A high-precision sea fog image recognition and prediction method based on ensemble learning

Abstract

The present invention relates to the field of sea fog image recognition and discloses a high-precision sea fog image recognition and prediction method based on ensemble learning. The purpose is to improve the efficiency and rational utilization of multi-source data fusion and enhance the stability of the model, thereby achieving more accurate sea fog recognition and more precise prediction of spatiotemporal evolution. The present invention constructs a three-dimensional feature space through visible light wavelet texture analysis, infrared dynamic temperature correction, and microwave liquid water path inversion. The image branch uses an improved DeepLabV3+ to perform fog area semantic segmentation. The combination of thermodynamic phase change criteria and fluid mechanics diffusion model realizes fog area recognition and trajectory prediction, ultimately forming a high-precision sea fog image recognition and prediction system. By replacing the black box model with physical equations, the shortcomings of traditional methods such as poor interpretability and insufficient real-time performance are addressed.

Description

Technical Field

[0001] The present invention relates to the field of sea fog image recognition, and in particular to a high-precision sea fog image recognition and prediction method based on ensemble learning. Background Art

[0002] Sea fog is a common weather phenomenon at sea and poses a great threat to the safety of maritime operations. Therefore, accurately identifying and predicting sea fog images is of great significance for ensuring the safety of maritime activities.

[0003] Currently, sea fog image recognition methods primarily include threshold segmentation, feature extraction, and machine learning. Threshold segmentation is simple and straightforward, but is significantly affected by external factors such as lighting and weather. Feature extraction relies on manual labor, which is time-consuming and labor-intensive, and can easily miss key information. Sea fog forecasting methods primarily include meteorological methods, numerical forecasting methods, and machine learning. Meteorological methods primarily rely on the forecaster's subjective experience and expertise, making accuracy difficult to guarantee in complex weather conditions. Numerical forecasting methods offer high accuracy for large-scale sea fog forecasts, but their accuracy is low for microscale sea fog forecasts, which are affected by complex local factors.

[0004] In recent years, machine learning methods have been gradually applied in the field of sea fog identification and forecasting, improving classification accuracy by automatically learning features. However, these methods have the following limitations:

[0005] Data dependency: A large number of labeled samples are required to train the model, but sea fog samples are difficult to obtain. Small sample sizes can easily lead to overfitting or the "curse of dimensionality."

[0006] Poor interpretability: Black-box models are difficult to relate to physical mechanisms and fail to reveal the key drivers of sea fog formation;

[0007] Insufficient real-time performance: Complex models are time-consuming to calculate, making it difficult to meet minute-level forecast requirements and relying on high computing resources.

[0008] Therefore, we propose a high-precision sea fog image recognition and prediction method based on ensemble learning to solve the above problems. Summary of the Invention

[0009] The present invention provides a high-precision sea fog image recognition and prediction method based on ensemble learning. Through the ternary architecture of physical mechanism modeling-multi-source hard fusion-dynamic feedback calibration, it provides a high-precision, low-dependence path for sea fog prediction.

[0010] The first aspect of the present invention provides a high-precision sea fog image recognition and prediction method based on ensemble learning, the high-precision sea fog image recognition and prediction method based on ensemble learning comprising: obtaining visible light satellite remote sensing images and infrared band data, performing multi-scale texture analysis on the visible light satellite remote sensing images, extracting texture features of fog area edges, generating visible light texture feature maps, calculating brightness temperature differences for the infrared band data, and generating infrared fog area probability maps in combination with sea surface temperature; obtaining sea surface temperature and air temperature data measured by ocean buoys, obtaining critical humidity thresholds, and generating a thermodynamic phase change criterion matrix in combination with edge sharpness quantification values ​​in the visible light texture feature maps; obtaining Microwave radiometer data is obtained, and the liquid water path is inverted based on the microwave brightness temperature data. Logical fusion is performed based on the infrared fog area probability map, the thermodynamic phase change criterion matrix and the liquid water path threshold to generate a multi-source fusion fog area identification map; based on the initial distribution of the fog area provided by the multi-source fusion fog area identification map, combined with wind field data and boundary layer height data, the horizontal diffusion rate and direction of the fog area are obtained, the future motion trajectory is simulated, and a fog belt diffusion vector field is generated; the multi-source fusion fog area identification map and the fog belt diffusion vector field are superimposed to generate a time series prediction result. When it is predicted that the fog area covers the port or waterway and the coverage probability is ≥ the preset value, a graded alarm signal is triggered to generate a comprehensive sea fog prediction alarm map.

[0011] Optionally, in a first implementation method of the first aspect of the present invention, it includes: performing a two-dimensional discrete wavelet transform on the visible light satellite remote sensing image, decomposing it to the third layer, extracting the energy value of the high-frequency component, and generating a wavelet high-frequency energy map; obtaining the sum of the high-frequency energy of each pixel point based on the wavelet high-frequency energy map, mapping the energy value to the [0,1] interval, and generating a visible light texture feature map; obtaining the brightness temperature difference BTD based on the channel brightness temperature data and the sea surface temperature SST, performing piecewise linear correction on the BTD according to the SST, and generating a corrected brightness temperature difference map; mapping the brightness temperature difference BTD to the probability of fog existence through an S-type function, and generating an infrared fog area probability map.

[0012] Optionally, in the second implementation of the first aspect of the present invention, the method includes: measuring the sea surface temperature SST and the air temperature 2 meters above the sea surface in real time based on the ocean buoy , and get the sea temperature difference: ; According to the positive and negative of ΔT, the thermodynamic state is divided and the sea-temperature difference state table is obtained; Based on the sea-temperature difference state table, the critical relative humidity threshold under different ΔT is obtained: when ΔT>0, ; When ΔT≤0, ; Get the dynamic critical humidity threshold table; Based on the visible light texture feature map and the dynamic critical humidity threshold table, perform weighted superposition of the edge sharpness quantification value and the critical relative humidity threshold: If the edge sharpness is ≥ 0.7, then Downgrade by 5%; if the edge sharpness is <0.7, then Maintain the original value; obtain the thermodynamic criterion matrix of the fused edge features; compare the real-time relative humidity RH with the corrected critical relative humidity threshold pixel by pixel based on the thermodynamic criterion matrix of the fused edge features and the real-time relative humidity field :When RH≥ When RH< When , it is determined to be a non-fog area; the thermodynamic phase change criterion matrix is ​​obtained.

[0013] Optionally, in a third implementation method of the first aspect of the present invention, it includes: performing radiation calibration and geolocation correction on the brightness temperature data, eliminating outliers affected by sea surface roughness anomalies, and obtaining a corrected microwave brightness temperature map; based on the corrected microwave brightness temperature map, based on the microwave radiation transmission model, inverting the liquid water path to obtain a liquid water path distribution map; resampling the liquid water path distribution map to 1 km resolution through bilinear interpolation, processing it according to the infrared fog area probability map, the thermodynamic criterion matrix and the liquid water path distribution map, and obtaining a preliminary screening fusion fog area distribution map; in the preliminary screening fog area, if the edge sharpness in the visible light texture feature map is ≥0.7, it is determined to be a reliable fog area, and the false detection areas with edge sharpness <0.7 are eliminated to generate a multi-source fusion fog area identification map.

[0014] Optionally, in a fourth implementation method of the first aspect of the present invention, it includes: converting the multi-source fusion fog area identification map into a rasterized concentration field, generating an initial concentration field of the fog area, adjusting the wind field intensity according to the boundary layer height: for every 100 meters the boundary layer height decreases, the wind field speed decays by 5%, and generating a corrected wind field vector diagram; based on the initial concentration field of the fog area and the corrected wind field vector diagram, with a time step of 10 minutes, iteratively calculating the movement trajectory of the fog area with the wind field in the next 3 hours, recording the front position of the fog area at each time step, and generating a fog area trajectory sequence; based on the fog area trajectory sequence, obtaining the fog area displacement of adjacent time steps, extracting the diffusion speed and direction, eliminating abnormal trajectory points caused by terrain obstruction, and generating a fog belt diffusion vector field.

[0015] Optionally, in a fifth implementation method of the first aspect of the present invention, it includes: temporally and spatially aligning the multi-source fusion fog area identification map with the fog band diffusion vector field, interpolating the fog area front position at intervals of 10 minutes, and generating a time series fog area prediction grid set; based on the time series fog area prediction grid set and the preset port and waterway geographic information database, for each port / waterway area, counting the proportion of the number of prediction grids covered by the fog area, and obtaining the regional coverage probability; when the coverage probability of the same area for three consecutive time steps is ≥70%, it is marked as a high-risk area, and a key area coverage probability table is obtained; based on the key area coverage probability table, setting the alarm threshold according to the risk level: yellow alarm: 70%≤probability<85%; orange alarm: 85%≤probability<95%; red alarm: probability≥95%; generating a graded alarm instruction set in combination with the regional priority weight; overlaying the time series fog area prediction grid set and the graded alarm instruction set on the electronic nautical chart base map, marking the fog area front time series and the alarm area, and generating a comprehensive sea fog prediction alarm map.

[0016] Optionally, in the sixth implementation method of the first aspect of the present invention, the following steps are also included: comparing the fog area range in the sea fog comprehensive forecast warning map with the visibility observation data transmitted back in real time by the ocean buoy to obtain the regional prediction error, the regional prediction error = predicted coverage - measured coverage; if the absolute value of the prediction error for three consecutive time steps is greater than 15%, calibration is triggered: adjusting the critical humidity threshold of the thermodynamic phase change criterion; correcting the wind field attenuation coefficient; and generating the optimized model parameter set.

[0017] The mechanism of the present invention is as follows: The present invention is based on a dynamic closed-loop prediction system coupled with multi-source physical characteristics, and replaces data-driven by physical model-driven to achieve breakthrough innovation in sea fog identification and prediction.

