A method for retrieving ocean surface current velocity based on satellite remote sensing data
By constructing an equatorial sea surface velocity inversion model based on a feedforward neural network and combining it with multi-element data, the problem of insufficient accuracy in sea surface velocity inversion in the equatorial sea area was solved, and higher inversion accuracy was achieved.
Patent Information
- Application Number
- CN202411121027.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2025-12-26
- Estimated Expiration
- 2044-08-15
AI Technical Summary
Existing technologies for retrieving sea surface current velocities in equatorial sea areas neglect key physical processes and important characteristics of seawater when it is not in equilibrium, resulting in insufficient retrieval accuracy, especially since high-temporal-resolution satellite remote sensing data cannot meet the accuracy requirements.
A feedforward neural network was used to construct an equatorial sea surface velocity inversion model. This model was combined with satellite remote sensing data, reanalysis data, and derived data, including sea surface temperature, salinity, and wave elements. The model was trained using a training dataset to obtain the sea surface velocity in the equatorial sea area.
It significantly improved the inversion accuracy of sea surface current velocity in the equatorial sea area, with the correlation coefficient of zonal current velocity increasing by 66% and the root mean square error decreasing by 69%, and the correlation coefficient of meridional current velocity increasing by 340% and the root mean square error decreasing by 52%.
Smart Images
Figure CN119066963B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to a method for retrieving ocean surface flow velocity based on satellite remote sensing data. BACKGROUND
[0002] The ocean surface flow velocity in the equatorial region is one of the key ocean variables in the local area, and retrieving the spatially and temporally continuous high-precision ocean surface flow velocity based on satellite remote sensing data is of great significance for climate change prediction, maritime navigation safety guarantee, and marine ecological and dynamic disaster warning. However, due to the special geographical location of the equator, the latitude is 0, and the Coriolis parameter is 0, the geostrophic balance theory and Ekman balance theory cannot be directly applied, and the retrieval of the equatorial sea surface flow velocity has always been a major problem in ocean data retrieval. The current technology for retrieving the equatorial sea surface flow velocity based on satellite remote sensing data mainly adds the geostrophic flow calculated by modifying the geostrophic balance in the equatorial sea area, i.e. the beta balance algorithm, and the wind-driven flow calculated by statistically modifying the Ekman balance, i.e. adding an amplitude coefficient, to obtain the sea surface flow velocity.
[0003] In order to verify the accuracy of the sea surface flow velocity obtained by the current retrieval algorithm, the sea surface flow velocity in the equatorial region is retrieved based on the current retrieval algorithm, the AVISO(Archiving, Validation, and Interpretation of Satellite Oceanographic Data) satellite remote sensing sea surface height data set, and the ERA5(ECMWF Reanalysis-5) ocean 10m wind speed data set, and is compared with the TAO(Tropical Atmosphere Ocean) ocean flow velocity observation data set in the region. The comparison results are shown in Figure 2 (a)-(b), and the results show that the effect of retrieving the flow velocity based on the algorithm is not ideal, wherein the correlation coefficient of the retrieved meridional flow velocity and the observed meridional flow velocity is only 0.38, and the root mean square error can reach 0.97m / s; the correlation coefficient of the retrieved zonal flow velocity and the observed zonal flow velocity is only 0.1, and the root mean square error reaches 0.44m / s.
