A sea fog forecasting method based on cloud life cycle

Abstract

The present application relates to the technical field of weather forecast, and particularly relates to a sea fog forecasting method based on cloud life cycle, comprising the following steps: carrying out sea fog identification to obtain a binary grid field, and dividing a sea fog single cloud life cycle stage; carrying out multi-scale optical flow calculation on the binary grid field, outputting a sea fog field per-grid optical flow vector field, and extrapolating to obtain a sea fog advection forecasting field per hour in the future 6 hours; constructing a sea fog appearance and disappearance probability forecasting model per hour in the future 1-6 hours; deeply fusing the sea fog advection forecasting field and the sea fog appearance and disappearance probability forecasting model to obtain a sea fog forecasting grid field, and carrying out secondary revision; the present application integrates the cloud life cycle into the whole process of the sea fog 6-hour extrapolation forecasting, solves the inherent defects of the traditional extrapolation method that is only translation and does not appear and disappear, prolongs the effective sea fog forecasting time from 2-3 hours to 6 hours, and the per-hour forecasting can greatly improve the 6-hour forecasting accuracy, and can accurately forecast the sudden new sea fog and the rapidly dissipating sea fog which cannot be captured by the traditional technology.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, specifically to a sea fog forecasting method based on cloud lifecycle. Background Technology

[0002] Sea fog is a hazardous weather phenomenon that occurs within the ocean-atmosphere boundary layer, resulting in horizontal visibility of less than 1 km at sea. It is characterized by its suddenness, rapid formation and dissipation, and wide range of impact. It is a core meteorological factor causing ship collisions at sea, port operation interruptions, near-shore flight diversions / cancellations, and offshore wind power operation and maintenance accidents.

[0003] Among these, hourly short-term forecasts for the next 0-6 hours are a core requirement for marine meteorological operations, as forecasts in this timeframe directly determine the emergency dispatch, risk management, and safety assurance capabilities for maritime operations. However, current sea fog short-term forecasting technology faces significant technical bottlenecks and cannot meet the demands for high-precision operational forecasts.

[0004] Current mainstream short-term sea fog forecasting technologies are mainly divided into two categories, both of which have inherent flaws that cannot be resolved: Traditional extrapolation forecasting techniques (optical flow method, COTREC method, semi-Lagrange extrapolation, etc.): These techniques are based on continuous observations from geostationary satellites / radar, calculate the horizontal motion vector of sea fog, and obtain future forecast results through advection extrapolation. Their advantages include high computational efficiency, fast update frequency, and suitability for short-term forecasting scenarios. However, their core drawback is that they only consider the horizontal translational motion of sea fog, completely ignoring the physical processes of its formation and dissipation (only translation, no formation or dissipation). Sea fog undergoes significant formation, development, and dissipation processes within a 6-hour forecast window, not simply rigid body translation. This results in an effective forecast lead time of only 2-3 hours for these methods. After 3 hours, the forecast error increases sharply, and the ThreatScore (Threshold Success Index) for 6-hour forecasts is generally below 0.35, with false alarm and missed prediction rates exceeding 50%, completely failing to meet operational requirements.

[0005] Numerical model forecasting technology: This type of technology is based on atmospheric dynamic equations and simulates the formation and dissipation process of sea fog through boundary layer parameterization schemes, enabling sea fog forecasts of more than 6 hours. However, its core drawbacks are: long model spin-up times, high sensitivity to initial fields and boundary layer parameterization schemes, and severely insufficient forecasting capability for sudden sea fog events within 0-6 hours, generally resulting in missed predictions and forecast lags; at the same time, numerical models consume large computational resources, with update frequencies mostly between 6-12 hours, making them unsuitable for the operational requirements of high-frequency updates for short-term forecasts, and the critical success index for 6-hour forecasts is generally below 0.45.

[0006] Therefore, a cloud lifecycle-based sea fog forecasting method is needed that can provide hourly short-term forecasts of sea fog for the next 0-6 hours. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a sea fog forecasting method based on cloud life cycle that can provide hourly short-term forecasts of sea fog for the next 0-6 hours.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A sea fog forecasting method based on cloud lifecycle includes: S1. Based on multi-source data, sea fog is identified to obtain a binary grid field, and the life cycle stages of individual sea fog clouds are divided, including: initial stage, development stage, mature stage and dissipation stage; typical historical sea fog processes in the target sea area are selected to construct a sample dataset; S2. The three-layer pyramid LK optical flow method is used to perform multi-scale optical flow calculations on the binary grid field and output the grid-by-grid optical flow vector field of the sea fog field. Lifecycle correction terms: Individual cell area correction factor α, correction factor β, where β is used to determine the persistence probability of grid points currently identified as sea fog. Make corrections; Initial stage: α = 1 + R, where R is the current hourly area growth rate, β = 1.2. Extrapolation involves expanding the range of individual sea fog cells hourly based on α, and adjusting the core area upwards based on β. Development stage: α=1+0.5R, β=1.1. When extrapolating, the sea fog range is expanded hourly according to α, and the core area is adjusted upward according to β. Maturation stage: α=1, β=1.0, the sea fog range remains constant during extrapolation, and it is translated according to the optical flow vector field; Dissipation stage: α=1-S, where S is the current hourly area reduction rate, β=0.8, the sea fog range decreases hourly according to α during extrapolation, and the edge area is adjusted downward according to β. ; Environmental field physical constraints: Introduce sea-temperature difference A, boundary layer height B, and 925hPa relative humidity C for secondary calibration. When A > 2℃, B > 500m, and C < 90%, α × 0.8, β × 0.7; when A < 0℃, B < 300m, and C > 95%, α × 1.1, β × 1.2. Advection extrapolation forecast output: Based on the optical flow vector field, combined with the life cycle correction term and the environmental field constraint term, a semi-Lagrange advection scheme is used to extrapolate the hourly sea fog advection forecast field for the next 6 hours; S3. Using the life cycle of individual sea fog clouds, environmental meteorological characteristics, and geographical and temporal characteristics as inputs, the LightGBM binary classification algorithm is used to construct six sea fog formation and dissipation probability forecast models for each hour from 1 to 6 hours in the future. The models are trained and validated using a sample dataset. The model label is whether the corresponding grid point is sea fog. S4. The hourly sea fog advection forecast field from 1 to 6 hours is deeply fused with the sea fog formation and dissipation probability forecast model to obtain the sea fog forecast grid field, and then the second correction is performed through real-time observation and topographic data. S5. Forecast result verification and operational output.

