Embedded hyperspectral atmospheric correction processing method

Abstract

The invention discloses an embedded hyperspectral atmospheric correction processing method, which comprises the following steps of: establishing an atmospheric correction general lookup table by using an MODTRAN radiation transfer model; constructing a neural network model based on the lookup table, and training the neural network model to obtain an optimal weight parameter; observation information in the data image is obtained, and the apparent reflectivity is calculated according to the radiometric calibration coefficient; setting an atmospheric parameter initial value according to the remote sensing image data, and performing collaborative inversion by using an atmospheric correction network model to obtain atmospheric water vapor content and visibility; calling an MODTRAN radiation transfer model according to remote sensing image observation information and the calculated atmospheric parameters to calculate atmospheric correction coefficients of all wavebands in parallel; because the atmospheric parameters of the adjacent spatial positions are similar, the atmospheric parameter of the previous stage is used as the atmospheric correction initial value of the next stage; and repeating the process until the correction is finished. Therefore, the problems that in the atmospheric correction technology, parameter searching is complex, the real-time state is affected, and the inversion precision is low are solved.

Description

Technical Field

[0001] The invention relates to the technical field of quantitative optical on-board data processing, in particular to an embedded hyperspectral atmospheric correction processing method. Background Art

[0002] In the field of quantitative optical technology, with the increasing maturity of hyperspectral remote sensing technology and its great advantages in spectral resolution, hyperspectral remote sensing technology has been widely used in the fields of ocean, vegetation ecology, agriculture, geology, atmosphere, etc. In the process of satellite-borne or airborne hyperspectral quantitative detection payload imaging, due to the influence of atmospheric scattering or absorption and terrain factors, the surface reflected radiation energy received by the sensor will change, resulting in distortion of ground object radiation information, which greatly affects the accuracy of spectral information extraction. In order to conduct quantitative research on hyperspectral remote sensing images, it is necessary to perform atmospheric correction on hyperspectral remote sensing images to remove the effect of the atmosphere on the radiation signal and obtain the true information of the ground object.

[0003] At present, domestic and foreign scholars have conducted a lot of research on the atmospheric correction of hyperspectral images. The existing atmospheric correction methods can be divided into two categories: empirical statistical methods based on image features and physics-based radiation transfer model methods. The former does not need to consider the influence of atmospheric parameters, the method is simple, and the calculation efficiency is fast, but it needs to use prior knowledge to obtain a large amount of measured data, and the algorithm cost is high. The latter takes into account the influence of the atmosphere on radiation transfer, and can calculate a more realistic surface reflectance, with higher correction accuracy, but it is necessary to first invert atmospheric parameters such as water vapor content and visibility, which has a large calculation cost and relatively low efficiency. At present, the use of atmospheric radiation transfer models to establish lookup tables is widely used in the atmospheric correction of hyperspectral images. In order to ensure the correction accuracy of the lookup table, it is necessary to consider a variety of atmospheric influencing factors when constructing the lookup table, and each factor considers multiple situations, which easily leads to a large amount of lookup table data, poor portability, attitude and orbit parameters and imaging conditions, and irrelevant long strip data inversion work. Therefore, the research on atmospheric correction methods for hyperspectral remote sensing data has become of great theoretical significance and practical value. It is necessary to improve the key technical methods of hyperspectral atmospheric correction, improve the identification process, and improve the efficiency and accuracy of atmospheric correction. Summary of the invention

[0004] The present invention provides an embedded hyperspectral atmospheric correction processing method. Aiming at the problems of traditional correction algorithms' dependence on data and low precision, the method takes the atmospheric correction of multispectral and hyperspectral remote sensing images acquired onboard or airborne as the core, and solves the problems of complex parameter search, real-time state influence and low inversion precision in the existing atmospheric correction technology using radiation transmission models based on the atmospheric radiation transmission mechanism and the classical atmospheric radiation transmission model, thereby providing algorithm support for improving the on-board real-time atmospheric correction of multispectral and hyperspectral remote sensing images acquired onboard or airborne.