[0018] Beneficial effects: Visible light wavelet decomposition quantifies the edge texture of fog areas, breaking through the traditional image segmentation's reliance on color features. Segmented correction of infrared brightness temperature difference and sea surface temperature solves the problem of misjudging fog areas in low sea temperature areas. The microwave liquid water path inversion formula enhances vertical information perception in fog areas through physical radiation models.

[0019] Dynamically adjust the critical humidity threshold based on the sea-temperature difference to achieve real-time optimization of the thermodynamic phase change criterion, correct the wind field attenuation rate with the boundary layer height, and incorporate the vertical stratification effect into the horizontal diffusion model;

[0020] Using physical equations instead of data-driven methods solves the inherent problems of traditional methods that rely on labeled data and are unexplainable black boxes.

[0021] A closed loop is formed from feature extraction, diffusion prediction to real-time calibration, breaking through the limitations of fixed numerical model parameters. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of an embodiment of a high-precision sea fog image recognition and prediction method based on ensemble learning in an embodiment of the present invention;

[0023] Figure 2 2 is a schematic diagram of another embodiment of a high-precision sea fog image recognition and prediction method based on ensemble learning in an embodiment of the present invention;

[0024] Figure 3 Schematic diagram of an embodiment of a high-precision sea fog image recognition and prediction device based on ensemble learning in an embodiment of the present invention;

[0025] Figure 4 A schematic diagram of an embodiment of a high-precision sea fog image recognition and prediction device based on ensemble learning in an embodiment of the present invention;

[0026] Figure 5 This is the improved DeepLabV3+ network structure diagram in the present invention. DETAILED DESCRIPTION

[0027] Embodiments of the present invention provide a high-precision sea fog image recognition and prediction method based on ensemble learning. Through a ternary architecture of physical mechanism modeling, multi-source hard fusion, and dynamic feedback calibration, this method provides a high-precision, low-dependency path for sea fog prediction. Furthermore, the terms "comprising," "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units expressly listed, but may include other steps or units not expressly listed or inherent to such process, method, product, or apparatus.

[0028] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 In one embodiment of the present invention, a high-precision sea fog image recognition and prediction method based on ensemble learning includes:

[0029] 101. Feature extraction of multispectral satellite remote sensing images. Input data: Visible light satellite remote sensing images: from the VISSR sensor of the Fengyun-4A satellite, with a spatial resolution of 500 meters; Infrared band data: 10.8μm and 12.0μm channel brightness temperature data from the AHI sensor of the Himawari-8 satellite;

[0030] Multi-scale texture analysis was performed on visible light images to extract texture features at the edge of the fog area and generate a visible light texture feature map. The brightness temperature difference (brightness temperature difference between 10.8 μm and 12.0 μm) was calculated for infrared band data and combined with sea surface temperature to generate an infrared fog area probability map.

[0031] Output products: visible light texture feature map (including edge sharpness quantification value); infrared fog area probability map (resolution of 2 kilometers, pixel value represents the probability of fog presence, ranging from 0 to 1);

[0032] It is understood that the execution subject of the present invention can be a high-precision sea fog image recognition and prediction device based on ensemble learning, or a terminal or a server, which is not limited here. The embodiment of the present invention is described by taking the server as the execution subject as an example.

[0033] It should be noted that the input data are visible light satellite remote sensing images (spatial resolution 500 meters) acquired by the VISSR sensor of the Fengyun-4A satellite.

[0034] Operational Process: Multi-Scale Texture Analysis: Sliding Window Settings: Three sliding window sizes (3×3, 5×5, and 7×7) are used to extract texture features at different scales. Gray-Level Co-occurrence Matrix (GLCM) Calculation: For each window, texture parameters such as contrast (CON), entropy (ENT), and correlation (COR) are calculated. Fog edges typically exhibit low contrast (CON ≤ 5) and high entropy (ENT ≥ 2.5), reflecting the smoothness and randomness of the fog texture. Edge Sharpness Quantification: The gradient magnitude is calculated using the Sobel operator and combined with GLCM features to generate an edge sharpness index (ranging from 0 to 1, with higher values ​​indicating sharper edges).

[0035] Feature fusion and dimensionality reduction: Multi-scale texture parameters are input into principal component analysis (PCA), and the first principal component (contribution rate ≥ 85%) is extracted as the core indicator of the visible light texture feature map. The output resolution is maintained at 500 meters and unified to a 1 km grid through bilinear interpolation to facilitate subsequent fusion.

[0036] Output product: visible light texture feature map (1 km resolution, pixel value is the quantified value of edge sharpness). The edge sharpness of a certain area is 0.8, indicating that there is a significant fog boundary in the area.

[0037] Infrared fog area probability map generation, input data: 10.8μm and 12.0μm channel brightness temperature data of the AHI sensor of the Himawari-8 satellite (original resolution 2 km).

[0038] Operation process: Brightness temperature difference calculation: Calculate 10.8μm (BT 10.8 ) and 12.0μm (BT 12.0 ) Channel brightness temperature difference ( Sea surface temperature (SST) constraint: Combined with remote sensing inversion SST data (MODIS SST product), screen areas where the difference between SST and dew point temperature is ≤ 2°C (the thermodynamic condition for fog formation). Construct a logistic regression model: , where σ is the Sigmoid function, and the output probability range is 0 to 1. Spatial interpolation and resolution adjustment: Kriging interpolation is used to unify the probability values ​​from a 2 km resolution to a 1 km grid and align them with the visible light feature map space.

[0039] Output product: infrared fog area probability map (2 km resolution, probability value example: a pixel value of 0.75 in a certain sea area indicates that there is a 75% probability of sea fog there).

[0040] 102. Modeling of thermodynamic phase change criteria, input data: sea surface temperature and air temperature 2 meters above the sea surface measured by ocean buoys; visible light texture feature map generated in step 101;

[0041] The critical humidity threshold is calculated based on the thermodynamic equation and combined with the edge sharpness quantification value in the visible light texture feature map to generate the thermodynamic phase change criterion matrix.

[0042] Output product: thermodynamic phase change criterion matrix (resolution is 1 km, 0 means no fog, 1 means fog);

[0043] It should be noted that the measured data, including sea surface temperature (SST = 20°C) and air temperature 2 meters above the surface (Ta = 18°C), measured by ocean buoys, must be converted from discrete point data to a 1 km resolution grid using kriging interpolation. The visible light texture feature map (from step 101) has a resolution of 1 km and an edge sharpness range of 0.3 to 0.9 (edge ​​sharpness in fog areas is typically ≥ 0.6).

[0044] The critical humidity threshold is calculated based on the thermodynamic phase transition criterion. When the air humidity reaches saturation (i.e., the air temperature is close to the dew point), sea fog may form. The specific steps are as follows:

[0045] Dew point temperature (Td) calculation: Using the measured air temperature (Ta) and relative humidity (RH), the dew point temperature is calculated using the Magnus formula:

[0046] ;

[0047] The unit of 237.3 is ℃, and the actual RH measured by the buoy is 90%, so the calculated value is Td≈16.5℃.

[0048] Judgment conditions: If , the air is close to saturation, meeting the humidity conditions for fog formation. In this example, , meets the requirements.

[0049] Sea surface temperature (SST) constraint: When SST ≥ 15°C and the difference between SST and Ta ≤ 3°C, water vapor evaporates sufficiently, which is conducive to fog formation. In this example, SST = 20°C, , meets the requirements.

[0050] Visible light texture feature fusion, extracting edge sharpness values ​​from the visible light texture feature map (setting the sharpness of a certain area to 0.7). When the sharpness is ≥ 0.6, it indicates that there is a clear fog area boundary in the area.

[0051] Logical fusion rule: If a pixel satisfies both: ; ; If the edge sharpness is ≥ 0.6, it is marked as fog area (set to 1 in the criterion matrix), otherwise it is 0.

[0052] Output product example Table 1, thermodynamic phase transition criterion matrix (1 km resolution):

[0053] Table 1

[0054] Location coordinates Criterion value Key parameters (120°E, 30°N) 1 Ta=18℃, Td=16.5℃, sharpness=0.7 (120.1°E,30°N) 0 Ta=19℃, Td=14℃ (humidity conditions are not met)

[0055] If the measured RH fluctuates to 85% (Td≈15℃), then ℃, the judgment criteria are invalid and the model needs to be recalibrated; verified by the ROC curve, when the sharpness threshold is increased from 0.6 to 0.65, the false positive rate decreases by 5%, but the missed detection rate increases by 3%, which requires a trade-off.

[0056] 103. Multi-source data fusion and fog area identification. Input data: infrared fog area probability map generated in step 101; thermodynamic phase change criterion matrix generated in step 102; microwave radiometer data: 36.5 GHz vertically polarized brightness temperature data from the AMSR2 sensor of the GCOM-W1 satellite;

[0057] Liquid water paths are inverted based on microwave brightness temperature data, and the infrared fog area probability map, thermodynamic phase change criterion matrix, and liquid water path threshold are logically fused to generate a multi-source fusion fog area identification map.

[0058] Output: Multi-source fusion fog area identification map (resolution of 1 km, rasterized binary distribution, 0 indicates non-fog area, 1 indicates fog area);

[0059] It should be noted that the infrared fog area probability map (from step 101) has a resolution of 2 kilometers, the pixel value range of a certain sea area is 0.2 (clear sky) to 0.9 (high fog probability), and the infrared probability value of a certain area (coordinates 120.5°E, 32°N) is 0.8, indicating that there is an 80% probability of sea fog there.

[0060] Thermodynamic phase change criterion matrix (from step 102): resolution is 1 km, binary data (0 or 1), and the thermodynamic criterion value of the same area is 1, indicating that the humidity and temperature conditions for fog formation are met.