[0009] Preferably, sea fog identification based on multi-source data further includes: The multi-source data includes: high spatiotemporal resolution geostationary meteorological satellite data, uniformly resampled to 15-minute temporal resolution and 4km spatial resolution, including visible light, near-infrared, and infrared split-window channels; ground observation data, including hourly visibility, air temperature, dew point, wind speed, and wind direction data from coastal national meteorological stations, island automatic stations, and offshore buoys; radiosonde and reanalysis data, including routine 08:00 and 20:00 radiosonde data from coastal radiosonde stations and hourly ERA5 reanalysis data, including sea surface temperature, boundary layer height, inversion intensity, relative humidity, and vertical velocity; and geographic information data, including digital elevation model (DEM) and coastline data. Precise remote sensing identification of sea fog: Based on multi-channel geostationary satellite data, a sea fog identification model is constructed. A multi-channel threshold method is used to distinguish between sea fog, mid-to-high clouds, low clouds, and clear-sky sea surfaces. The core identification thresholds are as follows: Infrared split-window brightness temperature difference <0.5K, mid-to-high cloud cover excluded; The reflectance of the visible light channel is 0.1~0.8, and the ratio of near-infrared to visible light reflectance is <0.8, which can distinguish sea fog from clear sky sea surface and water surface flares; By combining the cloud top brightness temperature threshold and removing convective clouds and mid-to-high clouds, a binary grid field for sea fog identification is finally obtained every 15 minutes, where 1 = sea fog and 0 = non-sea fog. The accuracy is verified by real-world data with ground visibility <1km to ensure that the identification accuracy is ≥90%.

[0010] Preferably, dividing the life cycle stages of sea fog clouds further includes: Based on a continuous 3-hour remote sensing observation sequence of sea fog and synchronous meteorological elements, sea fog cells are divided into: Initial stage: The hourly growth rate of sea fog cell area meets R≥15%; the hourly decrease in cloud top brightness temperature is ≥0.5K; the hourly increase in cloud optical thickness is ≥0.3; the boundary layer inversion intensity continues to increase; the hourly decrease in ground visibility is ≥200m; Development stages: The hourly growth rate of sea fog cell area meets the requirement of 5% ≤ R < 15%; the cloud top brightness temperature change rate is within ±0.3K / hour; the cloud optical thickness continues to increase; the inversion layer thickness continues to increase; and the visibility remains stable below 1km. Mature stage: The hourly change rate of the area of ​​individual sea fog cells meets the condition -5%≤R<5%; the hourly change rates of cloud top brightness temperature, cloud optical thickness, and inversion intensity are all within ±5%; visibility is stable below 500m, and the stable maintenance time is ≥2 hours; During the dissipation phase: the hourly reduction rate of the area of ​​individual sea fog cells is ≥10%; the hourly increase in cloud top brightness temperature is ≥0.8K; the hourly decrease in cloud optical thickness is ≥0.3; the intensity of temperature inversion continues to weaken; and the hourly increase in ground visibility is ≥200m.

[0011] Preferably, the division of the core area and the edge area includes: The region is divided using a single-grid morphological etching method: the internal grid points retained after etching are the core region, and the outer ring grid points after subtracting the core region from the original sea fog cell are the edge region; if the sea fog cell is too small to be completely empty after etching, then the entire sea fog cell is considered as the edge region.

[0012] Preferably, =min(1,β× ), The baseline maintenance probability is determined jointly by the current sea fog identification results and the local environmental field.

[0013] Preferably, typical historical sea fog processes in the target sea area are selected to construct a sample dataset. Each sample includes: satellite sea fog identification sequence for the first 3 hours, sea fog cell cloud life cycle stage labels, synchronous meteorological element field, and hourly sea fog data for the next 6 hours; and is divided into training set, validation set, and test set in a ratio of 8:1:1. The LightGBM binary classification algorithm was used to construct six hourly probability forecast models for the formation and dissipation of sea fog, covering the next 1-6 hours. The models were trained using a training set with the following core hyperparameters: learning rate 0.05, tree depth 8, number of leaf nodes 64, number of iterations 1000, and number of early stopping rounds 50. The hyperparameters were optimized using a validation set, and the model performance was evaluated using a test set to ensure that the AUC value of all six models was ≥0.92.

[0014] Preferably, the life cycle of a single sea fog cloud includes: initial stage marker, development stage marker, mature stage marker, dissipation stage marker, area change rate over the past 3 hours, cloud top brightness temperature change rate, cloud optical thickness change rate, and visibility change rate. Environmental meteorological characteristics include: sea surface temperature difference, sea surface temperature, air temperature at 2m, dew point temperature at 2m, relative humidity at 2m, wind speed at 10m, wind direction at 10m, boundary layer height, inversion intensity, inversion layer thickness, relative humidity at 925hPa, vertical velocity at 925hPa, air temperature at 925hPa, u-wind component at 925hPa, v-wind component at 925hPa, and air temperature-dew point difference; Geographic and temporal characteristics include: distance from the shore, latitude, month, and hour.

