CN117932343A - Satellite high-frequency sea surface wind estimation method - Google Patents
Satellite high-frequency sea surface wind estimation method Download PDFInfo
- Publication number
- CN117932343A CN117932343A CN202410120101.6A CN202410120101A CN117932343A CN 117932343 A CN117932343 A CN 117932343A CN 202410120101 A CN202410120101 A CN 202410120101A CN 117932343 A CN117932343 A CN 117932343A
- Authority
- CN
- China
- Prior art keywords
- data
- amv
- neural network
- connected neural
- sea surface
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 40
- 238000013528 artificial neural network Methods 0.000 claims abstract description 76
- 238000012549 training Methods 0.000 claims abstract description 18
- 230000004069 differentiation Effects 0.000 claims abstract description 15
- 238000007781 pre-processing Methods 0.000 claims abstract description 8
- 238000012360 testing method Methods 0.000 claims description 17
- 230000000694 effects Effects 0.000 claims description 14
- 238000012417 linear regression Methods 0.000 claims description 13
- 238000004364 calculation method Methods 0.000 claims description 10
- 241000713838 Avian myeloblastosis virus Species 0.000 claims description 7
- 210000002569 neuron Anatomy 0.000 claims description 7
- 230000001364 causal effect Effects 0.000 claims description 5
- 230000001419 dependent effect Effects 0.000 claims description 5
- 230000004913 activation Effects 0.000 claims description 4
- 230000001902 propagating effect Effects 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 3
- 239000000523 sample Substances 0.000 description 19
- 230000006870 function Effects 0.000 description 13
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 238000003062 neural network model Methods 0.000 description 5
- 238000005070 sampling Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 3
- 230000000644 propagated effect Effects 0.000 description 3
- 230000006978 adaptation Effects 0.000 description 2
- 238000012937 correction Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000002093 peripheral effect Effects 0.000 description 2
- 230000009897 systematic effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008033 biological extinction Effects 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 238000011065 in-situ storage Methods 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000007935 neutral effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 239000013074 reference sample Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000003746 surface roughness Effects 0.000 description 1
- 230000005570 vertical transmission Effects 0.000 description 1
Landscapes
- Aerodynamic Tests, Hydrodynamic Tests, Wind Tunnels, And Water Tanks (AREA)
Abstract
The invention provides a satellite high-frequency sea surface wind estimation method, which relates to the field of sea surface wind inversion methods, and comprises the following steps: acquiring AMV data, and preprocessing the AMV data to obtain AMV sample data; referencing the physical frame according to the AMV sample data to obtain AWM feature data; inputting AMV characteristic data into a fully-connected neural network, and training the fully-connected neural network; the input data are transmitted to each layer of the fully-connected neural network, gradient descent is realized through numerical differentiation, and network parameters are updated; and estimating by using the fully-connected neural network to obtain an estimation result. The invention provides a satellite high-frequency sea surface wind estimation method, which solves the problems that the model physical interpretation is weak, the cost function constraint is single, and the model parameter iteration is difficult to be guided comprehensively under a physical framework.
Description
Technical Field
The invention relates to the field of sea surface wind inversion methods, in particular to a satellite high-frequency sea surface wind estimation method.
Background method
Atmospheric wind guiding (AMV; atmospheric Motion Vector) provides high-value, high-frequency and large-range atmospheric wind vector distribution information for the ocean area; sea Surface Wind (SSW) is an important physical parameter in the Sea and the atmosphere, directly influences the motion of the upper layer of the Sea, has irreplaceable effects in hydrologic and energy circulation and global and local climates, is directly related to ship navigation, influences the ship navigation, and has important significance in social, economic and other aspects; the Sea Surface Wind obtained by AMV calculation is called Sea Surface Wind based on atmospheric Wind guide (ASWind; AMV-based Sea-Surface Wind), ASWind can provide Sea Surface Wind field information of tropical cyclone spiral cloud atmospheric Wind for us in more detail; .
The arrival of the big data age enables the deep learning method based on the artificial neural network to be widely applied to various disciplinary fields, is particularly suitable for solving the nonlinear problem, is suitable for the deep learning method due to the characteristics of large quantity, multiple types and high precision of satellite remote sensing data, is commonly used for satellite remote sensing image processing by using the convolutional neural network (CNN; convolutional Neural Network) as a representative, and provides a powerful method means for cloud identification, typhoon intensity prediction, sea surface flow, sea surface wind field inversion and the like in the satellite remote sensing field; although the CNN and a series of deep learning models derived from the CNN can fully consider the spatial characteristics of the image data to realize the 'face-to-face' calculation, the CNN is not suitable for highly discrete data, and is a choice for the traditional fully-connected neural network.
