response characteristics of low-cloud or foggy targets and can be used as one of the candidate conditions for low-cloud and fog; when this condition is not met, it is not considered as a priority candidate area for low-cloud and fog.

[0165] Meanwhile, to reduce misjudgments of snow accumulation, ice clouds, and highly reflective surfaces, this invention combines visible light brightness, 1.6μm reflectivity, and long-wave infrared brightness temperature for comprehensive evaluation. Areas with high visible light brightness but significantly abnormal infrared brightness temperature, insufficient temporal stability, or that do not conform to the slow-changing characteristics of fog areas are removed from the candidate fog areas.

[0166] S33. Visible light reflectance auxiliary constraint: Read the reflectance data of the B02 channel. Considering that the visible light reflectance is affected by the solar altitude angle, surface type, sea and land background and fog thickness, the visible light high reflectance threshold T_vis is set to 0.20~0.45. When the reflectance of the B02 channel is greater than T_vis, the pixel is regarded as a candidate pixel for high reflectance.

[0167] In the visible light reflectance-assisted constraint, in addition to the brightness, color, and texture features based on RGB images, this invention can also utilize the blue light channel reflectance as a candidate constraint for visible light. In the embodiments, when the blue light channel reflectance is greater than the visible light high reflectance threshold T_vis, the pixel is considered to have strong visible light reflectance characteristics, consistent with the high reflectance performance of fog or low clouds in the visible light band; when the blue light channel reflectance is not greater than the visible light high reflectance threshold T_vis, it is not considered a preferred candidate fog region.

[0168] By using auxiliary constraints of 1.38μm, 1.6μm and blue light channel reflectance thresholds, the impact of cirrus clouds, bright ground surfaces and weakly reflective backgrounds on fog recognition results can be further reduced.

[0169] S4. Long-wave infrared thermal feature extraction: Based on the long-wave infrared channel data of FY4B GHI, brightness temperature threshold features and infrared local smoothing features are extracted to remove low-temperature high-cloud and thick-cloud interference targets.

[0170] In this step, the long-wave infrared channel reflects the thermal radiation characteristics of clouds or the ground surface, serving as a crucial basis for distinguishing fog, high clouds, thick clouds, and the ground background. Compared to high clouds and thick clouds, fog is typically located near the ground, and its brightness temperature is generally closer to the ground surface or near-surface air temperature; while high clouds and thick clouds, due to their higher cloud tops, usually have significantly lower infrared brightness temperatures. Therefore, this embodiment uses the FY4B GHI long-wave infrared brightness temperature as an important constraint for fog identification.

[0171] In this embodiment, the specific steps for extracting long-wave infrared thermal features in step S4 are as follows:

[0172] S41. Long-wave infrared brightness temperature preprocessing: The brightness temperature data of channel B07 is resampled to a unified visible light grid, and a 3×3 window median filter is used to smooth local isolated abnormal brightness temperatures, and invalid and outlier brightness temperature values ​​are removed.

[0173] During long-wave infrared brightness temperature preprocessing, the FY4B GHI long-wave infrared channel data is first read, and the original observations are converted into brightness temperature values ​​according to the calibration table. Since the spatial resolution of the long-wave infrared channel is usually different from that of the visible light channel, this invention resamples the long-wave infrared brightness temperature data to a spatial grid consistent with the visible light candidate fog region. After resampling, each visible light pixel corresponds to a long-wave infrared brightness temperature value, enabling it to participate in pixel-by-pixel comprehensive judgment.

[0174] After spatial matching is completed, invalid and outlier brightness temperature values ​​are removed. For isolated local abnormal brightness temperatures, a 3×3 window median filter can be used for smoothing to reduce the impact of single-point noise on fog recognition results.

[0175] S42. Low-temperature target removal: Read the brightness temperature data of channel B07. Considering that the brightness temperature of fog areas is closely related to season, land and sea background, terrain height, near-surface air temperature, diurnal variation, and fog thickness, the brightness temperature threshold T_bt is set to 270-275K. When the brightness temperature of channel B07 is less than or equal to T_bt, the pixel is marked as a low-temperature high-cloud or thick-cloud interference pixel and removed from the candidate fog area. When the brightness temperature of channel B07 is greater than T_bt, it is identified as meeting the thermal characteristics of near-surface fog or low-cloud fog targets and continues to participate in subsequent comprehensive judgment.