[0004] Current ocean current inversion algorithm mainly has the following problems: (1) The formula for inverting geostrophic and wind-driven current is based on the assumption that seawater is in equilibrium state, ignoring key physical processes such as advection and mixing, and is derived by using a simplified seawater motion equation, which is only applicable to the inversion of sea surface flow velocity with a time resolution of more than one month and a large spatial scale that meets the geostrophic equilibrium. However, the time resolution of current satellite remote sensing data has reached one day or even higher, and the inversion calculation method based on the equilibrium assumption cannot meet the current inversion requirements; (2) Only two elements of sea surface height and sea surface wind speed are considered in the flow velocity inversion process, while the sea surface flow velocity in nature is also affected by sea surface temperature, salinity and sea wave, and ignoring these elements will greatly reduce the accuracy of the inversion result. SUMMARY
[0005] The purpose of the present application is to provide a method for inverting ocean surface flow velocity based on satellite remote sensing data, which mainly aims at the problem of poor inversion effect of equatorial sea surface flow velocity caused by the assumption that seawater is in equilibrium state and the lack of important features in traditional algorithms, and integrates sea surface temperature, salinity and wave elements into artificial intelligence neural network to determine the nonlinear relationship between sea surface flow velocity and satellite remote sensing features.
[0006] To achieve the above purpose, the present application provides the following scheme:
[0007] A method for inverting ocean surface flow velocity based on satellite remote sensing data, comprising:
[0008] Obtaining satellite remote sensing data and reanalysis data of the equatorial sea area, and obtaining derivative data of the satellite remote sensing data and the reanalysis data;
[0009] Inputting the satellite remote sensing data, the reanalysis data and the derivative data into an equatorial sea surface flow velocity inversion model to obtain equatorial sea surface flow velocity and perform inverse normalization processing to obtain final equatorial sea surface flow velocity, wherein the equatorial sea surface flow velocity inversion model is constructed based on a feedforward neural network and obtained by training a training data set, and the training data set includes historical satellite remote sensing data, historical reanalysis data, historical derivative data and corresponding field observation flow velocity data set.
[0010] Optionally, the reanalysis data includes reanalyzed sea surface temperature, salinity, wave elements, wind stress data and seawater surface density.
[0011] Optionally, the derivative data includes the latitude and longitude gradients of the satellite remote sensing data, the latitude and longitude gradients of the reanalysis data, and the wave direction cosine and sine.
[0012] Optionally, obtaining the training data set comprises:
[0013] The historical satellite remote sensing data, the historical reanalysis data and the historical derived data are interpolated with the spatiotemporal grid of the field observation flow rate dataset as a reference;
[0014] The field observation flow rate dataset is taken as an inversion label, and the historical satellite remote sensing data, the historical reanalysis data, the historical derived data and the corresponding inversion label are taken as the training dataset, and the training dataset is normalized.
[0015] Optionally, constructing the equatorial sea surface flow rate inversion model based on the feedforward neural network comprises:
[0016] According to the characteristics of the Pacific surface flow rate, the Pacific east-west basins are divided by using preset division longitudes, and the equatorial sea surface flow rate inversion model is constructed by using different structures of the feedforward neural network according to the division result, wherein a smoothing coefficient is added in a preset range of the preset division longitude.
[0017] Optionally, the different structures of the feedforward neural network are different hidden layer numbers and different hidden layer neuron numbers.
[0018] Optionally, training the equatorial sea surface flow rate inversion model comprises:
[0019] The historical satellite remote sensing data, the historical reanalysis data and the historical derived data are taken as inputs of the equatorial sea surface flow rate inversion model, the corresponding field observation flow rate data are taken as outputs, when a preset training stop condition is reached, the equatorial sea surface flow rate inversion model training is completed, and the trained equatorial sea surface flow rate inversion model is obtained.
[0020] Optionally, in the process of training the equatorial sea surface flow rate inversion model, the hidden layer activation function adopts an asymmetric Sigmoid function activation, and the output layer adopts linear activation.
[0021] The present application has the following beneficial effects:
[0022] The present application proposes a method for using a neural network to invert the equatorial sea surface flow rate, which reasonably avoids the assumption that seawater is in a balanced state in the traditional algorithm, avoids the large error of flow rate inversion caused by the unreasonable approximation of the fluid motion equation in the equatorial sea area, and effectively improves the inversion accuracy of the equatorial sea surface flow rate.