[0061] Microwave radiometer data (AMSR2 sensor on the GCOM-W1 satellite): 36.5 GHz vertically polarized brightness temperature data (original resolution 10 km), downscaled to a 1 km grid using bilinear interpolation, and the liquid water path (LWP) inverted to 200 g / m 2 (The threshold is set to ≥150g / m 2 is judged as fog).

[0062] Multi-source data logical fusion, fusion rules: Infrared probability threshold: Pixel value ≥ 0.6 (probability ≥ 60%) is considered a candidate fog area. Thermodynamic criterion: Matrix value is 1 (meets thermodynamic conditions). Liquid water path threshold: LWP ≥ 150g / m 2 (Lower limit of liquid water content in fog area).

[0063] Example scenario:

[0064] Parameters for a 1 km grid (coordinates 120.5°E, 32°N): infrared probability: 0.8; thermodynamic criterion: 1; LWP: 200 g / m 2 ;

[0065] Fusion result: All three conditions are met at the same time, and the mark 1 indicates a fog area.

[0066] Another grid (coordinates 120.6°E, 32°N) parameters: Infrared probability: 0.7; Thermodynamic criterion: 0 (insufficient humidity); LWP: 180g / m 2 ;

[0067] Fusion result: The thermodynamic criterion is not satisfied, and the mark 0 indicates the non-fog area.

[0068] Spatial alignment and resolution unification: infrared probability maps: interpolated from 2 km to 1 km grid, using bicubic interpolation to reduce edge blur; microwave data: downscaled from 10 km to 1 km, combined with the weighted average of the brightness temperature values ​​of neighboring pixels to ensure consistency with the thermodynamic matrix space.

[0069] Output product example Table 2: Multi-source fusion fog area identification map: 1 km resolution, binary distribution.

[0070] Table 2

[0071] Location coordinates Criterion value Key parameters (IR probability / thermodynamics / LWP) (120.5°E, 32°N) 1 0.8 / 1 / 200g / m2 (120.6°E, 32°N) 0 0.7 / 0 / 180g / m2

[0072] Parameter sensitivity analysis:

[0073] LWP threshold adjustment: If the threshold is reduced to 100g / m 2 , the fog coverage rate may increase by 15%, but the misjudgment rate will increase (such as misjudging thin clouds as fog).

[0074] Infrared probability threshold optimization: Verified by the ROC curve, a threshold of 0.6 balances the missed detection rate (8%) and the false detection rate (12%), making it suitable for most ocean scenarios.

[0075] 104. Fluid mechanics diffusion trajectory prediction, input data: multi-source fusion fog area identification map generated in step 103; 850hPa wind field data and boundary layer height data provided by numerical weather forecast;

[0076] The horizontal diffusion rate and direction of the fog area are calculated based on the fluid mechanics equation, and the movement trajectory in the next three hours is simulated to generate the fog belt diffusion vector field.

[0077] Output product: fog band diffusion vector field (including the spatiotemporal distribution of diffusion speed and direction, with speed unit in m / s);

[0078] It should be noted that the multi-source fusion fog area identification map (from step 103) has a resolution of 1 km, binary raster data (0 for non-fog area, 1 for fog area), and a sea area grid (coordinates 120.5°E, 32°N) is marked as 1, indicating that there is currently a fog area.

[0079] Numerical Weather Forecast Data: 850hPa Wind Data: Horizontal wind speed is 12m / s, northeasterly (45°). Boundary layer height: The current boundary layer height is 800 meters, which affects the vertical diffusion limit of the fog area.

[0080] The fluid mechanics diffusion model is constructed based on the Navier-Stokes equation and the advection-diffusion equation to calculate the horizontal diffusion rate and direction of the fog area:

[0081] Calculation of advection term: The wind field dominates the advection direction. Based on the 850hPa wind field speed (12m / s) and direction (45°), the hourly advection displacement of the fog area is calculated:

[0082] ;

[0083] ;

[0084] The direction of advection is consistent with the wind field, with a cumulative displacement of about 91.2 kilometers every 3 hours.

[0085] Calculation of turbulent diffusion term: Based on the boundary layer height (800 meters) and atmospheric stability (neutral conditions), the eddy diffusion coefficient K=50m is used 2 / s, calculate the horizontal diffusion rate:

[0086] ;

[0087] Comprehensive motion trajectory simulation: advection and diffusion are superimposed to generate a fog belt diffusion vector field. The predicted coordinates of the center point of a fog area (120.5°E, 32°N) after 3 hours are (121.3°E, 32.5°N), the diffusion speed is 12m / s, the direction is 45°, and the diffusion range radius is 1.8 kilometers.

[0088] Boundary layer height constraint: The boundary layer height limits vertical turbulent mixing, the fog area only spreads below 800 meters, and the vertical velocity is negligible.

[0089] Terrain correction: If the predicted path encounters the coastline (Yangtze River Estuary), the diffusion direction is adjusted to follow the terrain shear direction (deflection 20°).

[0090] See Table 3 for examples of output products:

[0091] Fog band diffusion vector field: 1 km resolution spatiotemporal distribution, each grid contains speed and direction.

[0092] Table 3

[0093] Location coordinates Speed ​​(m / s) Direction (°) Key parameters (advection / diffusion) (120.5°E, 32°N) 12 45 Wind farm dominated, diffusion radius 1.8km (121.0°E,32.3°N) 10 50 Terrain shear correction, diffusion radius 1.5km

[0094] Verification method: Compare the actual fog area position of the buoy 3 hours later with the predicted trajectory. If the error is ≤5 kilometers, it is considered valid.

[0095] Parameter sensitivity: If the boundary layer height drops to 500 meters, the diffusion radius shrinks to 1.2 kilometers, with the velocity direction unchanged. If the wind speed increases to 15 meters per second, the advection displacement increases to 113 kilometers per three hours, expanding the diffusion range.

[0096] 105. Comprehensive prediction and alarm generation, input data: multi-source fusion fog area identification map generated in step 103; fog band diffusion vector field generated in step 104;

[0097] The multi-source fusion fog area identification map and the fog band diffusion vector field are superimposed to generate time series prediction results. When the fog area is predicted to cover the port or waterway with a coverage probability of ≥70%, a graded alarm signal is triggered and a comprehensive sea fog forecast alarm map is generated.

[0098] Output: Comprehensive sea fog forecast warning map (including geographic coordinates, impact time and warning level);

[0099] It should be noted that the multi-source fusion fog area identification map (from step 103) is binary raster data with a resolution of 1 km (0 is a non-fog area, 1 is a fog area). A certain sea area (coordinates 120.3°E, 36.1°N) currently has a fog area (pixel value 1), covering an area of ​​approximately 50 square kilometers.

[0100] The fog band diffusion vector field (from step 104) includes the diffusion speed (12 m / s) and direction (northeast wind 45°). It is predicted that the fog area will move approximately 91.2 km southeast in the next three hours, and the diffusion radius will expand to 1.8 km.

[0101] Time series forecast: The fog area identification map is superimposed on the diffusion vector field, and the fog area position is simulated every 30 minutes for the next three hours based on the advection-diffusion model:

[0102] Current time (T0): The center coordinates of the fog area are 120.3°E, 36.1°N, and the coverage area is 50km. 2 .

[0103] T0+1 hour: The center moves to 120.8°E, 36.5°N, and the coverage area expands to 58km 2 (The diffusion radius increases by 0.6km).

[0104] T0+3 hours: The center reaches 121.6°E, 37.1°N, covering an area of ​​75 km 2 , covering part of the waterway of Qingdao Port.

[0105] Coverage probability calculation: Based on the uncertainty of the diffusion vector field (wind speed error ±2m / s), Monte Carlo simulation was used to generate 100 sets of trajectories. The fog coverage probability of the port area (120.5°E-121.5°E, 36.5°N-37.5°N) was calculated. The coverage probability of the Qingdao Port channel at T0+3 hours was 75% (exceeding the threshold of 70%).

[0106] Alarm triggering conditions:

[0107] Level 1 alarm (red): Coverage probability ≥ 90% and impact time ≥ 6 hours.

[0108] Level 2 alarm (orange): Coverage probability ≥ 70% and impact duration 3-6 hours.

[0109] Level 3 alarm (yellow): Coverage probability ≥ 50% and impact time ≤ 3 hours.

[0110] Example scenario:

[0111] The Qingdao Port waterway is predicted to have a coverage probability of 75%, with an impact time of 08:00-11:00 (3 hours), triggering an orange alert. Alert information includes:

[0112] Geographical coordinates: 120.8°E-121.6°E, 36.5°N-37.1°N.

[0113] Impact time: 2025-05-09 08:00 to 11:00.

[0114] Alert level: Level 2 (orange), it is recommended that the port suspend the entry and exit of ships and activate fog navigation control.

[0115] Spatial consistency test: Compare the visibility data of buoys measured during the same period (for example, the visibility of a certain buoy drops to 300 meters) to verify that the position error in the fog area is ≤ 3 kilometers.

[0116] Parameter sensitivity: If the diffusion rate error exceeds 15%, the alert level may be downgraded to yellow (coverage probability corrected to 65%).