[0176] In this step, when eliminating low-temperature targets, high clouds and thick clouds typically have high cloud tops and low long-wave infrared brightness temperatures. To avoid misclassifying high clouds or thick clouds as fog, this invention uses long-wave infrared brightness temperature for low-temperature target elimination. A long-wave infrared brightness temperature greater than a brightness temperature threshold T_bt is used as the initial temperature condition for near-surface fog or low cloud candidates. When the long-wave infrared brightness temperature of a pixel is greater than the brightness temperature threshold T_bt, it is considered that it does not exhibit obvious low-temperature high cloud or thick cloud characteristics and can continue to participate in subsequent fog identification; when the brightness temperature is less than or equal to the brightness temperature threshold T_bt, it is considered that it is more likely to be a high cloud, thick cloud, cold cloud top, or other low-temperature target, and it is eliminated from the candidate fog area or its fog identification confidence is reduced.

[0177] S43. Infrared Local Smoothing Feature Extraction: Using each pixel as the center, calculate the local standard deviation of brightness temperature using a 5×5 window, and retain only pixels with a local standard deviation less than a set threshold as candidate pixels for thermal smoothing.

[0178] When extracting local smoothing features in infrared imaging, fog areas typically exhibit a layered distribution with relatively smooth local thermal structures. In addition to the brightness temperature threshold, this embodiment also calculates local variation features for long-wave infrared brightness temperature images. Specifically, a 5×5 window is used to calculate the local mean and local mean square value centered on each pixel, and the local standard deviation is then calculated. When the local standard deviation is less than a set threshold, the thermal feature variation in that area is considered small, consistent with the smoothing characteristics of fog or layered low clouds; when the local standard deviation is large, the area is considered to be a cloud boundary, complex surface, or other non-fog target.

[0179] In the existing code's physical threshold branch, a 5×5 window is used to calculate the local standard deviation for the 0.47μm channel, with a value less than 0.01 used as the smoothness criterion. Extending this approach to the long-wavelength infrared channel, the same window statistics method can be used to calculate the local standard deviation of brightness temperature, and the infrared smoothing threshold can be determined by combining sample validation.

[0180] S5. Multi-temporal stability and motion characteristic analysis: Using FY4B GHI minute-level continuous fast scan images, the visible light temporal stability characteristics and optical flow motion characteristics of the target area are calculated to screen out the target areas that are persistent and move slowly.

[0181] In this step, the FY4B GHI possesses continuous fast-scan observation capabilities, enabling it to acquire multiple images of the same area within a short time interval. This invention utilizes this feature to analyze the persistence, stability, and motion velocity of a target within continuous images. In this embodiment, the specific steps of multi-temporal stability and motion characteristic analysis in step S5 are as follows:

[0182] S51. Visible light temporal stability analysis: Establish a time buffer queue with a length of 5 frames to store grayscale brightness images for 5 consecutive time intervals; calculate the temporal standard deviation of the grayscale brightness value of each pixel in 5 consecutive frames, and mark pixels with a temporal standard deviation greater than 0 and less than 8 as temporally stable pixels.

[0183] In the visible light temporal stability analysis, a time buffer queue is first established to store visible light RGB images or grayscale brightness images from multiple consecutive time intervals. In this embodiment, the length of the time buffer queue can be set to 5-10 frames; in this invention, it is set to 5 frames.

[0184] For each pixel, its grayscale brightness value is extracted from several consecutive frames of imagery, and the temporal standard deviation is calculated. If the temporal standard deviation is less than 8 and greater than 0, it indicates that the pixel brightness remains relatively stable over a short period of time, consistent with the characteristics of persistent and slow-changing fog. If the temporal standard deviation is greater than or equal to 8, it indicates that the pixel changes rapidly, possibly due to fast-moving clouds, shadow changes, ground reflections, or other unstable targets. This step yields a visible light temporal stability feature mask.

[0185] S52. Calculation of Optical Flow Motion Features: Using the current image and images from the previous 4 minutes as input, the Farneback dense optical flow algorithm is employed to calculate the pixel-by-pixel horizontal displacement component u and the vertical displacement component v, according to the formula... Calculate the synthetic motion velocity V, and mark pixels with a motion velocity less than 2.0 pixels as slowly moving pixels.

[0186] To identify the slow movement of fog areas during optical flow motion feature calculation, this invention employs the Farneback optical flow method to calculate the motion velocity between consecutive images. The optical flow method is used to estimate the pixel displacement of a target in an image from one moment to the next, reflecting the target's direction and speed of movement.