[0015] Preferably, the deep fusion of the hourly sea fog advection forecast field (1-6 hours) with the sea fog formation and dissipation probability forecast model to obtain the sea fog forecast grid field further includes: For grid points marked as sea fog in the advection extrapolation field, if the dissipation probability at the corresponding time is ≥70%, the grid point is corrected to not be sea fog. For grid points in the advection extrapolation field that are marked as non-sea fog, if the generation probability at the corresponding time is ≥60%, then the grid point is corrected to sea fog. For grid points with uncertainties between 40% and 60% of the probability of occurrence and dissipation, the original results of the sea fog advection forecast field are maintained; The second revision includes: Real-time observation bias correction: Using the fused sea fog forecast grid field as the background field, the effective visibility observation within a 50km range is used as the observation field. Gaussian weights are constructed according to the distance between the observation point and the grid point to be corrected. The closer the distance, the greater the weight. First, calculate the deviation between the measured visibility at each observation point and the interpolation of the background field at that observation point. Then, perform a weighted average of the observation deviations according to the distance Gaussian weight to obtain the analysis increment of the grid point to be corrected. Finally, superimpose the analysis increment onto the corresponding background grid point to obtain the corrected forecast result. If there are no effective observations within 50km of the grid point to be corrected, retain the original forecast value after fusion. Topography and coastline correction: Based on the DEM digital elevation model and coastline data, sea fog forecast results are removed for land areas with an elevation > 50m; for open ocean areas with a distance > 500km from the shore, secondary calibration is performed in conjunction with sea surface temperature gradient.

[0016] Preferably, the verification of forecast results further includes: The forecast results after secondary correction were quantitatively verified using common verification indicators in the meteorological forecasting industry. The core indicators include: critical success index, hit rate (POD) and false alarm rate (FAR). The results were compared and verified with the forecast results of traditional LK optical flow extrapolation method and ECMWF numerical model.

[0017] Preferably, the operational output further includes: outputting hourly gridded sea fog forecasts for the next 6 hours, visibility graded forecast products, sea fog impact range forecast maps, and warning information for the affected areas; the product update frequency is 15 minutes / time.

[0018] The beneficial effects of this invention are as follows: By systematically integrating cloud lifecycle theory into the entire process of 6-hour extrapolation forecasting of sea fog, it solves the inherent defect of traditional extrapolation methods that "only shift without generating or dissipating," extending the effective forecast lead time of sea fog from 2-3 hours to 6 hours, and significantly improving the accuracy of 6-hour forecasts through hourly forecasting. Independent sample validation shows that the average TS score of the 6-hour forecast of this invention is ≥0.68, which is 112% higher than the traditional optical flow method and 51% higher than numerical models; the hit rate is ≥0.82, and the false alarm rate is ≤0.21, far superior to existing mainstream technologies; through the dual fusion of physical mechanisms and data-driven approaches, the forecast exhibits strong stability and versatility. A dual-driven forecasting framework of "multi-scale optical flow extrapolation + lifecycle generation and dissipation probability" is constructed. This framework retains the high frequency and efficiency advantages of extrapolation while avoiding overfitting issues inherent in purely data-driven models through cloud lifecycle physical mechanisms and environmental field constraints. The quantitative lifecycle classification standard is adaptable to various high-occurrence sea fog areas globally (such as the Yellow Sea, Bohai Sea, East China Sea, South China Sea, west coast of North America, and Sea of ​​Japan in China). Localized deployment can be achieved simply by adjusting the threshold, demonstrating strong versatility. It boasts high computational efficiency, strong operational adaptability, and low deployment costs. Based on geostationary satellite remote sensing data and conventional meteorological observation data, this invention does not rely on high-performance computing clusters. A single server can complete operational rolling forecasts every 15 minutes, with a computation time of ≤30 seconds per forecast, far lower than that of numerical models. It can seamlessly integrate with existing marine meteorological operational systems, providing accurate short-term early warning support for industries such as port shipping, offshore wind power, and near-shore aviation, with broad application prospects. It solves the challenge of forecasting sudden sea fog events, demonstrating significant value in disaster prevention and mitigation. This invention, through a probability model of formation and dissipation, can accurately predict sudden formation and rapid dissipation of sea fog that cannot be captured by traditional technologies, significantly reducing the missed and false alarm rates, and effectively reducing maritime safety accidents caused by sea fog. It has both social and economic value. Attached Figure Description

[0019] Fig. 1 The satellite schematic representation of the life cycle stage characteristics of a sea fog cell based on cloud life cycle in a specific embodiment of the present invention is shown below; (the white semi-transparent area represents the main body of the sea fog and its boundary. Data source: NSMC / CMA Fengyun-4A satellite open-domain true-color composite image (FY-4A AGRI) base map) Fig. 2This diagram illustrates a comparison between a cloud lifecycle-based sea fog forecasting method of this invention and actual conditions, as well as the traditional LK extrapolation method. (Each column of the base map corresponds to the actual time points T0, T0+3h, and T0+6h, respectively; the orange dashed line represents the outline of the traditional LK extrapolation method, which only represents advection displacement; the green solid line represents the outline of the method of this invention, which simultaneously represents translation and formation / dissipation evolution. It is evident that the method of this invention is closer to actual conditions. Data source: NSMC / CMA Fengyun-4A satellite public domain true-color composite image (FY-4A AGRI) base map compilation.) Fig. 3 A color bar chart comparing the 6-hour forecast performance of a cloud lifecycle-based sea fog forecasting method and a traditional method, as shown in a specific embodiment of the present invention. Detailed Implementation

[0020] To explain in detail the technical content, objectives, and effects of the present invention, the following description is provided in conjunction with the embodiments and accompanying drawings.

[0021] Please refer to Figs. 1 to 3 A sea fog forecasting method based on cloud lifecycle, characterized by comprising: Step 1: Construction of Dataset for Sea Fog Identification and Cloud Lifecycle Stage Segmentation This step forms the basis of the invention. Its core is to achieve accurate remote sensing identification of sea fog, establish quantitative standards for dividing the life cycle stages of sea fog clouds, and construct standardized training and validation datasets.

[0022] Multi-source data acquisition: ① High spatiotemporal resolution geostationary meteorological satellite data (including data from my country's Fengyun-4 series geostationary meteorological satellites FY-4A / FY-4B and Japan's Himawari-8 / 9 geostationary meteorological satellites; original observation products were acquired according to the operational temporal resolution of each satellite and uniformly resampled to a 15-minute temporal resolution and a 4km spatial resolution, including visible light, near-infrared, and infrared split-window channels); ② Ground observation data (hourly visibility, air temperature, dew point, wind speed, and wind direction data from coastal national meteorological stations, island automatic stations, and offshore buoys); ③ Radiosonde and reanalysis data (conventional 08:00 and 20:00 radiosonde data from coastal radiosonde stations, and hourly ERA5 reanalysis data, including sea surface temperature, boundary layer height, inversion intensity, relative humidity, and vertical velocity); ④ Geographic information data (Digital Elevation Model (DEM) and shoreline data).

