Rapid atmospheric correction method for turbid water

Abstract

The present invention belongs to the field of photogrammetry and remote sensing technology, and relates to a method for rapid atmospheric correction of turbid water bodies, comprising: 1) using a 6SV radiation transmission code to simulate a CZI sensor receiving signal affected by the atmosphere to obtain a data set; 2) constructing a fully connected neural network model affected by weather according to the data set obtained in step 1); 3) performing atmospheric correction on measured remote sensing data according to the fully connected neural network model affected by weather constructed in step 2). The present invention provides a method for rapid atmospheric correction of turbid water bodies that can quickly and accurately obtain atmospheric correction parameters, improve the generalization ability of the atmospheric correction method, and greatly improve the efficiency and accuracy of atmospheric correction.

Description

Technical Field

[0001] The invention belongs to the technical field of photogrammetry and remote sensing, and relates to a method for rapid atmospheric correction of turbid water bodies, and in particular to a method for rapid atmospheric correction of turbid water bodies based on a multi-layer fully connected neural network. Background Art

[0002] Atmospheric correction (AC) is the first step in the ocean color algorithm. 80% of the water body signal acquired by the satellite comes from atmospheric noise, and the water body signal accounts for less than 20%. Ideally, the atmosphere should remove the atmospheric radiation contribution (including the radiation contribution of air molecules and aerosols) and retain the surface reflectance of the water body measured by the satellite (remote sensing reflectivity). In turbid water bodies, standard atmospheric correction algorithms usually show large uncertainties, which may cause negative values ​​of off-water reflectivity or remote sensing reflectivity, and there is a large deviation from the measured data. This is because standard atmospheric correction faces the challenges of complex meteorological and surface conditions when processing turbid water images, especially the lack of sensors in the shortwave infrared band. Therefore, the development of high-precision atmospheric correction algorithms for turbid water bodies is of great significance for water color research.

[0003] Traditional atmospheric correction algorithms usually use manual design of features and model parameters, usually have limited data processing capabilities, have poor generalization capabilities, and use manual adjustment of model parameters for training, based on statistical or geometric theories, which usually requires a lot of manual feature engineering. Obviously, the main problems of traditional atmospheric correction algorithms are: high computational complexity, high sensitivity to input parameters and assumptions, poor generalization capabilities, and large storage space for lookup table data.

[0004] The neural network method uses automatic learning features and parameters, has strong data processing capabilities, can handle high-dimensional and large-scale data, and has good generalization capabilities and can adapt to new data. The neural network method is based on artificial intelligence theory and uses automatic learning for training, which can automatically extract features through automatic learning.

[0005] In recent years, with the rapid development of computer software technology and neural network technology, more and more people are using neural networks instead of traditional computing methods. Neural network algorithms can directly learn the complex mapping relationship between surface reflectance and remote sensing images from raw data without manually designing features or assuming physical models, thus realizing an end-to-end learning and correction process. Compared with linear or approximate linear modeling of traditional methods, neural network algorithms can capture more complex nonlinear relationships and more accurately model the complex relationship between surface reflectance and atmospheric influence. At present, domestic research institutions, universities and enterprises are actively conducting relevant research on the optimization of atmospheric correction algorithms, involving theoretical research, algorithm optimization, engineering applications and other aspects. Summary of the invention

[0006] In order to solve the above-mentioned technical problems existing in the background technology, the present invention provides a method for rapid atmospheric correction of turbid water bodies, which can quickly and accurately obtain atmospheric correction parameters, improve the generalization ability of the atmospheric correction method, and greatly improve the efficiency and accuracy of atmospheric correction.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A method for rapid atmospheric correction of turbid water, characterized in that the method comprises the following steps:

[0009] 1) Use the 6SV radiation transfer code to simulate the CZI sensor receiving signal affected by the atmosphere and obtain the data set;

[0010] 2) constructing a fully connected neural network model affected by weather based on the data set obtained in step 1);

[0011] 3) Performing atmospheric correction on the measured remote sensing data based on the weather-affected fully connected neural network model constructed in step 2).

[0012] Preferably, the factors affected by the atmosphere in step 1) are the factors affected by Rayleigh scattering rate, the factors affected by aerosol scattering reflectivity and the factors affected by aerosol scattering transmittance.

[0013] Preferably, the specific implementation of step 1) is:

[0014] 1.1) Add the spectral response function of CZI to the 6SV code;

[0015] 1.2) The CZI sensor receiving signals affected by Rayleigh scattering rate, aerosol scattering reflectivity and aerosol scattering transmittance are simulated in the 6SV code.