[0023] The present application constructs an inversion method for the equatorial sea surface flow rate with multiple element characteristics inputs, uses a reasonable and effective combination of multiple satellite remote sensing parameters, and makes up for the deficiency of the traditional algorithm which only uses two parameters of sea surface height and wind stress. BRIEF DESCRIPTION OF DRAWINGS
[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed in the embodiments will be briefly introduced as follows. Obviously, the accompanying drawings in the following description only are some embodiments of the present application, and for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0025] Figure 1 A flow chart of a method for retrieving ocean surface current velocity based on satellite remote sensing data according to an embodiment of the present application;
[0026] Figure 2 Effect diagrams of zonal and meridional current velocities retrieved by a traditional algorithm and the present method, wherein (a) is an effect diagram of zonal current velocity retrieved by the traditional algorithm, (b) is an effect diagram of meridional current velocity retrieved by the traditional algorithm, (c) is an effect diagram of zonal current velocity retrieved by the present method, and (d) is an effect diagram of meridional current velocity retrieved by the present method;
[0027] Figure 3 Time-latitude diagrams of zonal retrieved current velocity and zonal wind stress according to an embodiment of the present application, wherein (a) is a diagram of zonal current velocity in the equatorial ocean region retrieved by the present method and the variation after 150-day low-pass filtering, and (b) is a diagram of zonal wind stress and the variation after 150-day low-pass filtering;
[0028] Figure 4 An effect diagram of 13-month moving average zonal wind stress according to an embodiment of the present application;
[0029] Figure 5 Time-latitude diagrams of meridional retrieved current velocity and meridional wind stress according to an embodiment of the present application, wherein (a) is a diagram of meridional current velocity in the equatorial ocean region retrieved by the present method and the variation after 150-day low-pass filtering, and (b) is a diagram of meridional wind stress and the variation after 150-day low-pass filtering;
[0030] Figure 6 An effect diagram of 13-month moving average meridional wind stress according to an embodiment of the present application;
[0031] Figure 7 A diagram of 5-month moving average Nino3.4 index according to an embodiment of the present application. DETAILED DESCRIPTION
[0032] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, and all other embodiments obtained by those skilled in the art based on the embodiments in the present application without creative effort are within the scope of the present application.
[0033] In order to make the above objectives, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.
[0034] As shown in the embodiment, a method for retrieving ocean surface flow velocity based on satellite remote sensing data is provided, comprising: Figure 1
[0035] Satellite remote sensing data and reanalysis data of the equatorial sea area are obtained, and derivative data of the satellite remote sensing data and the reanalysis data are obtained, wherein the satellite remote sensing data is sea surface height data;
[0036] The satellite remote sensing data, the reanalysis data and the derivative data are input into an equatorial sea surface flow velocity retrieval model to obtain equatorial sea surface flow velocity and perform inverse normalization processing to obtain final equatorial sea surface flow velocity, wherein the equatorial sea surface flow velocity retrieval model is constructed based on a feedforward neural network and obtained by training a training data set, and the training data set includes historical satellite remote sensing data, historical reanalysis data, historical derivative data and a corresponding field observation flow velocity data set.
[0037] Specifically, a feedforward neural network is selected as the basic architecture of the retrieval algorithm, the interpolated satellite remote sensing data and reanalysis data and the derivative data thereof are taken as input features, and the corresponding flow velocity label is taken as output to form the basic structure of the equatorial sea surface flow velocity retrieval algorithm.
[0038] Specifically, the inverse normalization processing includes: performing inverse normalization on the obtained normalized equatorial sea surface flow velocity to retrieve the equatorial sea surface flow velocity, i.e. to obtain the real zonal and meridional flow velocities, i.e. to convert the dimensionless normalized flow velocity into the dimensional sea surface flow velocity (unit: m / s), and the specific method of inverse normalization is: wherein is the equatorial flow velocity vector, u is the zonal flow velocity, v is the meridional flow velocity, i is the index of the data point, is the unit vector in the longitude direction, is the unit vector in the latitude direction; is the normalized dimensionless equatorial flow velocity vector, is the maximum flow velocity, is the minimum flow velocity.
