Morning and dusk atmosphere correction method and terminal equipment based on VIIRS satellite

Abstract

The present invention provides a method and terminal device for atmospheric correction at dawn and dusk based on VIIRS satellite, which belongs to the field of marine remote sensing technology and solves the problem that the existing technology is not suitable for the observation environment of solar zenith angle at dawn and dusk. The method includes: obtaining the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in the target area in the dawn and dusk data of VIIRS satellite, and determining the Rayleigh scattering radiance of each pixel point based on the Rayleigh scattering lookup table; obtaining the total radiance, flare radiance, and white hat radiance received by the VIIRS satellite, and then correcting the Rayleigh scattering radiance of each pixel point to obtain the Rayleigh corrected radiance of each pixel point; inputting the solar zenith angle, observation zenith angle, relative azimuth angle, and Rayleigh corrected radiance of each pixel point into the machine learning atmospheric correction model to obtain the aerosol scattering radiance of each pixel point; obtaining the remote sensing reflectance of each pixel point according to the Rayleigh corrected radiance of each pixel point and the aerosol radiance of each pixel point.

Description

Technical Field

[0001] The present invention relates to the field of ocean remote sensing technology, and in particular to a dawn and dusk atmosphere correction method and terminal equipment based on a VIIRS satellite. Background Art

[0002] The Visible Infrared Imaging Radiometer (VIIRS) is a multidisciplinary instrument installed on the Joint Polar Satellite System (JPSS)-Suomi National Polar Partnership (S-NPP) satellite, which was launched in October 2011. The existing atmospheric correction methods mainly include the spectral iteration method based on extrapolated inversion of aerosol contribution, the atmospheric correction method based on aerosol lookup table, the empirical linear method based on measured spectra, and the atmospheric correction method based on machine learning.

[0003] The spectral iteration method based on extrapolation inversion of aerosol contribution determines the optical properties of aerosols according to the proportional structure of the aerosols in the study area, and then calculates variables such as transmittance and aerosol path radiation reflectivity according to the single scattering theory, and uses the multiple scattering theory to correct the single scattering calculation results. In the face of the large solar zenith angle observation environment at dawn and dusk, the atmospheric correction method based on extrapolation inversion of aerosol contribution has different degrees of overestimation of the aerosol contribution in the visible light band due to the large intraday variation of offshore absorbing aerosols, uneven spatial distribution and variable aerosol properties.

[0004] The atmospheric correction method based on aerosol lookup table mainly uses the radiation transfer model to analyze the properties of different types of aerosols, and processes the data according to the lookup table during atmospheric correction. When facing the large solar zenith angle observation environment at dawn and dusk, the atmospheric correction method based on aerosol lookup table cannot accurately simulate the aerosol contribution because the solar zenith angle is large during the dawn and dusk period, which increases the influence of the earth's curvature. There are also large errors when processing satellite observation data.

[0005] The empirical linear method based on measured spectra is a statistical method that assumes that the DN value recorded by the sensor and the measured reflectance value of the corresponding area satisfy a linear relationship, thereby establishing a relationship between the DN value and the reflectance data. When facing the large solar zenith angle observation environment at dawn and dusk, the empirical linear method based on measured spectra requires a large amount of measured spectral data at dawn and dusk. Since measured data is very rare, the algorithm has large regional limitations and is therefore not suitable for observation conditions in high-latitude sea areas.

[0006] Atmospheric correction can be viewed as a nonlinear function approximation of the input spectrum, and machine learning methods can perform this work efficiently and accurately. When faced with the large solar zenith angle observation environment at dawn and dusk, the atmospheric correction method based on machine learning has no ability to process satellite data in the dawn and dusk observation environment because the training data set comes from the radiation transfer model or products of different satellites and the data set only covers data at medium and low solar zenith angles.

[0007] In summary, the construction and verification process of the algorithm used in the existing atmospheric correction method based on the VIIRS satellite is mainly aimed at the observation environment of medium and low solar zenith angles, so it is difficult to cope with the observation environment of large solar zenith angles at dawn and dusk. The solar zenith angle of the VIIRS satellite dawn and dusk data is large, and the light signal is weak. The existing atmospheric correction model fails in this observation environment, and the water color data is seriously missing. Summary of the invention

[0008] In view of the above analysis, an embodiment of the present invention aims to provide a dawn and dusk atmosphere correction method and terminal device based on the VIIRS satellite, so as to solve the problem that the prior art is not suitable for the dawn and dusk solar zenith angle observation environment.

[0009] On the one hand, an embodiment of the present invention provides a morning and evening atmosphere correction method based on the VIIRS satellite, comprising the following steps:

[0010] S1. Obtain the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in the target area in the VIIRS satellite dawn and dusk data, and determine the Rayleigh scattering radiance of each pixel point based on the high-precision Rayleigh scattering lookup table; high precision means that the deviation of the Rayleigh scattering radiance in the table does not exceed 1.07%;

[0011] S2. Obtain the total radiance, flare radiance, and white-hat radiance received by the VIIRS satellite, and then correct the Rayleigh scattering radiance of each pixel point to obtain the Rayleigh corrected radiance of each pixel point;

[0012] S3. Input the solar zenith angle, observation zenith angle, relative azimuth angle, and Rayleigh-corrected radiance of each pixel into the machine learning atmospheric correction model to obtain the aerosol scattering radiance of each pixel;

[0013] S4. The remote sensing reflectance of each pixel point is obtained according to the Rayleigh-corrected radiance of each pixel point and the aerosol radiance of each pixel point.