[0023] Precise remote sensing identification of sea fog: Based on multi-channel geostationary satellite data, a sea fog identification model is constructed. A multi-channel threshold method is used to distinguish between sea fog, mid-to-high clouds, low clouds, and clear-sky sea surfaces. The core identification thresholds are as follows: Infrared split window brightness temperature difference (BT10.8μm - BT12.0μm) < 0.5K, mid-to-high cloud cover is excluded; The reflectance of the visible light channel is 0.1~0.8, and the ratio of near-infrared to visible light reflectance is <0.8, which can distinguish sea fog from clear sky sea surface and water surface flares; By combining the cloud top brightness temperature threshold (TBB > 273K) and eliminating convective clouds and mid-to-high clouds, a binary grid field for sea fog identification every 15 minutes is finally obtained (1 = sea fog, 0 = non-sea fog). The accuracy is verified by real-world data with ground visibility < 1km to ensure that the identification accuracy is ≥ 90%.

[0024] Quantitative Classification of Sea Fog Cloud Life Cycle Stages: Based on a continuous 3-hour remote sensing observation sequence of sea fog and synchronous meteorological elements, sea fog cells are divided into four life cycle stages: initial stage, development stage, mature stage, and dissipation stage. Quantifiable and reproducible classification thresholds are established (optimized based on historical case statistics from high-incidence areas of sea fog globally, and can be adjusted to suit different sea areas), as detailed below: Initial stage: The hourly growth rate of sea fog cell area meets R≥15%; the hourly decrease in cloud top brightness temperature is ≥0.5K; the hourly increase in cloud optical thickness is ≥0.3; the boundary layer inversion intensity continues to increase; the hourly decrease in ground visibility is ≥200m; Development stages: The hourly growth rate of sea fog cell area meets the requirement of 5% ≤ R < 15%; the cloud top brightness temperature change rate is within ±0.3K / hour; the cloud optical thickness continues to increase; the inversion layer thickness continues to increase; and the visibility remains stable below 1km. Mature stage: The hourly change rate of the area of ​​individual sea fog cells meets the condition -5%≤R<5%; the hourly change rates of cloud top brightness temperature, cloud optical thickness, and inversion intensity are all within ±5%; visibility is stable below 500m, and the stable maintenance time is ≥2 hours; During the dissipation phase: the hourly reduction rate of the area of ​​individual sea fog cells is ≥10%; the hourly increase in cloud top brightness temperature is ≥0.8K; the hourly decrease in cloud optical thickness is ≥0.3; the intensity of temperature inversion continues to weaken; and the hourly increase in ground visibility is ≥200m.

[0025] Standardized dataset construction: Select typical historical sea fog processes in the target sea area to construct a sample dataset. Each sample includes: satellite sea fog identification sequence for the first 3 hours, sea fog life cycle stage labels, synchronous meteorological element field, and hourly sea fog data for the next 6 hours. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1 for subsequent model training and optimization.

[0026] Step 2: Construct a multi-scale optical flow extrapolation model that integrates cloud lifecycle characteristics This step is one of the core aspects of the invention. It breaks through the shortcomings of the traditional optical flow method, which only considers advection, and incorporates the life cycle characteristics as the core correction term into the extrapolation model to achieve coupled extrapolation of the advection movement of sea fog and its generation and dissipation trends.

[0027] Multi-scale optical flow field calculation: The three-layer pyramid Lucas-Kanade (LK) optical flow method is adopted to perform multi-scale optical flow calculation on the binary grid field of sea fog identification in three consecutive frames (15 minutes apart). The output is the grid-by-grid optical flow vector field of the sea fog field (u and v components, corresponding to the east-west and north-south movement velocities, respectively). This solves the problem of large calculation error of traditional optical flow methods for large-scale displacement and accurately captures the advection movement characteristics of sea fog. The detailed calculation process is further explained in the specific embodiment.

[0028] Lifecycle correction term construction: For each sea fog cell, based on its lifecycle stage, a corresponding area correction coefficient α and a sea fog persistence intensity correction coefficient β are constructed. β is used to determine the persistence probability of grid points currently identified as sea fog. Perform multiplicative correction, where This indicates the probability that the grid point will still be identified as sea fog in the next forecast. =min(1,β×p0), where p0 is the basic maintenance probability determined jointly by the current sea fog identification results and the local environmental field; β is not the inversion intensity, which is input separately as an environmental meteorological element. Furthermore, a morphological erosion method with a single grid point (4km) is used to divide the region: the internal grid points retained after erosion are the core region, and the outer ring grid points after subtracting the core region from the original monomer are the edge region; if the monomer scale is too small, resulting in an empty area after erosion, then the entire monomer is considered an edge region. The core rules are as follows: Initial stage: α = 1 + R (R is the current hourly area growth rate), β = 1.2, extrapolation is performed by expanding the range of sea fog cells hourly according to the α coefficient, and according to... =min(1,β×p0) Simultaneously increase the probability of sea fog persistence in the core area; Development stage: α = 1 + 0.5R, β = 1.1. Extrapolation follows the principle of α = 1 + 0.5R, with the sea fog extent expanding hourly, the expansion increment being 50% of the initial stage, and following... =min(1,β×p0) moderately increases the probability of sea fog persistence in the core area, with the increase being less than that in the initial stage; Mature stage: α=1, β=1.0. During extrapolation, the sea fog extent remains essentially unchanged, only undergoing advection translation based on the optical flow vector of the grid-by-grid optical flow vector field, and maintaining... =p0, no additional upward or downward adjustment is made; Dissipation phase: α = 1 - S (S is the current hourly area reduction rate), β = 0.8. Extrapolation is performed by decreasing the sea fog area hourly according to the α coefficient, and according to... =min(1,β×p0) synchronously lowers the probability of sea fog persistence in the edge region (i.e., the outermost grid-width ring of the sea fog cell).