[0187] First, the current image and the previous reference image are selected from the time buffer queue. In this embodiment, the reference interval is set to 5 frames; when the image time interval is 1 minute, it is equivalent to comparing the current image with the image from the previous 4 minutes. Then, the two RGB images are converted into grayscale images. The grayscale image from the previous reference time is used as input image one, and the grayscale image from the current time is used as input image two. The Farneback dense optical flow algorithm is used to calculate the horizontal and vertical displacement components pixel by pixel. Based on the calculated horizontal displacement component (u) and vertical displacement component (v), the pixel motion velocity is calculated. ;in, This represents the composite moving velocity of the pixel between two time intervals, where u is the horizontal displacement component and v is the vertical displacement component.

[0188] If the movement speed of a pixel is less than 2.0 pixels, it indicates that its movement is slow, which is consistent with the characteristics of slow movement or relative stability in a fog area; if the movement speed is greater than or equal to 2.0 pixels, it indicates that the pixel may be a fast-moving cloud or other dynamic target. This step yields the motion feature mask.

[0189] S6. Comprehensive Fog Area Judgment: Perform pixel-by-pixel joint logical judgment on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features, and multi-temporal spatiotemporal features to generate initial candidate fog areas.

[0190] In this step, after extracting visible light spatial features, near-infrared assisted identification, long-wave infrared thermal features, and multi-temporal spatiotemporal features, this invention comprehensively judges various features to generate candidate fog areas. The comprehensive judgment follows the steps of "first eliminating obvious interfering targets, then confirming fog area features." First, pixels that do not meet the fog identification conditions are eliminated, including invalid pixels, filled value pixels, cirrus cloud interference pixels, obviously low temperature high cloud pixels, suspected snow or ice cloud pixels, and suspected strong reflective surface pixels. Second, fog area features are confirmed in the remaining pixels.

[0191] In this embodiment, in step S6, the pixels determined to be initial candidate fog areas must simultaneously meet all of the following conditions:

[0192] (1) The luminance value enhanced by CLAHE is greater than the 75th percentile of the effective pixel luminance;

[0193] (2) The standard deviation of the RGB three-channel components is less than 10;

[0194] (3) The Laplacian texture intensity value is less than 1;

[0195] (4) The time standard deviation of visible light grayscale brightness is greater than 0 and less than 8;

[0196] (5) The motion velocity calculated based on the Farneback optical flow method is less than 2.0 pixels;

[0197] (6) The reflectivity of channel B05 is less than the cirrus removal threshold T_cirrus, where T_cirrus is set to 0.02 to 0.04;

[0198] (7) The reflectivity of the B06 channel is greater than the near-infrared candidate threshold T_nir, where T_nir is set to 0.2 to 0.4;

[0199] (8) The reflectivity of the B02 channel is greater than the visible light high reflectivity threshold T_vis, where T_vis is set to 0.20 to 0.45;

[0200] (9) The brightness temperature of channel B07 is greater than the brightness temperature threshold T_bt, where T_bt is set to 270~275K;

[0201] (10) The reflectivity or brightness temperature change within a local 5×5 window meets the preset smoothness condition.

[0202] S7. Post-processing: The initial candidate fog regions are sequentially processed by spatial morphology, connected component filtering, and temporal persistence verification, and the final fog recognition result is output.

[0203] In this step, after the candidate fog areas are generated, this embodiment further performs spatial post-processing and temporal persistence verification to improve the integrity and stability of the results. Spatial post-processing includes morphological closing operations and connected component filtering. First, a 5×5 structuring element is used to perform closing operations on the candidate fog areas to fill small holes inside the fog areas and connect local broken areas; then, connected component labeling is performed on the processed fog areas, the number of pixels in each patch is counted, patches with an area of ​​less than or equal to 50 pixels are removed, and only fog patches with an area of ​​more than 50 pixels are retained.

[0204] Temporal persistence verification involves constructing a historical fog area sequence of a certain length and counting the number of times the same pixel is identified as fog across multiple consecutive time periods. If a pixel is identified as fog at least three times in the images of the most recent consecutive time periods, it is determined to be a final fog area; if it only appears briefly in a single or a few time periods, it is considered to be noise, short-term glare, or other abnormal targets and is removed. The final output includes a fog identification binary raster product. The output results can be used for heavy fog monitoring, sea fog identification, traffic meteorological support, airport meteorological services, and other low-visibility weather monitoring operations.

[0205] In this embodiment, the specific steps of post-processing in step S7 are as follows:

[0206] S71. Spatial morphological processing: A 5×5 square structuring element is used to perform morphological closing operations on the initial candidate fog area to fill the small holes inside the fog area and connect the local broken areas.