[0039] Further, the reanalysis data includes: reanalyzed sea surface temperature, salinity, sea wave elements, wind stress, and sea surface density data.
[0040] Specifically, the sea surface density is calculated as follows:
[0041]
[0042] where, p(S, T, P) is the density of seawater under the conditions of salinity S, temperature T and pressure P, p(S, T, P) is the density of seawater under the pressure of 1 standard atmosphere (P=0), and K(S, T, P) is the reciprocal bulk modulus of seawater. The seawater density can be obtained by using the temperature and salinity data.
[0043] Further, the derived data includes: the longitude and latitude gradients of satellite remote sensing data, the longitude and latitude gradients of reanalysis data, and the wave direction cosine.
[0044] Specifically, the derived data of both the satellite remote sensing data and the reanalysis data is calculated, and the derived data mainly includes the longitude and latitude gradients and the wave direction cosine, and the longitude and latitude gradients are calculated as follows:
[0045]
[0046] where, f is a physical quantity, is the latitude gradient, is the longitude gradient, is the unit vector in the longitude direction, is the unit vector in the latitude direction, x is the longitude, and y is the latitude, is the Hamiltonian operator.
[0047] The wave direction cosine is calculated as follows:
[0048] wd x = cos(wd) (3)
[0049] wd y = sin(wd) (4)
[0050] where, wd x and wd y are the latitude and longitude components of the wave direction respectively, cos and sin are the cosine and sine functions respectively, and wd is the wave direction.
[0051] Further, the construction of the training data set includes: taking the spatiotemporal grid of the field observation flow rate data set as a reference, and interpolating the historical satellite remote sensing data, historical reanalysis data and historical derived data; taking the field observation flow rate data set as an inversion label, obtaining the training data set corresponding to the historical satellite remote sensing data, historical reanalysis data and historical derived data and the inversion label, and performing normalization processing on the training data set.
[0052] Specifically, a multi-source large-area data set and a sea surface flow velocity data set of an equatorial sea area are constructed, historical satellite remote sensing data and historical reanalysis sea surface temperature, salinity, sea wave elements, wind stress and sea surface density data of the Pacific equatorial sea area are collected to form a multi-source large-area data set and a sea surface observation flow velocity (TAO) data set. After quality control of the data set, a large data set is formed, and the reanalysis data is downloaded from the Copernicus Data Center; the quality control process of the data set includes removal of meaningless extreme values of field observation flow velocity and removal of instrument abnormal state observation data.
[0053] Then, based on the spatial and temporal grid points of the field observation flow velocity data set, the historical satellite remote sensing data and the historical reanalysis data and the historical derivative data thereof are interpolated, the field observation flow velocity is taken as the inversion label, and the historical satellite remote sensing data and the historical reanalysis data and the historical derivative data thereof are taken as the features, to form a mapping data set with one-to-one correspondence between the features and the labels.
[0054] All variables of the mapping data set are normalized, and the normalization method is: Wherein, X i is a dimensioned physical quantity before normalization, i is the index of the data point, nin(X i ) is the minimum value of the dimensioned physical quantity, max(X i ) is the maximum value of the dimensioned physical quantity, X norm is a dimensionless physical quantity normalized to [-1, 1], and is re-divided according to the observation station position. Finally, the mapping data sets of different stations are divided according to 7:3, wherein 70% of the data is used for training and 30% of the data is used for testing the algorithm performance.
[0055] Further, the equatorial sea surface flow velocity inversion model is constructed based on the feedforward neural network, which includes:
[0056] According to the characteristics of the Pacific surface flow velocity, the Pacific east-west basins are divided by using a preset division longitude, and the equatorial sea surface flow velocity inversion model is constructed by using different structures of feedforward neural network according to the division result, wherein a smoothing coefficient is added within the preset range of the preset division longitude.