[0016] Preferably, the factors affected by the Rayleigh scattering rate include atmospheric mode and observation geometry; the atmospheric mode includes tropical, mid-latitude summer, mid-latitude winter and near-polar summer; the observation geometry includes the solar zenith angle θ 0 、Sensor zenith angle θ v and relative azimuth

[0017] The factors affected by the aerosol scattering reflectivity include aerosol type, observation geometry, and aerosol optical thickness at 550nm; the aerosol type includes marine aerosol, urban aerosol, and continental aerosol; the observation geometry includes the solar zenith angle θ 0 、Sensor zenith angle θ v and relative azimuth

[0018] The factors affected by the aerosol scattering transmittance include aerosol type, observation geometry, and aerosol optical thickness at 550nm; the aerosol type includes marine aerosol, urban aerosol, and continental aerosol; the observation geometry includes the solar zenith angle θ 0 and the sensor zenith angle θ v .

[0019] Preferably, the fully connected neural network model affected by weather in step 2) includes a Rayleigh scattering reflectivity fully connected neural network model, an aerosol scattering reflectivity fully connected neural network model and an aerosol scattering transmittance fully connected neural network model.

[0020] Preferably, the Rayleigh scattering reflectivity fully connected neural network model is constructed in the following manner:

[0021] a.1) Obtaining the input parameters of Rayleigh scattering, wherein the input parameters of Rayleigh scattering include the solar zenith angle θ 0 、Sensor zenith angle θ v and relative azimuth

[0022] a.2) The input parameter of Rayleigh scattering obtained in step a.1) is the solar zenith angle θ 0 (0-88), sensor zenith angle θ v (0-84) and relative azimuth (0-180) divided by the maximum value (θ 0 ,θ v , The data were normalized by dividing by 88, 84 and 180 respectively to obtain the normalized data set;

[0023] a.3) Divide the normalized data set obtained in step a.2) into a training set and a test set according to a ratio of 0.7 / 0.3;

[0024] a.4) using the training set obtained by step a.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=2000, BETCH_SIZE=128, to obtain a trained fully connected neural network model;

[0025] a.5) Using the test set obtained in step a.3) to verify the trained fully connected neural network model obtained in step a.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, the network converges and a fully connected neural network model of Rayleigh scattering reflectivity is obtained.

[0026] Preferably, the aerosol scattering reflectivity fully connected neural network model is constructed in the following manner:

[0027] b.1) Obtaining input parameters of aerosol scattering reflectivity, wherein the input parameters of aerosol scattering reflectivity include the solar zenith angle θ 0 、Sensor zenith angle θ v , relative azimuth and 550nm aerosol optical depth;

[0028] b.2) The input parameter of the aerosol scattering reflectivity obtained in step b.1) is the solar zenith angle θ 0 (0-88), sensor zenith angle θ v (0-84) and relative azimuth (0-180) and the aerosol optical thickness divided by the maximum value (θ 0 / 88,θ v / 84, Aerosol optical thickness (AOD / 1.2) was normalized to obtain a normalized data set;

[0029] b.3) Divide the normalized data set obtained in step b.2) into a training set and a test set according to a ratio of 0.7 / 0.3;

[0030] b.4) using the training set obtained by step b.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=1000, BETCH_SIZE=256, to obtain a trained fully connected neural network model;

[0031] b.5) ​​Using the test set obtained in step b.3) to verify the trained fully connected neural network model obtained in step b.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, relative deviation is less than 10 -5 The network convergence is completed when , and the aerosol scattering reflectivity fully connected neural network model is obtained.

[0032] Preferably, the aerosol scattering transmittance fully connected neural network model is constructed in the following manner:

[0033] c.1) Obtaining input parameters of aerosol scattering transmittance, wherein the input parameters of aerosol scattering transmittance include the solar zenith angle θ 0 、Sensor zenith angle θ v and 550nm aerosol optical depth;

[0034] c.2) The solar zenith angle θ is the input parameter of the aerosol scattering transmittance obtained in step c.1) 0 (0-88), sensor zenith angle θ v (0-84) and aerosol optical depth (θ 0 / 88,θ v / 84, aerosol optical depth AOD / 1.2) is divided by the maximum value to obtain a normalized data set;

[0035] c.3) Divide the normalized data set obtained in step c.2) into a training set and a test set according to a ratio of 0.8 / 0.2;

[0036] c.4) using the training set obtained in step c.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=2000, BETCH_SIZE=128, to obtain a trained fully connected neural network model;

[0037] c.5) Using the test set obtained in step c.3) to verify the trained fully connected neural network model obtained in step c.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, the network converges and a fully connected neural network model of aerosol scattering transmittance is obtained.

[0038] Preferably, the specific implementation of step 3) is:

[0039] 3.1) According to the weather-affected fully connected neural network model constructed in step 2), predict the Rayleigh scattering reflectivity ρ under the specified atmospheric type r , aerosol reflectivity ρ under specified aerosol type a and transmittance t;

[0040] 3.2) Calculate the reflectivity ρt of the top atmosphere;

[0041] 3.3) Combining the results of step 3.1) and step 3.2), atmospheric correction is performed on the measured remote sensing data.

[0042] Preferably, the specific implementation method of the top atmosphere reflectivity ρt in step 3.2) is:

[0043]

[0044] in:

[0045] L t is the CZI radiance data;

[0046] F 0is the solar irradiance in the outer atmosphere;

[0047] θ 0 is the solar zenith angle.

[0048] The advantages of the present invention are:

[0049] Aiming at the shortcomings of the current atmospheric correction algorithm of satellite remote sensing images, such as large data storage volume, slow calculation speed, high calculation complexity, and poor model generalization ability, the present invention proposes a method for rapid atmospheric correction of turbid water bodies, including: 1) the training data set is generated by 6SV (Second Simulation of a Satellite Signal in the Solar Spectrum Vector) radiation transmission code simulation. Atmospheric correction requires obtaining Rayleigh scattering reflectivity, aerosol scattering reflectivity and transmittance under the sensor observation geometry. According to the sensor parameter settings and observation geometry of the satellite, the absorption, scattering and transmission characteristic parameters of the solar radiation in the process of passing through the atmosphere and entering the satellite sensor are simulated under different atmospheric modes, different aerosol types and different aerosol optical thicknesses, and the air molecules, aerosols and the mixed absorption, scattering and transmission of the two are mixed; 2) The simulated data is used as a training sample, and three neural networks are trained respectively to fit the Rayleigh scattering reflectivity, aerosol scattering reflectivity and aerosol scattering transmittance, so as to quickly and accurately obtain the above three atmospheric correction parameters; 3) Obtain satellite remote sensing data, calculate the apparent reflectivity through radiation calibration, and finally obtain the surface reflectivity data using the above three neural networks, and verify it using 6SV simulation and field measured data. The present invention is directly based on the atmospheric correction parameters simulated by the 6SV model, and network training can be directly performed. Specifically, it relates to a method that can quickly and accurately obtain atmospheric correction parameters and save computer memory space compared to traditional atmospheric correction methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is the spectral response curve of different wavelengths;

[0051] Figure 2 It is the process of adding band response functions;

[0052] Figure 3 It is the 6SV simulation result;

[0053] Figure 4 It is a mathematical model of a multi-layer fully connected neural network commonly used in the prior art;

[0054] Figure 5 It is a fully connected neural network model commonly used in the prior art;

[0055] Figure 6This is a comparison chart between the Rayleigh scattering predicted by the neural network and the 6SV simulation results;

[0056] Figure 7 This is a comparison chart of aerosol scattering predicted by the neural network and the 6SV simulation results;

[0057] Figure 8 This is a comparison chart of the aerosol transmission predicted by the neural network and the 6SV simulation results;

[0058] Figure 9-12 They are comparisons of four sets of field-measured water remote sensing reflectances collected from Axe Lake and the remote sensing reflectances after atmospheric correction during synchronous transit.

[0059] Fig.13 The present invention provides a flow chart of a method for rapid atmospheric correction of turbid water bodies. DETAILED DESCRIPTION

[0060] See also Fig.13 The present invention provides a method for rapid atmospheric correction of turbid water, which specifically comprises the following steps:

[0061] Step 1: Use the 6SV radiative transfer code to simulate the atmospheric absorption, scattering and transmission characteristics. The specific steps are as follows:

[0062] (1) Since the 6SV radiation transfer code (https: / / salsa.umd.edu / 6spage.html) does not support domestic sensors, the spectral response function of CZI will be added first (the spectral response function refers to the response degree or sensitivity of an optical sensor to light of different wavelengths within a specific spectral range. This spectral response function describes the sensor's ability to perceive light of different wavelengths and is usually expressed in the form of a spectral response curve. See Figure 1 , B1, B2, B3 and B4 are blue, green, red and near red bands respectively) are added to the 6SV code (Second Simulation of a Satellite Signal in the Solar Spectrum Vector) to simulate the influence of the atmosphere on the CZI sensor receiving signal. Steps to add the spectral response function: Open the 6SV2.1 folder in the Linux system, create the HY1.f file, and add the response functions of the four CZI bands to the file, see Figure 2, is the process of adding the band response function. The specific process is as follows: First, the comment line describes the starting wavelength and the ending wavelength, and the following is the specific spectral response value at each wavelength. Note that each row has 1501 columns, which cannot be modified. Because the spectral range in 6s is 0.25-4um, and the interval is 0.0025um, there are (4-0.25) / 0.0025+1=1501 wavelengths. The first item 67*0 in the 7-row data indicates that the first 67 numbers in this row of data are 0; the last item 1395*0 indicates that the last 1395 numbers are 0. There are 39 spectral response values ​​in the middle, and there are a total of 64+1395+39=1501 bands. The specific calculation method is: assuming that the range of the spectral response is (a, b), the calculation method for the first 0 is (a-250) / 2.5, the calculation method for the non-zero in the middle is (ba) / 2.5+1; the 0 behind is (4000-b) / 2.5. Continue to complete the remaining 3 bands using the above method.

[0063] (2) Rayleigh scattering reflectivity simulation: Rayleigh scattering is an optical phenomenon and a type of scattering. It is also called "molecular scattering". When the particle size is much smaller than the wavelength of the incident light (less than one tenth of the wavelength), the intensity of the scattered light in different directions is different. The intensity is inversely proportional to the fourth power of the wavelength of the incident light. This phenomenon is called Rayleigh scattering. Factors affecting Rayleigh scattering reflectivity: atmospheric model and observation geometry. The atmospheric model includes tropical, mid-latitude summer, mid-latitude winter and near-polar summer; the observation geometry includes the solar zenith angle θ 0 、Sensor zenith angle θ v , relative azimuth (The solar azimuth angle and the sensor azimuth angle are subtracted and the absolute value is taken). When the 6SV model is used to simulate Rayleigh scattering, the observation geometry input parameters are as follows: θ 0 (solar zenith angle, 0-88°, interval 2°); θ v (sensor zenith angle, 0-84°, interval 2°); (Relative azimuth, 0-180°, interval 5°). Output results Rayleigh scattering reflectivity of CZI 4 bands under different observation geometries under the above 4 atmospheric modes. After comparison and verification, it was found that the influence of different atmospheric modes on Rayleigh scattering reflectivity was very small, with the deviation within 1%, so it was considered to use only one (mid-latitude summer) instead of the other two.

[0064] In the Linux system, using the debugged 6SV simulation, according to my country's geographical distribution from south to north, the atmospheric mode is set to tropical, mid-latitude summer and mid-latitude winter; the change of observation angle may affect the intensity of the observed Rayleigh scattering. Under different atmospheric modes, the value of each band will also be different. Therefore, the observation geometry sets the solar zenith angle, sensor zenith angle and relative azimuth, and outputs the Rayleigh scattering reflectance of the blue, green, red and near-red bands under the three atmospheric modes of tropical, mid-latitude summer and mid-latitude winter. The output method is shown in Figure 3 The simulated Rayleigh scattering reflectivity data is divided into a training set and a test set, which can be used for the training and verification of a multi-layer fully connected neural network.

[0065] (3) Aerosol scattering reflectivity simulation: Factors affecting aerosol scattering reflectivity include: aerosol type, observation geometry and aerosol optical thickness at 550nm. Aerosol types include marine aerosol, urban aerosol, continental aerosol, etc.; observation geometry includes solar zenith angle, sensor zenith angle, and relative azimuth angle.

[0066] When the 6SV model is used to simulate the aerosol scattering reflectivity, the detailed parameters are set as follows: aerosol type: marine aerosol, urban aerosol, and continental aerosol; the aerosol optical thickness at 550 nm is set to 0.05 nm, 0.1 nm, 0.2 nm, 0.3 nm, 0.4 nm, 0.5 nm, 0.6 nm, 0.7 nm, 0.8 nm, 0.9 nm, and 1.2 nm, respectively; θ 0 (solar zenith angle, 0-88, interval 4). θ v (Sensor zenith angle, 0-84, interval 4). (Relative azimuth, 0-180, interval 10). The output results are the aerosol scattering reflectivity of the four bands of CZI under different aerosol types.

[0067] In the Linux system, using the debugged 6SV simulation, different aerosol types will affect the value of aerosol scattering reflectivity under the same observation geometry. Set the aerosol type: continental, oceanic and urban. The observation geometry includes the solar zenith angle, sensor zenith angle and relative azimuth. The aerosol optical thickness is set to 0.05nm, 0.1nm, 0.2nm, 0.3nm, 0.4nm, 0.5nm, 0.6nm, 0.7nm, 0.8nm, 0.9nm, 1.2nm, and output the aerosol scattering reflectivity of three or four bands. The simulated aerosol scattering reflectivity data is divided into training sets and test sets, which can be used for the training and verification of multi-layer fully connected neural networks.

[0068] (4) Aerosol scattering transmittance simulation: The total aerosol transmittance includes the uplink transmittance and the downlink transmittance. The total transmittance is directly obtained by 6SV simulation. Factors affecting the total aerosol transmittance include aerosol type, observation geometry and aerosol optical thickness. The aerosol type setting is the same as the setting for simulating aerosol scattering reflectivity. The detailed parameter settings are as follows: Aerosol type: marine aerosol, urban aerosol, continental aerosol. The aerosol optical thickness at 550nm is set to 0.05nm, 0.1nm, 0.2nm, 0.3nm, 0.4nm, 0.5nm, 0.6nm, 0.7nm, 0.8nm, 0.9nm and 1.2nm respectively. θ 0 (Solar zenith angle, 0-88°, interval 4°). θ v (Sensor zenith angle, 0-84°, interval 4°). The output result is the total aerosol scattering transmittance of the four bands of CZI under different aerosol types. After comparison and verification, it was found that the influence of different aerosol types on transmittance was very small, and the deviation was within 1%, so it was considered to use only one aerosol type instead of the other two.

[0069] In the Linux system, using the debugged 6SV simulation, different aerosol types will affect the value of aerosol scattering transmittance under the same observation geometry. Set the aerosol type: continental, oceanic and urban, and the observation geometry includes the solar zenith angle and the sensor zenith angle. The aerosol optical thickness is set to 0.05nm, 0.1nm, 0.2nm, 0.3nm, 0.4nm, 0.5nm, 0.6nm, 0.7nm, 0.8nm, 0.9nm, 1.2nm, and output the aerosol scattering reflectivity of three or four bands. The simulated aerosol scattering transmittance data is divided into training sets and test sets, which can be used for the training and verification of multi-layer fully connected neural networks.

[0070] Step 2: Establish multi-layer fully connected neural network models for Rayleigh scattering, aerosol scattering and transmission respectively, and use model = Sequential() sequential model, that is, build a deep neural network through layer-by-layer neural network connection, and superimpose some network layers through model.add() to quickly and accurately predict the above atmospheric correction parameters.

[0071] A fully connected neural network (Fully Connected Neural Network), also known as a densely connected neural network (Densely Connected Neural Network), is a common artificial neural network structure. In a fully connected neural network, each neuron in each layer is connected to all neurons in the previous layer, and each connection has a corresponding weight. A fully connected neural network structure usually consists of an input layer, several hidden layers, and an output layer. Each hidden layer and output layer consists of multiple neurons. During the training process, the weight of each connection in the network is adjusted through optimization algorithms such as backpropagation and gradient descent to minimize the loss function, so that the network can learn the mapping relationship between input and output. Figure 4 as well as Figure 5 , which are respectively a mathematical model of a multi-layer fully connected neural network commonly used in the prior art and a fully connected neural network model commonly used in the prior art.

[0072] Different neural networks have different node and layer number settings. The present invention sets different input layer nodes, hidden layer nodes and output layer nodes for the three networks according to the relationship and scale of the data.

[0073] (1) Network 1: Rayleigh scattering model. The parameter setting of this network model (referring to the multi-layer fully connected neural network mentioned above) includes an input layer, four hidden layers and an output layer. The input layer has three nodes; the number of nodes in each hidden layer is set to 16, 32, 64 and 8 respectively; the output layer has four nodes. The input parameter of this model is θ 0 (solar zenith angle), θ v (sensor zenith angle) and (relative azimuth), output the Rayleigh scattering reflectivity of four bands; the training set / test set division ratio is 0.7 / 0.3; learning rate is 0.01; EPOCHES=2000; BETCH_SIZE=128. Network convergence verification: determination coefficient R of test data set and network prediction results 2 Equal to 0.99, the verification results are as follows Figure 6 shown.

[0074] (2) Network 2: Aerosol scattering model. The parameter setting of this network model includes an input layer, three hidden layers and an output layer. The input layer has four nodes; the hidden layer nodes are set to 32, 64 and 128 respectively; the output layer has twelve nodes. The input parameter of this model is θ 0 (solar zenith angle), θ v (sensor zenith angle), (relative azimuth) and aerosol optical depth; the training set / test set division ratio is 0.7 / 0.3; the learning rate is 0.01; EPOCHES=1000; BATCH_SIZE=256. The output results are the aerosol scattering reflectivity of 4 bands under different aerosol optical depths for continental, urban and marine aerosol types. The output results here are the output results of training the multi-layer fully connected neural network, which are used to compare with the data of the test set to analyze and evaluate the accuracy of the prediction of the multi-layer fully connected neural network model. Network convergence verification: R of the training data set and the prediction data set 2 Equal to 0.99, relative deviation less than 10 -5 , the verification result is as follows Figure 7 shown.

[0075] (3) Network 3: Aerosol transmission model. The parameter setting of this network model includes an input layer, three hidden layers and an output layer. The input layer has three nodes; the number of nodes in the hidden layer is 16, 32, and 64 respectively; and the output layer has twelve nodes. The input parameter of this model is θ 0 (solar zenith angle), θ v (sensor zenith angle) and aerosol optical thickness; training set / test set split ratio is 0.7 / 0.3; learning rate is 0.01; EPOCHES=2000; BATCH_SIZE=128. Network convergence verification: determination coefficient R of test data set and network prediction results 2 Equal to 0.99, the verification results are as follows Figure 8 The output results are the aerosol scattering transmittance of four bands at different aerosol optical depths for continental, urban and marine aerosol types.

[0076] Step 3: Perform atmospheric correction of remote sensing data based on the multi-layer fully connected neural network model determined in step 2).

[0077] Based on the CZI radiance data L t (Downloaded from China Ocean Satellite Data Service System https: / / osdds.nsoas.org.cn / DataRetrieval ) is calculated to obtain the top atmospheric reflectivity ρ t

[0078]

[0079] F0 represents the solar irradiance in the outer atmosphere. The National Meteorological Information Center can provide solar radiation data from China's meteorological satellites for domestic meteorological and climate research. 0 Represents the solar zenith angle, obtained from: http: / / data.cma.cn / dataService / cdcindex / datacode / SK.0728.001 / show_value / normal.html .

[0080] Then, the solar zenith angle, sensor zenith angle and relative azimuth are extracted from the CZI brightness product, and the 550nm aerosol optical depth is obtained through the MOD04 aerosol product (MOD04 is aerosol data obtained by MODIS, which is an important device carried on Terra (morning star, descending node local time 10:30, MOD) and Aqua (afternoon star, ascending node local time 1:30, MYD). The commonly used download source of MODIS data: LAADS DAAC searchhttps: / / ladsweb.modaps.eosdis.nasa.gov / search / );

[0081] The solar zenith angle, sensor zenith angle, relative azimuth angle and 550nm aerosol optical depth are input into three neural networks, and the Rayleigh scattering reflectivity ρ under the specified atmospheric type is predicted by the three corresponding neural networks established above. r , aerosol reflectivity ρ under specified aerosol type a and transmittance t.

[0082] Step 4: Verification of atmospheric correction results.

[0083] (1) Use the 6SV model to simulate and verify data.

[0084] The validation set data uses 6SV to simulate new data, with specific parameter settings: solar zenith angle (0-88°, interval 3°), sensor zenith angle (0-84°, interval 3°), relative azimuth angle (0-180°, interval 16°). Aerosol optical thickness is 0.15, 0.25, 0.35, 0.45, 0.55, 0.65. Atmosphere type: mid-latitude summer; Aerosol type: urban type.

[0085] The atmospheric correction results were verified using synchronously measured (ground feature spectrometer) turbid water remote sensing reflectance (see Table 1 and Table 2 below).

[0086] The field spectroradiometer measures remote sensing reflectance under low wind speed and clear sky conditions. The calculation formula of remote sensing reflectance (Rrs) is Rrs = (Lw-r*Lsky) / (Lp*π / ρ), where Lw is the radiation of water; Lsky is the radiation of the sky; Lp is the brightness of the gray panel; ρ is the diffuse reflectance coefficient of the panel; r is the Fresnel reflectance of the water surface.

[0087] Table 1: Measured data and atmospheric correction data

[0088]

[0089] Table 2: Site information table

[0090]

[0091] It can be seen that the geometric parameters and aerosol optical depth input in the test data set are completely different from those in the training data set, so they can be used to verify the accuracy of the three networks and test the generalization ability of the networks. Figure 6 , Figure 7 as well as Figure 8 , where the results obtained by 6SV simulation are regarded as the “true value”, and the prediction results of these three networks are different from the “true value” R 2 All are greater than 0.99, and the RMSE are 0.0026, 0.0376, and 0.0028, respectively. It is worth noting that as the aerosol scattering reflectivity increases, the deviation between the predicted value and the true value also increases. Only when the solar zenith angle and the sensor zenith angle exceed 70° at the same time, the Rayleigh scattering reflectivity may be greater than 1, and this imaging geometry usually does not occur in the real scene where the sensor acquires data. Therefore, the increase in the large deviation of the aerosol scattering reflectivity will basically not affect the atmospheric correction accuracy of subsequent remote sensing data.

[0092] Four sets of water remote sensing reflectance collected at Axe Lake in Wuhan on October 18, 2022 were used to verify the atmospheric correction accuracy of the synchronously acquired CZI data, where the difference between the on-site collection time and the satellite transit time did not exceed ±2 hours.

[0093] According to the latitude and longitude of the field data, the corresponding spatial position in the CZI image is matched, and then the average reflectance of the 5*5 window is calculated and compared with the measured remote sensing reflectance. The results are as follows: Figure 9-12 shown.

[0094] The CZI data is atmospherically corrected to obtain the surface reflectance, which is converted to remote sensing reflectance by dividing it by pi (pi). Compared with the measured remote sensing reflectance, the atmospheric correction result predicted by the neural network is overestimated, which is consistent with the current research results of atmospheric correction. 2 The value of the atmospheric correction is 0.72, and the RMSE is 0.01, which shows that the two are in good agreement and strictly verifies the accuracy of the atmospheric correction results.

[0095] The present invention uses a multi-layer neural network to simulate the satellite to obtain the surface reflected radiation affected by the atmospheric absorption, scattering and transmission. Taking the domestic satellite HaiYang-1C carrying CZI sensor (abbreviated as HY-1C CZI) as an example, the CZI sensor is set with four bands, namely blue, green, red and near red. The atmospheric correction calculation method for each band is as follows:

[0096] L t (λ) = L r (λ)+L a (λ)+T(λ)L g (λ)+t(λ)L wc(λ)+t(λ)L w (λ)

[0097] λ is the wavelength, L t is the total radiation received by the sensor, L r is the radiation contributed by Rayleigh scattering of air molecules, L a is the aerosol scattered radiation, L w is the surface reflected radiation, and for water bodies it is the water-leaving radiance, T and t are the direct transmittance and diffuse transmittance of the atmosphere, respectively. g is the radiation from solar flares (glints) on the water surface, L wc It is the white cap radiation of the water surface. Non-water pixels in remote sensing images do not need to consider these two radiation parameters.

[0098] The purpose of atmospheric correction is to obtain the total radiation L t The water body radiance L is calculated w . Solar flare over water g and White Hat Radiation L wc , the sensor needs to be set to a specific band for evaluation, which is usually ignored for multispectral sensors such as CZI, so the above atmospheric correction formula becomes:

[0099] ρ t (λ)=ρ r (λ)+ρ a (λ)+t(λ)ρ w (λ)

[0100] The radiance L is replaced by the reflectivity ρ. t is the reflectivity of the top layer of the atmosphere. Rayleigh scattering reflectivity ρ r The deviation is less than 0.5% obtained through radiation transfer simulation. Therefore, the focus of atmospheric correction is to obtain the aerosol scattering reflectivity ρ during sensor imaging. a Common methods include dark pixels and radiation transfer simulation. Dark pixels are assumed to be water pixels in certain specific bands (usually near-infrared to short-wave infrared bands, and short-wave infrared bands are mainly used when correcting turbid water bodies). w =0, then ρ a (λ)=ρ t (λ)–ρ r (λ), and then the aerosol scattering reflectivity of other bands such as blue, green, and red is inferred. However, for sensors like CZI that do not have short-wave infrared sensors, dark pixels will no longer be applicable. In summary, for sensors like CZI that have blue, green, red, and near-red bands, the Rayleigh scattering and aerosol scattering data are generated through radiation transfer simulation to establish a lookup table, and then the atmospheric correction parameters are obtained by matching the lookup table pixel by pixel.

Claims

1. A method for rapid atmospheric correction of turbid water, characterized in that: The turbid water body rapid atmospheric correction method comprises the following steps: 1) Use the 6SV radiation transfer code to simulate the CZI sensor receiving signal affected by the atmosphere and obtain the data set; 2) constructing a weather-affected fully connected neural network model based on the data set obtained in step 1), wherein the weather-affected fully connected neural network model includes a Rayleigh scattering reflectivity fully connected neural network model, an aerosol scattering reflectivity fully connected neural network model, and an aerosol scattering transmittance fully connected neural network model; the input parameters of Rayleigh scattering in the Rayleigh scattering reflectivity fully connected neural network model include the solar zenith angle θ0, the sensor zenith angle θ v and relative azimuth The input parameters of the aerosol scattering reflectivity in the aerosol scattering reflectivity fully connected neural network model include the solar zenith angle θ0, the sensor zenith angle θ v , relative azimuth and 550nm aerosol optical thickness; the input parameters of aerosol scattering transmittance in the aerosol scattering transmittance fully connected neural network model include the solar zenith angle θ0, the sensor zenith angle θ v and 550nm aerosol optical depth; 3) Performing atmospheric correction on the measured remote sensing data based on the weather-affected fully connected neural network model constructed in step 2), specifically: 3.1) According to the weather-affected fully connected neural network model constructed in step 2), predict the Rayleigh scattering reflectivity ρ under the specified atmospheric type r , aerosol reflectivity ρ under specified aerosol type a and transmittance t; 3.2) Calculate the reflectivity ρt of the top atmosphere; 3.3) Combining the result of step 3.1) and the result of step 3.2), the atmospheric correction is performed on the measured remote sensing data. The calculation method of atmospheric correction for the measured remote sensing data in step 3.3) is: r t (λ)=ρ r (l)+r a (λ)+t(λ)ρ w (l); in: λ is the wavelength; The w is the water surface reflectivity; The water surface reflectivity ρ is calculated according to the atmospheric correction method w , completing rapid atmospheric correction for turbid water bodies.

2. The method for rapid atmospheric correction of turbid water according to claim 1, characterized in that: The factors affected by the atmosphere in step 1 are the Rayleigh scattering rate, the aerosol scattering reflectivity and the aerosol scattering transmittance.

3. The method for rapid atmospheric correction of turbid water according to claim 2, characterized in that: The specific implementation method of step 1) is: 1.1) Add the spectral response function of CZI to the 6SV code; 1.2) The CZI sensor receiving signals affected by Rayleigh scattering rate, aerosol scattering reflectivity and aerosol scattering transmittance are simulated in the 6SV code.

4. The method for rapid atmospheric correction of turbid water according to claim 3, characterized in that: The factors affected by the Rayleigh scattering rate include atmospheric mode and observation geometry; the atmospheric mode includes tropical, mid-latitude summer, mid-latitude winter and near-polar summer; the observation geometry includes the solar zenith angle θ0, the sensor zenith angle θ v and relative azimuth The factors affected by the aerosol scattering reflectivity include aerosol type, observation geometry, and aerosol optical thickness at 550nm; the aerosol type includes marine aerosol, urban aerosol, and continental aerosol; the observation geometry includes the solar zenith angle θ0, the sensor zenith angle θ v and relative azimuth The factors affected by the aerosol scattering transmittance include aerosol type, observation geometry, and aerosol optical thickness at 550nm; the aerosol type includes marine aerosol, urban aerosol, and continental aerosol; the observation geometry includes the solar zenith angle θ0 and the sensor zenith angle θ v .

5. The method for rapid atmospheric correction of turbid water according to any one of claims 1 to 4, characterized in that: The Rayleigh scattering reflectivity fully connected neural network model is constructed as follows: a.1) Obtaining the input parameters of Rayleigh scattering; a.2) normalizing the input parameters of the Rayleigh scattering obtained in step a.1) by dividing them by the maximum value to obtain a normalized data set; a.3) Divide the normalized data set obtained in step a.2) into a training set and a test set according to a ratio of 0.7 / 0.3; a.4) using the training set obtained by step a.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=2000, BETCH_SIZE=128, to obtain a trained fully connected neural network model; a.5) Using the test set obtained in step a.3) to verify the trained fully connected neural network model obtained in step a.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, the network converges and the fully connected neural network model of Rayleigh scattering reflectivity is obtained.

6. The method for rapid atmospheric correction of turbid water according to any one of claims 1 to 4, characterized in that: The aerosol scattering reflectivity fully connected neural network model is constructed in the following way: b.1) Obtain input parameters of aerosol scattering reflectivity; b.2) normalizing the input parameter of aerosol scattering reflectivity obtained in step b.1) by dividing it by the maximum value to obtain a normalized data set; b.3) Divide the normalized data set obtained in step b.2) into a training set and a test set according to a ratio of 0.7 / 0.3; b.4) using the training set obtained by step b.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=1000, BETCH_SIZE=256, to obtain a trained fully connected neural network model; b.5) ​​Using the test set obtained in step b.3) to verify the trained fully connected neural network model obtained in step b.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, relative deviation is less than 10 -5 The network convergence is completed when , and the aerosol scattering reflectivity fully connected neural network model is obtained.

7. The method for rapid atmospheric correction of turbid water according to any one of claims 1 to 4, characterized in that: The aerosol scattering transmittance fully connected neural network model is constructed as follows: c.1) Obtaining input parameters of aerosol scattering transmittance; c.2) normalizing the input parameter of aerosol scattering transmittance obtained in step c.1) by dividing it by the maximum value to obtain a normalized data set; c.3) Divide the normalized data set obtained in step c.2) into a training set and a test set according to a ratio of 0.8 / 0.2; c.4) using the training set obtained in step c.3), training the fully connected neural network model under the conditions of learning rate 0.01, EPOCHES=2000, BETCH_SIZE=128, to obtain a trained fully connected neural network model; c.5) Using the test set obtained in step c.3) to verify the trained fully connected neural network model obtained in step c.4), the determination coefficient R of the prediction results of the test set and the trained fully connected neural network model is 2 =0.99, the network converges and a fully connected neural network model of aerosol scattering transmittance is obtained.

8. The method for rapid atmospheric correction of turbid water according to any one of claims 1 to 4, characterized in that: The reflectivity of the top layer of the atmosphere in step 3.2) t The specific implementation method is: in: L t is the CZI radiance data; F0 is the solar irradiance in the outer layer of the atmosphere; θ0 is the solar zenith angle.

Images