[0207] S72. Connected Component Filtering: Mark connected components in the morphologically processed binary image, count the number of pixels in each connected component, and remove small patches with an area of ​​less than or equal to 50 pixels.

[0208] S73. Time persistence verification: Count the number of times the same pixel is identified as fog in the most recent 5 consecutive time intervals, and retain only the pixels that are identified as fog at least 3 times as the final fog area;

[0209] S74. Output a fog recognition binary raster product with a uniform spatial resolution, where 1 represents a fog area and 0 represents a non-fog area.

[0210] The fog area identification method based on the fusion of multispectral and spatiotemporal features is applicable to remote sensing monitoring of land fog and sea fog. The output fog identification results can be used for dynamic tracking of heavy fog processes, highway traffic meteorological services, aviation operation support, port and shipping scheduling, and identification of the impact of low cloud and fog in ecological and environmental remote sensing operations.

[0211] It should be understood that although the above description follows a certain order, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, some steps in this embodiment may include multiple steps or multiple stages, which are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of steps or stages in other steps.

[0212] This invention also provides a fog area recognition system based on multispectral and spatiotemporal feature fusion, used to execute the aforementioned fog area recognition method based on multispectral and spatiotemporal feature fusion. This fog area recognition system is implemented based on GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, and includes:

[0213] The data reading and preprocessing module is used to read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid.

[0214] The visible light spatial feature extraction module is used to extract the highlight features, color consistency features and texture smoothness features of the fog area based on the visible light channel data of FY4B GHI. The module performs pixel-by-pixel joint judgment on the three types of features to generate visible light candidate fog area masks.

[0215] The near-infrared cloud phase and interference suppression module is used to eliminate cirrus cloud interference targets using the cirrus cloud channel and near-infrared channel data of FY4B GHI, and to help distinguish low clouds, snow, ice clouds and highly reflective ground surfaces to obtain candidate fog areas after interference suppression.

[0216] The long-wave infrared thermal feature extraction module is used to extract brightness temperature threshold features and infrared local smoothing features from long-wave infrared channel data based on FY4B GHI, and to remove low-temperature high-cloud and thick-cloud interference targets.

[0217] The multi-temporal stability and motion feature analysis module is used to calculate the visible light temporal stability features and optical flow motion features of the target area using FY4B GHI minute-level continuous fast scan images, and to screen out the target areas that are persistent and move slowly.

[0218] The fog area comprehensive judgment module is used to perform pixel-by-pixel joint logical judgment on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features and multi-temporal spatiotemporal features to generate initial candidate fog areas.

[0219] The post-processing module is used to perform spatial morphological processing, connected component filtering, and temporal persistence verification on the initial candidate fog areas in sequence, and output the final fog recognition result.

[0220] In this embodiment, the number of consecutive time intervals read by the data reading and preprocessing module is 5-10, and the time interval between adjacent time intervals is 1 minute; the channels read include B01 panchromatic channel, B02 0.445~0.495μm blue light channel, B03 0.52~0.57μm green light channel, B04 0.62~0.67μm red light channel, B05 1.371~1.386μm cirrus channel, B06 1.58~1.64μm near-infrared channel, and B07 10.3~12.5μm long-wave infrared channel.

[0221] In this embodiment, the data reading and preprocessing module is specifically used to perform the following operations:

[0222] Read the original observation data and corresponding calibration table of each channel, convert the original digital values ​​of the reflectivity channels B01-B06 into reflectivity data, and convert the original digital values ​​of the long-wave infrared channel B07 into brightness temperature data.

[0223] Based on the data quality identifier and physical value range, fill values, invalid values, saturated values, and abnormal pixels that exceed the reasonable range of atmospheric and surface temperatures are set to null values;

[0224] Using the visible light channel identification grid as a unified processing grid, long-wave infrared brightness temperature data is resampled to the unified processing grid through nearest neighbor interpolation or bilinear interpolation.

[0225] The multi-channel data from multiple consecutive time periods are arranged in chronological order of observation time to form a time series dataset with dimensions of [number of time periods, number of channels, image height, and image width].

[0226] In this embodiment, the visible light spatial feature extraction module is specifically used to perform the following operations:

[0227] RGB true-color images were constructed by using channels B02, B03, and B04 as R, G, and B components, respectively. After normalization to the range of 0-255, the images were converted into grayscale images. The contrast-limited adaptive histogram equalization (CLAHE) method was used to enhance the local contrast of the grayscale images. The 75th percentile of the brightness of all effective pixels in the enhanced image was calculated as the adaptive brightness threshold. Pixels with brightness values ​​greater than this threshold were marked as bright pixels.