The self-adaptive, self-learning and nonlinear characteristics of the Fully connected neural network (FCNN; fully-Connected Neural Network) enable the Fully connected neural network to establish a highly nonlinear relationship between the low-level AMV and the SSW, but the establishment of the neural network between the AMV and the SSW only causes weak physical interpretation of the model, single constraint of a cost function and difficulty in comprehensively guiding parameter iteration of the model under a physical framework.
Disclosure of Invention
The invention provides a satellite high-frequency sea surface wind estimation method, which solves the problems that the model physical interpretation is weak, the cost function constraint is single, and the model parameter iteration is difficult to be guided comprehensively under a physical framework.
In order to solve the problems of the method, the scheme of the method is as follows:
A method of satellite high frequency sea surface wind estimation, the method comprising:
acquiring AMV data, and preprocessing the AMV data to obtain AMV sample data;
referencing the physical frame according to the AMV sample data to obtain AWM feature data;
inputting AMV characteristic data into a fully-connected neural network, and training the fully-connected neural network;
the input data are transmitted to each layer of the fully-connected neural network, gradient descent is realized through numerical differentiation, and network parameters are updated;
and estimating by using the fully-connected neural network to obtain an estimation result.
Further, acquiring AMV data and preprocessing the AMV data to obtain AMV sample data, including:
AMV data generated using ABI FD mode;
Performing time matching on the AMV data, and screening out paired samples with the time difference of more than 15 minutes;
Carrying out space matching on AMV data, and mutually matching two nearest sample points with the distance of less than 0.05 DEG in the equal longitude and latitude projection grid;
Performing height matching on the AMV data, and converting buoy anemometry of different heights to 10m height;
And taking the data after the AMV data and the ERA5 data are subjected to space-time matching as a sample data set.
Further, referring to the physical frame to obtain AWM feature data from the AMV sample data, comprising:
describing the variation of near-formation stroke with altitude using a logarithmic distribution;
reference is made to the distribution of stroke over height in the Ekkman layer;
and taking the wind speed, wind direction, altitude and latitude information representing the Coriolis force parameters, which have strong causal relationship in the AMV, as characteristic data.
Further, inputting the AMV feature data into the fully connected neural network and training the fully connected neural network, comprising:
Constructing a fully connected neural network;
taking the wind speed, the wind speed u-v component, the altitude and the latitude of the AMV as input characteristics;
All the neurons between every two layers of the fully connected neural network are connected, and nonlinear fitting capacity is introduced by using an activation function HARDSWISH;
Calculating MSE between three outputs of the fully connected neural network and corresponding components of ERA5 by taking MSE as a cost function, and taking wind speed, things thereof and north-south components as labels;
The average of all the labels MSE is taken as the final cost function.
Further, sampling the sample data set at intervals of 50, taking 25% of the sampled samples as a test set for testing the training effect and generalization capability of the model, taking 75% as the training set, and iteratively updating the model parameters.
Further, the input data is propagated to each layer of the fully connected neural network, gradient descent is realized through numerical differentiation, and network parameters are updated, including:
describing the relationship between the predicted value of the dependent variable and the independent variable by using a linear regression model;
When the linear regression model bias is 0, the linear regression model is changed to an empirical model, and the typhoon zone ASWind is estimated.
Further, the method for transmitting the input data to each layer of the fully connected neural network, realizing gradient descent through numerical differentiation, updating network parameters, and further comprises the following steps:
interpolation of AMV and ASWind into the equal warp and weft grids using an inverse distance weighted average interpolation method;
the inverse distance weighted average calculation method is as follows:
Wherein x i、yi represents the longitude and latitude of the original data, x 0、y0 represents the longitude and latitude of the target lattice point of the original data, z i represents the original data, z 0 represents the data of the original data after interpolation to the target lattice point, d i represents the distance from the original data to the target lattice point, and p is a constant, and the common value is 2.
Further, the method for transmitting the input data to each layer of the fully connected neural network, realizing gradient descent through numerical differentiation, updating network parameters, and further comprises the following steps:
Evaluating the calculation result of the model ASWind by using a Root Mean Square Error (RMSE), a Bias and a correlation coefficient (R2) model;
The root mean square error RMSE is:
bias is:
the correlation coefficient R2 is:
wherein, And x n represents the model predictive value and the label value in the test set,/>, respectivelyAnd/>Respectively/>And x n, N represents the test set data amount.