[0117] See Table 4 for an output example:

[0118] Table 4

[0119] Geographic coordinate range Impact time Coverage probability Alarm level Recommended measures 120.8°E-121.6°E,36.5°N-37.1°N 08:00-11:00 75% orange color Restricted access to waterways, activate fog lights for navigation

[0120] In the embodiment of the present invention, 500-meter-level visible light texture features, 2-kilometer-level infrared brightness temperature probability maps, and 10-kilometer-level microwave liquid water path data are integrated for the first time, and 1-kilometer-level grid alignment is achieved through bilinear interpolation and Kriging interpolation, breaking through the resolution limit of a single sensor; the thermodynamic phase change criterion and the visible light texture features are logically integrated to construct a criterion matrix that combines physical drive and data drive to reduce the misjudgment rate; 3×3, 5×5, and 7×7 sliding windows are used to extract GLCM parameters, and the Sobel operator is combined to generate an edge sharpness index to quantify the ambiguity of the fog area boundary; the advection term and turbulent diffusion term are simulated based on the Navier-Stokes equation to realize the fog belt trajectory prediction, and terrain shear correction is introduced; 36.5GHz vertical polarization brightness temperature data are used to extract GLCM parameters. Invert LWP and establish a dynamic threshold adjustment strategy; build a two-level warning system based on coverage probability and impact time, and use Monte Carlo simulation of wind speed error to quantify the uncertainty of coverage probability; reduce the confusion between mist and low clouds through edge sharpness quantification and infrared probability map dual verification; combine SST-Ta difference and dew point temperature difference constraints to eliminate interference from non-fog areas; calculate the advection displacement and diffusion radius based on the 850hPa wind field and boundary layer height, with a verification error of ≤5 kilometers, which is better than the prediction error of traditional numerical models; trigger an orange warning when the coverage probability is ≥70%, support graded control of ports and waterways, and reduce ship detention time; adapt to Fengyun-4A (FY-4A), Himawari-8 (Himawari-8), and GCOM-W1 satellite data, and be compatible with MODIS SST products and numerical weather prediction (NWP) wind fields to reduce deployment costs.

[0121] See also Figure 2 Another embodiment 2 of the high-precision sea fog image recognition and prediction method based on ensemble learning in the embodiment of the present invention includes:

[0122] 201. Feature extraction of multispectral satellite remote sensing images. Input data: Visible light satellite remote sensing images: from the VISSR sensor of the Fengyun-4A satellite, with a spatial resolution of 500 meters; Infrared band data: 10.8μm and 12.0μm channel brightness temperature data from the AHI sensor of the Himawari-8 satellite;

[0123] Multi-scale texture analysis was performed on visible light images to extract texture features at the edge of the fog area and generate a visible light texture feature map. The brightness temperature difference (brightness temperature difference between 10.8 μm and 12.0 μm) was calculated for infrared band data and combined with sea surface temperature to generate an infrared fog area probability map.

[0124] Output products: visible light texture feature map (including edge sharpness quantification value); infrared fog area probability map (resolution of 2 kilometers, pixel value represents the probability of fog presence, ranging from 0 to 1).

[0125] Specifically, sub-step S1-1: Input data: satellite remote sensing images in the visible light band (0.45-0.90 μm) acquired by the VISSR sensor of the Fengyun-4A satellite, with an original resolution of 500 meters; perform a two-dimensional discrete wavelet transform on the visible light image using the Daubechies-4 wavelet basis function and decompose it to the third layer; extract the energy values ​​of the high-frequency components (horizontally, vertically, and diagonally) to generate a wavelet high-frequency energy map; output product: wavelet high-frequency energy map (resolution 500 meters, quantifying the edge sharpness of the fog area);

[0126] Sub-step S1-2: Quantification and normalization of edge sharpness, input data: wavelet high-frequency energy map generated by sub-step S1-1; calculate the sum of high-frequency energy of each pixel; map the energy value to the interval [0,1] through linear normalization to generate a visible light texture feature map with a resolution of 500 meters, where the pixel value represents edge sharpness); product use: as input to step 202, used for thermodynamic criterion modeling.

[0127] Sub-step S1-3: Calculation of infrared brightness temperature difference and sea surface temperature correction, input data: 10.8μm and 12.0μm channel brightness temperature data (resolution 2 km) from the AHI sensor of the Himawari-8 satellite; sea surface temperature (SST, resolution 5 km) provided by the ERA5 reanalysis data; calculation of brightness temperature difference ( ); Perform piecewise linear correction on BTD according to SST: when SST < 285K, the correction coefficient α = 0.8; when SST ≥ 285K, the correction coefficient α = 1.2; generate a corrected brightness temperature difference map (resolution 2 km); product use: as input for sub-steps S1-4.

[0128] Sub-step S1-4: infrared fog area probability mapping, input data: the corrected brightness temperature difference map generated in sub-step S1-3; set the threshold range: when the correction When , it is determined to be a potential fog area; the brightness temperature difference is mapped to the probability of fog existence through the S-type function to generate an infrared fog area probability map (resolution 2 kilometers, pixel value ∈ [0,1]); product use: as the input of step 203, participate in multi-source data fusion.

[0129] It should be noted that the following is an example of a visible light image (500-meter resolution) of the East China Sea region acquired by the VISSR sensor of the Fengyun-4A satellite at 08:00 UTC on May 10, 2025:

[0130] The Daubechies-4 wavelet basis function is used to perform a three-layer decomposition, and the energy of the high-frequency components (horizontal, vertical, and diagonal directions) of each layer is calculated: the high-frequency energy of the third layer of a certain foggy edge pixel (coordinate E123°45', N30°15') is 12.5, 14.3, and 11.7 W / m, respectively. 2 sr μm, total 38.5 W / m 2 ·sr·μm;

[0131] The energy value is mapped to the interval [0,1] by linear normalization. Assume that the maximum high-frequency energy of the whole image is 200 W / mW / m 2 sr μm, the normalized value of this pixel is 38.5 / 200=0.192, reflecting the edge sharpness feature;

[0132] Infrared brightness temperature difference correction and fog area probability mapping are based on the AHI sensor data of the Himawari-8 satellite during the same period:

[0133] The original brightness temperature values ​​of a certain sea area (E125°00', N31°20') are: BT10.8μm=285K, BT12.0μm=283K, and initial BTD=2K;

[0134] Combined with ERA5 sea surface temperature data (SST = 282K < 285K), the correction factor α = 0.8 is applied. ;

[0135] Mapping probabilities through sigmoid functions: When the probability , indicating that there is a 78.5% probability that the pixel exists in a fog area;

[0136] The ERA5 5km SST data were downscaled to 2km resolution using bilinear interpolation and aligned with the AHI infrared data. At one interpolation point (E124°30', N30°45'), the original SST was 286K (5km grid), while the interpolated SST at the 2km grid was 285.3K, triggering a correction factor of α = 1.2.

[0137] After correction, the brightness temperature difference map is resampled to 500-meter resolution using the nearest neighbor method and spatially aligned with the visible light feature map to ensure subsequent fusion accuracy;

[0138] In the visible light texture feature map, continuous high-sharpness areas (>0.7) were detected in the waters around the Zhoushan Islands, with an 87% agreement with the visibility <1km area reported by the actual Automatic Identification System (AIS);

[0139] The infrared fog probability map shows that there is a high probability zone in the waters off the Yangtze River estuary (P>0.8), and the liquid water path in the microwave inversion of the GCOM-W1 satellite is >0.2kg / m 2 The regional spatial overlap rate reaches 92%.

[0140] 202. Modeling of thermodynamic phase change criteria, input data: sea surface temperature and air temperature 2 meters above the sea surface measured by ocean buoys; visible light texture feature map generated in step 201;

[0141] The critical humidity threshold is calculated based on the thermodynamic equation and combined with the edge sharpness quantification value in the visible light texture feature map to generate the thermodynamic phase change criterion matrix (resolution of 1 km, 0 indicates non-fog, 1 indicates fog);

[0142] Specifically, sub-step S2-1: Dynamic calculation of the temperature difference between the sea and air interface, input data: sea surface temperature (SST) measured in real time by ocean buoys; air temperature 2 meters above the sea surface (Tair); calculate the sea-air temperature difference: ; Thermodynamic state is divided according to the positive and negative value of ΔT (ΔT>0 is dominated by evaporation, ΔT≤0 is dominated by cooling); Output product: Sea-temperature difference state table (including ΔT value and thermodynamic state mark of each floating point); Product use: as input of sub-step S2-2.

[0143] Sub-step S2-2: Calculate the critical humidity threshold. Input data: the sea temperature difference state table generated in sub-step S2-1; calculate the critical relative humidity threshold under different ΔT: when ΔT>0, ; When ΔT≤0, ; Output product: dynamic critical humidity threshold table (spatial resolution 1 km, generated by spatial interpolation of buoy data); Product use: as input for sub-step S2-3.

[0144] Sub-step S2-3: Visible light edge sharpness fusion, input data: visible light texture feature map generated in step 201 (including edge sharpness quantization value); dynamic critical humidity threshold table generated above; weighted superposition of edge sharpness quantization value and critical humidity threshold: if edge sharpness ≥ 0.7, then threshold Lower by 5%; if the edge sharpness is <0.7, the threshold Maintain the original value; output product: thermodynamic criterion matrix of fused edge features (resolution of 1 km, including the corrected critical humidity threshold); product use: as input for sub-step S2-4.

[0145] Sub-step S2-4: Binarization fog area determination, input data: thermodynamic criterion matrix of fused edge features generated in sub-step; real-time relative humidity field (resolution 1 km) provided by ERA5 reanalysis data; pixel-by-pixel comparison of relative humidity with the corrected critical threshold: when RH≥ When RH< When , it is determined to be a non-fog area (marked as 0); output product: thermodynamic phase change criterion matrix (resolution of 1 km, binary fog area distribution); product use: as input to step 203, participate in multi-source data fusion;

[0146] It should be noted that the following is a specific implementation case of the thermodynamic phase change criterion modeling in the East China Sea on May 10, 2025 (focusing on step 202):

[0147] Input data: SST = 290K measured at buoy A (E122.5°, N30.2°) in Zhoushan waters, and T_air = 288K at 2 meters above the sea surface;

[0148] calculate (evaporation-dominated state), mark the thermodynamic state as "evaporation-dominated";

[0149] Output: Generates a table of sea-temperature differences for 20 buoys, with a ΔT range of [-1K, +3K] and spatially interpolated to form a 1km resolution grid.