[0228] Calculate the standard deviation σ_RGB of the three components R, G, and B for each pixel. Pixels with a standard deviation σ_RGB less than 10 are marked as color-uniform pixels.

[0229] A 3×3 Laplacian operator is used to perform convolution operation on the grayscale brightness image. The absolute value of the convolution response value is taken as the texture intensity value. Pixels with a texture intensity value less than 1 are marked as texture smooth pixels.

[0230] A pixel-by-pixel logical AND operation is performed on the masks of bright pixels, uniform color pixels, and smooth texture pixels to obtain visible light candidate haze masks.

[0231] In this embodiment, the parameters of the contrast-limited adaptive histogram equalization method in the visible light spatial feature extraction module are set as follows: the grayscale image is divided into an 8×8 local grid, and the contrast limit parameter is 2.0.

[0232] In this embodiment, the near-infrared cloud phase and interference suppression module is specifically used to perform the following operations:

[0233] Read the reflectance data of channel B05. Considering the influence of observation angle and regional atmospheric background, the cirrus cloud removal threshold T_cirrus is set to 0.02~0.04. When the reflectance of channel B05 is greater than or equal to T_cirrus, the pixel is marked as a cirrus cloud interference pixel and removed from the candidate fog area.

[0234] The reflectance data of channel B06 is read. Considering the regional and seasonal differences in the response of low clouds, fog, snow, ice clouds and bright surfaces in the near-infrared channel, the near-infrared candidate threshold T_nir is set to 0.2 to 0.4. When the reflectance of channel B06 is greater than T_nir, the pixel is regarded as a candidate pixel for low cloud and fog.

[0235] The reflectance data of the B02 channel is read. Considering that the visible light reflectance is affected by the solar altitude angle, surface type, sea and land background and fog thickness, the visible light high reflectance threshold T_vis is set to 0.20~0.45. When the reflectance of the B02 channel is greater than T_vis, the pixel is regarded as a candidate pixel for high reflectance.

[0236] In this embodiment, the long-wave infrared thermal feature extraction module is specifically used to perform the following operations:

[0237] The brightness temperature data of channel B07 was resampled to a unified visible light grid, and a 3×3 window median filter was used to smooth local isolated abnormal brightness temperatures, eliminating invalid and outlier brightness temperature values.

[0238] Brightness temperature data from channel B07 is read. Considering that brightness temperature in fog areas is closely related to season, land-sea background, terrain elevation, near-surface air temperature, diurnal variation, and fog thickness, the brightness temperature threshold T_bt is set to 270–275 K. When the brightness temperature of channel B07 is less than or equal to T_bt, the pixel is marked as a low-temperature high-cloud or thick-cloud interference pixel and removed from the candidate fog area. When the brightness temperature of channel B07 is greater than T_bt, it is identified as meeting the thermal characteristics of near-surface fog or low-cloud fog targets and continues to participate in subsequent comprehensive judgment.

[0239] Centered on each pixel, the local standard deviation σ_T of the brightness temperature is calculated using a 5×5 window. Only pixels with a local standard deviation less than a set threshold (1.5K) are retained as candidate pixels for thermal smoothing.

[0240] In this embodiment, the multi-temporal stability and motion characteristic analysis module is specifically used to perform the following operations:

[0241] Establish a time buffer queue with a length of 5 frames to store grayscale brightness images for 5 consecutive time intervals; calculate the time standard deviation of grayscale brightness values ​​of each pixel in 5 consecutive frames, and mark pixels with a time standard deviation greater than 0 and less than 8 as time-stable pixels.

[0242] Using the current image and images from the previous 4 minutes as input, the Farneback dense optical flow algorithm is employed to calculate the pixel-by-pixel horizontal displacement component u and the vertical displacement component v, according to the formula... Calculate the synthetic motion velocity V, and mark pixels with a motion velocity less than 2.0 pixels as slowly moving pixels.

[0243] In this embodiment, in the fog area comprehensive determination module, the pixels determined as initial candidate fog areas must simultaneously meet all of the following conditions:

[0244] (1) The luminance value enhanced by CLAHE is greater than the 75th percentile of the effective pixel luminance;

[0245] (2) The standard deviation of the RGB three-channel components is less than 10;

[0246] (3) The Laplacian texture intensity value is less than 1;

[0247] (4) The time standard deviation of visible light grayscale brightness is greater than 0 and less than 8;

[0248] (5) The motion velocity calculated based on the Farneback optical flow method is less than 2.0 pixels;

[0249] (6) The reflectivity of channel B05 is less than the cirrus removal threshold T_cirrus, where T_cirrus is set to 0.02 to 0.04;

[0250] (7) The reflectivity of the B06 channel is greater than the near-infrared candidate threshold T_nir, where T_nir is set to 0.2 to 0.4;

[0251] (8) The reflectivity of the B02 channel is greater than the visible light high reflectivity threshold T_vis, where T_vis is set to 0.20 to 0.45;

[0252] (9) The brightness temperature of channel B07 is greater than the brightness temperature threshold T_bt, where T_bt is set to 270~275K;

[0253] (10) The reflectivity or brightness temperature change within a local 5×5 window meets the preset smoothness condition.