Further, estimating sea surface wind by using the fully connected neural network to obtain an estimation result, including:
Comparing AMV and ASWind with ERA5 wind speed to obtain a comparison result;
AMVs and ASWind were compared to ERA5 wind direction to obtain comparison results.
Further, estimating sea surface wind by using the fully connected neural network to obtain an estimation result, and further comprising:
Comparing the AMV and ASWind with the NDBC buoy data to obtain a comparison result;
The AMVs and ASWind are compared with NDBC buoy data for wind direction to obtain comparison results.
The scheme of the invention at least comprises the following beneficial effects:
According to the scheme, wind speed, wind direction, height and latitude information in the AMV are taken as characteristics to be input into the fully-connected neural network, and the fully-connected neural network can fully consider the influence of parameters on ASWind by utilizing the self-adaption, self-learning and nonlinear characteristics of the neural network, so that ASWind is calculated; and a high nonlinear relation is established between the low-level AMV and the SSW, so that the expression capacity and the estimation accuracy of the model are improved.
Drawings
FIG. 1 is a flow chart of a satellite high frequency sea surface wind estimation method according to an embodiment of the present invention;
Fig. 2 is a flowchart of an implementation of a satellite high-frequency sea surface wind estimation method according to an embodiment of the present invention.
Detailed Description
Exemplary embodiments of the present disclosure will be described more closely below with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
As shown in fig. 1, an embodiment of the present invention proposes a satellite high-frequency sea surface wind estimation method, which includes:
step 11, acquiring AMV data, and preprocessing the AMV data to obtain AMV sample data;
step 12, referring to the physical frame according to the AMV sample data to obtain AWM characteristic data;
step 13, inputting AMV characteristic data into a fully-connected neural network, and training the fully-connected neural network;
step 14, the input data are transmitted to each layer of the fully-connected neural network, gradient descent is realized through numerical differentiation, and network parameters are updated;
and 15, estimating the fully-connected neural network to obtain an estimation result.
In the embodiment of the invention, the information of sea surface wind is estimated by using AMV data acquired by satellites, and high-frequency sea surface wind observation data is provided; acquiring AMV data and preprocessing the AMV data to obtain processed AMV sample data, and providing input for the subsequent steps; according to the AMV sample data, referring to the physical frame, AWM characteristic data can be obtained, namely, characteristic data related to sea surface wind is extracted by analyzing and understanding the AMV data; inputting AMV characteristic data into a fully-connected neural network for training, and establishing a mapping relation between the AMV data and sea surface wind through learning and optimization of the neural network; gradient descent is realized through numerical differentiation, network parameters are updated, so that the neural network can estimate sea surface wind more accurately, and the accuracy and performance of the model are improved; and estimating the fully connected neural network to obtain an estimation result, and evaluating the accuracy and performance of the model.
As shown in fig. 1, step 11, acquiring AMV data and preprocessing the AMV data to obtain AMV sample data, includes:
step 111, AMV data generated by using ABI FD mode;
step 112, performing time matching on the AMV data, and screening out paired samples with the time difference of more than 15 minutes;
step 113, performing space matching on AMV data, and mutually matching two nearest sample points with a distance of less than 0.05 DEG in the equal longitude and latitude projection grid;
114, performing height matching on AMV data, and converting buoy anemometry of different heights to 10m height;
and step 115, using the data after space-time matching of the AMV data and the ERA5 data as a sample data set.
In the embodiment of the invention, the GOES-16ABI AMV product has a mesoscale low-layer wind guiding detection capability, and VIS AMV data of 2021, 1 month, 1 day and 11 months and 30 days are used, so that an ocean overhead AMV data set is increased as much as possible, and AMV data in the study are generated by an FD mode of ABI; because the error between the VIS AMV and sea surface wind is gradually increased along with the increase of the height, and most of the VIS AMVs are below 950hPa, in order to enable the models to more efficiently acquire the correlation information between the wind guide and the sea wind, a low-layer AMV training model below 925hPa and within an atmospheric boundary layer is researched and selected, and when the applicability of the test ASWind in a wind area is tested, in order to enlarge the effect of a dataset and the test model at a higher height, all height VIS AMV data are used;
Integrating data from different sources to enable the data to be matched with each other in time and space, and further completing research; because the wind field can change along with time, if the time difference between two samples from different data sets is too large and the representativeness of the samples is weak, more difficulties are brought to the experiment, more errors are brought to the model, and in order to make the experiment smoothly proceed, the model has higher precision, the embodiment can give up the paired samples with the time difference of more than 15 minutes to be brought into the model for training; the wind field also has the characteristic of continuous change in space, in order to reduce the error brought by the space change, two nearest sample points from different data sets with the distance below 0.05 degrees in the equal longitude and latitude projection grid are matched with each other to be used as a group of samples; NDBC buoy station data due to wind speed Buoy anemometry is switched to 10m altitude, where h represents buoy in-situ observed wind field altitude, and s 10 and s h represent wind speeds at 10m and hm altitude, respectively.