[0150] Dynamic critical humidity threshold calculation, when buoy A is located at grid point ΔT=+2K, apply the formula ;

[0151] When ΔT=-0.5K near buoy B (E123.0°, N30.5°), ;

[0152] A 1km resolution threshold map was generated through Kriging interpolation, and the threshold distribution in the Zhoushan port area was 83.2-85.6%;

[0153] Visible light edge feature fusion, extracting the pixel edge sharpness value of 0.75 in the visible light texture feature map (E122.8°, N30.3°);

[0154] Because the sharpness is ≥0.7, this point A decrease of 5% from the interpolated baseline value of 84.3% to 79.8%;

[0155] The adjacent area (E122.9°, N30.4°) with edge sharpness <0.7 maintained the original threshold of 85.1%;

[0156] Binary fog area determination, calling ERA5 real-time relative humidity field, Zhoushan Port offshore (E122.7°, N30.25°) RH=82%;

[0157] After comparison correction =79.8%, meeting RH≥ Condition, marked as fog area (value 1);

[0158] At a certain point off the Yangtze River Estuary (E123.2°, N31.0°), RH = 78%, which is lower than the threshold of 82.1%, and is marked as a non-fog area (value 0);

[0159] Verified by AIS ship visibility reports, the accuracy of the judgment matrix in the Zhoushan port area reached 89%; the dynamic adjustment of the critical threshold increased the recognition rate of fog areas in the northern East China Sea by 12% (compared with the static threshold method); the edge sharpness fusion mechanism successfully eliminated three misjudgment areas caused by cloud shadows.

[0160] 203. Multi-source data fusion and fog area identification. Input data: infrared fog area probability map generated in step 201; thermodynamic phase change criterion matrix generated in step 202; microwave radiometer data: 36.5 GHz vertically polarized brightness temperature data from the AMSR2 sensor of the GCOM-W1 satellite;

[0161] Liquid water paths are inverted based on microwave brightness temperature data. The infrared fog area probability map, thermodynamic phase change criterion matrix, and liquid water path threshold are logically fused to generate a multi-source fusion fog area identification map (with a resolution of 1 km and a rasterized binary distribution, where 0 indicates non-fog and 1 indicates fog).

[0162] Specifically, sub-step S3-1: preprocessing of microwave brightness temperature data, input data: 36.5 GHz vertically polarized brightness temperature data collected by the AMSR2 sensor of the GCOM-W1 satellite (original resolution 25 km); radiometric calibration and geolocation correction of the brightness temperature data, and removal of outliers affected by abnormal sea surface roughness; output product: corrected microwave brightness temperature map (resolution 25 km, valid data area after removal of outliers); product use: as input for sub-step S3-2.

[0163] Sub-step S3-2: Liquid water path physical inversion, input data: the corrected microwave brightness temperature map generated in sub-step S3-1; based on the microwave radiation transmission model, invert the liquid water path (LWP) and obtain the liquid water path distribution map (resolution 25 km, mark LWP ≥ 0.15 kg / m 2 product use: as input to sub-step S3-3.

[0164] Sub-step S3-3: Multi-source logical fusion operation, input data: infrared fog area probability map generated in step 201 (resolution 2 km); thermodynamic phase change criterion matrix generated in step 202 (resolution 1 km); liquid water path distribution map generated in sub-step S3-2 (resolution 25 km); resample the liquid water path distribution map to 1 km resolution through bilinear interpolation; perform the infrared fog area probability map (P ≥ 0.6), thermodynamic criterion matrix (value 1) and LWP ≥ 0.15 kg / m 2 Perform logical AND operation on the area; output product: preliminary screening and fusion fog area distribution map (resolution 1 km, binary marking of potential fog areas); product use: as input of sub-step S3-4.

[0165] Sub-step S3-4: Spatial consistency check, input data: the initial screening fusion fog area distribution map generated by sub-step S3-3; the visible light texture feature map (resolution 500 meters) generated in step 201; in the initial screening fog area, if the edge sharpness in the visible light texture feature map is ≥0.7, it is determined to be a reliable fog area; eliminate the false detection areas with edge sharpness <0.7, and generate a multi-source fusion fog area identification map (resolution 1 kilometer, rasterized binary distribution, 0 / 1 represents non-fog / fog); product use: as input to step 204, driving fluid mechanics diffusion prediction.

[0166] It should be noted that the following is a specific implementation case of multi-source data fusion and fog area identification in the East China Sea region on May 10, 2025 (focusing on step 203):

[0167] Input data: 36.5 GHz vertically polarized brightness temperature data (E125°-128°, N30°-32°) acquired by the AMSR2 sensor on the GCOM-W1 satellite, with a native resolution of 25 km. A single pixel (E126.5°, N31.2°) exhibited an anomaly in the native brightness temperature (TB = 280 K, with the adjacent region's mean value being 160 K ± 10 K), which was determined to be an interference signal caused by abnormal sea surface roughness.

[0168] After radiometric calibration, the brightness temperature data were spatially aligned with the ERA5 wind field data through geographic correction, and outliers with a standard deviation of more than 3σ were removed. After processing, the effective data coverage rate reached 98.7%;

[0169] Based on the microwave radiation transmission model. In a certain sea area (E127.3°, N30.8°), LWP=0.75kg / m 2 , exceeding the threshold of 0.15kg / m 2 ;

[0170] The liquid water path distribution map shows that there is LWP>0.5kg / m in the southeast of Zhoushan Islands. 2In the high-value area, the spatial overlap with the area where the buoy's measured visibility is <500m is 85%;

[0171] The probability value P=0.8 of a certain grid (E126°-128°, N30.5°-31.5°) in the infrared fog area probability map (2km resolution) is improved to 1km resolution through bicubic interpolation;

[0172] The thermodynamic criterion matrix shows that 80% of the pixels in the area are marked as fog (value 1). After bilinear interpolation of the liquid water path distribution, 85% of the area has LWP ≥ 0.15 kg / m 2 ;

[0173] After logical AND operation, the initial screening fog area is generated: the area that meets the three conditions of P ≥ 0.6, thermodynamic criterion = 1, and LWP ≥ 0.15 accounts for 72% of the total area of ​​the region;

[0174] The edge sharpness values ​​of the visible light texture feature map were extracted in the initial fog screening area (E127.1°, N31.0°): the sharpness of the pixel points in the waters off Zhoushan Port was 0.75 (≥0.7), which was determined to be a reliable fog area; the sharpness of a certain area in the Yangtze River Estuary was 0.6 (<0.7), and combined with the AIS ship report, it was confirmed that the visibility in this area was >1km, so it was eliminated;

[0175] The final multi-source fusion fog area identification map has a resolution of 1km. Compared with the vertical profile observations of the CALIPSO satellite, the fog area identification accuracy rate reaches 89%, and the false detection rate is reduced to 7%.

[0176] 204. Fluid mechanics diffusion trajectory prediction, input data: multi-source fusion fog area identification map generated in step 203; 850hPa wind field data and boundary layer height data provided by numerical weather forecast;

[0177] The horizontal diffusion rate and direction of the fog area are calculated based on fluid mechanics equations, and the movement trajectory for the next three hours is simulated to generate a fog band diffusion vector field (including the spatiotemporal distribution of diffusion speed and direction, with speed units in m / s).

[0178] Specifically, sub-step S4-1: construct the initial field of the fog area, input data: the multi-source fusion fog area identification map generated in step 203 (resolution 1 km, binary fog area distribution); convert the fog area identification map into a rasterized concentration field (fog area is marked as 1, non-fog area is marked as 0), and generate the initial concentration field of the fog area (resolution 1 km, rasterized concentration value 0 / 1); product use: as input for sub-step S4-2.

[0179] Sub-step S4-2: Dynamic correction of the wind field. Input data: 850hPa wind field data (u / v components, time resolution of 1 hour) provided by ECMWF numerical weather forecast; boundary layer height data (derived from CALIPSO satellite vertical profile inversion); adjust the wind field intensity according to the boundary layer height: for every 100-meter decrease in boundary layer height, the wind field speed decreases by 5%; generate a corrected wind field vector diagram (including horizontal wind speed and direction, resolution of 1 km, wind speed unit is m / s); product use: as input to sub-step S4-3.

[0180] Sub-step S4-3: Advection-diffusion trajectory simulation, input data: initial concentration field of the fog area generated by sub-step S4-1; corrected wind field vector diagram generated by sub-step S4-2; based on the advection-diffusion equation, with a time step of 10 minutes, iteratively calculate the movement trajectory of the fog area along the wind field in the next 3 hours; record the front position of the fog area at each time step, and generate a fog area trajectory sequence containing the coordinates of the fog area boundary every 10 minutes); product use: as input for sub-step S4-4.

[0181] Sub-step S4-4: Vector field generation and verification. Input data: fog area trajectory sequence generated in sub-step S4-3. Calculate the fog area displacement at adjacent time steps and extract the diffusion speed and direction. Eliminate abnormal trajectory points caused by terrain occlusion and generate a fog band diffusion vector field (resolution 1 km, including speed magnitude and direction angle). Product use: As input to step 205, generate a comprehensive forecast warning map.

[0182] It should be noted that the following is a specific implementation case of the prediction of the fog diffusion trajectory in the East China Sea region on May 10, 2025 (focusing on step 204):

[0183] Input data: The multi-source fusion fog area identification map generated in step 203 shows that there is an area of ​​about 12,000 km in the central part of the East China Sea. 2 The binary raster is converted into a concentration field, with fog areas marked as 1 (concentration value 100%) and non-fog areas as 0, forming the initial concentration field grid.