[0057] Specifically, the basin division is optimized, the Pacific east-west basins are divided according to the characteristics of the Pacific surface flow velocity, different basins are respectively used for different forms of algorithm, such as the east-west basins are divided by 165°E in this embodiment, and a smoothing coefficient is added within the preset range of the division longitude of the east basin and the west basin, so that the flow velocity can be smoothly transitioned within the preset range of the division longitude, and no singular value is generated.
[0058] The specific smoothing process is that the flow rate is multiplied by a smoothing coefficient. For example, smoothing from 165 °E eastward by 5 °, the flow rate at 165 °E is multiplied by a smoothing coefficient a, and the flow rate at 165 °E eastward by one data resolution unit is multiplied by a smoothing coefficient b, wherein a+b=1. At 165 °E, a=1, and gradually decreases to 0 as the smoothing eastward.
[0059] Further, the feedforward neural networks with different structures are feedforward neural networks with different numbers of hidden layers and different numbers of hidden layer neurons.
[0060] Specifically, the adjustment algorithm includes the following specific structures: different numbers of hidden layers (the number of hidden layers is from 1 to 3) and different numbers of hidden layer neurons (the number of neurons is from 1 to 10) are selected respectively, different structures of algorithms are trained, in the training, the hidden layer activation function adopts an asymmetric Sigmoid function activation, and the output layer adopts a linear activation. The initial guess parameters are determined by using the Levenberg-Marquardt method. In order to avoid overfitting of the algorithm, the following exit conditions are used: (1) the root mean square error is not reduced for 20 consecutive optimizations; (2) the root mean square error reaches 5x10 -4 ; (3) the number of iterations exceeds 1000 times. The training results are brought into the test data set, the effects of different structure algorithms are tested, and the structure and parameters with the best effect are selected as the equatorial sea surface flow speed inversion algorithm, wherein the best effect is the correlation coefficient corresponding to the feedforward neural network. Different structures of the sea basin algorithm may be different, but the input data is the same, that is, the input is satellite remote sensing data, reanalysis data and derived data of the two, and finally one equatorial sea surface flow speed inversion model corresponds to one sea basin region.
[0061] Based on the above method, the embodiment establishes and verifies the intelligent inversion algorithm effect of the equatorial sea surface flow speed. The effects of the inversion of the zonal and meridional flow speeds by the traditional algorithm and the method of the embodiment are shown in Figs. Figure 2 (a)-(d), in which the horizontal axis is the TAO observation data, the vertical axis is the inversion data, the filled color graph is the normalized probability density, the larger the value, the more the distribution data in the range, and the black dotted straight line is the perfect inversion straight line, that is, the inversion value is completely equal to the observation value. The closer the distribution of the inversion and observation flow speeds to the straight line, the better the inversion effect. Figure 2 (a)-(d) show that, compared with the traditional algorithm, the inversion algorithm of the embodiment effectively improves the inversion effect of the equatorial flow speed, wherein the correlation coefficient of the inverted zonal flow speed is increased from 0.38 of the traditional algorithm to 0.63, which is increased by 66%, and the root mean square error is decreased from 0.97 m / s to 0.3 m / s, which is decreased by 69%; the correlation coefficient of the inverted meridional flow speed is increased from 0.1 of the traditional algorithm to 0.44, which is increased by 340%, and the root mean square error is decreased from 0.44 m / s to 0.21 m / s, which is decreased by 52%.