As shown in fig. 1, step 12, referencing the physical frame to obtain AWM feature data from the AMV sample data, includes:
Step 121, describing the variation of stroke with altitude in the near-formation using a logarithmic distribution;
Step 122, consulting the distribution of strokes in the Ekkman layer with respect to height;
And step 123, taking the wind speed, wind direction, altitude and latitude information representing the Coriolis force parameters, which have strong causal relationship, in the AMV as characteristic data.
In the embodiment of the invention, the atmospheric boundary layer is in direct contact with the ground surface, has the thickness of about 1-1.5 km, has turbulent flow characteristic, and can be divided into a stratum, a near stratum and an Ekman Ekman layer from bottom to top;
Turbulent friction and air pressure gradient force play a main role in near stratum, and the vertical transmission of physical quantity flux and the wind direction hardly change with the height, but the wind speed changes with the height, such as a formula The variation of stroke with height in the near-stratum under the neutral layer junction is shown to conform to logarithmic distribution, wherein V is the wind speed at height z, kappa is the Karman constant, z 0 is the surface roughness, V * is the friction speed, constant in the near-stratum, and h S is the near-stratum top height;
In Ekman layers, under the assumption that the motion is stable, the horizontal flow inertial force is negligible relative to the coriolis force, and the horizontal air pressure gradient force does not change with the height, the rule of the stroke of the boundary layer along with the height can be described by the Ekkman spiral, and on the premise that the ground rotation has only east-west components, the distribution of the stroke of the Ekkman layer along with the height is as follows:
Where u, v represent the latitudinal and meridional wind speeds, respectively, at altitude z in Ekman layers, u g is the east-west component of the ground rotation, k z is the vertical turbulence coefficient, where k z is assumed to be constant, f is the coriolis parameter, Ω is the earth rotation angular velocity, As latitude, h B is boundary layer top height;
part of the experiments in this example contained free atmosphere with boundary layer up to 700hPa, in which the expression of the amount of thermal wind under the p coordinate system can be obtained according to the atmospheric ground wind equation and the high formula:
Wherein u t、vt represents the things of hot wind and north-south components respectively, R is a gas constant, p 1、p2 is two layers of atmospheric pressure at the same horizontal position respectively, phi 2-φ1 is the potential thickness between the two layers of atmosphere, and T m is the average temperature of the gas column between the pressure surfaces of p 1、p2 and the like;
The wind speed, the wind direction, the height and the latitude information representing the Coriolis force parameters, which have strong causal relation in the AMV, are taken as characteristics, are input into a neural network, other unknown physical parameter information is deduced by utilizing the self-adaption, self-learning and nonlinear characteristics of the neural network, and ASWind can be calculated from the AMV.
As shown in fig. 1, step 13, inputting AMV feature data into the fully connected neural network, and training the fully connected neural network, includes:
Step 131, constructing a fully connected neural network;
Step 132, taking the wind speed, the wind speed u-v component, the altitude and the latitude of the AMV as input characteristics;
step 133, fully connecting neurons between every two layers of the fully connected neural network, and introducing nonlinear fitting capability by using an activation function HARDSWISH;
step 134, adopting a mean square error MSE as a cost function, calculating MSE between three outputs of the fully connected neural network and corresponding components of ERA5, and taking wind speed, things and north-south components thereof as labels;
in step 135, the average of all the labels MSE is taken as the final cost function.