[0029] Construction of environmental field physical constraints: Three core physical elements—sea-air temperature difference, boundary layer height, and 925 hPa relative humidity—are introduced to construct environmental field constraints. The correction coefficients are then calibrated a second time to ensure that the extrapolation results conform to the physical mechanisms of sea fog formation and dissipation. When the sea-air temperature difference is greater than 2℃, the boundary layer height is greater than 500m, and the relative humidity at 925hPa is less than 90%, α is forcibly multiplied by 0.8 and β is forcibly multiplied by 0.7 to suppress the development of sea fog. When the sea-air temperature difference is <0℃, the boundary layer height is <300m, and the relative humidity at 925hPa is >95%, α is forcibly multiplied by 1.1 and β is forcibly multiplied by 1.2 to promote the development of sea fog.

[0030] Advection extrapolation forecast output: Based on the grid-by-grid optical flow vector field of the sea fog field calculated in step 2, combined with the lifetime correction term and the environmental field constraint term, a semi-Lagrange advection scheme is adopted to extrapolate the hourly sea fog advection forecast field for the next 6 hours. The spatial resolution is consistent with the satellite data (4km), and the temporal resolution is 1 hour.

[0031] Step 3: Construct a lifecycle-based probability forecasting model for sea fog formation and dissipation. This step is the second core of the invention, solving the problem that traditional extrapolation methods cannot predict "newly formed sea fog" and "complete dissipation of existing sea fog". It uses a machine learning model to accurately predict the probability of sea fog formation and dissipation hourly for the next 6 hours.

[0032] Feature engineering: 28 input features across 3 categories, comprehensively covering the core driving factors of sea fog formation and dissipation, as detailed below: Life cycle evolution characteristics (8 dimensions): ① Initial stage marker; ② Development stage marker; ③ Mature stage marker; ④ Dissipation stage marker; ⑤ Area change rate over the past 3 hours; ⑥ Cloud top brightness temperature change rate; ⑦ Cloud optical thickness change rate; ⑧ Visibility change rate.

[0033] Environmental meteorological characteristics (16 dimensions): ⑨ Sea-air temperature difference; ⑩ Sea surface temperature; 2m air temperature; 2m dew point temperature; 2m relative humidity; 10m wind speed; 10m wind direction; Boundary layer height; Temperature inversion intensity; Inversion layer thickness; 925 hPa relative humidity; Vertical velocity at 925 hPa; 925 hPa air temperature; 925hPa u wind component; 925hPa v wind component; The temperature and dew point are different.

[0034] Geographical and temporal characteristics (4 dimensions): Offshore distance; latitude; month; Hourly, used to capture the diurnal and seasonal variations of sea fog.

[0035] Construction and training of sea fog formation and dissipation probability models: The LightGBM binary classification algorithm was used to construct six hourly sea fog formation and dissipation probability forecast models for the next 1 hour, 2 hours, 3 hours, 4 hours, 5 hours, and 6 hours. The model label indicates whether the corresponding grid point is foggy (1 = foggy, 0 = no foggy). Core hyperparameters: learning rate 0.05, tree depth 8, number of leaf nodes 64, number of iterations 1000, and early stopping rounds 50. Hyperparameters were optimized using a validation set, and model performance was evaluated using a test set to ensure that the AUC (Area Under Curve) value of all six models was ≥0.92.

[0036] Generation and dissipation probability field output: Gridded probability field of sea fog generation and dissipation for the next 6 hours, with a probability range of 0-1. The spatial resolution and temporal resolution are perfectly matched with the advection extrapolation field, providing a basis for subsequent forecast field fusion.

[0037] Step 4: Fusion and correction of hourly sea fog forecast fields over a 6-hour period This step is the key part of the invention, which deeply integrates the sea fog advection forecast field with the generation and dissipation probability field, and performs secondary correction through real-time observation and topographic data to output the final high-precision forecast result.

[0038] Fusion rules between sea fog advection forecast field and occurrence / disappearance probability field: Based on hourly advection extrapolation field and occurrence / disappearance probability field, fusion correction is performed according to the following rules: For grid points marked as sea fog in the sea fog advection forecast field, if the dissipation probability (1 - generation and dissipation probability) at the corresponding time is ≥70%, the grid point is corrected to be non-sea fog, thus solving the problem of false alarms in traditional extrapolation. For grid points marked as non-sea fog in the sea fog advection forecast field, if the generation probability at the corresponding time is ≥60%, the grid point will be corrected to sea fog to solve the underreporting problem of traditional extrapolation. For grid points with uncertainties between 40% and 60% of occurrence and dissipation probability, the original results of the sea fog advection forecast field are maintained to ensure the stability of the forecast.

[0039] Real-time observation bias correction: The optimal interpolation (OI) method is used to correct the bias of the fused forecast field using real-time visibility data from shore-based meteorological stations, island stations, and offshore buoys. Specifically, the fused sea fog forecast grid field is used as the background field, and the effective visibility observations within a 50km range are used as the observation field. Gaussian weights are constructed according to the distance between the observation point and the grid point to be corrected, with the weight increasing as the distance increases. First, the bias between the measured visibility at each observation point and the interpolation of the background field at that observation point is calculated. Then, the biases of each observation are weighted and averaged according to the distance Gaussian weights to obtain the analysis increment of the grid point to be corrected. Finally, this analysis increment is superimposed on the corresponding background grid point to obtain the corrected forecast result. If there are no effective observations within 50km of the grid point to be corrected, the original fused forecast value is retained, thereby eliminating systematic biases and improving the forecast accuracy of nearshore areas without disrupting the overall spatial continuity.

[0040] Topography and coastline correction: Based on the DEM digital elevation model and coastline data, sea fog forecasts are removed for land areas with an elevation > 50m; for open ocean areas with a distance > 500km from the shore, secondary calibration is performed in conjunction with sea surface temperature gradient to reduce false alarms caused by topography and underlying surface.

