A method for FY-4 sea cloud base height correction considering the cloud height instrument of buoy

Abstract

The application provides a FY-4 sea cloud base height correction method considering a buoy cloud height instrument, and belongs to the technical field of sea cloud base height correction. The application collects the buoy cloud height instrument and FY-4 satellite cloud base height data, performs sea surface reflection pollution correction on the satellite data, identifies the cloud layer structure by using K-means clustering, extracts the multi-scale fractal dimension features of the buoy and satellite data and calculates the difference parameters, organizes the correction data into a four-dimensional correction tensor, extracts the low-rank structure by using CANDECOMP decomposition, inputs the decomposition result and the fractal dimension difference parameters into a cloud base height bidirectional optimization model, and realizes the dynamic balance between the accuracy improvement and the spatiotemporal stability maintenance in the model by the game mechanism of the accuracy optimization sub-model and the stability optimization sub-model, so that the technical problem that the sea cloud base height precision and the spatiotemporal stability of the satellite inversion are difficult to be considered is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of sea cloud base height correction, and in particular, relates to a FY-4 sea cloud base height correction method considering a buoy cloud height instrument. BACKGROUND

[0002] The FY-4 satellite inverts sea cloud base height through infrared radiation measurement, which is an important technical means for marine weather monitoring. The traditional satellite cloud base height inversion method is based on a radiation transfer model, and the satellite received radiation brightness is associated with the atmospheric profile parameter to calculate the cloud base height, but the sea surface mirror reflection and white hat foam in the sea environment interfere with the radiation signal, resulting in systematic deviation of the inversion result. The existing correction method uses the measured data of the buoy cloud height instrument to correct the satellite inversion result, but there is a significant difference in spatial scale between single-point buoy observation and wide-range satellite observation, and direct use of buoy data correction will lead to the improvement of the correction result at the buoy station and the occurrence of non-physical phenomena such as spatial and temporal discontinuity in the non-buoy coverage area. In the prior art, when pursuing the improvement of the correction accuracy, the spatial and temporal stability is destroyed, and when maintaining the spatial and temporal smoothness, the local accuracy is lost, and the two optimization objectives are mutually restricted. That is to say, in the prior art, there is a technical problem that the sea cloud base height inversion by the satellite is difficult to balance the accuracy and the spatial and temporal stability. SUMMARY

[0003] Therefore, the application provides a FY-4 sea cloud base height correction method considering a buoy cloud height instrument, which can solve the technical problem that the sea cloud base height inversion by the satellite is difficult to balance the accuracy and the spatial and temporal stability in the prior art.

[0004] The application is implemented in the following manner: the application provides a FY-4 sea cloud base height correction method considering buoy cloud height instruments, which collects cloud base height measured data of multiple buoy cloud height instruments in a target sea area and FY-4 satellite cloud base height data corresponding to the spatial and temporal positions, performs sea surface reflection pollution correction on the FY-4 satellite cloud base height data, identifies cloud layer structure and marks cloud layer number identification in combination with microwave radiometer temperature profile data, extracts multi-scale fractal dimension features of the cloud base height measured data and the FY-4 satellite cloud base height data by using a box counting method, calculates the difference between the buoy fractal dimension and the satellite fractal dimension as a fractal dimension difference parameter, organizes the cloud base height correction data into a four-dimensional correction tensor in the longitude dimension, the latitude dimension, the time dimension and the meteorological condition dimension, extracts the low-rank structure of the four-dimensional correction tensor by using a CANDECOMP decomposition method to obtain a correction core tensor and a correction factor matrix group, inputs the correction core tensor, the correction factor matrix group, the fractal dimension difference parameter and the cloud layer number identification into a cloud base height bidirectional optimization model for correction, sets precision optimization sub-models and stability optimization sub-models in the cloud base height bidirectional optimization model, and establishes parameter correlation and alternately optimizes the two sub-models through a coupling term to form a game balance state, and outputs the corrected FY-4 satellite cloud base height data.

[0005] Further, the sea surface reflection pollution correction comprises: establishing a sea surface bidirectional reflection distribution function model through sea surface wind speed data, and decoupling cloud layer radiation signals and sea surface reflection signals by using a radiation transfer equation.

[0006] Further, the sea surface bidirectional reflection distribution function model takes the solar zenith angle, the observation zenith angle, the relative azimuth angle and the sea surface wind speed data as inputs, and outputs the sea surface reflectivity distribution function values at different angles.

[0007] Further, the decoupling by using the radiation transfer equation comprises: decomposing the total radiation brightness received by the satellite into the sum of the cloud layer radiation contribution and the sea surface reflection contribution, and subtracting the sea surface reflection contribution from the total radiation brightness through iterative optimization.

[0008] Further, the cloud layer structure identification adopts a K-means clustering algorithm to separate the cloud layer radiation signals into single-layer cloud radiation signal sets, double-layer cloud radiation signal sets and multi-layer cloud radiation signal sets.

[0009] Further, the K-means clustering algorithm extracts temperature gradient mutation points in the microwave radiometer temperature profile data as cloud layer demarcation features, and sets the clustering number to 3.

[0010] Further, the box counting method sets an initial box size of 1 km, counts the minimum number of boxes required to cover the cloud base data, and gradually reduces the box size to 0.125 km.

[0011] Further, the multi-scale fractal dimension feature is calculated by using the linear regression slope of the logarithm of the box number and the logarithm of the reciprocal of the box size as the fractal dimension value.

[0012] Further, the statistical correlation between the fractal dimension difference parameter and the cloud base height deviation is established by using a cubic polynomial regression fitting.

[0013] Further, the CANDECOMP decomposition method uses an alternating least squares algorithm to optimize the decomposition, and fixes three correction factor matrices to optimize the fourth correction factor matrix to minimize the reconstruction error.

[0014] The alternating least squares algorithm calculates the Frobenius norm difference between the reconstructed tensor and the four-dimensional correction tensor as a convergence criterion, and stops iteration when the norm difference change rate is less than 0.001.

[0015] The accuracy optimization sub-model of the cloud base height bidirectional optimization model aims to maximize the correction accuracy, and the stability optimization sub-model aims to maximize the spatiotemporal stability.

[0016] The accuracy optimization sub-model outputs an accuracy weight vector through a three-layer fully connected network, and the stability optimization sub-model outputs a stability weight vector by extracting spatiotemporal continuity features through a convolutional neural network; the accuracy weight vector and the stability weight vector are combined into a final correction weight through a weighted fusion layer; the final correction weight is multiplied by the feature vector to generate the corrected FY-4 satellite cloud base height data; the objective function of the accuracy optimization sub-model calculates the root mean square error of the measured cloud base height data and the model output cloud base height multiplied by the exponential function of the fractal dimension difference parameter, and then divided by the square root of the sum of the cloud layer number identifier; the objective function of the stability optimization sub-model calculates the product of the standard deviation of the adjacent time cloud base height difference and the standard deviation of the adjacent spatial position cloud base height difference, and then divided by the sum of the logarithm values of the meteorological condition change rate.

[0017] Compared with the prior art, the present application has the following technical effects:

[0018] This invention constructs a two-way optimization model for cloud base height, setting up two parallel optimization paths within the model: a precision optimization sub-model and a stability optimization sub-model. The precision optimization sub-model aims to minimize the root mean square error between the corrected data and the buoy-measured data, while the stability optimization sub-model aims to minimize the standard deviation of the cloud base height difference between adjacent spatiotemporal locations. The two sub-models establish parameter sharing and gradient transfer channels through coupling terms, alternately optimizing their respective objective functions during training to form a dynamic game equilibrium. When the precision optimization sub-model overfits, leading to spatiotemporal discontinuities, the constraint effect of the stability optimization sub-model suppresses the overfitting tendency. When the stability optimization sub-model becomes overly smooth, causing a decrease in accuracy, the driving effect of the precision optimization sub-model enhances the local correction capability. The game mechanism enables the model to find the optimal balance between accuracy and stability in the parameter space, ensuring that the correction results meet the high accuracy requirements at the buoy station location while maintaining spatiotemporal continuity over a large area of ​​sea. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in simultaneously achieving accuracy and spatiotemporal stability in satellite-retrieved maritime cloud base height. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention.

[0020] Figure 2 Scatter plot of the distribution of fractal dimension difference parameters for different cloud types.

[0021] Figure 3 This is a schematic diagram of the weight fusion process of the game layer in the bidirectional optimization model for cloud base height. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0023] like Figure 1 The diagram shows a flowchart of a method for correcting the cloud bottom height at sea using the Fengyun-4 buoy, which incorporates a buoy ceilometer, according to the present invention. This method includes the following steps:

[0024] S01. Collect measured cloud base height data from multiple buoy cloud height meters in the target sea area and cloud base height data from the Fengyun-4 satellite at the corresponding time and space location, and simultaneously acquire microwave radiometer temperature profile data and sea surface wind speed data as auxiliary parameters.