In the embodiment of the invention, the fully-connected neural network is a basic artificial neural network structure, neurons between every two layers are all connected, for a group of samples which are matched in time and space, the wind speed u-v component, the height and the latitude of AMV can be input into the neural network by referring to the causal relationship among physical quantities in a boundary layer model, and the wind speed, the wind speed u-v component, the height and the latitude are transmitted to an output layer through the forward direction of the network, and the model error can be obtained by the output ASWind wind speed and the u-v component thereof and the corresponding component of ERA5 through a specified cost function, so that gradient descent is realized through a numerical differentiation method, and network parameters are updated;
The following table lists specific parameters of the fully-connected neural network of this embodiment:
The fully-connected neural network comprises an input layer, two hidden layers and an output layer, and an activation function HARDSWISH between each two layers enables the neural network to have nonlinear fitting capability; the input layer has 5 neurons corresponding to 5 input features; in order to ensure the model accuracy and avoid excessive model parameters, the number of neurons of the two hidden layers is 12; the 3 neurons of the output layer will output ASWind and its east-west, north-south components, respectively;
The mean square error MSE is taken as a cost function, expressed as:
wherein N is the single Batch data Size batch_Size of the input neural network, The nth data output by the neural network for a single batch of samples, y n is the nth tag data in the single batch;
The three outputs of the neural network are respectively used for solving MSE with ERA5 wind speed, things, south and north components, the final cost function is the average of all the MSE labels, and the final cost function is as follows:
wherein, And/>The nth scalar wind speed, the weft wind speed and the warp wind speed in the output sample are respectively represented, and the corresponding labels are V n、un and V n respectively;
According to the wind and hot wind components in the Ekkman layer, the AMV wind speed and the u-v component in the lower atmosphere are related to the altitude and the Coriolis force, and the multi-dimensional physical parameter information can describe the AMV more fully, can map the AMV to ASWind more accurately and further estimate the SSW, so that the model adopts the scalar speed of the AMV and things, the north-south component, the altitude and the latitude of the AMV as 5 characteristic input networks, and the training network obtains ASWind, wherein the altitude adopts logarithmic pressure.
In the embodiment of the invention, in order to train out a model with higher precision while reducing the calculation power consumption and avoiding overfitting, the data set interval 50 is sampled; in order to test the training effect and generalization capability of the model, 25% of the data set obtained by interval sampling is used as a test set, and the test set is not involved in network parameter updating and is only used for testing the training effect of the model; 75% is used as training set to update model parameters iteratively.
As shown in fig. 1, step 14, propagating the input data to each layer of the fully connected neural network, implementing gradient descent through numerical differentiation, and updating network parameters, including:
step 141, describing the relation between the predicted value of the dependent variable and the independent variable by using a linear regression model;
when the bias of the linear regression model is 0, the linear regression model is changed into an empirical model, and the typhoon zone ASWind is estimated;
the linear regression model expression based on the least square method is:
wherein, As a predicted value of a dependent variable, x i is an independent variable,/>The model becomes the average of the dependent and independent variables, respectively, with bias b of 0:
As shown in fig. 1, in step 14, the input data is propagated to each layer of the fully-connected neural network, gradient descent is implemented through numerical differentiation, and network parameters are updated, and the method further includes:
step 142, using inverse distance weighted average interpolation to interpolate AMV and ASWind into equal warp and weft grids;
the inverse distance weighted average calculation method is as follows:
Wherein x i、yi represents the longitude and latitude of the original data, x 0、y0 represents the longitude and latitude of the target lattice point of the original data, z i represents the original data, z 0 represents the data of the original data after interpolation to the target lattice point, d i represents the distance from the original data to the target lattice point, and p is a constant, and the common value is 2.
As shown in fig. 1, in step 14, the input data is propagated to each layer of the fully-connected neural network, gradient descent is implemented through numerical differentiation, and network parameters are updated, and the method further includes:
step 143, evaluating the calculation result of the model ASWind by using the root mean square error RMSE, the Bias and the correlation coefficient R2 model;
The root mean square error RMSE is:
bias is:
The correlation coefficient R2 is
Wherein,And x n represents the model predictive value and the label value in the test set,/>, respectivelyAnd/>Respectively/>And x n, N represents the test set data amount.
As shown in fig. 1, in step 15, estimating sea surface wind by using a fully connected neural network to obtain an estimation result includes:
Step 151, comparing AMV and ASWind with ERA5 wind speed to obtain a comparison result;
step 152, the AMVs ASWind are compared with the ERA5 wind direction to obtain a comparison result.
In the embodiment of the invention, GOES-16VIS AMV data of 2021, 1 st, 11 th, 30 th and ERA5 data are subjected to space-time matching;
The effects of AMV and ASWind and ERA5 wind speeds, including ASWind products obtained by using a linear regression model, a fully connected neural network and other models, are compared; the method comprises the steps of adopting two linear models of unbiased and biased, wherein the former is often used as an empirical model, and AMV products obtained by least square fitting are respectively called SASWind and LASWind for short; the ASWind output by the fully connected neural network is found to have the optimal Root Mean Square Error (RMSE) and linear correlation coefficient, and compared with a linear regression result, the fully connected neural network has a better fitting effect; the comparison and inspection of the single physical quantity ASWind and the ERA5 wind speed show that the fully connected neural network model can extract more effective information by receiving more effective physical parameters, so that a better fitting effect is obtained;
By using the fully-connected neural network model, the east-west and north-south components of ASWind can be obtained, the wind direction of ASWind is estimated, and the linear regression model is limited by model conditions and cannot obtain the wind direction; comparison of AMV and ASWind with ERA5 wind direction shows that ASWind wind direction has a certain correction effect, and compared with lower AMV, the root mean square error of ASWind wind direction is reduced.