[0014] The beneficial effects of the above technical solution are as follows: a high-precision dawn and dusk atmospheric correction method for VIIRS satellite dawn and dusk observations is provided, which can generate effective and reliable remote sensing reflectance products. The difficulties in atmospheric correction, namely the calculation problems of Rayleigh scattering radiance and aerosol scattering radiance, are solved by using a high-precision Rayleigh scattering lookup table and a machine learning atmospheric correction model. Using the above atmospheric correction method, the VIIRS satellite dawn and dusk data with a large solar zenith angle observation environment can be effectively processed to obtain remote sensing reflectance products. The processing efficiency is 1 times higher than that of NASA's official processing software SeaDAS, and the amount of effective data obtained by processing is increased by more than 60%. This method can provide high-precision satellite observation data support, and can also provide support for my country's independent water color satellite data processing and application.

[0015] Based on the further improvement of the above method, step S1 further includes:

[0016] S11. Based on the preset spherical Monte Carlo radiation transfer model, obtain a high-precision Rayleigh scattering lookup table for the VIIRS satellite;

[0017] S12. Obtain the solar zenith angle, relative azimuth angle and observation zenith angle of each pixel point in the VIIRS satellite morning and evening data;

[0018] S13. According to the solar zenith angle, relative azimuth angle and observation zenith angle of each pixel point, the above-mentioned high-precision Rayleigh scattering lookup table is searched to determine the Rayleigh scattering radiance of each pixel point.

[0019] Further, step S2 further comprises:

[0020] S21. Process the VIIRS satellite morning and evening data using Seadas software to obtain the total radiance L received by the VIIRS satellite t (λ), flare radiance T(λ)*L g (λ), white hat radiance t'*L wc (λ);

[0021] S22. Based on the total radiance L received by the VIIRS satellite t (λ), flare radiance T(λ)*L g (λ), white hat radiance t'*L wc (λ), the Rayleigh scattering radiance of each pixel point is corrected by the following formula to obtain the Rayleigh corrected radiance L of each pixel point: rc (λ):

[0022] L rc (λ) = L t (λ)-T(λ)*L g (λ)-t'*L wc (λ)-L r (λ)

[0023] Where, L r (λ) is the Rayleigh scattering radiance of each pixel.

[0024] Further, the machine learning atmosphere correction model adopts a neural network atmosphere correction model, and step S3 further includes:

[0025] S31. Establish a neural network atmospheric correction model and train it;

[0026] S32. Input the solar zenith angle, observation zenith angle, relative azimuth, and Rayleigh-corrected radiance of each pixel into the trained neural network atmospheric correction model to obtain the aerosol scattering radiance of each pixel.

[0027] Further, the neural network atmosphere correction model includes the NN-V neural network atmosphere correction model in the following formula:

[0028]

[0029] In the formula, f(y i ) is the aerosol scattering radiance of a certain pixel point, y i is the remote sensing reflectance of each characteristic band of the pixel point, x 1 is the solar zenith angle, x 2 is the relative azimuth, x 3 is the observation zenith angle, x 4 is the Rayleigh-corrected radiance of the pixel at 412 nm, x 5 is the Rayleigh-corrected radiance of the pixel at 443nm; 6 is the Rayleigh-corrected radiance of the pixel at 486 nm, x 7 is the Rayleigh-corrected radiance of the pixel at 551 nm, x 8 is the Rayleigh-corrected radiance of the pixel at 671 nm, a i , b i 、c i 、e i 、f i , g i 、h i 、i i 、j i , k i , l i 、m i 、n i , o i 、p i ,q i are model parameters obtained during model training.

[0030] Furthermore, in step S31, the NN-V neural network atmospheric correction model is further trained through the following sub-steps:

[0031] S311. Acquire multiple morning and evening data images with a valid data volume of more than 20% of the VIIRS satellite, with a time span of one year;

[0032] S312. For each dawn-dusk data image, the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in a set high-latitude sea area in the VIIRS satellite dawn-dusk data are obtained, and the Rayleigh scattering radiance of each pixel point is determined based on a high-precision Rayleigh scattering lookup table;

[0033] S313. For each dawn-dusk data image, the total radiance, flare radiance, and white-hat radiance received by the VIIRS satellite are obtained, and then the Rayleigh scattering radiance of each pixel point is corrected to obtain the Rayleigh corrected radiance of each pixel point;

[0034] S314. Acquire multiple observation data of the polar-orbiting satellite in the set high-latitude sea area, each of which includes the solar zenith angle and aerosol radiance of each pixel point in the observation;

[0035] S315. Identify multiple observation data of the same geographical location in the set high-latitude sea area within a single day, and select multiple observation data of higher quality for the geographical location based on the effective data amount of the pixel frame of the location in the multiple observations being greater than a first set value and the coefficient of variation being less than a second set value;

[0036] S316. For the selected multiple observation data in a single day at the same geographical location, the aerosol radiance corresponding to the minimum solar zenith angle data is linked with the Rayleigh-corrected radiance, solar zenith angle, relative azimuth angle and observation zenith angle corresponding to the maximum solar zenith angle data as a set of training data;

[0037] S317. Train the NN-V neural network atmospheric correction model using the above training data to obtain a trained NN-V neural network atmospheric correction model.