[0041] Step 5: Forecast Result Verification and Operational Output Quantitative verification of forecast effectiveness: The forecast results of this invention are quantitatively verified using common verification indicators in the meteorological forecasting industry. The core indicators include: Threat Score (TS = number of hits / (number of hits + number of missed reports + number of false alarms), with a value of 0 to 1, the larger the value, the better), hit rate (POD), and false alarm rate (FAR). The results are compared and verified with the forecast results of the traditional LK optical flow extrapolation method and the ECMWF numerical model.

[0042] Operational product outputs include: hourly sea fog forecast grid fields for the next 6 hours, visibility classification forecast products (<200m, 200-500m, 500-1000m, >1000m), sea fog impact range forecast maps, and warning information for affected areas; the product update frequency is 15 minutes / time, fully adaptable to operational short-term forecasting needs.

[0043] As can be seen from the above description, the beneficial effects of the present invention are as follows: This invention overcomes the core bottlenecks of traditional technologies, significantly improving the accuracy of 6-hour forecasts. For the first time, it systematically integrates cloud lifecycle theory into the entire process of 6-hour extrapolation forecasting of sea fog, solving the inherent defect of traditional extrapolation methods that "only shift without generating or dissipating," thus extending the effective forecast lead time for sea fog from 2-3 hours to 6 hours. Independent sample validation shows that the average TS score of the 6-hour forecast is ≥0.68, an improvement of 112% compared to traditional optical flow methods and 51% compared to numerical models; the hit rate is ≥0.82, and the false alarm rate is ≤0.21, far superior to existing mainstream technologies.

[0044] The fusion of physical mechanisms and data-driven approaches results in strong forecast stability and versatility. A dual-driven forecasting framework of "multi-scale optical flow extrapolation + lifecycle occurrence and extinction probability" is constructed. This framework retains the high frequency and efficiency advantages of extrapolation while avoiding overfitting issues inherent in purely data-driven models through cloud lifecycle physical mechanisms and environmental field constraints. The quantitative lifecycle classification standard is adaptable to various high-occurrence sea fog regions globally (Yellow Sea, Bohai Sea, East China Sea, South China Sea, west coast of North America, Sea of ​​Japan, etc.), requiring only threshold adjustments for localized deployment, demonstrating exceptional versatility.

[0045] It boasts high computational efficiency, strong business adaptability, and low deployment cost. Driven by geostationary satellite remote sensing data and conventional meteorological observation data, this invention does not rely on high-performance computing clusters. A single server can complete operational rolling forecasts every 15 minutes, with a computation time of ≤30 seconds per forecast, far lower than that of numerical models. It can seamlessly integrate with existing marine meteorological operational systems, providing accurate short-term early warning support for industries such as port shipping, offshore wind power, and near-shore aviation, with broad application prospects.

[0046] This invention addresses the challenge of forecasting sudden sea fog, offering significant value for disaster prevention and mitigation. Through a probability model of formation and dissipation, it can accurately forecast sudden, newly forming sea fog and rapidly dissipating sea fog, which are often missed by traditional techniques. This significantly reduces the rates of missed and false alarms, effectively minimizing maritime safety accidents caused by sea fog and possessing extremely high social and economic value.

[0047] Example 1 A sea fog forecasting method based on cloud lifecycle includes: Implementation Procedure for 6-Hour Extrapolation Forecast of Sea Fog in the Yellow and Bohai Seas 1. Data Source and Dataset Construction 1.1 Data Source Acquisition: Satellite data: Data from the Fengyun-4 series geostationary meteorological satellites, with FY-4A used for 2018–2021 and FY-4B used for 2022–2023; raw observation products were acquired at operational temporal resolution and uniformly resampled to 15-minute intervals, with a spatial resolution of 4 km. The channels used include: 0.65 μm visible light channel, 1.6 μm near-infrared channel, and 10.8 μm and 12.0 μm infrared split-window channels. Real-time observation data: Hourly visibility, temperature, dew point, 10m wind speed and direction data from 23 national benchmark meteorological stations, 12 island automatic stations, and 8 marine buoys along the Yellow and Bohai Seas; routine 08:00 and 20:00 radiosonde data from 3 radiosonde stations in Yantai, Dalian, and Qingdao; Reanalysis data: ERA5 hourly reanalysis data with a spatial resolution of 0.25°×0.25°, collecting sea surface temperature, 2m air temperature, 2m dew point, 10m wind field, boundary layer height, inversion intensity, 925hPa relative humidity, vertical velocity and other elements; Geographic data: 30m resolution DEM digital elevation model, China's near-shore coastline data.

[0048] 1.2 Sea Fog Identification and Lifecycle Division: A sea fog recognition model for the Yellow and Bohai Seas was constructed based on a multi-channel threshold method. After verification under ground visibility conditions, the sea fog recognition accuracy reached 92.3%. Based on historical statistics of 128 typical sea fog events in the Yellow and Bohai Seas, the threshold for life cycle stage division was optimized, which is consistent with the threshold in step 1 of this invention. Dataset construction: A total of 126,000 valid samples were constructed, which were divided into training set (100,800 sets), validation set (12,600 sets), and test set (12,600 sets) in an 8:1:1 ratio.

[0049] 2. Model Training and Construction 2.1 Multi-scale optical flow extrapolation model integrating lifecycle: A three-layer pyramid LK optical flow method was adopted to calculate the optical flow vector field based on three consecutive frames of satellite sea fog identification fields with a time interval of 15 minutes. Specifically, the three consecutive frames of sea fog identification binary fields were first Gaussian smoothed and a three-layer image pyramid was constructed at a scale of 1:2:4. Then, the initial displacement field was obtained by local window least squares method in the coarsest layer, and the displacement results were incrementally upsampled to the high-resolution layer. Finally, the optical flow results of each layer were weighted and fused, and the 15-minute displacement was converted into u and v components for each hour to obtain the grid-by-grid horizontal motion vector field used for semi-Lagrange extrapolation.

[0050] Write code for lifecycle correction terms and environmental field constraint terms, embed a semi-Lagrange advection extrapolation scheme, and realize hourly extrapolation calculation.