[0184] Key features: The concentration gradient at the edge of the fog area near the Zhoushan Islands (E122.8°, N30.3°) reaches 0.8 / km, reflecting a clear diffusion interface.

[0185] ECMWF provides an initial wind speed of 850hPa (southeast wind) of 8m / s, with a temporal resolution of 1 hour. Combined with boundary layer height data retrieved from the CALIPSO satellite, the boundary layer height in the central East China Sea was 800 meters at 08:00 on the same day, a decrease of 120 meters from the previous period.

[0186] Corrected according to the "5% attenuation for every 100 meters" rule: The wind direction was adjusted to 15° south-east. A 1km resolution corrected wind field vector diagram was generated, and the wind speed gradient off Zhoushan was 7.2-7.8m / s.

[0187] Based on the advection-diffusion equation, an iterative calculation with a time step of 10 minutes was performed. The fog front was initially located at E125.2°, N31.0°. Driven by a wind speed of 7.6 m / s in the first 10 minutes, it displaced 4.56 km northwestward.

[0188] After 18 iterations (3 hours), the coordinate sequence of the fog area boundary was recorded. At 08:50, the fog front arrived 30 km outside Zhoushan Port (E122.5°, N30.2°), with a horizontal diffusion rate of 0.85 m / s, which agrees 91% with the liquid water path movement trend measured by the microwave radiometer.

[0189] The displacement of adjacent time steps was extracted: the average diffusion velocity from 08:00 to 09:00 was 0.78 m / s (direction 22° northwest), and five abnormal trajectory points around the Zhoushan Islands were eliminated after terrain correction (the velocity anomaly was greater than 2σ due to island occlusion).

[0190] The output 1km resolution vector field shows that the fog band in the central East China Sea spreads fan-shaped, with the maximum spreading velocity at E124.5°, N30.8° (0.92m / s) and the minimum velocity at the mouth of Hangzhou Bay (0.65m / s). Field measurements conducted by the Zhoushan VTS system show that the positioning error at the leading edge of the fog zone at 09:00 was less than 1.5km.

[0191] 205. Comprehensive prediction and alarm generation, input data: multi-source fusion fog area identification map generated in step 203; fog band diffusion vector field generated in step 204;

[0192] The multi-source fusion fog area identification map and the fog band diffusion vector field are superimposed to generate time series prediction results. When the fog area is predicted to cover the port or waterway with a coverage probability of ≥70%, a graded alarm signal is triggered and a comprehensive sea fog forecast alarm map (including geographic coordinates, impact time and alarm level) is generated.

[0193] Specifically, sub-step S5-1: spatiotemporal superposition and trajectory interpolation, input data: multi-source fusion fog area identification map generated in step 203 (resolution 1 km, binary fog area distribution); fog band diffusion vector field generated in step 204 (resolution 1 km, including speed and direction); align the fog area identification map and the diffusion vector field in time and space, and interpolate the front position of the fog area at intervals of 10 minutes; generate a time series fog area prediction grid set (including the fog area distribution every 10 minutes in the next 3 hours, with a time resolution of 10 minutes and a spatial resolution of 1 km); product use: as input for sub-step S5-2.

[0194] Sub-step S5-2: Calculation of coverage probability of key areas, input data: time series fog area prediction grid set generated by sub-step S5-1; preset port and waterway geographic information database (including coordinate range and priority weight); for each port / waterway area, count the proportion of prediction grids covered by fog area and calculate the regional coverage probability; when the coverage probability of the same area for three consecutive time steps (30 minutes) is ≥70%, it is marked as a high-risk area; output product: key area coverage probability table (including area name, coordinates, probability value and risk level); product use: as input for sub-step S5-3.

[0195] Sub-step S5-3: Dynamic graded alarm triggering, input data: key area coverage probability table generated by sub-step S5-2; set alarm threshold according to risk level: yellow alarm: 70% ≤ probability < 85%; orange alarm: 85% ≤ probability < 95%; red alarm: probability ≥ 95%; combine regional priority weights (international waterway weight × 1.2, local port × 1.0) to generate a graded alarm instruction set (including alarm area, trigger time, level and recommended measures); product use: as input for sub-step S5-4.

[0196] Sub-step S5-4: Multimodal warning map synthesis, input data: time series fog area prediction grid set generated by sub-step S5-1; graded warning instruction set generated by sub-step S5-3; superimpose the prediction grid set and warning instruction on the electronic nautical chart base map, mark the fog area front time series and warning area; generate a standardized format of sea fog comprehensive prediction warning map (supports GeoTIFF / Shapefile format, including geographic coordinates, impact time, warning level and visualization layer); product use: as the final output result, transmitted to the shipping management terminal.

[0197] It should be noted that the following is a specific implementation case of comprehensive forecasting and warning generation in the East China Sea region on May 10, 2025 (focusing on step 205):

[0198] Input data: Multi-source fusion fog area identification map shows that the area off Zhoushan Port is approximately 1,200 km 2 The diffusion vector field shows a southeasterly wind of 7.8 m / s and a diffusion rate of 0.85 m / s in the banded fog area (E122.5°-123.8°, N29.8°-30.5°).

[0199] After spatial and temporal alignment, bicubic interpolation was used to generate the fog boundary coordinates every 10 minutes for the next three hours. At 08:50, the fog front was predicted to arrive at E122.3°, N30.2°, shortening the distance to the main channel of Zhoushan Port to 18 km.

[0200] The preset database includes the coordinate ranges of six key ports, including Zhoushan Port (E122.1°, N30.0°) and Yangshan Port (E122.1°, N30.6°);

[0201] Statistics for the Zhoushan Port area (5×5 km grid): the coverage probabilities for three consecutive time periods between 08:50 and 09:20 were 72%, 78%, and 83%, respectively;

[0202] Applying priority weights: Zhoushan Port as an international shipping hub (weight × 1.2), the calculated weighted coverage probability reaches 89.6%;

[0203] According to the threshold setting: an orange alert (85% ≤ probability < 95%) is triggered at 09:10. Recommended measures include initiating Level 2 navigation control and issuing hourly visibility advisories.

[0204] The probability of consecutive times in the Yangshan Port area is 69%, 75%, and 82%. Because it does not meet the standard of ≥70% for three consecutive times, only a single yellow alarm is triggered;

[0205] Overlay elements: 1 km resolution fog area prediction raster (GeoTIFF format), warning area polygon (Shapefile), port thermal layer;

[0206] Visualization: A red warning zone (15 km radius) is marked in the Zhoushan Port area, and the timeline shows the expansion animation of the fog zone from 09:00 to 12:00.

[0207] 206. Dynamic parameter calibration and feedback optimization based on real-time observations. Input data: real-time visibility observation data transmitted back by ocean buoys (updated at the minute level); comprehensive sea fog forecast warning map generated in step 205; multi-source fusion fog area identification map generated in step 203;

[0208] Dynamic error detection: Compare the fog area range in the forecast warning map with the actual visibility data of the buoy, and calculate the regional forecast error (error = predicted coverage - measured coverage);

[0209] If the absolute value of the prediction error for three consecutive time steps (30 minutes) is greater than 15%, calibration is triggered: the critical humidity threshold of the thermodynamic phase change criterion in step S2 is adjusted ( ±2%); correct the wind field attenuation coefficient in step S4 (adjust the attenuation rate per 100 meters of boundary layer height from 5% to 4% or 6%);

[0210] Generate optimized model parameter set (including the corrected critical humidity threshold and wind field attenuation coefficient);

[0211] Product usage: Real-time feedback to steps 202 and 204, replacing the original fixed parameters and driving the model calculation of the next prediction cycle.

[0212] It should be noted that the following is a specific implementation case of dynamic parameter calibration and feedback optimization in the East China Sea region on May 10, 2025 (focusing on step 206):

[0213] Input data: The coverage rate of the measured visibility < 1 km at buoy A (E122.5°, N30.2°) outside Zhoushan Port from 08:30 to 09:00 is 65%, while the forecast warning map generated in step 205 shows that the forecast coverage rate for this area is 85%;

[0214] The absolute values ​​of the errors for three consecutive times (08:40 / 08:50 / 09:00) were 18%, 22%, and 19%, respectively, all exceeding the 15% threshold;

[0215] Thermodynamic Model Calibration: Critical Humidity Threshold The humidity determination criteria for the Zhoushan Port area (E122°-123.5°, N29.8°-30.5°) have been relaxed, allowing the low cloud areas that were previously misidentified as fog to be corrected.

[0216] Fluid dynamics correction: CALIPSO satellite monitoring indicates that the boundary layer height has dropped from 800 to 650 meters. Based on the "5% attenuation for every 100 meters drop" rule, the wind attenuation coefficient has been adjusted from 5% to 6%. The initial wind speed of 8 m / s in a certain sea area (E125.3°, N31.1°) has been corrected to 8 × (1-150 / 100 × 6%) = 7.28 m / s.

[0217] After the parameter update, the forecast coverage rate for 09:10-09:40 outside Zhoushan Port dropped from 85% to 68%, and the error from the buoy's measured value of 63% was reduced to 5%.

[0218] After the wind field correction in the Yangshan Port area (E122.1°, N30.6°), the predicted error of fog diffusion speed decreased from 0.25m / s to 0.12m / s, and the matching degree with the measured AIS ship trajectory increased to 91%.

[0219] After parameter adjustment, three low cloud areas (E123.8°, N30.7°, etc.) that were originally misjudged were eliminated, with an area of ​​120km. 2 The accuracy of locating the edge of the actual fog zone in the Yangtze River Estuary has been improved, and the visible light texture feature map (500 meters) shows that the proportion of reliable fog areas with a sharpness greater than 0.7 has increased from 78% to 85%.