[0038] Further, step S317 further includes:

[0039] Machine learning was performed using the Levenberg-Marquardt back-propagation algorithm, and the k-fold method was used for cross-validation to avoid overfitting, thus obtaining a trained NN-V neural network atmospheric correction model.

[0040] Further, the first set value is 66%, and the second set value is 0.15.

[0041] Further, step S4 further comprises:

[0042] The Rayleigh-corrected radiance of each pixel point is subtracted from the aerosol radiance of each pixel point to obtain the remote sensing reflectance of each pixel point.

[0043] On the other hand, the present invention also discloses a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the morning and evening atmosphere correction method of the above-mentioned VIIRS satellite when executing the computer program.

[0044] This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key features or essential features of the invention, nor is it intended to limit the scope of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0046] Figure 1 A schematic diagram of the steps of the morning and evening atmosphere correction method based on the VIIRS satellite in Example 1 is shown;

[0047] Figure 2 The technical roadmap of the morning and evening atmosphere correction method based on the VIIRS satellite in Example 1 is shown;

[0048] Figure 3 A schematic diagram showing the accuracy of the overall training results of the machine learning atmospheric correction model in Example 2 in different bands of VIIRS is shown;

[0049] Figure 4 A schematic diagram showing the test results of the high-precision VIIRS satellite morning and evening atmosphere correction method using a measured data set in Example 2 is shown;

[0050] Figure 5 It is a scatter plot comparing the atmospheric correction method of Example 2 in the Arctic and nearby high-latitude regions with the multi-year monthly average chlorophyll concentration products released by NASA. DETAILED DESCRIPTION

[0051] Embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to fully convey the scope of the present invention to those skilled in the art.

[0052] As used herein, the term "including" and its variations mean open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "based at least in part on". The terms "an example embodiment" and "an embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one further embodiment". The terms "first", "second", etc. may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0053] The principle involved in the present invention will be described in detail below.

[0054] The purpose of water color remote sensing is to quantitatively obtain and analyze the material content in the water body through the water surface signal obtained by the water color satellite, that is, to establish an inversion algorithm between the water component content and the remote sensing reflectivity. The total signal of the top layer of the atmosphere received by the water color satellite can be expressed by the following formula:

[0055] L t (λ) = L r (λ)+L a (λ)+L ra (λ)+T(λ)*L g (λ)+t(λ)*L wc (λ)+t(λ)*L w (λ) where L r (λ) is the Rayleigh scattered radiance, L a (λ) is the aerosol radiance, L ra (λ) is the radiance produced by the interaction between Rayleigh and aerosol. These three items constitute the atmospheric radiation; L g (λ) is the radiance produced by the solar flare, which is composed of the specular reflection of direct sunlight by the water body; L wc (λ) is the radiance caused by the white cap; L w (λ) is the water-off radiance, which is the final amount to be obtained by atmospheric correction; t(λ) and T(λ) are the atmospheric diffuse transmittance and direct transmittance, respectively, which are related to the components in the atmosphere, such as ozone concentration, water vapor content, and optical thickness of aerosols. Modern ocean color satellites can avoid solar flare radiance by deflecting the measurement angle, while the contribution of white cap radiance to the total radiance is very small and can be ignored when the wind speed is not large. t(λ) and T(λ) can be calculated by the following formula:

[0056]

[0057]

[0058] In the formula, τ r (λ) is the Rayleigh scattering optical thickness, τ a(λ) is the aerosol optical depth, τ oz (λ) is the ozone optical density.

[0059] There are two problems with the existing atmospheric correction methods.

[0060] 1. Calculation of Rayleigh scattering radiance

[0061] The existing Rayleigh scattering lookup table generated by the radiation transfer model can find the Rayleigh scattering radiance according to the information such as wind speed, air pressure, wavelength, etc. However, most of the current radiation transfer models are based on the parallel plane layering assumption, that is, based on the assumption of a flat atmosphere that extends infinitely in the horizontal direction, without considering the influence of the earth's curvature / atmosphere sphericity. This limits the generated Rayleigh scattering lookup table to high-precision calculations only under medium-low solar zenith angle (<70°) detection conditions. However, the existing VIIRS satellite generates a large amount of data with a solar zenith angle exceeding 70° when detecting in high-latitude sea areas and at dawn and dusk, so it is urgent to develop a targeted Rayleigh scattering lookup table.