[0025] S02. Correct sea surface reflection pollution for cloud base height data of Fengyun-4 satellite. Establish a two-way reflection distribution function model of sea surface using sea surface wind speed data. Decouple cloud radiation signal and sea surface reflection signal using radiative transfer equation. Use iterative optimization to remove sea surface reflection signal.

[0026] S03, identify the cloud layer structure by combining the microwave radiometer temperature profile data, use the K-means clustering algorithm to separate the cloud layer radiation signal into single-layer cloud radiation signal set, double-layer cloud radiation signal set and multi-layer cloud radiation signal set, and record the cloud layer number identification corresponding to each cloud layer radiation signal set;

[0027] S04, extract the multi-scale fractal dimension features of the cloud base height measured data and the FY-4 satellite cloud base height data by using the box counting method, calculate the difference between the fractal dimension of the buoy and the fractal dimension of the satellite as the fractal dimension difference parameter, and establish the statistical correlation between the fractal dimension difference parameter and the cloud base height deviation;

[0028] S05, organize the cloud base height correction data into a four-dimensional correction tensor according to the longitude dimension, latitude dimension, time dimension and meteorological condition dimension, extract the low-rank structure of the four-dimensional correction tensor by using the CANDECOMP decomposition method, and obtain the correction core tensor and correction factor matrix group by optimizing the decomposition through the alternating least squares algorithm;

[0029] S06, input the correction core tensor, correction factor matrix group, fractal dimension difference parameter and cloud layer number identification into the cloud base height bidirectional optimization model for correction optimization, the cloud base height bidirectional optimization model outputs the corrected FY-4 satellite cloud base height data, judges whether the root mean square error between the corrected FY-4 satellite cloud base height data and the cloud base height measured data is less than the threshold value 50m, if less than, output the corrected FY-4 satellite cloud base height data, if greater than, adjust the multi-head attention weight parameter of the cloud base height bidirectional optimization model and execute the step again.

[0030] The sea surface bidirectional reflection distribution function model describes the reflection characteristic distribution law of the sea surface to the incident radiation, and the input is the solar zenith angle, the observation zenith angle, the relative azimuth angle and the sea surface wind speed data, and the output is the sea surface reflectance distribution function value at different angles. The sea surface bidirectional reflection distribution function model is used to quantify the interference intensity of sea surface specular reflection and white cap foam on satellite radiation measurement.

[0031] The step of decoupling the cloud layer radiation signal and the sea surface reflection signal by the radiative transfer equation is specifically: decomposing the total radiation brightness received by the satellite into the sum of the cloud layer radiation contribution item and the sea surface reflection contribution item, calculating the theoretical value of the sea surface reflection contribution item by using the sea surface bidirectional reflection distribution function model, solving the cloud layer radiation contribution item by iteration through the radiative transfer equation, and obtaining the cloud layer radiation signal by deducting the sea surface reflection contribution item from the total radiation brightness.

[0032] The step of separating the cloud layer radiation signal by the K-means clustering algorithm is specifically: extracting a temperature gradient mutation point in the microwave radiometer temperature profile data as a cloud layer demarcation feature, calculating a brightness temperature difference value sequence of each height layer, inputting the brightness temperature difference value sequence into the K-means clustering algorithm, setting the number of clusters to 3 corresponding to three types of single-layer cloud, double-layer cloud and multi-layer cloud, optimizing the cluster center position through iteration until convergence, and marking the number of cloud layers according to the cluster center category to which the sample belongs.

[0033] The step of extracting a multi-scale fractal dimension feature by the box counting method is specifically: performing grid division on a cloud base height measured data time sequence and a spatial distribution map of the FY-4 satellite cloud base height data respectively, setting an initial box scale to 1 km, counting the minimum number of boxes required to cover the cloud base data, gradually reducing the box scale to one half of the initial scale until 0.125 km, recording a box number sequence under different scales, calculating a linear regression slope of a logarithm of the box number and a logarithm of a reciprocal of the box scale as a fractal dimension value, and taking a fractal dimension value corresponding to the cloud base height measured data as a buoy fractal dimension value and a fractal dimension value corresponding to the FY-4 satellite cloud base height data as a satellite fractal dimension value. The fractal dimension difference value parameter represents a difference degree of geometric complexity of the cloud base data spatial structure.

[0034] The step of establishing a statistical correlation between the fractal dimension difference value parameter and the cloud base height deviation is specifically: calculating a difference between the buoy fractal dimension and the satellite fractal dimension to obtain a fractal dimension difference value parameter, and simultaneously calculating a difference between the cloud base height measured data and the FY-4 satellite cloud base height data to obtain a cloud base height deviation, taking the fractal dimension difference value parameter of all matched samples as an independent variable and the cloud base height deviation as a dependent variable to perform polynomial regression fitting, and establishing a cubic polynomial relationship equation to describe a statistical correlation rule between the two.

[0035] The CANDECOMP decomposition method decomposes the four-dimensional correction tensor into a product form of a correction core tensor and four correction factor matrices. The correction core tensor is a scalar value representing the strength of multi-dimensional interaction, and the four correction factor matrices correspond to the main variation mode eigenvectors of the longitude dimension, the latitude dimension, the time dimension and the meteorological condition dimension respectively. The four correction factor matrices constitute a correction factor matrix group.

[0036] The step of optimizing and decomposing by the alternating least squares algorithm is specifically: initializing the four correction factor matrices as random orthogonal matrices, fixing three correction factor matrices to optimize the fourth correction factor matrix to minimize the reconstruction error, and sequentially optimizing each correction factor matrix. The Frobenius norm difference value between the reconstructed tensor and the four-dimensional correction tensor is taken as a convergence criterion. When the norm difference value change rate is less than 0.001, the iteration is stopped and the final correction core tensor and the correction factor matrix group are output.

[0037] The specific structure of the cloud base height bidirectional optimization model is: the input layer receives the corrected core tensor element value, the principal component value of the correction factor matrix group, the fractal dimension difference value parameter and the cloud layer quantity identifier, the input is mapped to a 128-dimensional feature vector through a feature fusion layer, and then enters the game layer for bidirectional optimization. The game layer includes an accuracy optimization sub-model with the goal of maximizing correction accuracy and a stability optimization sub-model with the goal of maximizing spatiotemporal stability. The accuracy optimization sub-model outputs an accuracy weight vector through a three-layer fully connected network, and the stability optimization sub-model outputs a stability weight vector by extracting spatiotemporal continuity features through a convolutional neural network. The accuracy weight vector and the stability weight vector are combined into a final correction weight through a weighted fusion layer. After the final correction weight is multiplied by the 128-dimensional feature vector, the corrected Fengyun-4 satellite cloud base height data is generated through the output layer.

[0038] The target function of the accuracy optimization sub-model is inputted with the cloud base height measured data, the Fengyun-4 satellite cloud base height data, the fractal dimension difference value parameter and the cloud layer quantity identifier. The target function calculates the root mean square error of the cloud base height measured data and the model output cloud base height multiplied by the exponential function of the fractal dimension difference value parameter, and then divided by the square root of the cloud layer quantity identifier, as the accuracy loss value. The smaller the accuracy loss value, the higher the correction accuracy. The target function of the stability optimization sub-model is inputted with the adjacent time cloud base height difference value, the adjacent spatial position cloud base height difference value and the meteorological condition change rate. The target function calculates the product of the standard deviation of the adjacent time cloud base height difference value and the standard deviation of the adjacent spatial position cloud base height difference value, and then divided by the sum of the logarithmic values of the meteorological condition change rate, as the stability loss value. The smaller the stability loss value, the higher the spatiotemporal stability. The two target functions are associated through the dot product of the accuracy weight vector and the stability weight vector as a coupling term, which constrains the synergistic relationship between accuracy optimization and stability optimization.

[0039] The steps of establishing the training data set of the cloud base height bidirectional optimization model include: selecting the cloud height instrument measured data of 12 buoy stations near China's coast and the corresponding Fengyun-4 satellite data from 2020 to 2024, screening out samples with cloud cover greater than 30% and data quality identifier as excellent, eliminating abnormal samples with precipitation or haze weather conditions, dividing the data into training set, validation set and test set according to the ratio of 7:2:1, performing data enhancement processing on the training set including adding 5% Gaussian noise and time shift disturbance, and labeling the cloud base height measured data of each sample as true value label.

[0040] The step of training the cloud base height bidirectional optimization model specifically comprises: setting the initial learning rate as 0.001 by using the Adam optimizer, setting the batch size as 32, setting the maximum training round as 200, optimizing the parameters of the precision optimization sub-model and the stability optimization sub-model simultaneously in the training process, updating the game layer weight every 10 batches, calculating the gradient of the model parameters by gradient back propagation, using gradient clipping to prevent gradient explosion, triggering the early stopping mechanism to end the training when the validation set loss value does not decrease for 20 consecutive rounds, and saving the model parameters at the time when the validation set loss value is the smallest as the final training result.