[0062] To further verify the inversion effect of the equatorial sea surface current velocity, the inverted equatorial sea surface current velocity is applied to the analysis of the seasonal and interannual variation of the equatorial current velocity. The analysis results are shown in FIGS. 10, 11, 12 and 13, wherein, Figures 3-7 Figure 3 (a), Figure 3 (b), and Figure 5 (a), Figure 5 (b) are the results of the 150-day low-pass filtering of the current velocity, indicating the annual variation of the inverted current, Figure 7 The light-colored part of (a) is less than -0.5°C, i.e., La Nina occurs, and the dark-colored part is greater than 0.5°C, i.e., El Nino occurs. The results show that: (1) for the zonal current velocity, the positive anomaly of the current velocity generated in the eastern sea area of 170°W will propagate westward at a speed of about 0.88 m / s, and the magnitude of the propagation speed is comparable to the wind-generated Rossby wave. In the El Nino year, the Kelvin wave excited by the west wind outbreak in the eastern sea basin will propagate eastward at a speed of 2.63 m / s to about 140°W. (2) For the meridional current velocity, its seasonal variation is stronger in the eastern sea basin, and the seasonal variation is stronger in the La Nina year and weaker in the El Nino year. In the western sea basin, the positive current velocity anomaly propagates eastward to about 160°E, which is consistent with the seasonal outbreak of the south wind. In the eastern sea basin, the current velocity anomaly propagates westward at a speed of about 0.84 m / s, and this westward propagation is associated with the south wind anomaly and is weakened in the El Nino year.
[0063] The above-described embodiments are merely descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art shall fall within the protection scope of the present application as defined by the claims.
Claims
1. A method for retrieving ocean surface current velocity based on satellite remote sensing data, characterized in that, include: Acquire satellite remote sensing data and reanalysis data of the equatorial sea area, and acquire derived data from the satellite remote sensing data and the reanalysis data; The reanalysis data includes: reanalyzed sea surface temperature, salinity, wave elements, wind stress data, and sea surface density; The derived data includes: the latitude and longitude gradients of the satellite remote sensing data, and the latitude and longitude gradients and wave direction sine and cosine of the reanalysis data; The satellite remote sensing data, the reanalysis data, and the derived data are input into the equatorial sea surface velocity inversion model to obtain the sea surface velocity in the equatorial sea area and perform inverse normalization processing to obtain the final sea surface velocity in the equatorial sea area. The equatorial sea surface velocity inversion model is constructed based on a feedforward neural network and trained on a training dataset. The training dataset includes historical satellite remote sensing data, historical reanalysis data, historical derived data, and corresponding field observation velocity datasets. Obtaining the training dataset includes: Using the time-space points of the in-situ observed flow velocity dataset as a reference, the historical satellite remote sensing data, the historical reanalysis data, and the historical derived data are interpolated; The in-situ observed flow velocity dataset is used as the inversion label, and the historical satellite remote sensing data, the historical reanalysis data, the historical derived data, and the corresponding inversion label are used as the training dataset, and the training dataset is normalized. The equatorial sea surface current velocity inversion model constructed based on the feedforward neural network includes: Based on the surface current characteristics of the Pacific Ocean, the eastern and western basins of the Pacific Ocean are delineated by a preset division of longitude. Based on the division results, the feedforward neural network with different structures is used to construct the equatorial sea surface current inversion model. A smoothing coefficient is added within a preset range of the preset division of longitude.
2. The method for retrieving ocean surface current velocity based on satellite remote sensing data according to claim 1, characterized in that, The feedforward neural networks with different structures are feedforward neural networks with different numbers of hidden layers and different numbers of neurons in the hidden layers.
3. The method for retrieving ocean surface current velocity based on satellite remote sensing data according to claim 1, characterized in that, Training the equatorial sea surface current inversion model includes: The historical satellite remote sensing data, the historical reanalysis data, and the historical derived data are used as inputs to the equatorial sea surface current velocity inversion model, and the corresponding field observation velocity data are used as outputs. When the preset training stop condition is reached, the training of the equatorial sea surface current velocity inversion model is completed, and the trained equatorial sea surface current velocity inversion model is obtained.
4. The method for retrieving ocean surface current velocity based on satellite remote sensing data according to claim 3, characterized in that, During the training of the equatorial sea surface current inversion model, the hidden layer activation function uses the asymmetric Sigmoid function, while the output layer uses linear activation.
Citation Information
Patent Citations
Sea surface flow velocity inversion method based on synthetic aperture radar
CN112162282A