[0062] 2. Calculation of aerosol scattering radiance

[0063] The key and difficulty of atmospheric correction for water color remote sensing is the calculation of aerosol radiance. However, due to the complex and changeable properties of aerosols in the large solar zenith angle observation environment, it is difficult to accurately estimate the aerosol contribution. The influence of the earth's curvature increases, and the radiation transfer model based on the parallel plane assumption cannot make an accurate simulation of the aerosol contribution. Due to the problem of aerosol estimation, the existing atmospheric correction models in the world can only effectively process observation data with a solar zenith angle of less than 70 degrees. In high-latitude areas, there is a serious lack of effective data.

[0064] Example 1

[0065] One embodiment of the present invention discloses a morning and evening atmosphere correction method based on the VIIRS satellite, such as Figure 1 As shown, the following steps are included:

[0066] S1. Obtain the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in the target area in the VIIRS satellite dawn and dusk data, and determine the Rayleigh scattering radiance of each pixel point based on the high-precision Rayleigh scattering lookup table.

[0067] Specifically, high precision refers to the deviation of the Rayleigh scattering radiance in the table not exceeding 1.07%. Deviation is also called apparent error, which refers to the difference between an individual measured value and the measured average value, and can be used to measure the precision of the measurement result. The high-precision Rayleigh scattering lookup table is not limited to the high-precision Rayleigh scattering lookup table of Example 2, and actual measured values ​​can also be used.

[0068] S2. Obtain the total radiance, flare radiance, and white-hat radiance received by the VIIRS satellite, and then correct the Rayleigh scattering radiance of each pixel point to obtain the Rayleigh corrected radiance of each pixel point.

[0069] Specifically, the method of correcting the Rayleigh scattered radiance of each pixel point according to the total radiance, flare radiance, and white hat radiance is not limited to the method described in Example 2. A fitting formula can be used to obtain it more accurately and the coefficients of the fitting formula can be calibrated in advance.

[0070] S3. Input the solar zenith angle, observation zenith angle, relative azimuth, and Rayleigh-corrected radiance of each pixel into the machine learning atmospheric correction model to obtain the aerosol scattering radiance of each pixel.

[0071] Specifically, the machine learning atmospheric correction model is not limited to the NN-V neural network atmospheric correction model described in Example 2, and other existing learning networks may also be implemented, such as other existing neural networks.

[0072] S4. The remote sensing reflectance of each pixel point is obtained according to the Rayleigh-corrected radiance of each pixel point and the aerosol radiance of each pixel point.

[0073] Specifically, the method for calculating the remote sensing reflectivity is not limited to the subtraction method described in Example 2, and a fitting formula may also be used to calibrate the coefficients of the fitting formula in advance to obtain more accurate results.

[0074] When implemented, the technical route of this method is as follows Figure 2 As shown, step S2 uses the Rayleigh scattered radiance obtained in step S1 to generate the Rayleigh corrected radiance; step S3 uses the Rayleigh corrected radiance generated in step S2 to build a model and calculate the aerosol scattered radiance; step S4 uses the Rayleigh corrected radiance and aerosol scattered radiance generated in steps S2 and S3 to finally generate a remote sensing reflectance product.

[0075] Compared with the prior art, this embodiment provides a high-precision dawn and dusk atmospheric correction method for VIIRS satellite dawn and dusk observations, which can generate effective and reliable remote sensing reflectance products. The difficulties in atmospheric correction, namely the calculation problem of Rayleigh scattering radiance and the calculation problem of aerosol scattering radiance, are solved by using a high-precision Rayleigh scattering lookup table and a machine learning atmospheric correction model. Using the above-mentioned atmospheric correction method, the dawn and dusk data of VIIRS satellites with a large solar zenith angle observation environment can be effectively processed to obtain remote sensing reflectance products. The processing efficiency is 1 times higher than that of NASA's official processing software SeaDAS, and the amount of effective data obtained by processing is increased by more than 60%. This method can provide high-precision satellite observation data support, and can also provide support for my country's independent water color satellite data processing and application.

[0076] Example 2

[0077] An improvement is made on the basis of Example 1, wherein high precision means that the deviation of the Rayleigh scattered radiance within the range of 70° of the solar zenith angle does not exceed 1.07%.

[0078] Preferably, step S1 further includes the following sub-steps S11 to S13:

[0079] S11. Based on the preset spherical Monte Carlo radiation transfer model, a high-precision Rayleigh scattering lookup table for the VIIRS satellite is obtained.

[0080] Specifically, the spherical Monte Carlo radiation transfer model refers to Xu Fei's doctoral thesis published in 2019 - "Research on Three-Dimensional Spherical Monte Carlo Radiation Transfer Model and Atmospheric Correction under Complex Conditions"; the spherical Monte Carlo radiation transfer model simulates the main parameters of the propagation process of photons at the ocean-atmosphere coupling interface: during the simulation, the number of photons is set to 1000000, the symmetry factor is set to 0.0, the depolarization factor is set to 0.029444999992847, the water refractive index is set to 1.34, the atmospheric stratification is set to 32 layers, the number of solar zenith angles is set to 45, from 0 to 90°, the number of observation zenith angles is set to 41, from 0 to 90°, and the number of relative azimuth directions is set to 19, from 0 to 180°.