[0051] 2.2 Probability forecasting model for sea fog formation and dissipation: Complete the extraction and standardization of 28-dimensional input features, and construct 6 hourly LightGBM models; After completing model training and hyperparameter optimization, the final AUC values ​​of the six models on the test set were: 1h (0.96), 2h (0.95), 3h (0.94), 4h (0.93), 5h (0.92), and 6h (0.92), which met the forecast accuracy requirements.

[0052] 3. Forecast field fusion and correction The hourly advection forecast field for the next 6 hours is output based on the optical flow extrapolation model and fused with the hourly probability field output by the generation and dissipation probability model according to the fusion rules of this invention. The optimal interpolation method is used to correct the bias using the current real-time observation data: the fused forecast field is used as the background field, and effective visibility observations of shore base stations, island stations and marine buoys within a 50km range are collected. The incremental data is calculated and analyzed according to the distance Gaussian weight, and the incremental data is written back to the corresponding grid point. If there are no effective observations in a local area, the original fused forecast results are retained.

[0053] Topographic correction was performed based on DEM data, and sea fog forecast results for land areas with an altitude > 50m were removed.

[0054] 4. Forecast effectiveness verification Twenty independent sea fog events (including 12 large-scale persistent sea fog events and 8 sudden sea fog events) in the Yellow and Bohai Seas in 2024 were selected to verify the forecast results of this invention. The comparison with the traditional LK optical flow method and ECMWF numerical model is shown in Table 1 below. Fig. 3 : Table 1: Comparison of 6-hour forecast performance between the method of this invention and the traditional method Verification results show that the 6-hour sea fog forecast accuracy of this invention is significantly better than that of existing mainstream technologies, and fully meets the needs of operational forecasting.

[0055] 5. Business-oriented output In this embodiment, the operational product is updated every 15 minutes, and the calculation time for a single forecast round is ≤25 seconds. The output products include: hourly sea fog area forecast map for the Yellow and Bohai Seas over the next 6 hours, visibility graded forecast products, gridded forecast dataset, and hourly sea fog warning information for coastal ports. It has been connected to the business systems of ports, maritime authorities, and offshore wind power companies along the Yellow and Bohai Seas to achieve operational operation.

[0056] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent modifications made based on the content of the present invention specification and drawings, or direct or indirect applications in related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A sea fog forecasting method based on cloud lifecycle, characterized in that, include: S1. Based on multi-source data, sea fog is identified to obtain a binary grid field, and the life cycle stages of individual sea fog clouds are divided, including: initial stage, development stage, mature stage and dissipation stage; typical historical sea fog processes in the target sea area are selected to construct a sample dataset; S2. The three-layer pyramid LK optical flow method is used to perform multi-scale optical flow calculations on the binary grid field and output the grid-by-grid optical flow vector field of the sea fog field. Lifecycle correction terms: Individual cell area correction factor α, correction factor β, where β is used to determine the persistence probability of grid points currently identified as sea fog. Make corrections; Initial stage: α = 1 + R, where R is the current hourly area growth rate, β = 1.

2. Extrapolation involves expanding the range of individual sea fog cells hourly based on α, and adjusting the core area upwards based on β. Development stage: α=1+0.5R, β=1.

1. When extrapolating, the sea fog range is expanded hourly according to α, and the core area is adjusted upward according to β. Maturation stage: α=1, β=1.0, the sea fog range remains constant during extrapolation, and it is translated according to the optical flow vector field; Dissipation stage: α=1-S, where S is the current hourly area reduction rate, β=0.8, the sea fog range decreases hourly according to α during extrapolation, and the edge area is adjusted downward according to β. ; Environmental field physical constraints: Introduce sea-temperature difference A, boundary layer height B, and 925hPa relative humidity C for secondary calibration. When A > 2℃, B > 500m, and C < 90%, α × 0.8, β × 0.7; when A < 0℃, B < 300m, and C > 95%, α × 1.1, β × 1.

2. Advection extrapolation forecast output: Based on the optical flow vector field, combined with the life cycle correction term and the environmental field constraint term, a semi-Lagrange advection scheme is used to extrapolate the hourly sea fog advection forecast field for the next 6 hours; S3. Using the life cycle of individual sea fog clouds, environmental meteorological characteristics, and geographical and temporal characteristics as inputs, the LightGBM binary classification algorithm is used to construct six sea fog formation and dissipation probability forecast models for each hour from 1 to 6 hours in the future. The models are trained and validated using a sample dataset. The model label is whether the corresponding grid point is sea fog. S4. The hourly sea fog advection forecast field from 1 to 6 hours is deeply fused with the sea fog formation and dissipation probability forecast model to obtain the sea fog forecast grid field, and then the second correction is performed through real-time observation and topographic data. S5. Forecast result verification and operational output.

2. The sea fog forecasting method based on cloud lifecycle as described in claim 1, characterized in that, Sea fog identification based on multi-source data further includes: The multi-source data includes: high spatiotemporal resolution geostationary meteorological satellite data, uniformly resampled to 15-minute temporal resolution and 4km spatial resolution, including visible light, near-infrared, and infrared split-window channels; ground observation data, including hourly visibility, temperature, dew point, wind speed, and wind direction data from coastal national meteorological stations, island automatic stations, and offshore buoys; radiosonde and reanalysis data, including routine 08:00 and 20:00 radiosonde data from coastal radiosonde stations and hourly ERA5 reanalysis data, including sea surface temperature, boundary layer height, inversion intensity, relative humidity, and vertical velocity; and geographic information data, including digital elevation model (DEM) and coastline data. Precise remote sensing identification of sea fog: Based on multi-channel geostationary satellite data, a sea fog identification model is constructed. A multi-channel threshold method is used to distinguish between sea fog, mid-to-high clouds, low clouds, and clear-sky sea surfaces. The core identification thresholds are as follows: Infrared split-window brightness temperature difference <0.5K, mid-to-high cloud cover excluded; The reflectance of the visible light channel is 0.1~0.8, and the ratio of near-infrared to visible light reflectance is <0.8, which can distinguish sea fog from clear sky sea surface and water surface flares; By combining the cloud top brightness temperature threshold and removing convective clouds and mid-to-high clouds, a binary grid field for sea fog identification is finally obtained every 15 minutes, where 1 = sea fog and 0 = non-sea fog. The accuracy is verified by real-world data with ground visibility <1km to ensure that the identification accuracy is ≥90%.