[0254] In this embodiment, the post-processing module is specifically used to perform the following operations:

[0255] A morphological closing operation was performed on the initial candidate fog region using a 5×5 square structuring element to fill the small holes inside the fog region and connect the local broken areas.

[0256] Connected component labeling is performed on the morphologically processed binary image, the number of pixels in each connected component is counted, and small patches with an area of ​​less than or equal to 50 pixels are removed.

[0257] The number of times the same pixel is identified as fog in the most recent 5 consecutive time intervals is counted, and only pixels that are identified as fog at least 3 times are retained as the final fog area;

[0258] Output a fog recognition binary raster product with a uniform spatial resolution, where 1 represents a fog area and 0 represents a non-fog area.

[0259] In this embodiment, the system also includes a result output and application module, which is used to push the final fog identification results to the fog process dynamic tracking system, highway traffic meteorological service platform, aviation operation support system, port shipping scheduling system and ecological environment remote sensing business platform to realize real-time identification and early warning of the impact of low cloud fog.

[0260] This invention uses the long-wave infrared brightness temperature of the 7th channel of FY4B GHI as the core thermal feature constraint, and constructs a multi-dimensional fusion criterion system that integrates visible light spatial details, near-infrared interference elimination, long-wave infrared thermal constraints, and minute-level spatiotemporal evolution. This significantly improves the ability to distinguish fog areas from easily confused targets such as high clouds, thin clouds, low clouds, snowfields, and bright ground surfaces, and reduces the false recognition rate.

[0261] To address the minute-level continuous observation characteristics of FY4B GHI, this invention designs a stability analysis based on 5-frame time series and an optical flow motion feature extraction method with 4-minute intervals. This effectively utilizes the advantage of the temporal continuity of the data, solves the problem of insufficient stability in single-frame identification, and reduces the missed identification rate. An adaptive brightness threshold based on image statistics is used instead of a fixed threshold, which reduces the impact of solar altitude angle, observation geometry, and regional background differences on the identification results. The adaptability of the algorithm in different seasons and regions is significantly improved.

[0262] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A fog area identification method based on multispectral and spatiotemporal feature fusion, implemented using GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, characterized in that... Includes the following steps: Data reading and preprocessing: Read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid. In data reading and preprocessing, the number of consecutive readings is 5-10, with a time interval of 1 minute between adjacent readings; the channels read include B01 panchromatic channel, B02 0.445~0.495μm blue light channel, B03 0.52~0.57μm green light channel, B04 0.62~0.67μm red light channel, B05 1.371~1.386μm cirrus channel, B06 1.58~1.64μm near-infrared channel, and B07 10.3~12.5μm long-wave infrared channel; Visible light spatial feature extraction: Based on the visible light channel data of FY4B GHI, the highlight features, color consistency features and texture smoothness features of the fog area are extracted respectively. The three types of features are jointly judged on a pixel-by-pixel basis to generate visible light candidate fog area masks. Near-infrared cloud facies and interference suppression feature extraction: Using cirrus cloud channel and near-infrared channel data from FY4B GHI, cirrus cloud interference targets are removed to help distinguish low clouds, snow, ice clouds, and highly reflective surfaces, thus obtaining candidate fog areas after interference suppression; In the near-infrared cloud facies and interference suppression feature extraction, cirrus cloud removal: The reflectance data of channel B05 is read. Considering the influence of observation angle and regional atmospheric background, the cirrus cloud removal threshold T_cirrus is set to 0.02~0.04; When the reflectance of channel B05 is greater than or equal to T_cirrus, the pixel is marked as a cirrus cloud interference pixel and removed from the candidate fog areas; Screening of low-cloud and high-reflectivity targets: Reflectance data from channel B06 is read. Considering the regional and seasonal differences in the near-infrared response of low clouds, fog, snow, ice clouds, and bright surfaces, the near-infrared candidate threshold T_nir is set to 0.2–0.

4. When the reflectance of channel B06 is greater than T_nir, the pixel is considered a candidate pixel for low-cloud and fog conditions. Visible light reflectance auxiliary constraint: Reflectance data from channel B02 is read. Considering that visible light reflectance is affected by solar altitude angle, surface type, sea and land background, and fog thickness, the visible light high reflectance threshold T_vis is set to 0.20–0.

45. When the reflectance of channel B02 is greater than T_vis, the pixel is considered a candidate pixel for high reflectance. Long-wave infrared thermal feature extraction: Based on FY4B GHI long-wave infrared channel data, brightness temperature threshold features and infrared local smoothing features are extracted to remove low-temperature high-cloud and thick-cloud interference targets. In long-wave infrared thermal feature extraction, the brightness temperature preprocessing involves resampling the brightness temperature data of the B07 channel to a unified visible light grid, using a 3×3 window median filter to smooth locally isolated abnormal brightness temperatures, and removing invalid and outlier brightness temperature values. Low-temperature target removal involves reading the brightness temperature data of the B07 channel, considering the close relationship between the brightness temperature in foggy areas and season, sea-land background, terrain height, near-surface air temperature, diurnal variation, and fog thickness. Relatedly, the brightness temperature threshold T_bt is set to 270-275K; when the brightness temperature of channel B07 is less than or equal to T_bt, the pixel is marked as a low-temperature high-cloud or thick-cloud interference pixel and removed from the candidate fog area; when the brightness temperature of channel B07 is greater than T_bt, it is identified as meeting the thermal characteristics of near-ground fog or low-cloud fog targets and continues to participate in subsequent comprehensive judgment; Infrared local smoothing feature extraction: with each pixel as the center, a 5×5 window is used to calculate the local standard deviation of the brightness temperature, and only pixels with a local standard deviation less than the set threshold are retained as thermal smoothing candidate pixels; Multi-temporal stability and motion feature analysis: Using FY4B GHI minute-level continuous fast-scan images, the visible light temporal stability and optical flow motion features of the target area are calculated to screen out target areas that are persistent and move slowly. In the multi-temporal stability and motion feature analysis, visible light temporal stability analysis: A time buffer queue of 5 frames is established to store grayscale brightness images from 5 consecutive time intervals; the temporal standard deviation of the grayscale brightness value of each pixel in the 5 consecutive frames is calculated, and pixels with a temporal standard deviation greater than 0 and less than 8 are marked as time-stable pixels; Optical flow motion feature calculation: The current image and the images from the previous 4 minutes are selected as input, and the Farneback dense optical flow algorithm is used to calculate the horizontal displacement component u and the vertical displacement component v per pixel, according to the formula... Calculate the synthetic motion velocity V, and mark pixels with a motion velocity of less than 2.0 pixels as slowly moving pixels; Comprehensive fog area determination: A pixel-by-pixel joint logical judgment is performed on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features, and multi-temporal spatiotemporal features to generate initial candidate fog areas; In the comprehensive fog area determination, the pixels determined as initial candidate fog areas must simultaneously meet all of the following conditions: the brightness value after CLAHE enhancement is greater than the 75th percentile of the effective pixel brightness; the standard deviation of the RGB three-channel components is less than 10. The Laplacian texture intensity value is less than 1; the temporal standard deviation of visible light grayscale brightness is greater than 0 and less than 8; the motion speed calculated based on Farneback optical flow is less than 2.0 pixels; the reflectance of channel B05 is less than the cirrus culling threshold T_cirrus, where T_cirrus is set to 0.02–0.04; the reflectance of channel B06 is greater than the near-infrared candidate threshold T_nir, where T_nir is set to 0.2–0.4; the reflectance of channel B02 is greater than the visible light high reflectance threshold T_vis, where T_vis is set to 0.20–0.45; the brightness temperature of channel B07 is greater than the brightness temperature threshold T_bt, where T_bt is set to 270–275K; the reflectance or brightness temperature variation within a local 5×5 window meets the preset smoothness conditions. Post-processing: Spatial morphological processing, connected component filtering, and temporal persistence verification are performed sequentially on the initial candidate fog regions to output the final fog recognition result.