[0081] S12. Obtain the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in the VIIRS satellite dawn and dusk data.

[0082] Specifically, the Seadas software was used to process the VIIRS satellite morning and evening data to obtain the solar zenith angle, relative azimuth angle, and observation zenith of each pixel point.

[0083] S13. According to the solar zenith angle, relative azimuth angle and observation zenith angle of each pixel point, the above-mentioned high-precision Rayleigh scattering lookup table is searched to determine the Rayleigh scattering radiance of each pixel point.

[0084] Specifically, the Rayleigh scattering radiance corresponding to each pixel point is found in a high-precision Rayleigh scattering lookup table according to the solar zenith angle, the relative azimuth angle, and the observation zenith angle.

[0085] Preferably, step S2 further includes the following sub-steps S21 to S22:

[0086] S21. Process the VIIRS satellite morning and evening data using Seadas software to obtain the total radiance L received by the VIIRS satellite t (λ), flare radiance T(λ)*L g (λ), white hat radiance t'*L wc (λ); λ is the wavelength and t' is the diffuse transmittance.

[0087] S22. Based on the total radiance L received by the VIIRS satellite t (λ), flare radiance T(λ)*L g (λ), white hat radiance t'*L wc (λ), the Rayleigh scattered radiance of each pixel point obtained in step S1 is corrected by the following formula to obtain the Rayleigh corrected radiance L of each pixel point: rc (λ):

[0088] L rc (λ)=L t (λ)-T(λ)*L g (λ)-t'*L wc (λ)-L r (λ)

[0089] Where, L r (λ) is the Rayleigh scattering radiance of each pixel.

[0090] Preferably, step S3 further comprises:

[0091] S31. Establish and train the NN-V (Neural network VIIRS) neural network atmospheric correction model;

[0092] S32. Input the solar zenith angle, observation zenith angle, relative azimuth, and Rayleigh-corrected radiance of each pixel into the trained NN-V neural network atmospheric correction model to obtain the aerosol scattering radiance of each pixel.

[0093] Preferably, the machine learning atmospheric correction model includes the NN-V neural network atmospheric correction model in the following formula:

[0094]

[0095] In the formula, f(y i ) is the aerosol scattering radiance of a certain pixel point, y i is the remote sensing reflectance of each characteristic band of the pixel point, y 1 is the remote sensing reflectance of the pixel at 412nm; 2 is the remote sensing reflectance of the pixel at 443nm; 3 is the remote sensing reflectance of the pixel at 486nm; 4 is the remote sensing reflectance of the pixel at 551nm; 5 is the 671nm band remote sensing reflectance of the pixel point, x 1 is the solar zenith angle, x 2 is the relative azimuth, x 3 is the observation zenith angle, x 4is the Rayleigh-corrected radiance of the pixel at 412 nm, x 5 is the Rayleigh-corrected radiance of the pixel at 443nm; 6 is the Rayleigh-corrected radiance of the pixel at 486 nm, x 7 is the Rayleigh-corrected radiance of the pixel at 551 nm, x 8 is the Rayleigh-corrected radiance of the pixel at 671 nm, a i , b i 、c i 、e i 、f i , g i 、h i 、i i 、j i , k i , l i 、m i 、n i , o i 、p i ,q i is the model parameter, obtained during model training, a i is the compensation coefficient, b i 、c i 、e i 、f i , g i 、h i 、i i 、j i , k i , l i 、m i 、n i , o i 、p i ,q i are the fitting coefficients of each dependent variable.

[0096] By training the above NN-V neural network atmospheric correction model, the model parameter a is obtained. i , b i 、c i 、e i 、f i , g i 、h i 、i i 、j i , k i , l i 、m i 、n i , o i 、p i ,q i .

[0097] Preferably, in step S31, the NN-V neural network atmospheric correction model is further trained through the following sub-steps:

[0098] S310. Obtain the remote sensing reflectance of each pixel point in the set high latitude (the area between 60 degrees north and south latitude on the earth's surface and the North and South Poles) sea area through NASA standard products as a set of input data in the training data;

[0099] S311. Acquire multiple morning and evening data images with a valid data volume of more than 20% of the VIIRS satellite, with a time span of one year;

[0100] S312. For each dawn-dusk data image, the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in a set high-latitude sea area in the VIIRS satellite dawn-dusk data are obtained, and the Rayleigh scattering radiance of each pixel point is determined based on a high-precision Rayleigh scattering lookup table;

[0101] S313. For each morning and evening data image, the total radiance, flare radiance, and white hat radiance received by the VIIRS satellite are obtained, and then the Rayleigh scattering radiance of each pixel point is corrected to obtain the Rayleigh corrected radiance of each pixel point; (the input data includes the Rayleigh corrected radiance of the pixel point, the solar zenith angle, the relative azimuth, the observation zenith angle, and the corresponding aerosol radiance after linking. Linking has not been performed here, so there is no aerosol radiance. For the specific linking method, see step S316)