[0220] In the embodiment of the present invention, the three-dimensional data of visible light texture features, infrared brightness temperature difference probability, and microwave liquid water path are integrated to break through the limitation of a single sensor. The three-layer decomposition of the visible light image is realized through the Daubechies-4 wavelet basis, and the high-frequency energy is extracted to quantify the edge sharpness of the fog area, forming a spatial scale complementarity with the infrared fog area probability map and the microwave inversion liquid water path; when the visible light texture sharpness is ≥0.7, the thermodynamic threshold is set. The forecast error of fog diffusion speed was reduced from 0.25m / s to 0.12m / s after correction based on CALIPSO satellite measured data. When the error between the buoy's measured visibility and the predicted coverage rate was greater than 15% for 30 consecutive minutes, the forecast error was automatically adjusted. The system uses the threshold and wind field attenuation coefficient to reduce the prediction error of Zhoushan Port from 22% to 5%; through multi-source data resampling and interpolation, it ultimately outputs a 1 km resolution prediction result, which is five times more accurate than traditional satellite products; the dynamic parameter calibration mechanism achieves minute-level updates and supports the issuance of hourly visibility advisories, meeting the real-time decision-making needs of international waterways; it outputs standardized GeoTIFF / Shapefile format products that can be directly superimposed on electronic nautical chart systems, reducing manual interpretation time by approximately 40%; when the orange / red alarm is triggered, it can recommend that ships adjust their speed (when the fog diffusion rate is 0.85 m / s, the recommended speed is ≤10 knots), reducing fuel consumption by approximately 15%; it deeply couples the thermodynamic phase change criterion with the fluid mechanics diffusion equation to achieve cross-scale simulation from microscopic droplet formation to macroscopic fog movement; it integrates Fengyun-4A, Himawari-8 (infrared), and GCOM-W1 (microwave) satellite data, combined with CALIPSO lidar vertical detection, to construct a three-dimensional observation network, representing the current forefront of marine meteorological monitoring technology.

[0221] When all other conditions are the same, a third embodiment of a high-precision sea fog image recognition and prediction method based on ensemble learning includes the following features, which are basically the same as the first embodiment, except that:

[0222] The improved DeepLabV3+ is used to perform semantic segmentation of fog areas based on visible light satellite remote sensing images to obtain a fog area probability map, and the final fog area report is obtained based on the sea fog comprehensive forecast warning map and the fog area probability map.

[0223] Reference Figure 5The image branch architecture requires an improved DeepLabV3+ model. Its backbone network adopts the EfficientNet-B5 architecture, which boasts efficient feature extraction capabilities, reducing computational effort while improving feature representation. An adaptive dilated convolution module is added to the decoder layer, with a dilated ratio dynamic range of (2, 6). By dynamically adjusting the dilated ratio, the model can capture feature information at different scales. A weighted combination of DiceLoss and FocalLoss is used as the loss function, with a weight coefficient of α = 0.7.

[0224] During the DeepLabV3+ model training process, the present invention continuously adjusts the model parameters to minimize the loss function until the fog area probability map output by the model is The coefficient (mean of intersection and union ratio) is ≥ 0.85. (MeanIntersectionoverUnion) is the prediction accuracy indicator of foggy area pixels, and is calculated as follows:

[0225] ;

[0226] in is the number of categories (fog / non-fog), is the number of correctly predicted fog area pixels.

[0227] It should be noted that this embodiment uses the visible light remote sensing image (resolution 500 meters) acquired by the FY-4A satellite VISSR sensor as input, performs fog area semantic segmentation through the improved DeepLabV3+ model, and generates the final fog area report by combining the thermodynamic criterion matrix, infrared fog area probability map, and diffusion prediction results.

[0228] Improve DeepLabV3+ model architecture and training, and optimize the backbone network:

[0229] EfficientNet-B5 is used as the backbone network, and its compound scaling factor (depth 1.2, width 2.2, resolution 456) significantly improves the feature extraction capability.

[0230] Adaptive dilated convolution module: An ASPP module (with a dilated ratio range of 2-6) that dynamically adjusts the dilation ratio is embedded in the decoder. Learnable parameters are used to automatically select the optimal dilation ratio based on the spatial distribution of the input feature map. A small dilation ratio (r=2) is used to capture details at the edges of foggy areas (such as coastlines), while a large dilation ratio (r=6) is used to extract global features in large foggy areas.

[0231] Loss function design: A combination of DiceLoss (weight 0.3) and FocalLoss (weight 0.7) is used to address the imbalance in the number of pixels between foggy and non-foggy areas. FocalLoss, with a γ of 2, focuses on difficult-to-classify pixels (those confused by light fog and low cloud).

[0232] Training data: 10,000 labeled foggy images (500-meter resolution) were used, including 8,000 training images and 2,000 validation images. Data augmentation included random rotation (±30°), horizontal flipping, and Gaussian noise injection (σ=0.05).

[0233] Training results: After the model converges, the test set Reaching 0.87, including fog area The predicted probability of fog area for a certain sample (coordinate 120.5°E, 32°N) is 0.92, and the accuracy of the measured visibility verification is 91%.

[0234] Fog area probability map generation and multi-source fusion, input and processing flow:

[0235] Visible light image preprocessing: The input image is normalized to [0, 1] and the contrast is enhanced by histogram equalization.

[0236] The model outputs a fog area probability map (resolution 500 meters, pixel value 0-1). For example, the fog area probability of a certain sea area (coordinates 121°E, 35°N) is 0.88, indicating that there is an 88% probability of sea fog in the area.

[0237] Multi-source data fusion rules:

[0238] Spatial alignment: The fog area probability map, thermodynamic criterion matrix (1 km), and infrared fog area probability map (1 km interpolation) are unified to a 1 km grid, and bilinear interpolation is used to ensure spatial consistency.

[0239] Logical fusion conditions:

[0240] Fog area probability map threshold: ≥0.7 (if it is lower than this value, it will be judged as non-fog even if other conditions are met).

[0241] Thermodynamic criterion matrix: must be 1 (meeting humidity and temperature conditions).

[0242] Infrared fog area probability map: ≥0.6 (probability ≥60%).

[0243] Example: The parameters of a grid (121.2°E, 35.5°N) are:

[0244] Fog area probability = 0.88, thermodynamic criterion = 1, infrared probability = 0.75 → the fusion result is fog area.

[0245] If the infrared probability = 0.55 (does not meet the threshold), it is judged as a non-fog area.

[0246] Final fog area report generation, time-space weighting and alarm triggering:

[0247] Time series prediction overlay: The 3-hour fog area position predicted by the diffusion vector field (the Qingdao Port waterway is expected to have a coverage probability of 75% in T+3 hours) is overlaid with the current fusion result to generate a spatiotemporal probability distribution map.

[0248] Alarm rules:

[0249] Red alert: Coverage probability ≥ 90% and impact duration ≥ 6 hours (e.g., persistent dense fog in a certain channel in the Yellow Sea).

[0250] Orange alert: Coverage probability ≥ 70% and impact duration 3-6 hours (Example: Qingdao Port waterway, probability 75%, triggering an orange alert).

[0251] Output format:

[0252] Geographical range: 120.8°E-121.6°E, 36.5°N-37.1°N;

[0253] Impact time: 2025-05-2008:00 to 11:00;

[0254] Alarm level: orange (level 2);

[0255] Recommended measures: restrict vessel entry and exit and activate fog navigation system.

[0256] The above describes the high-precision sea fog image recognition and prediction method based on ensemble learning in the embodiment of the present invention. The following describes the high-precision sea fog image recognition and prediction device based on ensemble learning in the embodiment of the present invention. Figure 3In one embodiment of the present invention, a high-precision sea fog image recognition and prediction device based on ensemble learning includes: an acquisition module 301 for acquiring visible light satellite remote sensing images and infrared band data, performing multi-scale texture analysis on the visible light satellite remote sensing images, extracting texture features of fog area edges, generating a visible light texture feature map, calculating brightness temperature difference for the infrared band data, and generating an infrared fog area probability map in combination with the sea surface temperature; a processing module 302 for acquiring sea surface temperature and air temperature data measured by ocean buoys, obtaining a critical humidity threshold, and generating a thermodynamic phase change criterion matrix in combination with the edge sharpness quantization value in the visible light texture feature map; and an identification module 303 for acquiring microwave radiation counts. According to the microwave brightness temperature data, the liquid water path is inverted, and the infrared fog area probability map, the thermodynamic phase change criterion matrix and the liquid water path threshold are logically integrated to generate a multi-source fusion fog area identification map; the simulation module 304 is used to obtain the horizontal diffusion rate and direction of the fog area based on the initial distribution of the fog area provided by the multi-source fusion fog area identification map, combined with the wind field data and the boundary layer height data, simulate the future motion trajectory, and generate a fog band diffusion vector field; the allocation module 305 is used to superimpose the multi-source fusion fog area identification map and the fog band diffusion vector field to generate a time series prediction result. When it is predicted that the fog area covers the port or waterway and the coverage probability is ≥ the preset value, a graded alarm signal is triggered to generate a sea fog comprehensive prediction alarm map.

[0257] In an embodiment of the present invention, the thermodynamic phase change conditions are converted into matrix operations to avoid the discretization error of traditional threshold comparison and improve real-time processing efficiency; microwave liquid water path inversion can capture tiny water droplets that are difficult to identify with visible light / infrared, thereby improving the accuracy of identifying low-visibility fog areas; simulation based on diffusion vector fields can realize the prediction of fog area trajectories in the next 6-12 hours, buying valuable time for channel control; through vector field superposition, the needs of refined port-scale management are met.