[0041] The multi-head attention weight parameters in the cloud base height bidirectional optimization model are dynamically adjusted according to the fractal dimension difference parameter, the cloud layer quantity identifier and the correction core tensor element value, and the adjustment mode is to normalize the fractal dimension difference parameter, the cloud layer quantity identifier and the correction core tensor element value, and then map them to the attention weight distribution ratio by using the softmax function. The greater the fractal dimension difference parameter is, the higher the corresponding attention weight is, indicating that stronger correction strength is needed. The greater the cloud layer quantity identifier is, the higher the corresponding attention weight is, indicating that more detailed processing is needed for the multi-cloud layer situation. The greater the correction core tensor element value is, the higher the corresponding attention weight is, indicating that the multi-dimensional interaction is more significant and needs more attention.

[0042] The principle of embedding the game layer is to build two mutually competitive and mutually restraining optimization objectives, namely the precision optimization sub-model and the stability optimization sub-model. The precision optimization sub-model pursues high consistency between the correction result and the cloud base height measured data, and the stability optimization sub-model pursues the smooth continuity of the correction result in the space-time dimension. The precision optimization sub-model and the stability optimization sub-model establish a parameter sharing and gradient transmission channel through a coupling term, and alternately optimize their own objective functions in the training process to form a dynamic game balance state. When the precision optimization sub-model overfits and leads to space-time discontinuity, the constraint effect of the stability optimization sub-model will inhibit the overfitting tendency. When the stability optimization sub-model is excessively smoothed and leads to precision decline, the driving effect of the precision optimization sub-model will enhance the local correction ability.

[0043] The specific implementation of the game layer in the cloud base height bidirectional optimization model structure is to branch out two parallel optimization paths after the feature fusion layer, which correspond to the precision optimization sub-model and the stability optimization sub-model respectively. Each optimization path independently performs forward propagation to calculate its own objective function loss value. In the back propagation phase, the parameter updates of the two optimization paths are related to each other through the coupling term gradient. The alternating gradient descent strategy is adopted to first fix the stability optimization sub-model parameters to optimize the precision optimization sub-model parameters, and then fix the precision optimization sub-model parameters to optimize the stability optimization sub-model parameters. The cycle iteration continues until the loss values of the precision optimization sub-model and the stability optimization sub-model converge simultaneously.

[0044] The technical effects brought by the game layer mechanism to the cloud bottom height two-way optimization model are that the contradiction between precision improvement and stability maintenance in the traditional correction method is fundamentally solved, the precision optimization sub-model is used to ensure that the correction result is highly consistent with the cloud bottom height measured data at the buoy station, the stability optimization sub-model is used to ensure that the correction result is physically reasonable and has time and space continuity in the non-buoy coverage area, the game interaction between the precision optimization sub-model and the stability optimization sub-model ensures that the model pursues local precision without producing non-physical phenomena of time and space mutation, and the model pursues global smoothness without losing the ability to depict local details, and the constraint of the coupling term makes the two optimization objectives find the optimal balance point in the parameter space, avoiding the overfitting or underfitting problem caused by single objective optimization.

[0045] The technical effects brought by the game layer mechanism to the whole FY-4 sea cloud bottom height correction scheme are that the adaptability and robustness of the correction result in different sea areas, different seasons and different cloud types are significantly improved, the corrected FY-4 satellite cloud bottom height data meets the high-precision requirement at the buoy station and the time and space consistency requirement in a large sea area, and more reliable cloud bottom height data products are provided for marine weather forecasting, navigation safety guarantee and marine environment monitoring.

[0046] The principle of the fractal dimension feature matching algorithm is to use the fractal geometric characteristics of the self-similarity of the spatial distribution and time evolution of the cloud field, count the minimum box number required to cover the cloud bottom data at different observation scales by the box counting method, establish the power law relationship between the box number and the observation scale, and the exponent of the power law relationship is the fractal dimension value. The fractal dimension value quantitatively describes the filling density and the degree of fluctuation of the cloud bottom height data in the spatial or time dimension. The cloud bottom height measured data reflects the real structure of the cloud layer observed vertically at a single point, and the FY-4 satellite cloud bottom height data reflects the overall shape of the cloud layer observed horizontally in a large range. The difference between the two observation angles and spatial resolutions causes a systematic deviation in the fractal dimension value. The systematic deviation and the cloud bottom height inversion error have an inherent physical correlation. The statistical model of the fractal dimension difference parameter and the cloud bottom height deviation is established to realize the mapping and conversion from the geometric topological features to the physical quantity error.

[0047] The rationality of applying the fractal dimension feature matching algorithm to the present scheme lies in that the spatial structure and evolution process of the offshore cloud system naturally has fractal characteristics, and the fractal dimension values of different cloud types are significantly different. The fractal dimension value of stratocumulus is usually between 1.2 and 1.4, showing a relatively uniform tiled structure, the fractal dimension value of cumulus is usually between 1.6 and 1.8, showing a clumpy aggregated structure, and the fractal dimension value of cirrus is usually between 1.4 and 1.6, showing a fibrous discrete structure. The fractal dimension value feature can fundamentally distinguish the geometric topological differences of different cloud types, providing a physical basis for classification and correction. Compared with the traditional cloud top temperature or reflectivity feature, the fractal dimension value feature has less dependence on observation angle and illumination condition, has stronger scale invariance and rotation invariance, and is suitable for the variable observation conditions at sea.

[0048] The specific implementation steps of the fractal dimension feature matching algorithm in the present scheme are as follows: firstly, the cloud base height measured data time series are preprocessed to eliminate abnormal values caused by instrument failure and precipitation interference, missing data are filled by using cubic spline interpolation, the time series are converted into equidistant sampling sequences, then the FY-4 satellite cloud base height data spatial distribution map is projected to the geographical area corresponding to the buoy station position, the cloud base height pixel values within a range of 50 km radius centered on the buoy are extracted, the spatial distribution map is resampled to a uniform 1 km grid resolution, next, the box counting method is applied to the time series and the spatial distribution map respectively. The one-dimensional box counting method is used for the time series, the box scale sequence is set as 16 minutes, 8 minutes, 4 minutes, 2 minutes and 1 minute, the minimum number of boxes required to cover the entire time series is counted under each scale, the two-dimensional box counting method is used for the spatial distribution map, the box scale sequence is set as 16 km, 8 km, 4 km, 2 km and 1 km, the minimum number of boxes required to cover all non-zero pixels is counted under each scale, then the scatter plot of the logarithm of the box number versus the logarithm of the reciprocal of the box scale is drawn for each scale sequence, and the linear regression fitting straight line slope is the fractal dimension value. The fractal dimension value of the cloud base height measured data time series is denoted as the buoy fractal dimension, and the fractal dimension value of the FY-4 satellite cloud base height data spatial distribution map is denoted as the satellite fractal dimension. The difference between the buoy fractal dimension and the satellite fractal dimension is calculated as the fractal dimension difference parameter. Finally, the statistical correlation between the fractal dimension difference parameter and the cloud base height deviation is established. The sample data of all buoy station positions at different times are collected, the fractal dimension difference parameter is taken as the independent variable, and the difference between the cloud base height measured data and the FY-4 satellite cloud base height data is taken as the dependent variable. A cubic polynomial function is used for nonlinear regression fitting to obtain the mapping function of the fractal dimension difference parameter to the cloud base height deviation. The mapping function is used to estimate the systematic deviation of the cloud base height according to the fractal dimension difference parameter in the subsequent correction process.

[0049] The fractal dimension feature matching algorithm provides a correction basis from the perspective of cloud field geometric topology, breaks through the strong dependence of traditional correction methods based on radiation transfer model on the accuracy of atmospheric profile parameters, and provides a unified benchmark for cloud base height data under different observation conditions.

[0050] The fractal dimension feature matching algorithm realizes geometric topology correction from buoy point observation to satellite surface observation by establishing a statistical model of fractal dimension difference parameters and cloud base height deviation, effectively compensating for the systematic differences in spatial scale and observation angle between the two observation methods. For the case of multiple cloud layers coexisting and cloud layer boundary being blurred, the fractal dimension value feature can comprehensively reflect the overall structural characteristics of the cloud field, avoiding the cumulative error of traditional layer-by-layer identification methods and improving the adaptability of the correction method to different cloud types and multiple cloud systems.

[0051] The game mechanism of the fractal dimension feature matching algorithm combined with the two-way optimization model of cloud base height makes the correction process pursue buoy station accuracy while maintaining the geometric structure consistency of large-scale sea areas, fundamentally improving the physical reasonableness and business application value of the correction results.