As shown in fig. 1, in step 15, the estimating method of the sea surface wind by using the fully connected neural network to obtain an estimated result further includes:
Step 153, comparing AMV and ASWind with NDBC buoy data to obtain a comparison result;
Step 154, the AMV and ASWind are compared with the NDBC buoy data in wind direction to obtain a comparison result.
In the embodiment of the invention, GOES-16VIS AMV from 1 month 1 day 2021 to 30 days 11 months 1 is subjected to space-time matching with NDBC buoy data;
ASWind is lower than the buoy wind speed by 0.2199m/s as a whole, but has the best Root Mean Square Error (RMSE) and correlation coefficient compared with the other two models; SASWind has systematic deviation in a middle-low wind speed region, and the overall wind speed is 0.3318m/s smaller than ERA 5; LASWind, while possessing the best Bias, overestimates occur in low wind speed regions and underestimates occur in high wind speed regions;
the AMV wind direction is larger at 80 degrees relative to the buoy, and ASWind corrects the wind direction at the position, so that the root mean square error of the wind direction is reduced slightly relative to the AMV.
Selecting 12 times of typhoons in 2021 to test the model scene application effect, wherein typhoons are listed in the following table in terms of example information:
| Typhoon name | Sampling time (UTC) | Sampling range |
| Kevin | 2021-08-08 18:00:00 | 5°N-25°N；100°W-120°W |
| Kevin | 2021-08-09 18:00:00 | 7°N-27°N；105°W-125°W |
| Kevin | 2021-08-10 18:00:00 | 10°N-30°N；105°W-125°W |
| Kevin | 2021-08-11 18:00:00 | 12°N-32°N；107°W-127°W |
| Marty | 2021-08-23 18:00:00 | 10°N-30°N；105°W-125°W |
| Victory | 2021-10-01 18:00:00 | 0°N-20°N；25°W-45°W |
| Sam | 2021-09-28 18:00:00 | 7N-27°N；45°W-65°W |
| Sam | 2021-09-29 18:00:00 | 10°N-30°N；45°W-65°W |
| Sam | 2021-09-30 18:00:00 | 15°N-35°N；48°W-68°W |
| Sam | 2021-10-01 18:00:00 | 20°N-40°N；50°W-70°W |
| Sam | 2021-10-02 18:00:00 | 25°N-45°N；50°W-70°W |
| Sam | 2021-10-03 18:00:00 | 27°N-47°N；45°W-65°W |
When the platform wind area ASWind is manufactured, the VIS AMV with all heights is input into the fully-connected neural network model for calculation; as a model comparison, in the aspect of a linear model, SASWind is obtained by using an empirical coefficient of 0.76 in a typhoon area to serve as a comparison of the effect of the fully-connected neural network ASWind product;
Because the buoy can collect limited typhoon samples, the AMV has fewer samples in the typhoon area, and ERA5 is used as a reference sample for ASWind product effects in the part; inputting AMV in the region into a fully connected neural network model to obtain ASWind by selecting a specific region, and performing space-time matching with ERA5 data to obtain a data set so as to complete effect analysis;
Typhoon Kelvin located in the southwest of the peninsula of california under pacific northeast at year 8, month 11, 18 (UTC) 2021 is in the extinction phase; the GOES-16ABI visible light channel recognizes a large number of AMVs in the northwest direction of the spiral cloud system; compared with an ERA5 10m wind field, AMV wind speed in the northeast part of the spiral cloud system is larger, and ASWind wind speed obtained through a fully connected neural network algorithm is more consistent with ERA5 in the region; ASWind also has better performance in the northwest of typhoon victoriy at 10 months 1 and 18 days (UTC) of the same year; ASWind of AMV estimation with larger relative ERA5 in typhoon Sam peripheral cloud system on the same day is also corrected;
The statistics result shows that compared with the ERA5, the typhoon area ASWind has obvious correction in wind speed, and the typhoon area AMV is in typhoon peripheral cloud system, high wind speed samples are not more, but ASWind can be reasonably estimated by combining the height, latitude, wind speed and wind speed components of the samples for the high altitude high wind speed sample fully connected neural network;
SASWind has better performance in a typhoon region RMSE, and particularly for a medium-high wind speed region, a high wind sample can be effectively estimated; however, in the wind speed region below 10m/s, SASWind has obvious systematic deviation, the average wind speed is lower than ERA5 by 0.6655m/s, the ASWind has better performance in the wind speed region, and the ASWIND RMSE performance is better than SASWind;
In terms of wind direction, the RMSE of the typhoon area AMV and ERA5 wind direction is 19.4686, and compared with the all-weather AMV below 925hPa, the RMSE is larger; although AMV wind direction accuracy is reduced in typhoon areas, ASWind estimated by the fully connected neural network model can still be modified to a small extent.