[0258] above Figure 3 The high-precision sea fog image recognition and prediction device based on ensemble learning in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The high-precision sea fog image recognition and prediction device based on ensemble learning in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0259] Figure 4 This is a schematic diagram of the structure of a high-precision sea fog image recognition and prediction device based on ensemble learning provided by an embodiment of the present invention. The high-precision sea fog image recognition and prediction device 400 based on ensemble learning may have relatively large differences due to different configurations or performances. The device 400 includes a transmitter 401, a receiver 402, and a processor 403. The processor 403 can also be a controller. Figure 4denoted as “controller / processor 403 ”. Optionally, the device 400 may further include a modem processor 405 , wherein the modem processor 405 may include an encoder 406 , a modulator 407 , a decoder 408 , and a demodulator 409 .

[0260] In one example, transmitter 401 conditions (e.g., performs analog-to-analog conversion, filtering, amplification, and frequency upconversion) the output samples and generates an uplink signal, which is transmitted via an antenna to an access network device. On the downlink, the antenna receives the downlink signal transmitted by the access network device. Receiver 402 conditions (e.g., performs filtering, amplification, frequency downconversion, and digitization) the signal received from the antenna and provides input samples. Within modem processor 405, encoder 406 receives traffic data and signaling messages to be transmitted on the uplink and processes them (e.g., formats, encodes, and interleaves them). Modulator 407 further processes (e.g., performs symbol mapping and modulation) the encoded traffic data and signaling messages and provides output samples. Demodulator 409 processes (e.g., demodulates) the input samples and provides symbol estimates. Decoder 408 processes (e.g., deinterleaves and decodes) the symbol estimates and provides decoded data and signaling messages for transmission to device 400. The encoder 406, modulator 407, demodulator 409, and decoder 408 can be implemented by the combined modem processor 405. These units perform processing based on the radio access technology (e.g., LTE and other evolved system access technologies) used by the radio access network. It should be noted that when the device 400 does not include the modem processor 405, the above functions of the modem processor 405 can also be performed by the processor 403.

[0261] Processor 403 controls and manages the actions of device 400, and is configured to execute the processing performed by device 400 in the above-described embodiments of the present disclosure. For example, processor 403 is also configured to execute the various steps of the sending device or receiving device in the above-described method embodiments, and / or other steps of the technical solutions described in the embodiments of the present disclosure.

[0262] Furthermore, the device 400 may further include a memory 404 , and the memory 404 is used to store program codes and data for the device 400 .

[0263] It is understandable that Figure 4 Only a simplified design of the device 400 is shown. In actual applications, the device 400 may include any number of transmitters, receivers, processors, modem processors, memories, etc., and all devices that can implement the embodiments of the present disclosure are within the scope of protection of the embodiments of the present disclosure.

[0264] The present invention also provides a high-precision sea fog image recognition and prediction device based on ensemble learning. The high-precision sea fog image recognition and prediction device based on ensemble learning includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the steps of the high-precision sea fog image recognition and prediction method based on ensemble learning in the above-mentioned embodiments.

[0265] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the high-precision sea fog image recognition and prediction method based on ensemble learning.

[0266] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0267] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0268] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-precision sea fog image recognition and prediction method based on ensemble learning, characterized in that: The high-precision sea fog image recognition and prediction method based on ensemble learning includes: Obtain visible light satellite remote sensing images and infrared band data, perform multi-scale texture analysis based on the visible light satellite remote sensing images, extract texture features of the fog area edge, generate a visible light texture feature map, calculate the brightness temperature difference based on the infrared band data, and generate an infrared fog area probability map in combination with the sea surface temperature; Obtaining sea surface temperature and air temperature data measured by ocean buoys to obtain a critical humidity threshold, and combining this with the edge sharpness quantification value in the visible light texture feature map to generate a thermodynamic phase change criterion matrix; Acquire microwave radiometer data, invert the liquid water path based on the microwave brightness temperature data, perform logical fusion based on the infrared fog area probability map, the thermodynamic phase change criterion matrix and the liquid water path threshold, and generate a multi-source fusion fog area identification map; Based on the initial distribution of the fog area provided by the multi-source fusion fog area identification map, combined with wind field data and boundary layer height data, the horizontal diffusion rate and direction of the fog area are obtained, the future movement trajectory is simulated, and the fog belt diffusion vector field is generated; The multi-source fusion fog area identification map and the fog band diffusion vector field are superimposed to generate a time series prediction result. When the fog area is predicted to cover the port or waterway and the coverage probability is ≥ the preset value, a graded alarm signal is triggered to generate a comprehensive sea fog prediction alarm map.

2. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 1 is characterized in that: include: Perform two-dimensional discrete wavelet transform on visible light satellite remote sensing images, decompose them to the third layer, extract the energy value of high-frequency components, and generate wavelet high-frequency energy maps; The sum of high-frequency energy of each pixel is obtained based on the wavelet high-frequency energy map to generate a visible light texture feature map; The brightness temperature difference (BTD) is obtained based on the channel brightness temperature data and the sea surface temperature (SST). The BTD is piecewise linearly corrected according to the SST to generate a corrected brightness temperature difference map. The brightness temperature difference (BTD) is mapped to the probability of fog presence through the S-type function to generate an infrared fog area probability map.

3. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 2 is characterized in that: include: Based on the real-time measurement of sea surface temperature (SST) and air temperature 2 meters above the sea surface by ocean buoys , and get the sea temperature difference: ； The thermodynamic state is divided according to the positive and negative nature of ΔT, and the sea-temperature difference state table is obtained; According to the sea-temperature difference state table, the critical relative humidity threshold under different ΔT is obtained: When ΔT>0, ; When ΔT≤0, ; Get the dynamic critical humidity threshold table; According to the visible light texture feature map and the dynamic critical humidity threshold table, the edge sharpness quantization value and the critical relative humidity threshold are weighted superimposed: If the edge sharpness is ≥ 0.7, then Down 5%; If the edge sharpness is <0.7, then Maintain the original value; Obtain the thermodynamic criterion matrix of fused edge features; According to the thermodynamic criterion matrix of the fused edge features and the real-time relative humidity field, the real-time relative humidity RH is compared with the corrected critical relative humidity threshold pixel by pixel. : When RH≥ When , it is determined to be a fog area; When RH< When , it is determined to be a non-fog area; The thermodynamic phase transition criterion matrix is ​​obtained.

4. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 3 is characterized in that: include: The brightness temperature data were radiometrically calibrated and geo-positioned to remove outliers affected by sea surface roughness anomalies, thus obtaining a corrected microwave brightness temperature map. According to the corrected microwave brightness temperature image and the microwave radiation transmission model, the liquid water path is inverted to obtain the liquid water path distribution map; The liquid water path distribution map was resampled to 1 km resolution through bilinear interpolation and processed according to the infrared fog area probability map, thermodynamic criterion matrix and liquid water path distribution map to obtain the initial screening fusion fog area distribution map; In the initial screening fog area, if the edge sharpness in the visible light texture feature map is ≥0.7, it is determined to be a reliable fog area. The false detection areas with edge sharpness <0.7 are eliminated to generate a multi-source fusion fog area identification map.

5. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 4 is characterized in that: include: The multi-source fusion fog area identification map is converted into a rasterized concentration field to generate the initial concentration field of the fog area. The wind field intensity is adjusted according to the boundary layer height to generate a corrected wind field vector map. Based on the initial concentration field of the fog area and the corrected wind field vector diagram, the movement trajectory of the fog area along the wind field in the next 3 hours is iteratively calculated with a time step of 10 minutes. The front position of the fog area at each time step is recorded to generate a fog area trajectory sequence. The fog area displacement at adjacent time steps is obtained based on the fog area trajectory sequence, the diffusion speed and direction are extracted, the abnormal trajectory points caused by terrain occlusion are eliminated, and the fog belt diffusion vector field is generated.

6. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 5 is characterized in that: include: The multi-source fusion fog area identification map is temporally and spatially aligned with the fog band diffusion vector field, and the fog area front position is interpolated at 10-minute intervals to generate a time series fog area prediction grid set. Based on the time series fog area prediction grid set and the preset port and waterway geographic information database, for each port / waterway area, the proportion of the predicted grids covered by fog is counted to obtain the regional coverage probability. When the coverage probability of the same area for three consecutive time steps is ≥70%, it is marked as a high-risk area, and a key area coverage probability table is obtained; Based on the key area coverage probability table, set the alarm threshold according to the risk level, and generate a hierarchical alarm instruction set based on the regional priority weight; The time series fog area prediction grid set and the graded alarm instruction set are superimposed on the electronic nautical chart base map, the fog area front time series and the alarm area are marked, and a comprehensive sea fog prediction and alarm map is generated.

7. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 6 is characterized in that: The following steps are also included: Compare the fog area range in the comprehensive sea fog forecast warning map with the visibility observation data transmitted back by ocean buoys in real time to obtain the regional forecast error, which is calculated as follows: regional forecast error = predicted coverage rate - measured coverage rate; If the absolute value of the prediction error for three consecutive time steps is greater than 15%, calibration is triggered: Adjust the critical humidity threshold of the thermodynamic phase transition criterion; Corrected wind field attenuation coefficient; Generate optimized model parameter set.

8. The high-precision sea fog image recognition and prediction method based on ensemble learning according to claim 1 is characterized in that: Based on the visible light satellite remote sensing image, the improved DeepLabV3+ is used to perform fog area semantic segmentation to obtain the fog area probability map. The final fog area report is obtained based on the sea fog comprehensive prediction warning map and the fog area probability map. The improved DeepLabV3+ model includes: the backbone network adopts the EfficientNet-B5 architecture; the decoder layer adds an adaptive void convolution module, and the void rate dynamic range is set to (2, 6); the loss function is a weighted combination of Dice Loss and Focal Loss, with a weight coefficient α=0.7.

Images