[0052] The principle of the tensor decomposition multi-dimensional correction algorithm is to regard the cloud base height correction problem as a multi-dimensional data analysis task, and to organize the multiple dimensional factors affecting the cloud base height deviation, including the longitude dimension and the latitude dimension of geographical position, the time dimension of time evolution, and the meteorological condition dimension of environmental conditions, into a four-dimensional correction tensor data structure. The four-dimensional correction tensor structure naturally retains the interaction and coupling effects between dimensions. The CANDECOMP decomposition method is used to decompose the four-dimensional correction tensor into a low-rank representation, extract the main variation mode and core interaction of each dimension, and the correction factor matrix group captures the dominant eigenvector within a single dimension, reflecting the independent contribution of the dimension to the cloud base height deviation. The correction core tensor represents the interaction strength between multiple dimensions, reflecting the synergistic effect of different dimensional factors. The optimal decomposition is iteratively solved by the alternating least squares optimization algorithm, realizing data compression and denoising while maintaining the main information of the original data. The low-rank decomposition naturally has a regularization effect that can suppress overfitting and noise interference.

[0053] The rationality of applying the tensor decomposition multi-dimensional correction algorithm to the scheme lies in that the satellite inversion error of the sea cloud base height is the result of the comprehensive action of multi-dimensional factors, the geographical position dimension influences the influence of the differences of the sea surface temperature, the water vapor content and the aerosol distribution in different sea areas on the radiation transmission, the time dimension influences the angle change of the solar radiation and the atmospheric stability change caused by the seasonal change and the daily change, the meteorological condition dimension influences the interference of the local meteorological elements such as the wind speed, the cloud amount and the cloud type on the cloud base height observation, the dimensional factors are not independent of each other but exist in a nonlinear coupling relationship, and the tensor decomposition method realizes the accurate description of the coupling relationship by correcting the core tensor to model the multi-dimensional interaction.

[0054] The specific implementation steps of the tensor decomposition multi-dimensional correction algorithm in the scheme are: first, a four-dimensional correction tensor data structure is constructed, the longitude dimension is divided into 0.5 degree grid from 105 degrees east to 135 degrees east, a total of 60 grid points, the latitude dimension is divided into 0.5 degree grid from 0 degrees north to 45 degrees north, a total of 90 grid points, the time dimension is divided into 1 hour time step from January 1, 2020 to December 31, 2024, a total of 43800 time steps, and the meteorological condition dimension is divided into cloud amount, wind speed, air temperature and relative humidity four elements, each element is divided into 10 levels, a total of 40 states, the cloud base height deviation value corresponding to each space-time position and meteorological state is filled into the corresponding position of the four-dimensional correction tensor to form a four-dimensional correction tensor of 60*90*43800*40, the missing data position is filled by adjacent interpolation or space-time Kriging interpolation, then the constructed four-dimensional correction tensor is subjected to CANDECOMP decomposition, the decomposition rank is set to 20 to represent the extraction of the first 20 dominant modes, the four correction factor matrices are initialized as random orthogonal matrices, and the alternating least squares algorithm is used for iterative optimization, in each iteration, three correction factor matrices are fixed to optimize the fourth correction factor matrix to minimize the Frobenius norm difference between the reconstructed tensor and the four-dimensional correction tensor, and the longitude correction factor matrix, the latitude correction factor matrix, the time correction factor matrix and the meteorological condition correction factor matrix are optimized in turn, the reconstruction error of the current iteration is calculated, and when the relative change rate of the reconstruction error is less than 0.001 or the maximum iteration number is reached, the optimization is stopped, and the final correction core tensor and correction factor matrix group are output, then the physical meaning of the correction factor matrix group is analyzed, each principal component of the longitude correction factor matrix reflects the cloud base height deviation pattern of different longitude zones, each principal component of the latitude correction factor matrix reflects the cloud base height deviation pattern of different latitude zones, each principal component of the time correction factor matrix reflects the seasonal period and daily period cloud base height deviation change law, and each principal component of the meteorological condition correction factor matrix reflects the influence weight of different meteorological element combinations on the cloud base height deviation, finally, the low rank representation obtained by decomposition is used for correction, for any given space-time position and meteorological condition, the corresponding correction factor matrix group element value and correction core tensor element value are extracted, the cloud base height deviation estimation value of the position is reconstructed by tensor product operation, and the cloud base height deviation estimation value is used as the correction amount to superimpose on the FY-4 satellite cloud base height data to obtain the corrected FY-4 satellite cloud base height data.

[0055] The tensor decomposition multi-dimensional correction algorithm brings the following technical effects to the entire FY-4 sea cloud base height correction scheme: the collaborative correction of multi-dimensional factors and the explicit modeling of interaction effects are realized, the limitations of traditional independent correction method by dimension are broken through, the dominant change pattern of the cloud base height deviation is naturally extracted by low rank tensor decomposition, the observation noise and random disturbance are filtered, and the stability and generalization ability of the correction result are enhanced.

[0056] The tensor decomposition multi-dimensional correction algorithm greatly reduces the number of model parameters by using the sparse representation characteristics of the correction factor matrix group. The original four-dimensional correction tensor contains more than 200 million elements, and the total parameter amount of the decomposed correction factor matrix group and correction core tensor is reduced to the order of hundreds of thousands, realizing efficient compression of data, significantly reducing storage requirements and computational overhead, and making statistical analysis of large-scale historical data and real-time business application possible.

[0057] The tensor decomposition multi-dimensional correction algorithm reveals the coupling law of cloud base height deviation under different spatiotemporal scales and meteorological conditions by capturing multi-dimensional interactions through the correction core tensor, providing a data-driven scientific basis for in-depth understanding of the physical mechanism of satellite inversion error.

[0058] The tensor decomposition multi-dimensional correction algorithm, the fractal dimension feature matching algorithm, and the cloud base height bidirectional optimization model form a complementary synergy. The tensor decomposition multi-dimensional correction algorithm provides macro multi-dimensional systematic correction, the fractal dimension feature matching algorithm provides micro geometric topological structure correction, and the cloud base height bidirectional optimization model provides fine local adaptive correction. The combination of the three constitutes a multi-level correction system from macro to micro, from statistics to physics, and from global to local, which comprehensively improves the accuracy and applicability of the Fengyun-4 satellite cloud base height data.

[0059] The Spark distributed computing framework is used for big data parallel processing, and the data sharding strategy is used to divide the full disc data into multiple sub-regions according to the longitude dimension and the latitude dimension. Each sub-region is allocated to an independent computing node for parallel processing, reducing the single machine computing pressure and processing time. The GPU acceleration is used for matrix operation and iterative solution process of the radiative transfer equation, which uses the parallel computing capability of the graphics processing unit to improve the calculation speed. The lookup table acceleration scheme is used to balance the calculation accuracy and efficiency. The radiative transfer eigenvalues under different combinations of atmospheric temperature profile, water vapor content, and aerosol optical depth are pre-calculated and stored as a multi-dimensional lookup table. During real-time processing, multi-dimensional linear interpolation is used for fast query to avoid repeated radiative transfer simulation calculation.

[0060] Optionally, the application also provides a Fengyun-4 sea cloud base height correction system formed by a computer implemented manner. The computer is provided with a storage medium, and the storage medium stores program instructions. The program instructions execute the above-mentioned Fengyun-4 sea cloud base height correction method considering the buoy cloud height instrument when running in the computer.

[0061] The specific implementation of the above steps is described in detail below.

[0062] The specific implementation of step S01 is to realize the synchronous acquisition of multi-source cloud base height data by establishing a sea-air-ground integrated data acquisition network. First, an array of buoy cloud height instruments is deployed in the target sea area. The station selection principle is to cover different latitude zones and different sea state characteristic areas. The station spacing is controlled within 100-200 km to ensure spatial representativeness. The buoy cloud height instrument records cloud base height data every minute using the laser pulse ranging principle and automatically uploads it to the data center. The cloud base height product data of the Fengyun-4 satellite scanning radiometer is synchronously acquired. According to the buoy position and observation time, the corresponding satellite pixel data is extracted. The time matching window is set to 15 minutes before and after to ensure the synchronization of the observation. The spatial matching window is set to a 5km by 5km area around the buoy station, and the average value is taken as the corresponding Fengyun-4 satellite cloud base height data. In addition, the microwave radiometer carried by the buoy is used to obtain brightness temperature data at different height layers and to invert temperature profile data. The vertical resolution of the temperature profile data is 50m from sea level to 10km height. At the same time, sea surface wind speed data is obtained from the satellite and the buoy for subsequent sea surface reflection correction. The purpose of the step is to establish the spatio-temporal correspondence between buoy point observation and satellite surface observation, and to provide paired sample data sets for subsequent statistical analysis and model training.