[0102] S314. Acquire multiple observation data of the polar-orbiting satellite in the set high-latitude sea area, each of which includes the solar zenith angle and aerosol radiance of each pixel point in the observation;

[0103] S315. Identify multiple observation data of the same geographical location in the set high-latitude sea area within a single day, and select multiple observation data of higher quality for the geographical location based on the effective data amount of the pixel frame of the location in the multiple observations being greater than a first set value and the coefficient of variation being less than a second set value;

[0104] S316. Link multiple observation data of polar-orbiting satellites in set high-latitude sea areas, that is, for multiple observation data of the same geographical location selected within a single day, record the aerosol radiance corresponding to the minimum solar zenith angle data and the Rayleigh-corrected radiance, solar zenith angle, relative azimuth and observation zenith angle corresponding to the maximum solar zenith angle data together as a set of training data; link the data of the past year to form a data set.

[0105] S317. Train the NN-V neural network atmospheric correction model using the training data in the above data set to obtain a trained NN-V neural network atmospheric correction model.

[0106] Preferably, step S317 further includes: performing machine learning through the Levenberg-Marquardt back propagation algorithm, and using the k-fold method for cross-validation to avoid overfitting, so as to obtain a trained NN-V neural network atmospheric correction model.

[0107] Among them, the Levenberg-Marquardt algorithm is an optimization algorithm that combines the advantages of gradient descent and Gauss-Newton method, with fast convergence and good stability. Specifically, the following steps are included: First, the weights and bias parameters of the neural network are randomly initialized. By inputting data, the output of the neural network is calculated. The output of each layer is calculated by the output of the previous layer through the activation function. The loss function is used to measure the gap between the model prediction value and the true value. From the output layer to the input layer, the gradient of the loss function for each parameter is calculated. This step uses the chain rule and the back propagation algorithm. The Levenberg-Marquardt algorithm is used to update the parameters based on the gradient information to minimize the loss function. The algorithm dynamically adjusts the learning rate to accelerate convergence and improve stability. Repeat the forward propagation, loss calculation, back propagation, and parameter update steps until the stop condition is reached (maximum number of iterations 1000 or loss function convergence).

[0108] Among them, k-fold cross validation is a model evaluation technique used to evaluate the performance and generalization ability of the model while reducing the dependence on a single data segmentation. Specifically, the following steps are included: Divide the data set into 10 mutually exclusive subsets, usually one of which is used as a validation set and the remaining subsets are used as training sets. For each subset, use it as a validation set and the remaining subsets as training sets, train the model and calculate performance indicators (relative deviation). For each validation set, use the trained model for prediction and calculate performance indicators. Average or weighted average the performance indicators obtained from k validations as the performance evaluation result of the model on the entire data set. According to the cross-validation results, adjust the hyperparameters of the model. Select the optimal model based on the cross-validation results, and usually select the model with the best performance on the validation set as the final model.

[0109] Preferably, the first set value is 66% and the second set value is 0.15.

[0110] Preferably, step S4 further comprises: subtracting the Rayleigh corrected radiance of each pixel point from the aerosol radiance of each pixel point to obtain the remote sensing reflectance of each pixel point.

[0111] During implementation, by inputting VIIRS satellite data, the remote sensing reflectance product after atmospheric correction can be obtained. Figure 3 and Figure 4This is the result of the accuracy test of the model using global ocean measured data, and it can be seen that the accuracy is relatively high.

[0112] Figure 3 This is a schematic diagram of the accuracy of the overall training results of the machine learning atmospheric correction model in different bands of VIIRS, where NN-V R rs (sr -1 ) is a remote sensing reflectivity product obtained by inversion using the NN-V model of the present invention; NASA R rs (sr -1 ) is a remote sensing reflectance product released by NASA. Figure 3 (a) is the comparison between the inversion result and the actual value in the 412nm band. Figure 3 (b) is the comparison between the inversion result and the actual value in the 443nm band. Figure 3 (c) is the comparison between the inversion result and the actual value in the 486nm band. Figure 3 (d) is the comparison between the inversion result and the actual value in the 551nm band. Figure 3 (e) in the figure is the comparison between the inversion result and the actual value in the 671nm band.

[0113] Figure 4 Schematic diagram of the accuracy of the overall inversion results of the machine learning atmospheric correction model in different bands of VIIRS using the global ocean measured data set, where NN-V R rs (sr -1 ) is the remote sensing reflectivity product obtained by inversion using the NN-V model of the present invention, OC-CCI R rs (sr -1 ) is the remote sensing reflectance product in the measured dataset OC-CCI. Figure 4 (a) is the comparison between the inversion result and the measured value in the 412nm band. Figure 4 (b) is the comparison between the inversion result and the measured value in the 443nm band. Figure 4 (c) is the comparison between the inversion result and the measured value in the 486nm band. Figure 4 (d) is the comparison between the inversion result and the measured value in the 551nm band. Figure 4 (e) is the comparison between the inversion result and the measured value in the 671nm band.