3. The sea fog forecasting method based on cloud lifecycle as described in claim 2, characterized in that, The life cycle of sea fog clouds is further divided into stages including: Based on a continuous 3-hour remote sensing observation sequence of sea fog and synchronous meteorological elements, sea fog cells are divided into: Initial stage: The hourly growth rate of sea fog cell area meets R≥15%; the hourly decrease in cloud top brightness temperature is ≥0.5K; the hourly increase in cloud optical thickness is ≥0.3; the boundary layer inversion intensity continues to increase; the hourly decrease in ground visibility is ≥200m; Development stages: The hourly growth rate of sea fog cell area meets the requirement of 5% ≤ R < 15%; the cloud top brightness temperature change rate is within ±0.3K / hour; the cloud optical thickness continues to increase; the inversion layer thickness continues to increase; and the visibility remains stable below 1km. Mature stage: The hourly change rate of the area of ​​individual sea fog cells meets the condition -5%≤R<5%; the hourly change rates of cloud top brightness temperature, cloud optical thickness, and inversion intensity are all within ±5%; visibility is stable below 500m, and the stable maintenance time is ≥2 hours; During the dissipation phase: the hourly reduction rate of the area of ​​individual sea fog cells is ≥10%; the hourly increase in cloud top brightness temperature is ≥0.8K; the hourly decrease in cloud optical thickness is ≥0.3; the intensity of temperature inversion continues to weaken; and the hourly increase in ground visibility is ≥200m.

4. The sea fog forecasting method based on cloud lifecycle as described in claim 2, characterized in that, The division between core and peripheral areas includes: The region is divided using a single-grid morphological etching method: the internal grid points retained after etching are the core region, and the outer ring grid points after subtracting the core region from the original sea fog cell are the edge region; if the sea fog cell is too small to be completely empty after etching, then the entire sea fog cell is considered as the edge region.

5. The sea fog forecasting method based on cloud lifecycle as described in claim 1, characterized in that, =min(1,β× ), The baseline maintenance probability is determined jointly by the current sea fog identification results and the local environmental field.

6. The sea fog forecasting method based on cloud lifecycle as described in claim 2, characterized in that, Typical historical sea fog processes in the target sea area were selected to construct a sample dataset. Each sample set includes: satellite sea fog identification sequence of the first 3 hours, life cycle stage label of sea fog cells, synchronous meteorological element field, and hourly sea fog data for the next 6 hours. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:

1. The LightGBM binary classification algorithm was used to construct six hourly probability forecast models for the formation and dissipation of sea fog, covering the next 1-6 hours. The models were trained using a training set with the following core hyperparameters: learning rate 0.05, tree depth 8, number of leaf nodes 64, number of iterations 1000, and number of early stopping rounds 50. The hyperparameters were optimized using a validation set, and the model performance was evaluated using a test set to ensure that the AUC value of all six models was ≥0.

92.

7. The sea fog forecasting method based on cloud lifecycle as described in claim 6, characterized in that, The life cycle of a single sea fog cloud includes: initial stage marker, development stage marker, mature stage marker, dissipation stage marker, area change rate over the past 3 hours, cloud top brightness temperature change rate, cloud optical thickness change rate, and visibility change rate. Environmental meteorological characteristics include: sea surface temperature difference, sea surface temperature, air temperature at 2m, dew point temperature at 2m, relative humidity at 2m, wind speed at 10m, wind direction at 10m, boundary layer height, inversion intensity, inversion layer thickness, relative humidity at 925hPa, vertical velocity at 925hPa, air temperature at 925hPa, u-wind component at 925hPa, v-wind component at 925hPa, and air temperature-dew point difference; Geographic and temporal characteristics include: distance from the shore, latitude, month, and hour.

8. The sea fog forecasting method based on cloud lifecycle as described in claim 2, characterized in that, The deep fusion of the hourly sea fog advection forecast field from 1 to 6 hours with the sea fog formation and dissipation probability forecast model yields the sea fog forecast grid field, which further includes: For grid points marked as sea fog in the advection extrapolation field, if the dissipation probability at the corresponding time is ≥70%, the grid point is corrected to not be sea fog. For grid points in the advection extrapolation field that are marked as non-sea fog, if the generation probability at the corresponding time is ≥60%, then the grid point is corrected to sea fog. For grid points with uncertainties between 40% and 60% of the probability of occurrence and dissipation, the original results of the sea fog advection forecast field are maintained; The second revision includes: Real-time observation bias correction: Using the fused sea fog forecast grid field as the background field, the effective visibility observation within a 50km range is used as the observation field. Gaussian weights are constructed according to the distance between the observation point and the grid point to be corrected. The closer the distance, the greater the weight. First, calculate the deviation between the measured visibility at each observation point and the interpolation of the background field at that observation point. Then, perform a weighted average of the observation deviations according to the distance Gaussian weight to obtain the analysis increment of the grid point to be corrected. Finally, superimpose the analysis increment onto the corresponding background grid point to obtain the corrected forecast result. If there are no effective observations within 50km of the grid point to be corrected, retain the original forecast value after fusion. Topography and coastline correction: Based on the DEM digital elevation model and coastline data, sea fog forecast results are removed for land areas with an elevation > 50m; for open ocean areas with a distance > 500km from the shore, secondary calibration is performed in conjunction with sea surface temperature gradient.

9. The sea fog forecasting method based on cloud lifecycle as described in claim 1, characterized in that, Further verification of forecast results includes: The forecast results after secondary correction were quantitatively verified using common verification indicators in the meteorological forecasting industry. The core indicators include: critical success index, hit rate and false alarm rate. The results were also compared and verified with the forecast results of traditional LK optical flow extrapolation method and ECMWF numerical model.

10. The sea fog forecasting method based on cloud lifecycle according to claim 1, characterized in that, The operational outputs further include: hourly sea fog forecast grid fields for the next 6 hours, visibility graded forecast products, sea fog impact range forecast maps, and warning information for the affected areas; the product update frequency is 15 minutes / time.

Images