[0063] The specific implementation of step S02 is to eliminate the systematic interference of sea surface reflection on satellite cloud base height inversion based on radiation transfer theory. First, a sea surface bidirectional reflection distribution function model is constructed using sea surface wind speed data. The model uses the Cox-Munk sea surface slope probability distribution function to describe the sea surface roughness characteristics caused by wind waves, and combines the Fresnel reflection law to calculate the sea surface reflectivity at different incident angles and reflection angles. The model parameters include the solar zenith angle, the observation zenith angle, the relative azimuth angle, and the sea surface wind speed data. When the wind speed is greater than 7m / s, the Lambertian reflection contribution of white foam needs to be considered additionally. Next, a radiation transfer equation is established to decompose the total radiation brightness received by the satellite into three parts: the radiation term of the incident radiation at the top of the atmosphere after atmospheric attenuation reaching the cloud top, the radiation term of the cloud layer itself emission and reflection, and the radiation term of the sea surface reflection back to the atmosphere. The sea surface bidirectional reflection distribution function model is used to calculate the sea surface reflection contribution and subtract it from the total radiation brightness. An iterative optimization algorithm is used to solve the radiation transfer equation. The initial guess value is set to ignore the sea surface reflection. The cloud layer radiation parameters are updated and the residual error between the calculated total radiation brightness and the observed value is recalculated in each iteration. When the root mean square value of the residual error is less than the threshold value 0.5K or the iteration number exceeds 20, the iteration is stopped and the pure cloud layer radiation signal is output. The purpose of the step is to separate the pollution of sea surface reflection on satellite radiation observation from the physical mechanism, and to improve the accuracy of subsequent cloud base height inversion.

[0064] The specific implementation of step S03 is to identify the vertical structure of the cloud layer and realize the layered marking by using the microwave radiometer temperature profile data. First, the temperature profile data is preprocessed, including removing abnormal jump points and smoothing filtering, calculating the temperature gradient sequence between adjacent height layers, and determining the potential cloud layer boundary point when the absolute value of the temperature gradient is greater than the threshold value 0.6℃ / 100m, because the temperature distribution characteristics inside and outside the cloud layer are significantly different. Next, the brightness temperature difference value of each channel of the microwave radiometer is extracted as the cloud layer recognition feature, and the absorption and scattering characteristics of different frequencies of microwave are different for different height cloud layers, and the cloud layer signal is extracted by analyzing the spatial distribution pattern of the multi-channel brightness temperature difference value. The extracted brightness temperature difference value sequence is input into the K-means clustering algorithm for classification, and the K-means clustering algorithm maximizes the similarity of samples in the same category and maximizes the difference between different categories through iterative optimization, and the clustering number K is set to 3, corresponding to single-layer cloud, double-layer cloud and multi-layer cloud three cloud layer structure types. When the algorithm is initialized, three samples are randomly selected as the initial clustering center, the Euclidean distance of all samples to each clustering center is calculated, and each sample is assigned to the nearest clustering center, then the centroid of each category is recalculated as the new clustering center, and the above process is repeated until the clustering center position no longer changes or the change amplitude is less than the threshold value 0.01. According to the clustering category to which the sample belongs, the cloud layer number is marked for each observation time, and the single-layer cloud is marked as 1, the double-layer cloud is marked as 2, and the multi-layer cloud is marked as 3. The purpose of the step is to distinguish different cloud layer structure types to provide a basis for subsequent classification correction.

[0065] The specific implementation of step S04 is to quantify the spatial structure complexity characteristics of the cloud base height data through fractal geometry theory. First, the quality control is performed on the time series of the cloud base height measured data, and the abnormal values caused by precipitation, haze or instrument failure are removed. The abnormal value determination standard is that the data point deviates from the 30-minute moving average value by more than 3 times the standard deviation. For missing data, the cubic spline interpolation method is used to fill in to ensure the continuity of the time series. The one-dimensional box counting method is applied to the processed time series, and the box scale sequence is set as 16 minutes, 8 minutes, 4 minutes, 2 minutes and 1 minute, a total of 5 scales. For each box scale, the time axis is divided into several intervals, and the minimum number of boxes required to cover the cloud base height change range is counted. The number of boxes and the box scale satisfy the power law relationship. At the same time, the spatial distribution map of the Fengyun-4 satellite cloud base height data is processed, the cloud base height pixel data within a range of 50 km from the center of the buoy station is extracted and resampled to a uniform grid with a resolution of 1 km, and the two-dimensional box counting method is applied. The box scale sequence is set as 16 km, 8 km, 4 km, 2 km and 1 km. For each box scale, the spatial plane is divided into a grid, and the minimum number of boxes required to cover all non-zero cloud base height pixels is counted. The scatter plot of the logarithm of the number of boxes and the logarithm of the reciprocal of the box scale of the time series and the spatial distribution map is drawn respectively, and the linear regression fitting is performed by the least squares method. The slope of the fitted straight line is the fractal dimension value. The fractal dimension value of the time series is denoted as the buoy fractal dimension, and the fractal dimension value of the spatial distribution map is denoted as the satellite fractal dimension. The difference between the two is the fractal dimension difference parameter. Finally, the statistical correlation between the fractal dimension difference parameter and the cloud base height deviation is established, the fractal dimension difference parameter and the corresponding cloud base height deviation of all paired samples are collected, and the nonlinear regression fitting is performed by using a cubic polynomial function to obtain the mapping function. The role of the step is to establish the structural similarity measurement of the buoy point observation and the satellite surface observation from the geometric topology angle, and to provide a quantitative index for identifying the systematic deviation under different observation scales.

[0066] The specific implementation of step S05 is to realize the low-rank representation and pattern extraction of the multi-dimensional cloud base height deviation by using the tensor algebra method. First, a four-dimensional correction tensor data structure is constructed. The longitude dimension is divided into 60 grids with an interval of 0.5 degrees from 105 degrees east to 135 degrees east, the latitude dimension is divided into 90 grids with an interval of 0.5 degrees from 0 degrees north to 45 degrees north, the time dimension is divided into 43800 time steps with an interval of 1 minute from January 1, 2020 to December 31, 2024, and the meteorological condition dimension includes cloud cover, wind speed, air temperature and relative humidity, each of which is divided into 10 levels to form a 40-dimensional state space. The difference between the measured cloud base height data corresponding to each space-time position and meteorological state and the Fengyun-4 satellite cloud base height data is filled into the corresponding position of the four-dimensional correction tensor. For the position with missing data, the distance weighted interpolation method is used to fill the values according to the adjacent non-missing positions. The four-dimensional correction tensor constructed is subjected to CANDECOMP decomposition. The decomposition method approximately represents the four-dimensional tensor as a linear combination of multiple rank-1 tensors, each of which is composed of the outer product of four vectors corresponding to the characteristic patterns of the four dimensions. The decomposition rank is set to 20, which means that the first 20 dominant patterns can retain more than 95% of the information of the original data. The alternating least squares algorithm is used to optimize the decomposition parameters. The algorithm alternately fixes three correction factor matrices to optimize the fourth correction factor matrix to minimize the Frobenius norm difference between the reconstructed tensor and the original four-dimensional correction tensor. The Frobenius norm is the square root of the sum of the squares of all elements of the tensor. In the iterative optimization process, the longitude correction factor matrix, the latitude correction factor matrix, the time correction factor matrix and the meteorological condition correction factor matrix are updated in turn. When the relative change rate of reconstruction error of two consecutive iterations is less than 0.001, it is determined to be converged. The final correction core tensor and correction factor matrix group are output to represent the main variation law of the cloud base height deviation in different dimensions and the interaction effect between dimensions. The purpose of the step is to realize data dimension reduction and denoising by low-rank tensor decomposition, and extract the multi-dimensional cooperative variation pattern of the cloud base height deviation.

[0067] The specific implementation of step S06 is to construct a deep learning model with a game mechanism to realize adaptive fine correction of cloud base height. First, the correction core tensor element value, the principal component value of the correction factor matrix group, the fractal dimension difference value parameter and the cloud layer number identifier are input into the input layer of the model. The input layer normalizes the data to uniformly map the numerical range to 0 to 1. High-dimensional feature representation is extracted through the feature fusion layer. The feature fusion layer uses a fully connected neural network structure to map the input to a 128-dimensional feature vector. The feature vector enters the game layer for bidirectional optimization. The game layer includes two parallel branches: an accuracy optimization sub-model and a stability optimization sub-model. The accuracy optimization sub-model calculates the degree of fit between the measured cloud base height data and the model predicted value through a three-layer fully connected network and outputs an accuracy weight vector. The stability optimization sub-model extracts the continuity features of adjacent spatio-temporal positions through a two-dimensional convolutional neural network and outputs a stability weight vector. The accuracy weight vector and the stability weight vector are linearly combined by the weighting fusion layer according to the learnable weight coefficients to generate the final correction weight. After the final correction weight and the 128-dimensional feature vector are multiplied element by element, the corrected FY-4 satellite cloud base height data is obtained through the linear transformation of the output layer. The root mean square error of the corrected FY-4 satellite cloud base height data and the measured cloud base height data is calculated. When the root mean square error is less than the threshold value of 50 m, it is determined that the correction effect meets the requirements and the result is output. When the root mean square error is greater than the threshold value of 50 m, the multi-head attention weight parameter is adjusted. The adjustment method is to redistribute the attention weight according to the relative size of the fractal dimension difference value parameter, the cloud layer number identifier and the correction core tensor element value, and to re-execute this step until the accuracy requirement is met. The role of the step is to realize high-precision adaptive correction of cloud base height through the powerful fitting ability of the deep learning model and the dynamic balance characteristics of the game mechanism.

[0068] It should be noted that one of the key technical ideas of the present application is that the fractal dimension feature matching algorithm establishes a quantitative correlation between buoy point observation and satellite surface observation from the perspective of geometric topology, breaking through the limitation of traditional methods that rely only on radiation features for correction. Fractal dimension, as a scale-invariant geometric quantity, can stably describe the spatial complexity of the cloud field and is not affected by observation time and lighting conditions, providing a unified and comparable benchmark for different observation methods. By establishing a statistical model of the fractal dimension difference and the cloud base height deviation, the mapping from geometric structure to physical deviation is realized, especially for multi-cloud layers and boundary ambiguity, which can comprehensively reflect the overall structural characteristics and avoid cumulative errors of layer-by-layer identification.

[0069] The second key technical idea is to organize the multi-dimensional factors affecting the cloud base height bias, such as geographical location, time evolution and meteorological conditions, into a four-dimensional tensor structure through tensor decomposition. The low-rank decomposition naturally extracts the dominant change patterns in each dimension and the interaction effects between dimensions. Compared with the traditional independent correction method, the coupling relationship between dimensions can be explicitly modeled. The sparse representation of the factor matrix compresses the original data from the order of 200 million parameters to the order of hundreds of thousands, achieving efficient storage and fast calculation. The regularization characteristics of low-rank decomposition naturally suppress observation noise and random disturbances, improving the stability of the correction results.

[0070] The third key technical idea is the game layer mechanism embedded in the cloud base height two-way optimization model. By building a competitive and cooperative relationship between the precision optimization sub-model and the stability optimization sub-model, the contradiction between precision improvement and stability maintenance is fundamentally solved. The precision optimization sub-model drives the correction results to approach the buoy measured values, and the stability optimization sub-model constrains the correction results to maintain spatial and temporal continuity. The two sub-models establish a gradient transmission channel through the coupling term and dynamically balance during the training process. When one side is over-optimized, the coupling term will automatically adjust the optimization direction, so that the model can pursue local precision without producing non-physical mutations, and pursue global smoothness without losing detail description ability.

[0071] The synergistic effect of the above three key technical ideas forms a multi-level correction system from macro to micro, from statistics to physics, and from global to local. Tensor decomposition provides a macro correction basis for multi-dimensional systematic bias, fractal dimension characteristics provide a physical explanation of observation scale differences from the perspective of geometric topology, and the two-way optimization model realizes fine and adaptive correction considering multiple constraints. The three complement each other to overcome the limitations of a single method. The low-rank representation of tensor decomposition provides compact input features for the two-way optimization model, reducing the difficulty of model training. Fractal dimension characteristics provide physical prior knowledge for the model's attention mechanism, enhancing the model's interpretability. The game mechanism of the two-way optimization model ensures that the correction provided by tensor decomposition and fractal dimension characteristics is consistent at the global and local levels, ultimately achieving high-precision and stable correction of FY-4 sea cloud base height data in different sea areas, different seasons, and different cloud type conditions.

[0072] It needs to be explained that the present application also solves the technical problems that the sea cloud bottom height satellite inversion is affected by the sea surface reflection pollution. The specular reflection of the sea surface to the solar radiation and the atmospheric downward radiation and the diffuse reflection of the white cap foam are superimposed into the cloud layer radiation signal, so that the total radiation brightness received by the satellite contains the contributions of the cloud layer radiation and the sea surface reflection, and the traditional inversion algorithm is difficult to effectively separate the two, causing the cloud bottom height inversion result to have systematic deviation. The present application quantifies the reflection characteristics of the sea surface to the incident radiation by establishing a sea surface bidirectional reflection distribution function model, inputs the solar zenith angle, observation zenith angle, relative azimuth angle and sea surface wind speed data to calculate the sea surface reflectivity distribution at different angles, uses the radiation transfer equation to decompose the total radiation brightness received by the satellite into the cloud layer radiation contribution term and the sea surface reflection contribution term, calculates the theoretical value of the sea surface reflection contribution term by iterative optimization and deducts it from the total radiation brightness, realizes the decoupling separation of the cloud layer radiation signal and the sea surface reflection signal, and effectively eliminates the interference of the sea surface reflection pollution on the cloud bottom height inversion.

[0073] Specifically, the principle of the present application is that the root of the difficulty in balancing the precision and stability in cloud bottom height correction lies in the scale difference between single-point observation and area observation and the limitation of single-target optimization. The present application establishes a dynamic balance relationship between the precision target and the stability target through the game mechanism of the bidirectional optimization model, the precision optimization sub-model focuses on the local correction precision at the buoy station, the stability optimization sub-model focuses on the spatio-temporal continuity of the large-scale sea area, and the two sub-models realize parameter correlation and gradient transmission through the coupling term, and form a mutually restraining game state through alternating optimization in the training process. The coupling term constraint makes that the precision improvement will not cause spatio-temporal mutation, and the spatio-temporal smoothing will not lose local details, and the model automatically finds the optimal parameter configuration that the precision loss value and the stability loss value converge at the same time. The fractal dimension feature matching algorithm quantifies the scale difference between the buoy observation and the satellite observation from the geometric topology angle, the four-dimensional tensor decomposition extracts the dominant change mode and the interaction effect of multi-dimensional factors, and provides physical constraints and prior knowledge for the bidirectional optimization model, and the three cooperates to form a multi-level correction system, so that the correction result meets the physical rationality.

[0074] A specific embodiment 1 of the present application is provided below, and the specific implementation of step S01 in the embodiment 1 is the same as the foregoing, and will not be described in detail here. The specific implementation of other steps is described in detail as follows.

[0075] The specific implementation of step S02 is that the formula of the sea surface bidirectional reflection distribution function model is as follows:

[0076] ;

[0077] In the formula, is the normalized sea surface bidirectional reflection distribution function value, dimensionless; is the observed sea surface bidirectional reflection distribution function, unit is ; is the standard reference reflectivity, unit is , the experience value is 0.02 ; is the solar zenith angle, unit is degree; is the observation zenith angle, unit is degree; is the relative azimuth angle, unit is degree; is the sea surface wind speed, unit is .

[0078] The formula of decoupling the cloud radiation signal and the sea surface reflection signal by the radiative transfer equation is as follows:

[0079] ;

[0080] ;

[0081] In the formula, is the total radiation brightness received by the satellite, unit is ; is the cloud radiation contribution item, unit is ; is the sea surface reflection contribution item, unit is , which is obtained by calculating the sea surface bidirectional reflection distribution function model.

[0082] The specific implementation of step S03 is to first calculate the normalized brightness temperature difference value of each height layer, and the formula is as follows:

[0083] ;

[0084] In the formula, is the normalized brightness temperature difference value of the i-th height layer, dimensionless; is the brightness temperature of the i-th height layer, unit is ; is the brightness temperature of the i-th height layer, unit is ; is the brightness temperature of the i-th height layer, unit is ; is the reference temperature, usually 280 . .

[0085] The objective function formula of the K-means clustering algorithm for separating the cloud radiation signal is as follows:

[0086] ;

[0087] In the formula, is the clustering objective function value, dimensionless; is the total number of samples; is the i-th sample, unit is the normalized brightness temperature difference value of the i-th sample, dimensionless; , dimensionless; is the i-th cluster center, dimensionless; is the membership degree of the i-th sample belonging to the j-th cluster, dimensionless, ranging from 0 to 1; is the Euclidean distance; is the maximum value of the normalized brightness temperature difference value among all samples, dimensionless; is the minimum value of the normalized brightness temperature difference value among all samples, dimensionless.

[0088] The specific implementation of step S04 is that the formula for extracting the fractal dimension feature by the box counting method is as follows:

[0089] ;

[0090] In the formula, is the fractal dimension value, dimensionless; is the minimum number of boxes required to cover the cloud base data when the box size is ; is the reference box number, usually taking a value of 100; is the i-th box size, with the unit of ; is the reference box size, usually taking a value of 1 ; ; is the number of box sizes, which is set to 5 in this scheme.