[0114] Figure 5 It is a scatter plot comparing the atmospheric correction method of Example 2 in the Arctic and nearby high-latitude regions with the multi-year monthly average chlorophyll concentration products released by NASA, wherein NN-V Rrs(sr-1) is the remote sensing reflectance product obtained by inverting the NN-V model of the present invention, and NASA VIIRS Rrs(sr-1) is the remote sensing reflectance product of VIIRS released by NASA. Figure 5 The middle left picture shows the comparison data in October 2020, and the right picture shows the comparison data in January 2021. It can be seen that the results of the two are basically consistent in areas with lower latitudes, indicating that the accuracy of the model constructed by the present invention is reliable.

[0115] Compared with the prior art, the morning and evening atmosphere correction method based on the VIIRS satellite provided in this embodiment has the following beneficial effects:

[0116] 1. At present, most radiation transfer models are based on the assumption of a plane atmosphere that extends infinitely in the horizontal direction, and the influence of the earth curvature / atmosphere sphericity is not considered, which limits the generated Rayleigh scattering lookup table to be calculated only under the detection conditions of medium-low solar zenith angle (<70°). Step S1 uses a preset spherical Monte Carlo radiation transfer model to calculate the Rayleigh scattering lookup table, considering the influence of the earth curvature under a large solar zenith angle. The generated Rayleigh scattering lookup table has high precision and is easy to use. The spherical Monte Carlo radiation transfer model simulates the radiation transfer process, which can effectively avoid the integral solution of the three-dimensional radiation transfer equation. The spherical Monte Carlo radiation transfer model simulates the operation process of photons in the atmosphere and sea surface, including atmospheric molecular Rayleigh scattering, aerosol scattering, cross scattering of the two, and calm / rough sea surface specular reflection and sea surface white foam diffuse reflection. The earth curvature factor can also be introduced. The established radiation transfer model is closer to the actual photon operation process, without making the formula more complicated, and without introducing approximation errors. Therefore, a high-precision Rayleigh scattering lookup table suitable for VIIRS satellites can be simulated.

[0117] 2. Step S2 uses the accurate Rayleigh scattered radiance to perform calculations, thereby obtaining the accurate Rayleigh corrected radiance.

[0118] 3. Atmospheric correction can be regarded as a nonlinear function approximation of the input spectrum. The use of machine learning methods can perform this work efficiently and accurately. The key to whether the machine learning atmospheric correction model can successfully process satellite data lies in the construction of the machine learning training data set. This patent uses the selected high-quality satellite data set for machine learning, bypassing the calculation of aerosol scattering radiance to directly obtain the final remote sensing reflectance product. The pixel frame, effective data volume (greater than 66%) and coefficient of variation (less than 0.15) used when screening data during the construction of the machine learning atmospheric correction model training data set are the optimal values ​​determined after a large number of experiments. These values ​​ensure the consistency of the selected high-quality data within the time range, avoid rapid changes in water bodies caused by algal blooms or strong currents, and ensure the accuracy of the machine learning atmospheric correction model training set. The machine learning algorithm uses the Levenberg-Marquardt back propagation algorithm, and uses the k-fold method for cross-validation to avoid overfitting. The advantage of this method is that the randomly generated parts are repeatedly used for training and verification, and the results are verified each time. The machine learning atmospheric correction model constructed using this method is highly efficient in processing satellite data. The processing time for a single image is 1 times shorter than that of NASA's official processing software. In high-latitude sea areas or dawn and dusk observation environments, the amount of effective data is increased by more than 60%.

[0119] 4. Accurate remote sensing reflectance products can be obtained by using accurate Rayleigh scattering radiance and aerosol radiance for calculation. The preset spherical Monte Carlo radiation transfer model and machine learning atmospheric correction model solve the difficulties in atmospheric correction, namely the calculation of Rayleigh scattering radiance and aerosol scattering radiance.