[0005] In a first aspect, an embedded hyperspectral atmospheric correction processing method is provided, characterized by comprising:

[0006] Step 1, using the MODTRAN radiation transfer model, establish a general lookup table for atmospheric correction for specified bands, where the specified bands include aerosol inversion bands, water vapor absorption bands, and reference bands;

[0007] Step 2: construct a neural network model based on the atmospheric correction general lookup table, and obtain the optimal weight parameters of the neural network model through training;

[0008] Step 3: Set the atmospheric water vapor content WVC i-1 AOT with visibility i-1 , together with the observation information in the remote sensing image data, is input into the neural network model, and WVC is determined based on the output of the neural network model. i-1 The corresponding atmospheric transmittance T' is obtained by comparing T' with the atmospheric transmittance reference value T to adjust the inversion to obtain WVC i ; Determine AOT based on the output of the neural network model i-1 The corresponding simulated ground reflectivity ρ” is obtained by comparing ρ and ρ” to adjust the inversion to obtain AOT i ; Repeat the inversion process until the difference between T' and T is less than the preset error, and the difference between ρ and ρ" is less than the preset error, and determine the atmospheric water vapor content WVC and visibility AOT obtained by the final inversion; Step 4, based on the observation information in the remote sensing image data, and the atmospheric water vapor content WVC and visibility AOT obtained by the inversion, call MODTRAN to parallelly calculate the atmospheric correction coefficient x for all bands a ,x b ,x c , corresponding to L p , G b +G t , S, L p is the atmospheric radiation, G b , G t are the influence of direct atmospheric radiation and scattering, respectively, and S is the atmospheric albedo.

[0009] In combination with the first aspect, in some implementations of the first aspect, in step 1, the MODTRAN radiation transfer model is used to simulate the atmospheric radiation transfer process to obtain the atmospheric transmittance, and then the atmospheric path radiation L is calculated. p , large balloon surface albedo S and radiation correction parameter G b +G t ; In the simulation process, only the atmospheric transmission process of the aerosol inversion band, water vapor absorption band and reference band was simulated; Among them,

[0010]

[0011] Aerosol inversion bands include blue light band, red light band, and near-infrared band;

[0012] The water vapor absorption bands include 1117-1143nm, 935-955nm, and 810-830nm;

[0013] The upper reference band corresponding to the water vapor absorption band 1117-1143nm includes 1184-1210nm, and the lower reference band corresponding to the water vapor absorption band 1117-1143nm includes 1050-1067nm;

[0014] The upper reference band corresponding to the water vapor absorption band 935-955nm includes 870-890nm, and the lower reference band corresponding to the water vapor absorption band 935-955nm includes 995-1020nm;

[0015] The upper reference band corresponding to the water vapor absorption band 810-830 nm includes 850-870 nm, and the lower reference band corresponding to the water vapor absorption band 810-830 nm includes 770-790 nm.

[0016] In combination with the first aspect, in some implementations of the first aspect, in step 1, simulations are performed for three cases where the surface reflectivity ρ is 0, 0.25, and 0.5, respectively, and the atmospheric correction coefficient is solved by the following set of equations:

[0017]

[0018] The atmospheric correction factor is:

[0019]

[0020] Where L' 0.25 =L 0.25 -L0, L' 0.5 =L 0.5 -L0.

[0021] In combination with the first aspect, in certain implementations of the first aspect, in step 1, the process of building a table for MODTRAN based on a multi-core processor is:

[0022] Enumerate different observation conditions and atmospheric states, distribute each condition as a single case to multiple threads or processes through command distribution, and generate a Tape5 file;

[0023] Call MODTRAN in parallel through multi-core processors to build a lookup table;

[0024] In the process of processing hyperspectral remote sensing data, the parallel and for instructions in OpenMP are used to decompose the tasks and enable thread groups to process multiple situations simultaneously.

[0025] Execution ends: Read the data in the result file Tape7, calculate the atmospheric correction coefficient, and package the input conditions and calculation results into the database.

[0026] In combination with the first aspect, in some implementations of the first aspect, in step 2, a neural network model is constructed; the model input includes visibility AOT, atmospheric water vapor content WVC, observation zenith angle θ v 、Sun zenith angle θ s , relative azimuth Red band center wavelength λ red 、Blue band center wavelength λ blue , water vapor absorption band center wavelength λ absorb and reference bands; the output is the atmospheric correction coefficient x for the red, blue and water vapor absorption bands a ,x b ,x c , atmospheric transmittance, x a ,x b ,x c L p , G b +G t , S; The network structure of the neural network model includes:

[0027] 1) Convolutional layer Conv1d 3×1×32

[0028] 2) Activation layer Relu

[0029] 3) Convolutional layer Conv1d 3×32×32

[0030] 4) Activation layer Relu

[0031] 5)Maxpool1d

[0032] 6) Convolutional layer Conv1d 3×32×32

[0033] 7) Activation layer Relu

[0034] 8) Convolutional layer Conv1d 3×32×32

[0035] 9) Activation layer Relu

[0036] 10)Maxpool1d

[0037] 11) Convolutional layer Conv1d 3×32×10

[0038] 12) Activation layer Relu.

[0039] In combination with the first aspect, in certain implementations of the first aspect, in step 3, the normalized difference vegetation index NDVI and the apparent reflectance at 2100 nm are used to determine the dark pixels of the hyperspectral image data, wherein the dark pixel requirements are as follows:

[0040]

[0041] ρ' 860 , ρ' 660 , ρ' 2100 They are the apparent reflectance of the hyperspectral images at 860nm, 660nm, and 2100nm respectively; the atmospheric water vapor content WVC and visibility AOT are inverted through the screened dark pixels.

[0042] In combination with the first aspect, in certain implementations of the first aspect, in step 3, attitude and trajectory parameters and imaging conditions are obtained by real-time monitoring, and the initial values ​​of visibility and water vapor content are set to AOT0=AOT0 660 =AOT0 470 ,WVC0; Feedback the parameters to the neural network model and output the atmospheric correction coefficients of the red and blue bands and the water vapor absorption band and atmospheric transmittance T';

[0043] The simulated ground reflectivity ρ" is calculated based on the atmospheric correction coefficients and apparent radiance of the red and blue bands 660 ,ρ” 470 ;

[0044] If ρ” 660 and ρ 660 If the difference is not within the allowable range, adjust AOT i 660 , ρ 660 ≈ρ' 2100 / 2; if ρ” 660 Less than ρ 660 , then AOT i 660 Increase σ1 660 , if ρ” 660 Greater than ρ 660 , then AOT i 660 Reduce the step size σ1 660 , until the difference between the two is within the allowable error range;

[0045] If ρ” 470 and ρ 470 If the difference is not within the allowable range, adjust AOT i 470 , ρ470 ≈ρ' 2100 / 4; if ρ” 470 Less than ρ 470 , then AOT i 470 Increase σ1 470 , if ρ” 470 Greater than ρ 470 , then AOT i 470 Reduce the step size σ1 470 , until the difference between the two is within the allowable error range;

[0046] The average visibility obtained from the red and blue bands As a result of the visibility of this inversion AOT i ;

[0047] For the reference band corresponding to the water vapor absorption band, the atmospheric transmittance reference value T satisfies:

[0048]

[0049] In the formula, ρ(λ),ρ(λ up ),ρ(λ down ) are the apparent reflectances of the water vapor absorption band, the upper reference band, and the lower reference band, respectively; c1 and c2 are the reflectance weights of the reference bands, respectively; if T is less than the transmittance T' obtained by the neural network model, the atmospheric water vapor content is WVC i Increase the step size σ2, otherwise decrease it.

[0050] In combination with the first aspect, in some implementations of the first aspect, in step 4, based on the image observation information and the atmospheric parameters calculated in step 3, the hyperspectral image bands are grouped and MODTRAN is called to calculate the atmospheric correction coefficients L for all bands. p , G b +G t , S; the atmospheric radiation transfer equation is expressed as:

[0051]

[0052] Where L is the apparent radiance received by the sensor; L p is the atmospheric radiation; G b , G t are the influence of direct atmospheric radiation and scattering, respectively; S is the atmospheric albedo; ρ is the surface reflectivity or the true reflectivity of the ground object;

[0053] Combined with the atmospheric correction factor L p , G b +G t, S, and then substitute the apparent reflectivity obtained at the sensor into the above formula to obtain the surface reflectivity or the true reflectivity of the object, thereby completing the atmospheric correction.

[0054] In combination with the first aspect, in some implementations of the first aspect, the method further includes:

[0055] Step 5, repeat steps 3-4 to perform atmospheric correction for the next remote sensing image, wherein the inversion result of step 4 is used as input for the next stage to speed up the inversion.

[0056] In combination with the first aspect, in certain implementations of the first aspect, in step 5, for adjacent pixels having the same ground objects or similar surface reflectances, the surface reflectance of the atmospherically corrected adjacent pixels of the previous remote sensing image is subtracted from the atmospherically corrected pixel of the current remote sensing image as the dark pixel.

[0057] Compared with the prior art, the solution provided by the present invention includes at least the following beneficial technical effects:

[0058] 1. The present invention establishes a universal full-duty cycle lightweight lookup table based on the radiation transfer model, which can meet the atmospheric correction requirements in various imaging situations and will no longer be limited by the sensor.

[0059] 2. The present invention utilizes the powerful nonlinear mapping and generalization capabilities of neural networks to lightweight the atmospheric correction lookup table, construct an atmospheric correction network module, and enhance its portability.

[0060] 3. The attitude and orbit parameters and imaging conditions of the present invention are monitored in real time, and the joint influence of atmospheric water vapor content and visibility on radiation transmission is taken into account to improve the accuracy of atmospheric correction.

[0061] 4. The present invention adopts the idea of ​​rolling correction, utilizes the similarity of atmospheric parameters between adjacent pixels, uses the atmospheric parameters obtained from the previous stage inversion as the current inversion input value, reduces the number of iterations, accelerates the convergence of results, and eliminates the problem caused by the absence of dark pixels in this aerosol inversion; and uses a multi-core processor to call MODTRAN in parallel to realize the rapid calculation of atmospheric correction coefficients for hyperspectral data band by band and pixel by pixel, effectively improving the accuracy and efficiency of real-time atmospheric parameter inversion of long strip data. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 The figure is a flow chart of the method of the present invention.

[0063] Figure 2 Flowchart for constructing a full-cycle lightweight lookup table.

[0064] Figure 3 This is the flow chart of the coordinated inversion of water vapor and aerosol inversion. DETAILED DESCRIPTION

[0065] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] like Figure 1 As shown, the present invention provides an embedded hyperspectral atmospheric correction processing method, which mainly includes the following steps.

[0067] Step S1, using the MODTRAN radiation transfer model, establishing a general lookup table for atmospheric correction for a specified band, where the specified band includes an aerosol inversion band, a water vapor absorption band, and a reference band.

[0068] The imaging conditions and atmospheric conditions of hyperspectral remote sensing images are diverse. At the same time, the amount of hyperspectral remote sensing image data is huge, and it takes a long time to process them band by band. Among the absorption and scattering of gas molecules and aerosols that cause atmospheric effects, the main absorbers include water vapor, ozone, carbon dioxide, and oxygen. Atmospheric molecular scattering and absorption caused by gases such as ozone and oxygen are relatively easy to correct because the concentrations of these elements are relatively stable in time and space. The difficulty lies in the estimation of aerosol and water vapor parameters. Therefore, removing the influence of aerosol and water vapor has become the main content of atmospheric correction.

[0069] In step S1, when constructing a lightweight universal lookup table that can meet the full working cycle, the MODTRAN radiation transfer model is used to simulate the atmospheric radiation transfer process to obtain the atmospheric transmittance T, and then calculate the atmospheric path radiation L p , large balloon surface albedo S and radiation correction parameter G b +G t .

[0070] In order to avoid unnecessary computational overhead and ensure the lightness of the lookup table, only the atmospheric transmission process of the aerosol inversion band, water vapor absorption band and reference band was simulated during the simulation. The specific settings of the atmospheric correction coefficient lookup table are shown in Table 1.

[0071] Table 1

[0072]

[0073] According to the above table, different atmospheric states and observation conditions are set, and simulations are performed for three cases where the surface reflectivity ρ is 0, 0.25, and 0.5. The atmospheric correction coefficient can be solved by the following equations:

[0074]

[0075] The atmospheric correction factor is:

[0076]

[0077] Where L' 0.25 =L 0.25 -L0, L' 0.5 =L 0.5 -L0.

[0078] Reference Figure 2 The specific process of MODTRAN table building based on multi-core processors is as follows: enumerate all possibilities of different observation conditions and atmospheric states, distribute each situation as a single Case to multiple threads or processes through instruction distribution, and generate Tape5 files. Call MODTRAN in parallel through multi-core processors to build lookup tables to achieve the purpose of accelerating execution and shortening table building time. Combined with the parallel programming OpenMP (Open Multi-Processing) technology on shared storage computers, in the process of hyperspectral remote sensing data processing, the parallel instructions and for instructions in OpenMP are used to decompose tasks, and thread groups are opened to process multiple situations at the same time to achieve the purpose of accelerating execution. After the execution is completed, the data in the result file Tape7 is read, the atmospheric correction coefficient is calculated, and the input conditions and calculation results are packaged and stored in the database.

[0079] Step S2, based on the atmospheric correction universal lookup table obtained in step S1, a neural network model is constructed and trained to obtain optimized weight parameters.

[0080] After the lookup table is established, the atmospheric correction neural network model is constructed. The model input includes visibility AOT, atmospheric water vapor content WVC, observation zenith angle θ v 、Sun zenith angle θ s , relative azimuth Red band center wavelength λ red 、Blue band center wavelength λ blue , water vapor absorption band center wavelength λ absorb and reference bands; the output is the atmospheric correction coefficient x for the red, blue and water vapor absorption bands a ,x b ,x c , atmospheric transmittance T, x a ,x b ,x c L p , G b +G t , S. The network structure information of the neural network model is shown in Table 2.

[0081] Table 2

[0082]

[0083] Step S3, obtaining observation information in the data image, including the solar zenith angle θ s , observation zenith angle θ v , relative azimuth Image formation time, longitude and latitude, and radiation calibration coefficient, etc.; the apparent reflectivity ρ' is calculated based on the radiation calibration coefficient.

[0084] In step S3, the hyperspectral image that needs to be corrected for the atmosphere is preprocessed to extract the observation information of the image, including the solar zenith angle θ s , observation zenith angle θ v , relative azimuth Image imaging time, longitude and latitude, and radiation calibration coefficient; obtain the calibration coefficient, perform radiation calibration on the hyperspectral image data to obtain the apparent reflectance ρ'; extract the calibrated water vapor band, reference band, and aerosol inversion band images and central wavelength, and the wavelength selection comparison table is shown in Table 3.

[0085] Table 3

[0086]

[0087] Step S4, according to the remote sensing image data, the initial values ​​of the atmospheric parameters are set, and the atmospheric water vapor content WVC and the visibility AOT are obtained by collaborative inversion using the atmospheric correction network model.

[0088] In step S4, refer to Figure 3 The specific process of establishing a parallel architecture for real-time monitoring and real-time computing is as follows.

[0089] In step S4, the Normalized Difference Vegetation Index (NDVI) and the apparent reflectance at 2100 nm are used to determine the dark pixels of the hyperspectral image data, where the dark pixel requirements are as follows:

[0090]

[0091] Where ρ' 860 , ρ' 660 , ρ' 2100 They are the apparent reflectance of the hyperspectral image at 860nm, 660nm, and 2100nm. The near-infrared band is less affected by aerosols and can be ignored, so the apparent reflectance of the 2100 band is used to approximately replace the surface reflectance; for dark pixels, there is the following linear relationship between 2100nm and the red and blue light bands (660nm, 470nm):

[0092] ρ 660 ≈ρ' 2100 / 2

[0093] ρ 470 ≈ρ' 2100 / 4

[0094] Use real-time monitoring to obtain attitude and orbit parameters and imaging conditions, set the initial values ​​of visibility and water vapor content to AOT0, WVC0; feed the parameters back to the atmospheric correction neural network model, and output the atmospheric correction coefficients for the red and blue bands and the water vapor absorption band And the atmospheric transmittance T.

[0095] For the water vapor absorption band and the nearby reference band, the surface reflectivity changes are similar or approximately linear, and the band ratio method can be used to approximately calculate the atmospheric water vapor transmittance. The present invention uses a three-band ratio method to invert the atmospheric water vapor content:

[0096]

[0097] In the formula, ρ(λ),ρ(λ up ),ρ(λ down ) are the apparent reflectances of the water vapor absorption band and the reference band, respectively; T is the atmospheric transmittance; c1 and c2 are the reflectance weights of the reference band, respectively.

[0098] The simulated ground reflectivity ρ is calculated based on the atmospheric correction coefficients and apparent radiance of the red and blue bands. 660 ,ρ” 470 If ρ” 660 and ρ 660 If the difference is not within the allowable range, the AOT needs to be adjusted i 660 ; Among them, if ρ" 660 Less than ρ 660 , then AOT i 660 Increase σ1 660 , if ρ” 660 Greater than ρ 660 , then AOT i 660 Reduce the step size σ1 660 , until the difference between the two is within the allowable error range. If ρ” 470 and ρ 470 If the difference is not within the allowable range, adjust AOT i 470 , ρ 470 ≈ρ' 2100 / 4; if ρ” 470 Less than ρ 470 , then AOT i 470 Increase σ1 470 , if ρ” 470 Greater than ρ 470 , then AOT i 470 Reduce the step size σ1470 , until the difference between the two is within the allowable error range. The average visibility obtained by the red and blue bands As a result of the visibility of this inversion AOT i .

[0099] If T is less than the transmittance T' obtained by the atmospheric correction network model, the atmospheric water vapor content is WVC i Increase σ2, otherwise it decreases.

[0100] If the difference between the two results is within the allowable error range ε1, ε2, proceed to step S5, otherwise repeat this step.

[0101] Step S5, calling MODTRAN to calculate the atmospheric correction coefficients x for all bands in parallel based on the remote sensing image observation information and the atmospheric parameters obtained in step S4 a ,x b ,x c , corresponding to L p , G b +G t , S.

[0102] In step S5, based on the image observation information and the atmospheric parameters calculated in step S4, the hyperspectral image bands are grouped and MODTRAN is called to calculate the atmospheric correction coefficients L for all bands. p , G b +G t , S. The atmospheric radiation transfer equation can be expressed as:

[0103]

[0104] Where L is the apparent radiance received by the sensor; L p is the atmospheric radiation; G b , G t They are the influence of direct atmospheric radiation and scattering respectively; S is the atmospheric albedo; ρ is the surface reflectivity or the true reflectivity of the ground object.

[0105] The MODTRAN radiation transfer model is used to simulate atmospheric radiation transfer to obtain three atmospheric correction coefficients L p , G b +G t , S, and then substitute the apparent reflectivity obtained at the sensor into the above formula to obtain the surface reflectivity or the true reflectivity of the object, thereby completing the atmospheric correction.

[0106] Step S6, repeating steps S3-S5 to perform atmospheric correction for the next remote sensing image, wherein the inversion result of step S5 is used as input for the next stage to speed up the inversion speed.

[0107] In step S6, if the atmospheric modes and parameters of adjacent pixels of the hyperspectral image are similar, the results of the atmospheric parameter correction this time are used as the initial values ​​of the atmospheric correction in the next stage to accelerate the convergence speed of the collaborative inversion. In the present invention, it cannot be guaranteed that there are dark pixels in the hyperspectral image. For adjacent pixels, they have the same ground objects or similar surface reflectances. Therefore, the surface reflectance of the adjacent pixels corrected this time is subtracted from the atmospheric correction pixels in the next stage as the dark pixels.

[0108] The method of the present invention realizes high efficiency and high precision of atmospheric correction of hyperspectral images by designing rolling correction of adjacent pixels, and provides a technical method for real-time atmospheric correction of long-strip data or satellite-borne and airborne hyperspectral cameras.

[0109] Although the present invention is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art may make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope defined by the claims of the present invention.

Claims

1. An embedded hyperspectral atmospheric correction processing method, characterized in that: include: Step 1, using the MODTRAN radiation transfer model, establish a general lookup table for atmospheric correction for specified bands, where the specified bands include aerosol inversion bands, water vapor absorption bands, and reference bands; Step 2: construct a neural network model based on the atmospheric correction general lookup table, and obtain the optimal weight parameters of the neural network model through training; Step 3: Set the atmospheric water vapor content WVC i-1 AOT with visibility i-1 , together with the observation information in the remote sensing image data, is input into the neural network model, and WVC is determined based on the output of the neural network model. i-1 The corresponding atmospheric transmittance T' is obtained by comparing T' with the atmospheric transmittance reference value T to adjust the inversion to obtain WVC i ; Determine AOT based on the output of the neural network model i-1 The corresponding simulated ground reflectivity ρ” is obtained by comparing ρ and ρ” to adjust the inversion to obtain AOT i ; Repeat the inversion process until the difference between T' and T is less than the preset error, and the difference between ρ and ρ" is less than the preset error, and determine the atmospheric water vapor content WVC and visibility AOT obtained by the final inversion; Step 4: Based on the observation information in the remote sensing image data and the inverted atmospheric water vapor content WVC and visibility AOT, call MODTRAN to calculate the atmospheric correction coefficient x for all bands in parallel. a ,x b ,x c , corresponding to L p , G b +G t , S, L p is the atmospheric radiation, G b , G t are the influence of direct atmospheric radiation and scattering, respectively, and S is the atmospheric albedo.

2. The method according to claim 1, characterized in that In step 1, the MODTRAN radiation transfer model is used to simulate the atmospheric radiation transfer process to obtain the atmospheric transmittance, and then the atmospheric path radiation L is calculated. p , large balloon surface albedo S and radiation correction parameter G b +G t ; In the simulation process, only the atmospheric transmission process of the aerosol inversion band, water vapor absorption band and reference band was simulated; Among them, Aerosol inversion bands include blue light band, red light band, and near-infrared band; The water vapor absorption bands include 1117-1143nm, 935-955nm, and 810-830nm; The upper reference band corresponding to the water vapor absorption band 1117-1143nm includes 1184-1210nm, and the lower reference band corresponding to the water vapor absorption band 1117-1143nm includes 1050-1067nm; The upper reference band corresponding to the water vapor absorption band 935-955nm includes 870-890nm, and the lower reference band corresponding to the water vapor absorption band 935-955nm includes 995-1020nm; The upper reference band corresponding to the water vapor absorption band 810-830nm includes 850-870nm, and the lower reference band corresponding to the water vapor absorption band 810-830nm includes 770-790nm.

3. The method according to claim 1, characterized in that In step 1, simulations are performed for three cases where the surface reflectivity ρ is 0, 0.25, and 0.5, and the atmospheric correction coefficient is solved by the following set of equations: The atmospheric correction factor is: Where L' 0.25 =L 0.25 -L0, L' 0.5 =L 0.5 -L0.

4. The method according to claim 1, characterized in that: In step 1, the process of building a table for MODTRAN based on a multi-core processor is as follows: Enumerate different observation conditions and atmospheric states, distribute each condition as a single case to multiple threads or processes through command distribution, and generate a Tape5 file; Call MODTRAN in parallel through multi-core processors to build a lookup table; In the process of processing hyperspectral remote sensing data, the parallel and for instructions in OpenMP are used to decompose the tasks and enable thread groups to process multiple situations simultaneously. Execution ends: Read the data in the result file Tape7, calculate the atmospheric correction coefficient, and package the input conditions and calculation results into the database.

