A neural network-based atmospheric correction algorithm optimization method

Abstract

The application discloses a neural network-based atmospheric correction algorithm optimization method, which comprises the following steps: obtaining m measured spectrum data of a specific sea area and obtaining matched j effective satellite spectrum data, scoring the j effective satellite spectrum data and scoring the m measured spectrum data to obtain m+j effective training sample numbers; training a neural network; applying selected n atmospheric correction algorithms to perform pixel-by-pixel preprocessing on satellite data of the specific sea area, obtaining n spectrum data for each pixel point, inputting the spectrum data into the trained neural network to obtain corresponding scores, averaging scores of all pixel points, and the algorithm with the maximum score is the optimal algorithm. The method disclosed by the application can quickly judge the adaptability of various atmospheric correction algorithms in a specific sea area, accurately select the most suitable algorithm, and thus improve the precision of remote sensing reflectance R rs data and indirectly improve the precision of water color information inversion results.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of ocean remote sensing, and particularly relates to a neural network-based atmospheric correction algorithm optimization method. BACKGROUND

[0002] The ratio of the water-leaving radiance to the just above-water downward irradiance is defined as the remote sensing reflectance (R rs ) of the water body spectrum, and high-precision ocean remote sensing retrieval research requires accurate R rs data. The water body contribution to the water-leaving radiance in the total signal received by the ocean color satellite sensor is less than 10%, and we need to perform atmospheric correction on the received signal, which is one of the key technologies of ocean optical remote sensing. Accurate atmospheric correction is a prerequisite for obtaining high-precision R rs data and achieving high-precision water color information retrieval.

[0003] There are many internationally publicized and operational atmospheric correction algorithms at present, such as the NASA (National Aeronautics and Space Administration) standard near-infrared atmospheric correction algorithm, the MUMM (Management Unit of the North Sea Mathematical Models) algorithm, the KOSC GOCI GDPS2.0 standard atmospheric correction algorithm and the KOSC GOCI GDPS1.3 standard atmospheric correction algorithm published by KOSC mainly for stationary satellite GOCI, and the Acolite dark spectrum fitting algorithm and the exponential extrapolation algorithm mainly for high-resolution optical satellites, etc.

[0004] In the past, for the sea area to be studied, a certain atmospheric correction algorithm was usually selected according to some scientific research experience to perform atmospheric correction processing on satellite data. However, numerous studies have shown that different atmospheric correction algorithms are suitable for different sea areas and water body types, and how to quickly evaluate the suitable atmospheric correction algorithm for specific satellite data in specific sea areas from numerous atmospheric correction models is a problem to be solved. In the past, researchers have evaluated by error analysis with field data sets, but the actual evaluation significance of the single band comparison scatter plot cannot represent the overall situation of the complete R rs (λ) spectrum. Wei et al. proposed a quality assurance system for evaluating the quality of the spectrum in 2016, but the system has too large a calculation amount and is not suitable for use in the optimization process of multiple atmospheric correction algorithms. SUMMARY

[0005] To solve the above technical problems, the application provides a neural network-based atmospheric correction algorithm optimization method, which can quickly judge the adaptability of various atmospheric correction algorithms in a specific sea area, and more accurately optimize the most suitable model algorithm, thereby improving the accuracy of remote sensing reflectance R rs The accuracy of the data indirectly improves the accuracy of the water color information inversion result.

[0006] To achieve the above purpose, the technical scheme of the application is as follows:

[0007] A neural network-based atmospheric correction algorithm optimization method, comprising the following steps:

[0008] Step 1: Construct a spectral data set:

[0009] Obtain m measured spectral data in a specific sea area, and obtain j effective satellite spectral data matched according to the time and latitude and longitude information of the m measured spectral data. Score the j effective satellite spectral data, and score the m measured spectral data to obtain m+j effective training sample numbers;

[0010] Step 2: Train the neural network:

[0011] Divide the above m+j effective training samples into a training set and a validation set. The training set is used to train the neural network, and the input is the measured spectral data and the effective satellite spectral data, and the output is the respective score value. The validation set is used to verify the training effect;

[0012] Step 3: Atmospheric correction algorithm optimization:

[0013] Apply the selected n atmospheric correction algorithms to the satellite image data of the specific sea area for pixel-by-pixel preprocessing. Each pixel point obtains n spectral data, which is input into the trained neural network to obtain the corresponding score value. The average score value of all pixel points of the satellite image of the specific sea area is calculated, and the atmospheric correction algorithm with the maximum score value is selected as the optimal item. If multiple satellite data are used, the average of all score values of the multiple satellite data should be calculated. In the above scheme, in step 1, the measured spectral data refers to the measured remote sensing reflectance R rs (λ), and the wavelength λ covers k wavebands λ1, λ2,..., λ i ,..., λ k .

[0014] In the above scheme, in step 1, the m measured spectral data should cover data in spring, summer, autumn and winter, and m≥12.

[0015] Preferably, the m measured spectral data cover at least one data in clear sky in each month.

[0016] In the above scheme, in step one, the λ for each band can be obtained from m measured spectral data. i The upper limit of the spectral value R rs_max (λ i ) and lower limit R rs_min (λ i Then, based on the principle of a time scale within 3 hours and a spatial scale within 300m, j effective satellite spectral data points R are obtained from the time and latitude / longitude information of m measured spectral data stations. rs_star (λ).

[0017] In the above scheme, the method for scoring the j valid satellite spectral data in step one is as follows:

[0018]

[0019] Wherein, Sc(λ) i ) is the score of the i-th band, if R rs_star (λ i The value of ) in R rs_min (λ i ) and R rs_max (λ i Outside the boundary, the score Sc(λ) of this band i ) will be assigned the value 0, that is, Sc(λ) i ) = 0; otherwise Sc(λ) i ) = 1, k is the number of bands.

[0020] In the above scheme, in step one, the score assigned to each of the m measured spectral data is 1.

[0021] In the above scheme, in step two, the neural network is a BP neural network.

[0022] In the above scheme, in step two, the training set and the validation set account for 80% and 20% of the effective training samples, respectively.

[0023] Through the above technical solution, the atmospheric correction algorithm optimization method based on neural networks provided by the present invention has the following beneficial effects:

[0024] This invention provides a simple and effective method for selecting the optimal atmospheric correction algorithm, which can accurately select the best algorithm suitable for a specific sea area from among many operational atmospheric correction algorithms. This invention processes measured spectral data of a specific sea area to obtain matching effective satellite spectral data, scores each data point, and constructs a dataset. The dataset is then used to train a neural network, which is used to select the optimal atmospheric correction algorithm. This method achieves high accuracy and can quickly select the most suitable atmospheric correction algorithm for a specific sea area, thereby improving the remote sensing reflectance R. rs The accuracy of the data indirectly improves the results of water color information inversion.

[0025] It should be noted that once the model is trained, it will not be limited by the satellite sensors used in the training samples and can be applied to the atmospheric correction algorithm optimization process for other optical satellite sensor data. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0027] Figure 1 This is a schematic diagram of an atmospheric correction algorithm optimization process disclosed in an embodiment of the present invention;

[0028] Figure 2 R for four atmospheric correction algorithms rs (λ) Spatial distribution of score values; (a) Seadas-default; (b) Seadas-MUMM; (c) GDPS2.0; (d) GDPS1.3. Detailed Implementation

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0030] This invention provides a method for optimizing an atmospheric correction algorithm based on a neural network, comprising the following steps:

[0031] Step 1: Construct a spectral dataset:

[0032] Acquire m measured spectral data points from a specific sea area, and obtain j valid satellite spectral data points based on the time and latitude / longitude information of the m measured spectral data points. Score the j valid satellite spectral data points, and assign scores to the m measured spectral data points to obtain m+j valid training samples.

[0033] Among them, the measured spectral data refers to the measured remote sensing reflectance R. rs(λ), where band λ covers k bands λ1, λ2, ..., λ that are the same as the target optical satellite data bands. i ,...,λ k .

[0034] The *m* measured spectral data points must cover data from all four seasons (spring, summer, autumn, and winter), and *m* ≥ 12. Preferably, the *m* measured spectral data points cover at least one data point under clear skies for each month. The larger *m* is, the smaller the specific sea area selected, and the more accurate the final evaluation results.

[0035] λ for each band can be obtained from m measured spectral data. i The upper limit of the spectral value R rs_max (λ i ) and lower limit R rs_min (λ i Then, based on the principle of a time scale within 3 hours and a spatial scale within 300m, j effective satellite spectral data points R are obtained from the time and latitude / longitude information of m measured spectral data stations. rs_star (λ).

[0036] The method for scoring j valid satellite spectral data is as follows:

[0037]

[0038] Wherein, Sc(λ) i ) is the score of the i-th band, if R rs_star (λ i The value of ) in R rs_min (λ i ) and R rs_max (λ i Outside the boundary, the score Sc(λ) of this band i ) will be assigned the value 0, that is, Sc(λ) i ) = 0; otherwise Sc(λ) i ) = 1, k is the number of bands.

[0039] As can be seen from equation (1), the total score is in the range of [0,1]. The higher the score, the higher the quality of the spectral data and the more suitable the atmospheric correction algorithm is.

[0040] This yields m+j effective training samples, including m measured spectral data R. rs (λ), Score = 1; j valid satellite spectral data R rs_star (λ), the score value is calculated by equation (1).

[0041] Step 2, train the neural network:

[0042] The above m+j effective training samples are divided into a training set and a validation set, and the training set and the validation set account for 80% and 20% of the number of effective training samples, respectively. The training set is used to train the BP neural network, and the input is the measured spectrum data and the effective satellite spectrum data, and the output is the respective score; the validation set is used to verify the training effect;

[0043] Step three, the atmospheric correction algorithm is preferably:

[0044] According to the trained neural network, we can select the adaptability of n atmospheric correction algorithms (n≥2) in a specific sea area during the optical satellite data processing process, as shown in the following formula: Figure 1 The specific process is as follows:

[0045] The selected n atmospheric correction algorithms are applied to the satellite image data (x*y pixels) of the specific sea area for pixel-by-pixel preprocessing, and each pixel point obtains n spectrum data (R rs_1 (λ), R rs_2 (λ),... R rs_n (λ)) are input into the trained neural network to obtain corresponding n Score values (Score(R rs_1 (λ)), Score(R rs_1 (λ)),... Score(R rs_n (λ))), and the average of the Score values of all x*y pixels of the satellite image is taken, and the maximum average value corresponds to the optimal atmospheric correction algorithm selected by the neural network. If multiple satellite data are selected, the average of all Score values of the multiple satellite data should be calculated.

[0046] The above preprocessing refers to atmospheric correction of satellite data. Each atmospheric correction algorithm has its own model, and satellite data can be input to perform atmospheric correction to obtain atmospheric correction results-remote sensing reflectance spectrum data R rs (λ).

[0047] Example 1: GOCI atmospheric correction algorithm optimization in the Bohai Sea area

[0048] Step one, construct a spectrum data set:

[0049] First, we collected the in-situ measured R rs (λ) data m=76 (out-to-sea measurement, SeaBASS, and NOMAD) in the sea area, obtained the upper limit R rs_max (λ i ) and the lower limit R rs_min (λ iSee Table 1. Since GOCI I has 8 landscape measurement data points per day (8:00 AM to 3:00 PM) and GOCI II has 10 landscape measurement data points per day (7:00 AM to 4:00 PM), we will use the L2 R data provided on the official website. rs (λ) data is matched with field experimental data to obtain the final matched effective satellite R. rs_star (λ) Spectral data j = 297, using formula (1) to obtain m+j effective training samples (m on-site R rs (λ) Experimental data, Score = 1; j effective satellites R rs_star (λ) Spectral data, the score value is calculated by equation (1).

[0050] Table 1R rs (λ) Lower limit of spectral values ​​R rs_min (λ i ) and upper limit R rs_max (λ i )

[0051]

[0052] Step 2, train the neural network:

[0053] The BP neural network is trained using 80% of the obtained sample dataset. Input sample R rs (λ) data, outputting the Score value. The remaining 20% ​​of the sample dataset was used for testing and analysis, resulting in a correlation coefficient of 0.886 between the BP neural network output Score value and the sample value, with an average relative error of 23.2%.

[0054] Step 3, Atmospheric correction algorithm optimization:

[0055] This example uses the NASA (National Aeronautics and Space Administration) standard near-infrared atmospheric correction algorithm (Seadas-default), MUMM (Management Unit of the North Sea Mathematical Models, Seadas-MUMM), KOSC GOCI GDPS2.0 standard atmospheric correction algorithm (GDPS2.0), and KOSC GOCI GDPS1.3 standard atmospheric correction algorithm (GDPS1.3) to perform atmospheric correction preprocessing on selected GOCI satellite data. (The appropriate algorithm should be selected based on actual needs, such as time and latitude / longitude. This example uses GOCI data from the Bohai Sea on September 13, 2015.) The corrected remote sensing reflectance R for each pixel of each data point under each of the four atmospheric correction algorithms is obtained.rs (λ) data, which are respectively input into the trained neural network to score, and get the score value of the corresponding output Score(R rs_1 (λ)), Score(R rs_2 (λ)), Score(R rs_3 (λ)), Score(R rs_4 (λ)), and finally the spatial distribution of the GOCI data Score value under the four atmospheric correction algorithms is shown in Figure 2

[0056] The average Score of the sea area under the four atmospheric correction algorithms (Seadas-default, Seadas-MUMM, GDPS2.0, GDPS1.3) is 0.5985, 0.6220, 0.6883, and 0.9064, respectively. The maximum value is 0.9064. Therefore, the final screening shows that the KOSC GOCI GDPS1.3 standard atmospheric correction algorithm has the best performance in atmospheric correction of GOCI data in the sea area. The atmospheric correction algorithm model optimized by the model is consistent with the actual evaluation result, and the atmospheric correction model of GOCI satellite data in the study sea area is optimized simply.

[0057] Note that the trained neural network scoring model can not only be used for atmospheric correction model optimization of GOCI satellite data, but also can be applied to other optical satellite data, such as MODIS, Landsat OLI, MERIS, etc.

[0058] The above description of disclosed embodiments enables those skilled in the art to carry out or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.​

Claims

1. A neural network-based atmospheric correction algorithm optimization method, preferably characterized by, The method comprises the following steps: Step one, constructing a spectral data set: Obtain m measured spectral data of a specific sea area, and obtain j effective satellite spectral data matched according to the time and latitude and longitude information of the m measured spectral data, score the j effective satellite spectral data, and score the m measured spectral data to obtain m+j effective training sample numbers; Step two, training a neural network: Divide the above m+j effective training samples into a training set and a validation set, train the neural network using the training set, input the measured spectral data and the effective satellite spectral data, and output the respective scores; and verify the training effect using the validation set; Step three, optimal selection of an atmospheric correction algorithm: Apply the selected n atmospheric correction algorithms to perform pixel-by-pixel preprocessing on the satellite image data of the specific sea area, obtain n spectral data for each pixel point, input the n spectral data into the trained neural network to obtain corresponding scores, average the scores corresponding to all pixel points of the satellite image of the specific sea area, and select the atmospheric correction algorithm with the maximum score as the optimal item; if multiple satellite data are selected, the average of all scores of the multiple satellite data is obtained when the optimal item is selected.

2. The neural network-based atmospheric correction algorithm optimization method according to claim 1, characterized in that, In step one, the measured spectral data refers to the measured remote sensing reflectance R rs (λ), the waveband λ covers the same k wavebands λ1, λ2,..., λ i ,..., λ k .

3. The neural network-based atmospheric correction algorithm optimization method of claim 1, wherein In step one, the m measured spectral data need to meet the data under four seasons of spring, summer, autumn and winter, and m≥12.

4. The neural network-based atmospheric correction algorithm according to claim 1 or 3, characterized in that, In the m measured spectral data, the data under clear sky in each month at least includes one.

5. The neural network-based atmospheric correction algorithm optimization method according to claim 2, characterized in that, In step one, λ for each band can be obtained from m measured spectral data. i Upper limit of spectral values R rs_max (λ) i ) and lower limit R rs_min (λ) i Then, based on the principle of a time scale within 3 hours and a spatial scale within 300m, j effective satellite spectral data points are obtained by matching the time and latitude / longitude information of m measured spectral data stations. R rs_star (λ).

6. The neural network-based atmospheric correction algorithm optimization method according to claim 5, characterized in that, In step one, the scoring method of the j effective satellite spectral data is as follows: (1) Wherein, Sc(λ) i ) is the score of the i-th band, if R rs_star (λ i The value of ) is in R rs_min (λ) i )and R rs_max (λ) i Outside the boundary, the score Sc(λ) of this band i ) will be assigned the value 0, that is, Sc(λ) i ) = 0; otherwise Sc(λ) i )=1, k is the number of bands.

7. The neural network-based atmospheric correction algorithm optimization method according to claim 1, characterized in that, In step one, the scores of the m measured spectral data are all 1.

8. The neural network-based atmospheric correction algorithm optimization method according to claim 1, characterized in that, In step two, the neural network is a BP neural network.

9. The neural network-based atmospheric correction algorithm optimization method according to claim 1, characterized in that, In step two, the training set and the validation set account for 80% and 20% of the number of effective training samples, respectively.

Images