[0120] Embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The terms used herein are selected to best explain the principles of the embodiments, practical applications, or improvements over the prior art, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A morning and evening atmospheric correction method based on VIIRS satellite, characterized in that: Includes steps: S1. Obtain the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in the target area in the VIIRS satellite dawn and dusk data, and determine the Rayleigh scattering radiance of each pixel point based on a high-precision Rayleigh scattering lookup table; High precision means that the deviation of Rayleigh scattering radiance within the meter does not exceed 1.07%; S2. Obtain the total radiance, flare radiance, and white-hat radiance received by the VIIRS satellite, and then correct the Rayleigh scattering radiance of each pixel point to obtain the Rayleigh corrected radiance of each pixel point; S3. Input the solar zenith angle, observation zenith angle, relative azimuth angle, and Rayleigh-corrected radiance of each pixel into the machine learning atmospheric correction model to obtain the aerosol scattering radiance of each pixel; The machine learning atmosphere correction model adopts a neural network atmosphere correction model, and step S3 further includes: S31. Establish a neural network atmospheric correction model and train it; The neural network atmospheric correction model includes the NN-V neural network atmospheric correction model in the following formula: In the formula, f(y i ) is the aerosol scattering radiance of a certain pixel point, y i is the remote sensing reflectance of each characteristic band of the pixel point, x1 is the solar zenith angle, x2 is the relative azimuth, x3 is the observation zenith angle, x4 is the Rayleigh corrected radiance of the 412nm band of the pixel point, x5 is the Rayleigh corrected radiance of the 443nm band of the pixel point; x6 is the Rayleigh corrected radiance of the 486nm band of the pixel point, x7 is the Rayleigh corrected radiance of the 551nm band of the pixel point, x8 is the Rayleigh corrected radiance of the 671nm band of the pixel point, a i 、b i 、c i 、e i 、f i , g i 、h i 、i i 、j i , k i , l i 、m i 、n i , o i 、p i ,q i are model parameters, obtained during model training; In step S31, the NN-V neural network atmospheric correction model is further trained through the following sub-steps: S311. Acquire multiple morning and evening data images with a valid data volume of more than 20% of the VIIRS satellite, with a time span of one year; S312. For each dawn-dusk data image, the solar zenith angle, relative azimuth angle, and observation zenith angle of each pixel point in a set high-latitude sea area in the VIIRS satellite dawn-dusk data are obtained, and the Rayleigh scattering radiance of each pixel point is determined based on a high-precision Rayleigh scattering lookup table; S313. For each dawn-dusk data image, the total radiance, flare radiance, and white-hat radiance received by the VIIRS satellite are obtained, and then the Rayleigh scattering radiance of each pixel point is corrected to obtain the Rayleigh corrected radiance of each pixel point; S314. Acquire multiple observation data of the polar-orbiting satellite in the set high-latitude sea area, each of which includes the solar zenith angle and aerosol radiance of each pixel point in the observation; S315. Identify multiple observation data of the same geographical location in the set high-latitude sea area within a single day, and select multiple observation data of higher quality for the geographical location based on the effective data amount of the pixel frame of the location in the multiple observations being greater than a first set value and the coefficient of variation being less than a second set value; S316. For the selected multiple observation data in a single day at the same geographical location, the aerosol radiance corresponding to the minimum solar zenith angle data is linked with the Rayleigh-corrected radiance, solar zenith angle, relative azimuth angle and observation zenith angle corresponding to the maximum solar zenith angle data as a set of training data; S317. Train the NN-V neural network atmospheric correction model using the above training data to obtain a trained NN-V neural network atmospheric correction model; S32. Inputting the solar zenith angle, observation zenith angle, relative azimuth angle, and Rayleigh-corrected radiance of each pixel into the trained neural network atmospheric correction model to obtain the aerosol scattering radiance of each pixel; S4. The remote sensing reflectance of each pixel point is obtained according to the Rayleigh-corrected radiance of each pixel point and the aerosol radiance of each pixel point.

2. The method for morning and evening atmosphere correction based on VIIRS satellite according to claim 1, characterized in that: Step S1 further comprises: S11. Based on the preset spherical Monte Carlo radiation transfer model, obtain a high-precision Rayleigh scattering lookup table for the VIIRS satellite; S12. Obtain the solar zenith angle, relative azimuth angle and observation zenith angle of each pixel point in the VIIRS satellite morning and evening data; S13. According to the solar zenith angle, relative azimuth angle and observation zenith angle of each pixel point, the above-mentioned high-precision Rayleigh scattering lookup table is searched to determine the Rayleigh scattering radiance of each pixel point.

3. The method for morning and evening atmosphere correction based on VIIRS satellite according to claim 2, characterized in that: Step S2 further comprises: S21. Process the VIIRS satellite morning and evening data using Seadas software to obtain the total radiance L received by the VIIRS satellite t (λ), flare radiance T(λ)*L g (λ), white hat radiance t′*L wc (λ); S22. Correct the Rayleigh scattering radiance of each pixel point by the following formula to obtain the Rayleigh corrected radiance L of each pixel point: rc (λ): L rc (λ)=L t (λ)-T(λ)*L g (λ)-t′*L wc (l)-L r (l) Where, L r (λ) is the Rayleigh scattering radiance of each pixel.

4. The method for morning and evening atmosphere correction based on VIIRS satellite according to claim 1, characterized in that: Step S317 further includes: Machine learning was performed using the Levenberg-Marquardt back-propagation algorithm, and the k-fold method was used for cross-validation to avoid overfitting, thus obtaining a trained NN-V neural network atmospheric correction model.

5. The method for morning and evening atmosphere correction based on VIIRS satellite according to claim 4, characterized in that: The first setting value is 66%, and the second setting value is 0.

15.

6. The method for morning and evening atmosphere correction based on VIIRS satellite according to claim 5, characterized in that: Step S4 further comprises: The Rayleigh-corrected radiance of each pixel is subtracted from the aerosol radiance of the corresponding pixel to obtain the remote sensing reflectance of each pixel.

7. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for morning and evening atmosphere correction of the VIIRS satellite as described in any one of claims 1 to 6 is implemented.

Images