While the foregoing is directed to the preferred embodiments of the present invention, it will be appreciated by those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the present invention.
Claims (10)
1. A method for estimating satellite high frequency sea surface wind, the method comprising:
acquiring AMV data, and preprocessing the AMV data to obtain AMV sample data;
referencing the physical frame according to the AMV sample data to obtain AWM feature data;
inputting AMV characteristic data into a fully-connected neural network, and training the fully-connected neural network;
the input data are transmitted to each layer of the fully-connected neural network, gradient descent is realized through numerical differentiation, and network parameters are updated;
And estimating sea surface wind by using the fully connected neural network to obtain an estimation result.
2. The method of estimating sea surface wind at high frequency of a satellite according to claim 1, wherein acquiring AMV data and preprocessing the AMV data to obtain AMV sample data comprises:
AMV data generated using ABI FD mode;
Performing time matching on the AMV data, and screening out paired samples with the time difference of more than 15 minutes;
Carrying out space matching on AMV data, and mutually matching two nearest sample points with the distance of less than 0.05 DEG in the equal longitude and latitude projection grid;
Performing height matching on the AMV data, and converting buoy anemometry of different heights to 10m height;
And taking the data after the AMV data and the ERA5 data are subjected to space-time matching as a sample data set.
3. The method of estimating sea surface wind at high frequency of satellites according to claim 2, wherein referencing the physical frame based on the AMV sample data to obtain the AWM feature data comprises:
describing the variation of near-formation stroke with altitude using a logarithmic distribution;
reference is made to the distribution of stroke over height in the Ekkman layer;
and taking the wind speed, wind direction, altitude and latitude information representing the Coriolis force parameters, which have strong causal relationship in the AMV, as characteristic data.
4. A satellite high frequency sea surface wind estimation method according to claim 3, wherein inputting AMV characteristic data into the fully connected neural network and training the fully connected neural network comprises:
Constructing a fully connected neural network;
taking the wind speed, the wind speed u-v component, the altitude and the latitude of the AMV as input characteristics;
All the neurons between every two layers of the fully connected neural network are connected, and nonlinear fitting capacity is introduced by using an activation function HARDSWISH;
Calculating MSE between three outputs of the fully connected neural network and corresponding components of ERA5 by taking MSE as a cost function, and taking wind speed, things thereof and north-south components as labels;
The average of all the labels MSE is taken as the final cost function.
5. The method of estimating sea surface wind at high frequency of satellite according to claim 4, wherein the sample data set is sampled at intervals of 50%, 25% of the sampled samples are used as a test set for testing training effect and generalization ability of the model, and 75% are used as a training set for iteratively updating model parameters.
6. The method of estimating sea surface wind at high frequency of satellite according to claim 5, wherein the step of propagating the input data to each layer of the fully-connected neural network, and implementing gradient descent by numerical differentiation, and updating network parameters comprises:
describing the relationship between the predicted value of the dependent variable and the independent variable by using a linear regression model;
When the linear regression model bias is 0, the linear regression model is changed to an empirical model, and the typhoon zone ASWind is estimated.
7. The method of estimating sea surface wind at high frequency of satellite according to claim 6, wherein the step of propagating the input data to each layer of the fully-connected neural network, and the step of realizing gradient descent by numerical differentiation and updating network parameters, further comprises:
interpolation of AMV and ASWind into the equal warp and weft grids using an inverse distance weighted average interpolation method;
the inverse distance weighted average calculation method is as follows:
Wherein x i、yi represents the longitude and latitude of the original data, x 0、y0 represents the longitude and latitude of the target lattice point of the original data, z i represents the original data, z 0 represents the data of the original data after interpolation to the target lattice point, d i represents the distance from the original data to the target lattice point, and p is a constant, and the common value is 2.