[0091] The formula for the fractal dimension difference value parameter is as follows:

[0092] ;

[0093] In the formula, is the fractal dimension difference value parameter, dimensionless; is the buoy fractal dimension, dimensionless; is the satellite fractal dimension, dimensionless; is the reference fractal dimension, with an empirical value of 1.5.

[0094] The formula for the statistical correlation between the fractal dimension difference value parameter and the cloud base height deviation is as follows:

[0095] ;

[0096] In the formula, is the cloud base height deviation, with the unit of ; is the undetermined coefficient, dimensionless, obtained by regression fitting;​​​ is a height scale normalization coefficient, unit is , usually takes 1000 .

[0097] The specific implementation of step S05 is that the formula of the CANDECOMP decomposition method for decomposing the four-dimensional correction tensor is as follows:

[0098] ;

[0099] In the formula, is a four-dimensional correction tensor, and the element unit is ; is a decomposition rank, which is set to 20 in the present solution; is a correction core tensor scalar value of the mth mode, unit is ; is a longitude dimension correction factor vector of the mth , dimensionless; is a latitude dimension correction factor vector of the mth , dimensionless; is a time dimension correction factor vector of the mth , dimensionless; is a meteorological condition dimension correction factor vector of the mth , dimensionless; is a vector outer product operator.

[0100] The objective function formula of the alternating least squares algorithm optimization is as follows:

[0101] ;

[0102] In the formula, is a normalized reconstruction error function value, dimensionless; is a reconstruction tensor, and the element unit is ; is a Frobenius norm.

[0103] The specific implementation of step S06 is that the formula of the multi-head attention weight parameter is as follows:

[0104] ;

[0105] In the formula, is a multi-head attention weight parameter of the mth sample, dimensionless; is a weight coefficient, dimensionless, and the empirical values are 1.2, 0.8, and 1.5, respectively; is a fractal dimension difference value parameter of the mth sample, dimensionless; For the first The number of cloud layers per sample, dimensionless; For the first The value of the calibration core tensor element corresponding to each sample, in units of ; For the number of heads; This is the weight normalization scale, with units of . The value is usually 50. .

[0106] The objective function formula for the accuracy optimization sub-model is expressed as follows:

[0107] ;

[0108] In the formula, This is a dimensionless value representing the accuracy loss. The number of samples; For the first Measured cloud base height data for each sample, in units of ; For the first The corrected cloud base height output by each sample model, in units of ; is the fractal dimension weighting coefficient, dimensionless, with an empirical value of 2.5; For the first The fractal dimension difference parameter of each sample, dimensionless; For the first The number of cloud layers per sample, dimensionless; This is the height-scale normalization coefficient, in units of... The value is usually 1000. .

[0109] The objective function of the stability optimization sub-model is expressed as follows:

[0110] ;

[0111] In the formula, This is the stability loss value, which is dimensionless. The standard deviation of the cloud base height difference between adjacent time points, in units of ; The standard deviation of the cloud base height difference between adjacent spatial locations, in units of ; This is the meteorological condition weighting coefficient, dimensionless, with an empirical value of 0.8; The rate of change of meteorological conditions is dimensionless and represents the change in normalized meteorological parameters per unit time. This is the height-scale normalization coefficient, in units of... , usually taking the value of 1000 .

[0112] The formula of the coupling term loss function is as follows:

[0113] ;

[0114] wherein, is a coupling term loss value, dimensionless; is the i-th element of the precision weight vector, dimensionless; is the i-th element of the stability weight vector, dimensionless; is the i-th element of the stability weight vector, dimensionless; is the dimension of the weight vector. It should be noted that the variables involved in the present embodiment are explained in detail as shown in Table 1, Table 2 and Table 3.

[0115] Table 1 Variable Explanation Table (First Part)

[0116]

[0117] Table 2 Variable Explanation Table (Second Part)

[0118]

[0119] Table 3 Variable Explanation Table (Third Part)

[0120]

[0121]

[0122] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: In order to verify the effect of the present application, the technical personnel set up a test environment, through collecting the cloud height instrument observation data of 12 buoy stations in a target sea area from June to August in 2024 and the corresponding wind cloud four satellite cloud base height data, to carry out sea cloud base height correction test. The sea area mainly covers the northwest Pacific region, and the test period is mainly summer monsoon cloud system, and the cloud type includes mixed distribution of stratocumulus, cumulus and cirrus. The test data includes buoy cloud height instrument measured cloud base height, FY-4 satellite inversion cloud base height, microwave radiometer temperature profile and sea surface wind speed data.

[0123] ​The skilled person first performs sea surface reflection pollution correction on the FY-4 satellite cloud base height data. A sea surface bidirectional reflection distribution function model is established by collecting sea surface wind speed data, and the sea surface reflection distribution function value is calculated as 0.032 by inputting the solar zenith angle as 35 degrees, the observation zenith angle as 42 degrees, the relative azimuth angle as 78 degrees, and the measured sea surface wind speed as 8.5 m / s. The satellite received total radiation brightness is decomposed into cloud layer radiation contribution and sea surface reflection contribution by using the radiation transfer equation, and the cloud layer radiation contribution is iteratively solved. The sea surface reflection contribution is subtracted from the total radiation brightness to obtain the corrected cloud layer radiation signal. After 5 iterations of optimization, the sea surface reflection signal is effectively removed, and the signal-to-noise ratio of the cloud layer radiation signal is improved to 42 dB.

[0124] The skilled person identifies the cloud layer structure in combination with the microwave radiometer temperature profile data. The temperature gradient mutation points in the temperature profile are extracted as the cloud layer boundary features, and three significant temperature gradient mutation points are identified at heights of 1.2 km, 2.8 km and 5.6 km, respectively. The brightness temperature difference value sequence of each height layer is calculated. The brightness temperature difference value sequence is input into the K-means clustering algorithm, and the number of clusters is set to 3 corresponding to single-layer cloud, double-layer cloud and multi-layer cloud three types. The cluster center position is optimized by 15 iterations to achieve convergence. The single-layer cloud sample proportion in the test data set is 46%, the double-layer cloud sample proportion is 38%, and the multi-layer cloud sample proportion is 16%, as shown in Table 4. According to the cluster center category of sample attribution, the number of cloud layers is marked as an identifier, which provides cloud layer structure parameters for subsequent correction.

[0125] Table 4 Cloud type distribution statistics table

[0126]

[0127] The skilled person extracts the multi-scale fractal dimension features of the cloud base height measured data and the FY-4 satellite cloud base height data using the box counting method. The cloud base height measured data time series is divided into grids, and the initial box scale is set to 1 km. The minimum number of boxes required to cover the cloud base data is 158. The box scale is gradually reduced to 0.5 km, 0.25 km and 0.125 km, and the corresponding number of boxes is 312, 625 and 1247, respectively. The spatial distribution map of the FY-4 satellite cloud base height data is also subjected to box counting statistics. The initial box scale of 1 km corresponds to 192 boxes, and the number of boxes after reducing the box scale is 378, 741 and 1502, respectively. The linear regression slope of the logarithm of the number of boxes and the logarithm of the reciprocal of the box scale is calculated. The fractal dimension value of the ceilometer measured data is 1.58, the fractal dimension value of the FY-4 satellite data is 1.73, and the fractal dimension difference parameter is 0.15. As shown in FIG. 6, the fractal dimension difference parameter is 0.15. Figure 2As shown, the difference parameters of the fractal dimension of different cloud types have significant differences. The average value of the difference parameter of the fractal dimension of stratus cloud is 0.08, the average value of the difference parameter of the fractal dimension of cumulus cloud is 0.18, and the average value of the difference parameter of the fractal dimension of cirrus cloud is 0.13. The statistical correlation between the difference parameter of the fractal dimension and the deviation of the cloud base height is established, and a cubic polynomial function is used for nonlinear regression fitting, and the fitting determination coefficient (R^2) reaches 0.87, indicating that there is a strong correlation between the difference parameter of the fractal dimension and the deviation of the cloud base height.

[0128] The cloud base height correction data is organized by the technician into a four-dimensional correction tensor in the longitude dimension, latitude dimension, time dimension, and meteorological condition dimension. The longitude dimension is divided into 0.5 degree grids, a total of 60 grid points, the latitude dimension is divided into 0.5 degree grids, a total of 90 grid points, the time dimension is divided into 1 hour time steps, a total of 2208 time steps, and the meteorological condition dimension is divided into 40 states, constructing a 60x90x2208x40 four-dimensional correction tensor. The CANDECOMP decomposition method is used to extract the low-rank structure of the four-dimensional correction tensor, and the decomposition rank is set to 20. Initialize the four correction factor matrices as random orthogonal matrices, and optimize the decomposition through the alternating least squares algorithm. The reconstruction error of the first iteration is 0.342, and after 128 iterations, the reconstruction error converges to 0.018, and the relative change rate of the reconstruction error decreases to 0.0008, which is less than the threshold 0.001, and the iteration is stopped. Output the final correction core tensor and correction factor matrix set. Analyze the physical meaning of the correction factor matrix set, the first principal component of the longitude correction factor matrix reflects the difference in cloud base height deviation between the eastern and western sea areas, the first principal component of the latitude correction factor matrix reflects the change in cloud base height deviation between low and high latitudes, the first principal component of the time correction factor matrix reflects the influence of the day-night cycle and the seasonal cycle, and the first principal component of the meteorological condition correction factor matrix reflects the comprehensive effect of wind speed and cloud cover.

[0129] The technician inputs the correction core tensor, correction factor matrix set, fractal dimension difference parameter, and cloud layer number identifier into the cloud base height bidirectional optimization model for correction optimization. The input layer of the cloud base height bidirectional optimization model receives the correction core tensor element value 8.6, the correction factor matrix set principal component values 0.72, 0.58, 0.63, 0.81, respectively, the fractal dimension difference parameter 0.15, and the cloud layer number identifier 2, and maps them into a 128-dimensional feature vector through the feature fusion layer. Enter the game layer for bidirectional optimization, the precision optimization sub-model outputs a precision weight vector through a three-layer fully connected network, and the precision loss value is 0.035. The stability optimization sub-model extracts the spatiotemporal continuity feature through a convolutional neural network to output a stability weight vector, and the stability loss value is 0.028. As Figure 3As shown, the precision weight vector and the stability weight vector are synthesized into the final correction weight through the weighted fusion layer, and the weighted coefficients are 0.62 and 0.38 respectively. After the final correction weight is multiplied with the 128-dimensional feature vector, the corrected FY-4 satellite cloud bottom height data is generated through the output layer. The root mean square error of the corrected FY-4 satellite cloud bottom height data and the measured cloud bottom height data is 38m, which is less than the threshold 50m and meets the accuracy requirement. The test results show that the root mean square error of the corrected single-layer cloud scene is 32m, the root mean square error of the corrected double-layer cloud scene is 41m, and the root mean square error of the corrected multi-layer cloud scene is 47m, as shown in Table 5.

[0130] Table 5 Correction accuracy statistics table of different cloud layer types

[0131]

[0132] The progress of the present application relative to the traditional sea cloud bottom height correction method mainly lies in the following aspects. The traditional method mainly relies on the radiation transfer model and the atmospheric profile parameter for correction, and the accuracy of the atmospheric temperature and humidity profile is relatively high. When the profile data has errors or is missing, the correction effect will decrease significantly. The present application introduces a fractal dimension feature matching algorithm to establish a physical correlation between buoy observation and satellite observation from the geometric topological structure of the cloud field. The fractal dimension, as a scale-invariant geometric feature, is not affected by observation time and lighting conditions, providing a unified correction benchmark for data under different observation conditions. The traditional method usually uses a dimension-by-dimension independent correction strategy to handle multi-dimensional factors, ignoring the coupling effects between geographical location, time evolution and meteorological conditions. The present application uses a tensor decomposition multi-dimensional correction algorithm to organize multiple dimensional factors into a four-dimensional tensor structure. The CANDECOMP decomposition explicitly models the multi-dimensional interaction, and the coupling relationship captured by the correction core tensor reveals the collaborative rules of cloud bottom height bias at different spatiotemporal scales, achieving a deep understanding from statistical characteristics to physical mechanisms. The traditional method often leads to a decline in spatiotemporal continuity when pursuing correction accuracy, and loses the ability to depict local details when pursuing spatiotemporal smoothing. The present application builds a dynamic balance between the precision optimization sub-model and the stability optimization sub-model through the game layer mechanism of the cloud bottom height bidirectional optimization model. The precision optimization sub-model ensures high-precision fitting at the buoy station, and the stability optimization sub-model ensures physical rationality in the non-buoy coverage area. The two sub-models establish parameter sharing and gradient transmission channels through the coupling term, and find the optimal balance point between precision and stability in the parameter space, fundamentally solving the contradiction between precision improvement and stability maintenance.

[0133] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, characterized in that, Measured cloud base height data from multiple buoy ceilometers within the target sea area and corresponding Fengyun-4 satellite cloud base height data were collected. Sea surface reflectance pollution correction was applied to the Fengyun-4 satellite cloud base height data. Cloud structure was identified and cloud quantity markers were assigned using microwave radiometer temperature profile data. Multi-scale fractal dimension features of the measured cloud base height data and Fengyun-4 satellite cloud base height data were extracted using box counting. The difference between the buoy fractal dimension and the satellite fractal dimension was calculated as the fractal dimension difference parameter. The cloud base height correction data was organized into a four-dimensional correction tensor according to longitude, latitude, time, and meteorological conditions. The CANDECOMP decomposition method was used to extract the low-rank structure of the four-dimensional correction tensor to obtain the correction core tensor and correction factor matrix set. The following steps are performed: First, the values ​​of the core correction tensor elements, the principal component values ​​of the correction factor matrix group, the fractal dimension difference parameters, and the cloud layer quantity identifier are input into the input layer of the bidirectional optimization model for cloud base height. The input layer normalizes the data to map the numerical range to between 0 and 1. After the feature fusion layer extracts high-dimensional feature representations, the feature fusion layer uses a fully connected neural network structure to map the input into a 128-dimensional feature vector. feature The vectors are fed into a game theory layer for bidirectional optimization. This layer comprises two parallel branches: a precision optimization sub-model and a stability optimization sub-model. The precision optimization sub-model uses a three-layer fully connected network to calculate the degree of agreement between the measured cloud base height data and the model's predicted values, and outputs a precision weight vector. The stability optimization sub-model uses a two-dimensional convolutional neural network to extract continuous features of adjacent spatiotemporal locations and outputs a stability weight vector. The precision weight vector and the stability weight vector are linearly combined through a weighted fusion layer according to learnable weight coefficients to generate the final correction weights. The final correction weights are then multiplied element-wise by the 128-dimensional feature vector and linearly transformed by the output layer to obtain the corrected Fengyun-4 satellite cloud base height data. Next, the root mean square error (RMSE) between the corrected Fengyun-4 satellite cloud base height data and the measured cloud base height data is calculated. If the RMSE is less than the threshold of 50m, the correction effect is deemed satisfactory and the result is output. If the RMSE is greater than the threshold of 50m, the multi-head attention weight parameters are adjusted. The adjustment method involves redistributing the attention weights based on the fractal dimension difference parameter, the cloud layer number identifier, and the relative magnitude of the correction core tensor element values, and this step is repeated until the precision requirements are met.

2. The method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 1, is characterized in that... The sea surface reflection pollution correction includes: establishing a two-way reflection distribution function model of the sea surface using sea surface wind speed data, and decoupling cloud radiation signals and sea surface reflection signals using the radiative transfer equation.

3. The method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 2, is characterized in that... The inputs to the sea surface two-way reflectance distribution function model are the solar zenith angle, the observed zenith angle, the relative azimuth angle, and the sea surface wind speed data, and the outputs are the sea surface reflectance distribution function values ​​at different angles.

4. The method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 3, is characterized in that... The decoupling of the radiative transfer equation includes: decomposing the total radiance received by the satellite into the sum of cloud radiation contribution and sea surface reflection contribution, and subtracting the sea surface reflection contribution from the total radiance through iterative optimization.

5. A method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 4, is characterized in that... The cloud structure identification uses the K-means clustering algorithm to separate cloud radiation signals into single-layer cloud radiation signal sets, double-layer cloud radiation signal sets, and multi-layer cloud radiation signal sets.

6. The method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 5, is characterized in that... The K-means clustering algorithm extracts abrupt changes in temperature gradient from microwave radiometer temperature profile data as cloud boundary features, with the number of clusters set to 3.

7. A method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 6, is characterized in that... The box counting method sets the initial box size to 1km, counts the minimum number of boxes required to cover the cloud base data, and gradually reduces the box size to 0.125km.

8. The method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 7, is characterized in that... The multi-scale fractal dimension feature is calculated using the slope of a linear regression between the logarithm of the number of boxes and the inverse logarithm of the box scale as the fractal dimension value.

9. A method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 8, is characterized in that... The statistical correlation between the fractal dimension difference parameter and the cloud base height deviation was established using cubic polynomial regression fitting.

10. A method for correcting the cloud bottom height at sea using the Fengyun-4 buoy-based ceilometer, as described in claim 9, is characterized in that... The CANDECOMP decomposition method employs an alternating least squares algorithm to optimize the decomposition, fixing three correction factor matrices and optimizing the fourth correction factor matrix to minimize the reconstruction error.

Images