2. The fog area identification method based on multispectral and spatiotemporal feature fusion according to claim 1, characterized in that, The data undergoes scaling transformation, invalid and outlier removal, spatial resolution matching, and time series organization in sequence, including the following steps: Read the original observation data and corresponding calibration table of each channel, convert the original digital values ​​of the reflectivity channels B01-B06 into reflectivity data, and convert the original digital values ​​of the long-wave infrared channel B07 into brightness temperature data. Based on the data quality identifier and physical value range, fill values, invalid values, saturated values, and abnormal pixels that exceed the reasonable range of atmospheric and surface temperatures are set to null values; Using the visible light channel identification grid as a unified processing grid, long-wave infrared brightness temperature data is resampled to this unified grid through nearest neighbor interpolation or bilinear interpolation. The multi-channel data from multiple consecutive time periods are arranged in chronological order of observation time to form a time series dataset with dimensions of [number of time periods, number of channels, image height, and image width].

3. The fog area identification method based on multispectral and spatiotemporal feature fusion according to claim 2, characterized in that, Visible light spatial feature extraction, Includes the following steps: Highlight Feature Extraction: RGB true-color images are constructed by using channels B02, B03, and B04 as R, G, and B components, respectively. After normalization to the 0-255 range, the images are converted into grayscale images. The contrast-limited adaptive histogram equalization (CLAHE) method is used to locally enhance the contrast of the grayscale images. The 75th percentile of the brightness of all effective pixels in the enhanced image is calculated as the adaptive brightness threshold, and pixels with brightness values ​​greater than this threshold are marked as high-brightness pixels. Color consistency feature extraction: Calculate the standard deviation of the three components R, G, and B for each pixel. Pixels with a standard deviation less than 10 are marked as color-uniform pixels. Texture smoothing feature extraction: A 3×3 Laplacian operator is used to perform convolution operation on the grayscale brightness image, and the absolute value of the convolution response value is taken as the texture intensity value. Pixels with a texture intensity value less than 1 are marked as texture smoothing pixels. A pixel-by-pixel logical AND operation is performed on the masks of bright pixels, uniform color pixels, and smooth texture pixels to obtain visible light candidate haze masks.

4. The fog area identification method based on multispectral and spatiotemporal feature fusion according to claim 3, characterized in that, Post-processing includes the following steps: Spatial morphological processing: A 5×5 square structuring element is used to perform morphological closing operations on the initial candidate fog area to fill the small holes inside the fog area and connect the local broken areas. Connected component filtering: Connected components are labeled in the morphologically processed binary image, the number of pixels in each connected component is counted, and small patches with an area of ​​less than or equal to 50 pixels are removed. Temporal persistence verification: Count the number of times the same pixel is identified as fog in the most recent 5 consecutive time intervals, and only retain the pixels that are identified as fog at least 3 times as the final fog area; Output a fog recognition binary raster product with a uniform spatial resolution, where 1 represents a fog area and 0 represents a non-fog area.

5. A fog area recognition system based on multispectral and spatiotemporal feature fusion, characterized in that, A fog area identification method based on multispectral and spatiotemporal feature fusion as described in any one of claims 1-4 is used to perform the fog area identification system, which is implemented based on GHI multispectral fast scan data from the FY4B geostationary orbit rapid imager, comprising: The data reading and preprocessing module is used to read FY4B GHI region fast scan L1 HDF data from multiple consecutive time intervals, and perform calibration transformation, invalid and outlier removal, spatial resolution matching and time series organization on the data in sequence to obtain a multi-channel multi-temporal dataset with a unified spatial grid. The visible light spatial feature extraction module is used to extract the highlight features, color consistency features and texture smoothness features of the fog area based on the visible light channel data of FY4B GHI. The module performs pixel-by-pixel joint judgment on the three types of features to generate visible light candidate fog area masks. The near-infrared cloud phase and interference suppression module is used to eliminate cirrus cloud interference targets using the cirrus cloud channel and near-infrared channel data of FY4B GHI, and to help distinguish low clouds, snow, ice clouds and highly reflective ground surfaces to obtain candidate fog areas after interference suppression. The long-wave infrared thermal feature extraction module is used to extract brightness temperature threshold features and infrared local smoothing features from long-wave infrared channel data based on FY4B GHI, and to remove low-temperature high-cloud and thick-cloud interference targets. The multi-temporal stability and motion feature analysis module is used to calculate the visible light temporal stability features and optical flow motion features of the target area using FY4B GHI minute-level continuous fast scan images, and to screen out the target areas that are persistent and move slowly. The fog area comprehensive judgment module is used to perform pixel-by-pixel joint logical judgment on visible light spatial features, near-infrared interference suppression features, long-wave infrared thermal features and multi-temporal spatiotemporal features to generate initial candidate fog areas. The post-processing module is used to sequentially perform spatial morphological processing, connected component filtering, and temporal persistence verification on the initial candidate fog regions, and output the final fog recognition result.

Images