5. The method according to claim 1, characterized in that In step 2, a neural network model is constructed; the model input includes visibility AOT, atmospheric water vapor content WVC, observation zenith angle θ v 、Sun zenith angle θ s , relative azimuth Red band center wavelength λ red 、Blue band center wavelength λ blue , water vapor absorption band center wavelength λ absorb and reference bands; the output is the atmospheric correction coefficient x for the red, blue and water vapor absorption bands a ,x b ,x c , atmospheric transmittance, x a ,x b ,x c L p , G b +G t , S; The neural network model network structure includes: 1) Convolutional layer Conv1d 3×1×32 2) Activation layer Relu 3) Convolutional layer Conv1d 3×32×32 4) Activation layer Relu 5)Maxpool1d 6) Convolutional layer Conv1d 3×32×32 7) Activation layer Relu 8) Convolutional layer Conv1d 3×32×32 9) Activation layer Relu 10)Maxpool1d 11) Convolutional layer Conv1d 3×32×10 12) Activation layer Relu.

6. The method according to claim 1, characterized in that In step 3, the normalized difference vegetation index NDVI and the apparent reflectance at 2100nm are used to determine the dark pixels of the hyperspectral image data, where the dark pixel requirements are as follows: ρ' 860 , ρ' 660 , ρ' 2100 They are the apparent reflectance of the hyperspectral images at 860nm, 660nm, and 2100nm respectively; the atmospheric water vapor content WVC and visibility AOT are inverted through the screened dark pixels.

7. The method according to claim 1, characterized in that In step 3, the attitude and orbit parameters and imaging conditions are obtained by real-time monitoring, and the initial values ​​of visibility and water vapor content are set to Feed the parameters back to the neural network model and output the atmospheric correction coefficients for the red and blue bands and the water vapor absorption band and atmospheric transmittance T'; The simulated ground reflectivity ρ" is calculated based on the atmospheric correction coefficients and apparent radiance of the red and blue bands 660 ,ρ” 470 ; If ρ” 660 and ρ 660 If the difference is not within the allowable range, adjust AOT i 660 , ρ 660 ≈ρ' 2100 / 2; if ρ” 660 Less than ρ 660 , then AOT i 660 Increase If ρ” 660 Greater than ρ 660 , then AOT i 660 Reduce step size Until the difference between the two is within the allowable error range; If ρ” 470 and ρ 470 If the difference is not within the allowable range, adjust AOT i 470 , ρ 470 ≈ρ' 2100 / 4; if ρ” 470 Less than ρ 470 , then AOT i 470 Increase If ρ” 470 Greater than ρ 470 , then AOT i 470 Reduce step size Until the difference between the two is within the allowable error range; The average visibility obtained from the red and blue bands As a result of the visibility of this inversion AOT i ; For the reference band corresponding to the water vapor absorption band, the atmospheric transmittance reference value T satisfies: In the formula, ρ(λ),ρ(λ up ),ρ(λ down ) are the apparent reflectances of the water vapor absorption band, the upper reference band, and the lower reference band, respectively; c1 and c2 are the reflectance weights of the reference bands, respectively; if T is less than the transmittance T' obtained by the neural network model, the atmospheric water vapor content is WVC i Increase the step size σ2, otherwise decrease it.

8. The method according to claim 1, characterized in that In step 4, based on the image observation information and the atmospheric parameters calculated in step 3, the hyperspectral image bands are grouped and MODTRAN is called to calculate the atmospheric correction coefficients L for all bands. p , G b +G t , S; the atmospheric radiation transfer equation is expressed as: Where L is the apparent radiance received by the sensor; L p is the atmospheric radiation; G b , G t They are the influence of direct atmospheric radiation and scattering respectively; S is the atmospheric albedo; ρ is the surface reflectivity or the true reflectivity of the ground object; Combined with the atmospheric correction factor L p , G b +G t , S, and then substitute the apparent reflectivity obtained at the sensor into the above formula to obtain the surface reflectivity or the true reflectivity of the object, thereby completing the atmospheric correction.

9. The method according to claim 1, characterized in that: The method further comprises: Step 5, repeat steps 3-4 to perform atmospheric correction for the next remote sensing image, wherein the inversion result of step 4 is used as input for the next stage to speed up the inversion.

10. The method according to claim 1, characterized in that In step 5, for adjacent pixels with the same ground objects or similar surface reflectance, the surface reflectance of the adjacent pixels of the atmospheric correction of the previous remote sensing image is subtracted from the atmospheric correction pixel of the current remote sensing image as the dark pixel.

Images