8. The method of estimating sea surface wind at high frequency of satellite according to claim 7, wherein the step of propagating the input data to each layer of the fully-connected neural network, and the step of realizing gradient descent by numerical differentiation and updating network parameters further comprises:
Evaluating the calculation result of the model ASWind by using a Root Mean Square Error (RMSE), a Bias and a correlation coefficient (R2) model;
The root mean square error RMSE is:
bias is:
the correlation coefficient R2 is:
wherein, And x n represents the model predictive value and the label value in the test set,/>, respectivelyAnd/>Respectively/>And x n, N represents the test set data amount.
9. The method of estimating sea surface wind at high frequency of satellite as claimed in claim 8, wherein estimating sea surface wind using a fully connected neural network to obtain an estimated result comprises:
Comparing AMV and ASWind with ERA5 wind speed to obtain a comparison result;
AMVs and ASWind were compared to ERA5 wind direction to obtain comparison results.
10. The method of estimating sea surface wind at high frequency in a satellite of claim 9, wherein estimating sea surface wind using a fully connected neural network to obtain an estimated result, further comprising:
Comparing the AMV and ASWind with the NDBC buoy data to obtain a comparison result;
The AMVs and ASWind are compared with NDBC buoy data for wind direction to obtain comparison results.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410120101.6A CN117932343A (en) | 2024-01-29 | 2024-01-29 | Satellite high-frequency sea surface wind estimation method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410120101.6A CN117932343A (en) | 2024-01-29 | 2024-01-29 | Satellite high-frequency sea surface wind estimation method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117932343A true CN117932343A (en) | 2024-04-26 |
Family
ID=90764351
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410120101.6A Pending CN117932343A (en) | 2024-01-29 | 2024-01-29 | Satellite high-frequency sea surface wind estimation method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117932343A (en) |
-
2024
- 2024-01-29 CN CN202410120101.6A patent/CN117932343A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112905560B (en) | Air pollution prediction method based on multi-source time-space big data deep fusion | |
| Lermusiaux | Data Assimilation via Error Subspace Statistical Estimation.: Part II: Middle Atlantic Bight Shelfbreak Front Simulations and ESSE Validation | |
| CN109344865B (en) | Data fusion method for multiple data sources | |
| CN112213727A (en) | Precipitation correction method of satellite-borne radar based on active and passive microwave combined detection | |
| CN112946784B (en) | SuperDARN radar convection diagram short-term forecasting method based on deep learning | |
| CN115062527B (en) | Geostationary satellite sea temperature inversion method and system based on deep learning | |
| CN115795399B (en) | Multi-source remote sensing precipitation data self-adaptive fusion method and system | |
| Uchida et al. | Error and energy budget analysis of a nonhydrostatic stretched-grid global atmospheric model | |
| Yu et al. | Improvement of medium-range forecasts using the analog-dynamical method | |
| CN113139327B (en) | Ionized layer TEC single-point prediction method and system based on GRU network model | |
| Fan et al. | A comparative study of four merging approaches for regional precipitation estimation | |
| Espinosa-Gavira et al. | Cloud motion estimation from small-scale irradiance sensor networks: General analysis and proposal of a new method | |
| CN114595876A (en) | Regional wind field prediction model generation method and device and electronic equipment | |
| CN113984198A (en) | Short wave radiation prediction method and system based on convolutional neural network | |
| Xue et al. | The GoMOOS nowcast/forecast system | |
| Verkhoglyadova et al. | Estimation of winds from GPS radio occultations | |
| Harpham et al. | A Bayesian method for improving probabilistic wave forecasts by weighting ensemble members | |
| He et al. | Bayesian temporal tensor factorization-based interpolation for time-series remote sensing data with large-area missing observations | |
| CN116609859A (en) | Weather disaster high-resolution regional mode forecasting system and method | |
| CN117932343A (en) | Satellite high-frequency sea surface wind estimation method | |
| Stohl et al. | Diagnostic downscaling of large-scale wind fields to compute local-scale trajectories | |
| Zhao et al. | Artificial intelligence powered forecast of oceanic mesoscale phenomena: A typhoon cold wake case occurring in Northwest Pacific Ocean | |
| Sorkhabi et al. | Deep learning of ionosphere single-layer model and tomography | |
| Lu et al. | An objective tropical cyclone intensity estimation model based on digital IR satellite images | |
| Natras et al. | Machine learning ensemble approach for ionosphere and space weather forecasting with uncertainty quantification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |